diff --git a/.buildkite/check-wheel-size.py b/.buildkite/check-wheel-size.py index 0412c5f37952d..e29eb78a9f945 100644 --- a/.buildkite/check-wheel-size.py +++ b/.buildkite/check-wheel-size.py @@ -2,8 +2,11 @@ import sys import zipfile -# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 250 MB -VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 250)) +# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 300 MiB +# Note that we have 400 MiB quota, please use it wisely. +# See https://github.com/pypi/support/issues/3792 . +# Please also sync the value with the one in Dockerfile. +VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 300)) def print_top_10_largest_files(zip_file): diff --git a/.buildkite/nightly-benchmarks/scripts/nightly-annotate.sh b/.buildkite/nightly-benchmarks/scripts/nightly-annotate.sh index 686f70dbece6c..69b6b146b3549 100644 --- a/.buildkite/nightly-benchmarks/scripts/nightly-annotate.sh +++ b/.buildkite/nightly-benchmarks/scripts/nightly-annotate.sh @@ -43,7 +43,7 @@ main() { - # The figures should be genereated by a separate process outside the CI/CD pipeline + # The figures should be generated by a separate process outside the CI/CD pipeline # # generate figures # python3 -m pip install tabulate pandas matplotlib diff --git a/.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh b/.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh index 3f38cf5137535..32bd34c431c89 100644 --- a/.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh +++ b/.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh @@ -301,6 +301,104 @@ run_serving_tests() { kill_gpu_processes } +run_genai_perf_tests() { + # run genai-perf tests + + # $1: a json file specifying genai-perf test cases + local genai_perf_test_file + genai_perf_test_file=$1 + + # Iterate over genai-perf tests + jq -c '.[]' "$genai_perf_test_file" | while read -r params; do + # get the test name, and append the GPU type back to it. + test_name=$(echo "$params" | jq -r '.test_name') + + # if TEST_SELECTOR is set, only run the test cases that match the selector + if [[ -n "$TEST_SELECTOR" ]] && [[ ! "$test_name" =~ $TEST_SELECTOR ]]; then + echo "Skip test case $test_name." + continue + fi + + # prepend the current serving engine to the test name + test_name=${CURRENT_LLM_SERVING_ENGINE}_${test_name} + + # get common parameters + common_params=$(echo "$params" | jq -r '.common_parameters') + model=$(echo "$common_params" | jq -r '.model') + tp=$(echo "$common_params" | jq -r '.tp') + dataset_name=$(echo "$common_params" | jq -r '.dataset_name') + dataset_path=$(echo "$common_params" | jq -r '.dataset_path') + port=$(echo "$common_params" | jq -r '.port') + num_prompts=$(echo "$common_params" | jq -r '.num_prompts') + reuse_server=$(echo "$common_params" | jq -r '.reuse_server') + + # get client and server arguments + server_params=$(echo "$params" | jq -r ".${CURRENT_LLM_SERVING_ENGINE}_server_parameters") + qps_list=$(echo "$params" | jq -r '.qps_list') + qps_list=$(echo "$qps_list" | jq -r '.[] | @sh') + echo "Running over qps list $qps_list" + + # check if there is enough GPU to run the test + if [[ $gpu_count -lt $tp ]]; then + echo "Required num-shard $tp but only $gpu_count GPU found. Skip testcase $test_name." + continue + fi + + if [[ $reuse_server == "true" ]]; then + echo "Reuse previous server for test case $test_name" + else + kill_gpu_processes + bash "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/launch-server.sh" \ + "$server_params" "$common_params" + fi + + if wait_for_server; then + echo "" + echo "$CURRENT_LLM_SERVING_ENGINE server is up and running." + else + echo "" + echo "$CURRENT_LLM_SERVING_ENGINE failed to start within the timeout period." + break + fi + + # iterate over different QPS + for qps in $qps_list; do + # remove the surrounding single quote from qps + if [[ "$qps" == *"inf"* ]]; then + echo "qps was $qps" + qps=$num_prompts + echo "now qps is $qps" + fi + + new_test_name=$test_name"_qps_"$qps + backend=$CURRENT_LLM_SERVING_ENGINE + + if [[ "$backend" == *"vllm"* ]]; then + backend="vllm" + fi + #TODO: add output dir. + client_command="genai-perf profile \ + -m $model \ + --service-kind openai \ + --backend vllm \ + --endpoint-type chat \ + --streaming \ + --url localhost:$port \ + --request-rate $qps \ + --num-prompts $num_prompts \ + " + + echo "Client command: $client_command" + + eval "$client_command" + + #TODO: process/record outputs + done + done + + kill_gpu_processes + +} prepare_dataset() { @@ -328,12 +426,17 @@ main() { pip install -U transformers + pip install -r requirements-dev.txt + which genai-perf + # check storage df -h ensure_installed wget ensure_installed curl ensure_installed jq + # genai-perf dependency + ensure_installed libb64-0d prepare_dataset @@ -345,6 +448,10 @@ main() { # run the test run_serving_tests "$BENCHMARK_ROOT/tests/nightly-tests.json" + # run genai-perf tests + run_genai_perf_tests "$BENCHMARK_ROOT/tests/genai-perf-tests.json" + mv artifacts/ $RESULTS_FOLDER/ + # upload benchmark results to buildkite python3 -m pip install tabulate pandas python3 "$BENCHMARK_ROOT/scripts/summary-nightly-results.py" diff --git a/.buildkite/nightly-benchmarks/tests/genai-perf-tests.json b/.buildkite/nightly-benchmarks/tests/genai-perf-tests.json new file mode 100644 index 0000000000000..edbe9f2df0ce0 --- /dev/null +++ b/.buildkite/nightly-benchmarks/tests/genai-perf-tests.json @@ -0,0 +1,23 @@ +[ + { + "test_name": "llama8B_tp1_genai_perf", + "qps_list": [4,8,16,32], + "common_parameters": { + "model": "meta-llama/Meta-Llama-3-8B-Instruct", + "tp": 1, + "port": 8000, + "num_prompts": 500, + "reuse_server": false + }, + "vllm_server_parameters": { + "disable_log_stats": "", + "disable_log_requests": "", + "gpu_memory_utilization": 0.9, + "num_scheduler_steps": 10, + "max_num_seqs": 512, + "dtype": "bfloat16" + }, + "genai_perf_input_parameters": { + } + } +] \ No newline at end of file diff --git a/.buildkite/run-cpu-test.sh b/.buildkite/run-cpu-test.sh index 9925db7bea593..e19ace782feb5 100644 --- a/.buildkite/run-cpu-test.sh +++ b/.buildkite/run-cpu-test.sh @@ -83,6 +83,6 @@ function cpu_tests() { tests/lora/test_qwen2vl.py" } -# All of CPU tests are expected to be finished less than 25 mins. +# All of CPU tests are expected to be finished less than 40 mins. export -f cpu_tests -timeout 30m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE" +timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE" diff --git a/.buildkite/run-hpu-test.sh b/.buildkite/run-hpu-test.sh index 8f3b08212fd6a..1edcb1d2669e9 100644 --- a/.buildkite/run-hpu-test.sh +++ b/.buildkite/run-hpu-test.sh @@ -8,9 +8,17 @@ set -ex docker build -t hpu-test-env -f Dockerfile.hpu . # Setup cleanup +# certain versions of HPU software stack have a bug that can +# override the exit code of the script, so we need to use +# separate remove_docker_container and remove_docker_container_and_exit +# functions, while other platforms only need one remove_docker_container +# function. +EXITCODE=1 remove_docker_container() { docker rm -f hpu-test || true; } -trap remove_docker_container EXIT +remove_docker_container_and_exit() { remove_docker_container; exit $EXITCODE; } +trap remove_docker_container_and_exit EXIT remove_docker_container # Run the image and launch offline inference -docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic.py \ No newline at end of file +docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic.py +EXITCODE=$? diff --git a/.buildkite/run-neuron-test.sh b/.buildkite/run-neuron-test.sh index 189714ebb6d75..0590dad4f311f 100644 --- a/.buildkite/run-neuron-test.sh +++ b/.buildkite/run-neuron-test.sh @@ -25,8 +25,11 @@ if [ -f /tmp/neuron-docker-build-timestamp ]; then last_build=$(cat /tmp/neuron-docker-build-timestamp) current_time=$(date +%s) if [ $((current_time - last_build)) -gt 86400 ]; then + # Remove dangling images (those that are not tagged and not used by any container) docker image prune -f - docker system prune -f + # Remove unused volumes / force the system prune for old images as well. + docker volume prune -f && docker system prune -f + # Remove huggingface model artifacts and compiler cache rm -rf "${HF_MOUNT:?}/*" rm -rf "${NEURON_COMPILE_CACHE_MOUNT:?}/*" echo "$current_time" > /tmp/neuron-docker-build-timestamp diff --git a/.buildkite/test-pipeline.yaml b/.buildkite/test-pipeline.yaml index 74b287c7adbfa..d5d02fdeb7f4b 100644 --- a/.buildkite/test-pipeline.yaml +++ b/.buildkite/test-pipeline.yaml @@ -52,7 +52,6 @@ steps: - tests/worker - tests/standalone_tests/lazy_torch_compile.py commands: - - pip install git+https://github.com/Isotr0py/DeepSeek-VL2.git # Used by multimoda processing test - python3 standalone_tests/lazy_torch_compile.py - pytest -v -s mq_llm_engine # MQLLMEngine - pytest -v -s async_engine # AsyncLLMEngine @@ -77,7 +76,9 @@ steps: - tests/basic_correctness/test_basic_correctness - tests/basic_correctness/test_cpu_offload - tests/basic_correctness/test_preemption + - tests/basic_correctness/test_cumem.py commands: + - pytest -v -s basic_correctness/test_cumem.py - pytest -v -s basic_correctness/test_basic_correctness.py - pytest -v -s basic_correctness/test_cpu_offload.py - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py @@ -107,7 +108,7 @@ steps: source_file_dependencies: - vllm/ commands: - - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py + - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py --ignore=entrypoints/llm/test_collective_rpc.py - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process @@ -126,11 +127,15 @@ steps: - tests/distributed - tests/spec_decode/e2e/test_integration_dist_tp4 - tests/compile + - examples/offline_inference/rlhf.py commands: - pytest -v -s distributed/test_utils.py - pytest -v -s compile/test_basic_correctness.py - pytest -v -s distributed/test_pynccl.py - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py + # TODO: create a dedicated test section for multi-GPU example tests + # when we have multiple distributed example tests + - python3 ../examples/offline_inference/rlhf.py - label: Metrics, Tracing Test # 10min num_gpus: 2 @@ -178,7 +183,16 @@ steps: - vllm/ - tests/v1 commands: - - VLLM_USE_V1=1 pytest -v -s v1 + # split the test to avoid interference + - VLLM_USE_V1=1 pytest -v -s v1/core + - VLLM_USE_V1=1 pytest -v -s v1/engine + - VLLM_USE_V1=1 pytest -v -s v1/sample + - VLLM_USE_V1=1 pytest -v -s v1/worker + - VLLM_USE_V1=1 pytest -v -s v1/test_stats.py + - VLLM_USE_V1=1 pytest -v -s v1/test_utils.py + # TODO: accuracy does not match, whether setting + # VLLM_USE_FLASHINFER_SAMPLER or not on H100. + - VLLM_USE_V1=1 pytest -v -s v1/e2e - label: Examples Test # 25min working_dir: "/vllm-workspace/examples" @@ -462,7 +476,10 @@ steps: - vllm/worker/worker_base.py - vllm/worker/worker.py - vllm/worker/model_runner.py + - entrypoints/llm/test_collective_rpc.py commands: + - pytest -v -s entrypoints/llm/test_collective_rpc.py + - torchrun --nproc-per-node=2 distributed/test_torchrun_example.py - pytest -v -s ./compile/test_basic_correctness.py - pytest -v -s ./compile/test_wrapper.py - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed' @@ -471,7 +488,9 @@ steps: - pytest models/encoder_decoder/language/test_bart.py -v -s -m 'distributed(num_gpus=2)' - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m 'distributed(num_gpus=2)' - pytest models/decoder_only/vision_language/test_models.py -v -s -m 'distributed(num_gpus=2)' - - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py + # this test fails consistently. + # TODO: investigate and fix + # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/disagg_test.py @@ -509,7 +528,9 @@ steps: - vllm/engine - tests/multi_step commands: - - pytest -v -s multi_step/test_correctness_async_llm.py + # this test is quite flaky + # TODO: investigate and fix. + # - pytest -v -s multi_step/test_correctness_async_llm.py - pytest -v -s multi_step/test_correctness_llm.py - label: Pipeline Parallelism Test # 45min diff --git a/.buildkite/test-template.j2 b/.buildkite/test-template.j2 index 559da713ea14c..ce448836a8278 100644 --- a/.buildkite/test-template.j2 +++ b/.buildkite/test-template.j2 @@ -27,7 +27,7 @@ steps: depends_on: - "amd-build" agents: - queue: amd_rocm_gpu + queue: amd_gpu commands: - bash .buildkite/run-amd-test.sh "cd {{ (step.working_dir or default_working_dir) | safe }} ; {{ step.command or (step.commands | join(" && ")) | safe }}" env: diff --git a/.github/CODEOWNERS b/.github/CODEOWNERS index 3cb91fc0f8232..bc324d8b988b1 100644 --- a/.github/CODEOWNERS +++ b/.github/CODEOWNERS @@ -2,32 +2,35 @@ # for more info about CODEOWNERS file # This lists cover the "core" components of vLLM that require careful review -/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill -/vllm/core @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill -/vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill -/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill -/vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill -/vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill -/vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill +/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill +/vllm/core @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill +/vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill +/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill +/vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill +/vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill +/vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill +/vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth +/vllm/model_executor/guided_decoding @mgoin +/vllm/multimodal @DarkLight1337 @ywang96 CMakeLists.txt @tlrmchlsmth # vLLM V1 -/vllm/v1 @WoosukKwon @robertgshaw2-neuralmagic @njhill @ywang96 @comaniac @alexm-neuralmagic +/vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat # Test ownership -/tests/async_engine @njhill @robertgshaw2-neuralmagic @simon-mo +/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo /tests/test_inputs.py @DarkLight1337 @ywang96 -/tests/entrypoints @DarkLight1337 @robertgshaw2-neuralmagic @simon-mo +/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo /tests/models @DarkLight1337 @ywang96 /tests/multimodal @DarkLight1337 @ywang96 /tests/prefix_caching @comaniac @KuntaiDu /tests/spec_decode @njhill @LiuXiaoxuanPKU /tests/kernels @tlrmchlsmth @WoosukKwon -/tests/quantization @mgoin @robertgshaw2-neuralmagic +/tests/quantization @mgoin @robertgshaw2-redhat /.buildkite/lm-eval-harness @mgoin @simon-mo /tests/distributed/test_multi_node_assignment.py @youkaichao /tests/distributed/test_pipeline_parallel.py @youkaichao /tests/distributed/test_same_node.py @youkaichao -/tests/multi_step @alexm-neuralmagic @comaniac +/tests/multi_step @alexm-redhat @comaniac /tests/weight_loading @mgoin @youkaichao /tests/basic_correctness/test_chunked_prefill @rkooo567 @comaniac diff --git a/.github/workflows/actionlint.yml b/.github/workflows/actionlint.yml deleted file mode 100644 index e9f884caaa79f..0000000000000 --- a/.github/workflows/actionlint.yml +++ /dev/null @@ -1,43 +0,0 @@ -name: Lint GitHub Actions workflows -on: - push: - branches: - - "main" - - "develop" - - "llama_fp8_12062024" - paths: - - '.github/workflows/*.ya?ml' - - '.github/workflows/actionlint.*' - - '.github/workflows/matchers/actionlint.json' - pull_request: - branches: - - "main" - - "llama_fp8_12062024" - paths: - - '.github/workflows/*.ya?ml' - - '.github/workflows/actionlint.*' - - '.github/workflows/matchers/actionlint.json' - -env: - LC_ALL: en_US.UTF-8 - -defaults: - run: - shell: bash - -permissions: - contents: read - -jobs: - actionlint: - runs-on: ubuntu-latest - steps: - - name: "Checkout" - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 - with: - fetch-depth: 0 - - - name: "Run actionlint" - run: | - echo "::add-matcher::.github/workflows/matchers/actionlint.json" - tools/actionlint.sh -color diff --git a/.github/workflows/clang-format.yml b/.github/workflows/clang-format.yml deleted file mode 100644 index 7d9f8b8095a9e..0000000000000 --- a/.github/workflows/clang-format.yml +++ /dev/null @@ -1,57 +0,0 @@ -name: clang-format - -on: - # Trigger the workflow on push or pull request, - # but only for the main branch - push: - branches: - - main - - develop - - llama_fp8_12062024 - paths: - - '**/*.h' - - '**/*.cpp' - - '**/*.cu' - - '**/*.cuh' - - '.github/workflows/clang-format.yml' - pull_request: - branches: - - main - - develop - - llama_fp8_12062024 - paths: - - '**/*.h' - - '**/*.cpp' - - '**/*.cu' - - '**/*.cuh' - - '.github/workflows/clang-format.yml' - -jobs: - clang-format: - runs-on: ubuntu-latest - strategy: - matrix: - python-version: ["3.11"] - steps: - - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0 - with: - python-version: ${{ matrix.python-version }} - - name: Install dependencies - run: | - python -m pip install --upgrade pip - pip install clang-format==18.1.5 - - name: Running clang-format - run: | - EXCLUDES=( - 'csrc/moe/topk_softmax_kernels.cu' - 'csrc/quantization/gguf/ggml-common.h' - 'csrc/quantization/gguf/dequantize.cuh' - 'csrc/quantization/gguf/vecdotq.cuh' - 'csrc/quantization/gguf/mmq.cuh' - 'csrc/quantization/gguf/mmvq.cuh' - ) - find csrc/ \( -name '*.h' -o -name '*.cpp' -o -name '*.cu' -o -name '*.cuh' \) -print \ - | grep -vFf <(printf "%s\n" "${EXCLUDES[@]}") \ - | xargs clang-format --dry-run --Werror diff --git a/.github/workflows/codespell.yml b/.github/workflows/codespell.yml deleted file mode 100644 index d1b45bef6b542..0000000000000 --- a/.github/workflows/codespell.yml +++ /dev/null @@ -1,49 +0,0 @@ -name: codespell - -on: - # Trigger the workflow on push or pull request, - # but only for the main branch - push: - branches: - - main - - develop - - llama_fp8_12062024 - paths: - - "**/*.py" - - "**/*.md" - - "**/*.rst" - - pyproject.toml - - requirements-lint.txt - - .github/workflows/codespell.yml - pull_request: - branches: - - main - - develop - - llama_fp8_12062024 - paths: - - "**/*.py" - - "**/*.md" - - "**/*.rst" - - pyproject.toml - - requirements-lint.txt - - .github/workflows/codespell.yml - -jobs: - codespell: - runs-on: ubuntu-latest - strategy: - matrix: - python-version: ["3.12"] - steps: - - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0 - with: - python-version: ${{ matrix.python-version }} - - name: Install dependencies - run: | - python -m pip install --upgrade pip - pip install -r requirements-lint.txt - - name: Spelling check with codespell - run: | - codespell --toml pyproject.toml diff --git a/.github/workflows/doc-lint.yml b/.github/workflows/doc-lint.yml deleted file mode 100644 index e66e622ab2c9e..0000000000000 --- a/.github/workflows/doc-lint.yml +++ /dev/null @@ -1,34 +0,0 @@ -name: Lint documentation - -on: - push: - branches: - - main - - llama_fp8_12062024 - paths: - - "docs/**" - pull_request: - branches: - - main - - llama_fp8_12062024 - paths: - - "docs/**" - -jobs: - doc-lint: - runs-on: ubuntu-latest - strategy: - matrix: - python-version: ["3.12"] - steps: - - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0 - with: - python-version: ${{ matrix.python-version }} - - name: Install dependencies - run: | - python -m pip install --upgrade pip - pip install -r requirements-lint.txt - - name: Linting docs - run: tools/doc-lint.sh diff --git a/.github/workflows/matchers/ruff.json b/.github/workflows/matchers/ruff.json deleted file mode 100644 index f6d4479ee1996..0000000000000 --- a/.github/workflows/matchers/ruff.json +++ /dev/null @@ -1,17 +0,0 @@ -{ - "problemMatcher": [ - { - "owner": "ruff", - "pattern": [ - { - "regexp": "^(.+?):(\\d+):(\\d+): (\\w+): (.+)$", - "file": 1, - "line": 2, - "column": 3, - "code": 4, - "message": 5 - } - ] - } - ] - } diff --git a/.github/workflows/mypy.yaml b/.github/workflows/mypy.yaml deleted file mode 100644 index 376caa84ac099..0000000000000 --- a/.github/workflows/mypy.yaml +++ /dev/null @@ -1,55 +0,0 @@ -name: mypy - -on: - # Trigger the workflow on push or pull request, - # but only for the main branch - push: - branches: - - main - - develop - - llama_fp8_12062024 - paths: - - '**/*.py' - - '.github/workflows/mypy.yaml' - - 'tools/mypy.sh' - - 'pyproject.toml' - pull_request: - branches: - - main - - develop - - llama_fp8_12062024 - # This workflow is only relevant when one of the following files changes. - # However, we have github configured to expect and require this workflow - # to run and pass before github with auto-merge a pull request. Until github - # allows more flexible auto-merge policy, we can just run this on every PR. - # It doesn't take that long to run, anyway. - #paths: - # - '**/*.py' - # - '.github/workflows/mypy.yaml' - # - 'tools/mypy.sh' - # - 'pyproject.toml' - -jobs: - mypy: - runs-on: ubuntu-latest - strategy: - matrix: - python-version: ["3.9", "3.10", "3.11", "3.12"] - steps: - - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0 - with: - python-version: ${{ matrix.python-version }} - - name: Install dependencies - run: | - python -m pip install --upgrade pip - pip install mypy==1.11.1 - pip install types-setuptools - pip install types-PyYAML - pip install types-requests - pip install types-setuptools - - name: Mypy - run: | - echo "::add-matcher::.github/workflows/matchers/mypy.json" - tools/mypy.sh 1 ${{ matrix.python-version }} diff --git a/.github/workflows/png-lint.yml b/.github/workflows/png-lint.yml deleted file mode 100644 index 35dd5699f4426..0000000000000 --- a/.github/workflows/png-lint.yml +++ /dev/null @@ -1,39 +0,0 @@ -name: Lint PNG exports from excalidraw -on: - push: - branches: - - "main" - - "llama_fp8_12062024" - paths: - - '*.excalidraw.png' - - '.github/workflows/png-lint.yml' - pull_request: - branches: - - "main" - - "llama_fp8_12062024" - paths: - - '*.excalidraw.png' - - '.github/workflows/png-lint.yml' - -env: - LC_ALL: en_US.UTF-8 - -defaults: - run: - shell: bash - -permissions: - contents: read - -jobs: - actionlint: - runs-on: ubuntu-latest - steps: - - name: "Checkout" - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 - with: - fetch-depth: 0 - - - name: "Run png-lint.sh to check excalidraw exported images" - run: | - tools/png-lint.sh diff --git a/.github/workflows/pre-commit.yml b/.github/workflows/pre-commit.yml new file mode 100644 index 0000000000000..06564969dc778 --- /dev/null +++ b/.github/workflows/pre-commit.yml @@ -0,0 +1,19 @@ +name: pre-commit + +on: + pull_request: + push: + branches: [main] + +jobs: + pre-commit: + runs-on: ubuntu-latest + steps: + - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 + - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0 + with: + python-version: "3.12" + - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json" + - uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd # v3.0.1 + with: + extra_args: --all-files --hook-stage manual diff --git a/.github/workflows/ruff.yml b/.github/workflows/ruff.yml deleted file mode 100644 index 6c83f479d3936..0000000000000 --- a/.github/workflows/ruff.yml +++ /dev/null @@ -1,55 +0,0 @@ -name: ruff - -on: - # Trigger the workflow on push or pull request, - # but only for the main branch - push: - branches: - - main - - develop - - llama_fp8_12062024 - paths: - - "**/*.py" - - pyproject.toml - - requirements-lint.txt - - .github/workflows/matchers/ruff.json - - .github/workflows/ruff.yml - pull_request: - branches: - - main - - develop - # This workflow is only relevant when one of the following files changes. - # However, we have github configured to expect and require this workflow - # to run and pass before github with auto-merge a pull request. Until github - # allows more flexible auto-merge policy, we can just run this on every PR. - # It doesn't take that long to run, anyway. - #paths: - # - "**/*.py" - # - pyproject.toml - # - requirements-lint.txt - # - .github/workflows/matchers/ruff.json - # - .github/workflows/ruff.yml - -jobs: - ruff: - runs-on: ubuntu-latest - strategy: - matrix: - python-version: ["3.12"] - steps: - - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0 - with: - python-version: ${{ matrix.python-version }} - - name: Install dependencies - run: | - python -m pip install --upgrade pip - pip install -r requirements-lint.txt - - name: Analysing the code with ruff - run: | - echo "::add-matcher::.github/workflows/matchers/ruff.json" - ruff check --output-format github . - - name: Run isort - run: | - isort . --check-only diff --git a/.github/workflows/shellcheck.yml b/.github/workflows/shellcheck.yml deleted file mode 100644 index 1887f4cd69076..0000000000000 --- a/.github/workflows/shellcheck.yml +++ /dev/null @@ -1,41 +0,0 @@ -name: Lint shell scripts -on: - push: - branches: - - "main" - - "develop" - - "llama_fp8_12062024" - paths: - - '**/*.sh' - - '.github/workflows/shellcheck.yml' - pull_request: - branches: - - "main" - - "develop" - - "llama_fp8_12062024" - paths: - - '**/*.sh' - - '.github/workflows/shellcheck.yml' - -env: - LC_ALL: en_US.UTF-8 - -defaults: - run: - shell: bash - -permissions: - contents: read - -jobs: - shellcheck: - runs-on: ubuntu-latest - steps: - - name: "Checkout" - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 - with: - fetch-depth: 0 - - - name: "Check shell scripts" - run: | - tools/shellcheck.sh diff --git a/.github/workflows/yapf.yml b/.github/workflows/yapf.yml deleted file mode 100644 index b2702adb812e3..0000000000000 --- a/.github/workflows/yapf.yml +++ /dev/null @@ -1,38 +0,0 @@ -name: yapf - -on: - # Trigger the workflow on push or pull request, - # but only for the main branch - push: - branches: - - main - - develop - paths: - - "**/*.py" - - .github/workflows/yapf.yml - pull_request: - branches: - - main - - develop - - llama_fp8_12062024 - -jobs: - yapf: - runs-on: ubuntu-latest - strategy: - matrix: - python-version: ["3.12"] - steps: - - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0 - with: - python-version: ${{ matrix.python-version }} - - name: Install dependencies - run: | - python -m pip install --upgrade pip - pip install yapf==0.32.0 - pip install toml==0.10.2 - - name: Running yapf - run: | - yapf --diff --recursive . diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml new file mode 100644 index 0000000000000..ede092746c94d --- /dev/null +++ b/.pre-commit-config.yaml @@ -0,0 +1,92 @@ +default_stages: + - pre-commit # Run locally + - manual # Run in CI +repos: +- repo: https://github.com/google/yapf + rev: v0.32.0 + hooks: + - id: yapf + args: [--in-place, --verbose] + additional_dependencies: [toml] # TODO: Remove when yapf is upgraded +- repo: https://github.com/astral-sh/ruff-pre-commit + rev: v0.6.5 + hooks: + - id: ruff + args: [--output-format, github] +- repo: https://github.com/codespell-project/codespell + rev: v2.3.0 + hooks: + - id: codespell + exclude: 'benchmarks/sonnet.txt|(build|tests/(lora/data|models/fixtures|prompts))/.*|csrc/rocm/.*|csrc/gradlib/.*' +- repo: https://github.com/PyCQA/isort + rev: 5.13.2 + hooks: + - id: isort +- repo: https://github.com/pre-commit/mirrors-clang-format + rev: v18.1.5 + hooks: + - id: clang-format + exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))' + types_or: [c++, cuda] + args: [--style=file, --verbose] +- repo: https://github.com/jackdewinter/pymarkdown + rev: v0.9.27 + hooks: + - id: pymarkdown + files: docs/.* +- repo: https://github.com/rhysd/actionlint + rev: v1.7.6 + hooks: + - id: actionlint +- repo: local + hooks: + - id: mypy-local + name: Run mypy for local Python installation + entry: tools/mypy.sh 0 "local" + language: python + types: [python] + additional_dependencies: &mypy_deps [mypy==1.11.1, types-setuptools, types-PyYAML, types-requests] + stages: [pre-commit] # Don't run in CI + - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward + name: Run mypy for Python 3.9 + entry: tools/mypy.sh 1 "3.9" + language: python + types: [python] + additional_dependencies: *mypy_deps + stages: [manual] # Only run in CI + - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward + name: Run mypy for Python 3.10 + entry: tools/mypy.sh 1 "3.10" + language: python + types: [python] + additional_dependencies: *mypy_deps + stages: [manual] # Only run in CI + - id: mypy-3.11 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward + name: Run mypy for Python 3.11 + entry: tools/mypy.sh 1 "3.11" + language: python + types: [python] + additional_dependencies: *mypy_deps + stages: [manual] # Only run in CI + - id: mypy-3.12 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward + name: Run mypy for Python 3.12 + entry: tools/mypy.sh 1 "3.12" + language: python + types: [python] + additional_dependencies: *mypy_deps + stages: [manual] # Only run in CI + - id: shellcheck + name: Lint shell scripts + entry: tools/shellcheck.sh + language: script + types: [shell] + - id: png-lint + name: Lint PNG exports from excalidraw + entry: tools/png-lint.sh + language: script + types: [png] + - id: suggestion + name: Suggestion + entry: bash -c 'echo "To bypass pre-commit hooks, add --no-verify to git commit."' + language: system + verbose: true diff --git a/CMakeLists.txt b/CMakeLists.txt old mode 100644 new mode 100755 index 84194a2ff5116..9014e26648dd4 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -24,9 +24,6 @@ include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake) # Suppress potential warnings about unused manually-specified variables set(ignoreMe "${VLLM_PYTHON_PATH}") -# Prevent installation of dependencies (cutlass) by default. -install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS) - # # Supported python versions. These versions will be searched in order, the # first match will be selected. These should be kept in sync with setup.py. @@ -215,6 +212,31 @@ endif() # Define extension targets # +# +# cumem_allocator extension +# + +set(VLLM_CUMEM_EXT_SRC + "csrc/cumem_allocator.cpp") + +set_gencode_flags_for_srcs( + SRCS "${VLLM_CUMEM_EXT_SRC}" + CUDA_ARCHS "${CUDA_ARCHS}") + +if(VLLM_GPU_LANG STREQUAL "CUDA") + message(STATUS "Enabling cumem allocator extension.") + # link against cuda driver library + list(APPEND CUMEM_LIBS cuda) + define_gpu_extension_target( + cumem_allocator + DESTINATION vllm + LANGUAGE CXX + SOURCES ${VLLM_CUMEM_EXT_SRC} + LIBRARIES ${CUMEM_LIBS} + USE_SABI 3.8 + WITH_SOABI) +endif() + # # _C extension # @@ -287,7 +309,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA") # Only build Marlin kernels if we are building for at least some compatible archs. # Keep building Marlin for 9.0 as there are some group sizes and shapes that # are not supported by Machete yet. - cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0" ${CUDA_ARCHS}) + cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}") if (MARLIN_ARCHS) set(MARLIN_SRCS "csrc/quantization/fp8/fp8_marlin.cu" @@ -308,8 +330,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA") endif() # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require - # CUDA 12.0 or later (and only work on Hopper, 9.0/9.0a for now). - cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0;9.0a" "${CUDA_ARCHS}") + # CUDA 12.0 or later (and only work on Hopper, 9.0a for now). + cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0a" "${CUDA_ARCHS}") if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS) set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu") set_gencode_flags_for_srcs( @@ -363,7 +385,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA") # 2:4 Sparse Kernels # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor - # require CUDA 12.2 or later (and only work on Hopper, 9.0/9.0a for now). + # require CUDA 12.2 or later (and only work on Hopper, 9.0a for now). if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS) set(SRCS "csrc/sparse/cutlass/sparse_compressor_c3x.cu" "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu") @@ -463,6 +485,9 @@ if(VLLM_GPU_LANG STREQUAL "HIP") endif() message(STATUS "Enabling C extension.") +if(VLLM_GPU_LANG STREQUAL "CUDA") + list(APPEND VLLM_C_LIBS cuda) +endif() define_gpu_extension_target( _C DESTINATION vllm @@ -471,6 +496,7 @@ define_gpu_extension_target( COMPILE_FLAGS ${VLLM_GPU_FLAGS} ARCHITECTURES ${VLLM_GPU_ARCHES} INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR} + LIBRARIES ${VLLM_C_LIBS} USE_SABI 3 WITH_SOABI) @@ -570,7 +596,7 @@ if(VLLM_GPU_LANG STREQUAL "HIP") endif() # vllm-flash-attn currently only supported on CUDA -if (NOT VLLM_TARGET_DEVICE STREQUAL "cuda") +if (NOT VLLM_GPU_LANG STREQUAL "CUDA") return() endif () @@ -593,7 +619,7 @@ endif() # They should be identical but if they aren't, this is a massive footgun. # # The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place. -# To only install vllm-flash-attn, use --component vllm_flash_attn_c. +# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3). # If no component is specified, vllm-flash-attn is still installed. # If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading. @@ -605,43 +631,41 @@ if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR}) endif() if(VLLM_FLASH_ATTN_SRC_DIR) - FetchContent_Declare(vllm-flash-attn SOURCE_DIR ${VLLM_FLASH_ATTN_SRC_DIR}) + FetchContent_Declare( + vllm-flash-attn SOURCE_DIR + ${VLLM_FLASH_ATTN_SRC_DIR} + BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn + ) else() FetchContent_Declare( vllm-flash-attn GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git - GIT_TAG 96266b1111111f3d11aabefaf3bacbab6a89d03c + GIT_TAG d4e09037abf588af1ec47d0e966b237ee376876c GIT_PROGRESS TRUE # Don't share the vllm-flash-attn build between build types BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn ) endif() -# Set the parent build flag so that the vllm-flash-attn library does not redo compile flag and arch initialization. -set(VLLM_PARENT_BUILD ON) - -# Ensure the vllm/vllm_flash_attn directory exists before installation -install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" COMPONENT vllm_flash_attn_c) - -# Make sure vllm-flash-attn install rules are nested under vllm/ -install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" COMPONENT vllm_flash_attn_c) -install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c) -install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" COMPONENT vllm_flash_attn_c) # Fetch the vllm-flash-attn library FetchContent_MakeAvailable(vllm-flash-attn) message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}") -# Restore the install prefix -install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c) -install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" COMPONENT vllm_flash_attn_c) +# Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in +# case only one is built, in the case both are built redundant work is done) +install( + DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/ + DESTINATION vllm_flash_attn + COMPONENT _vllm_fa2_C + FILES_MATCHING PATTERN "*.py" +) -# Copy over the vllm-flash-attn python files install( - DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/ - DESTINATION vllm/vllm_flash_attn - COMPONENT vllm_flash_attn_c - FILES_MATCHING PATTERN "*.py" + DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/ + DESTINATION vllm_flash_attn + COMPONENT _vllm_fa3_C + FILES_MATCHING PATTERN "*.py" ) # Nothing after vllm-flash-attn, see comment about macros above diff --git a/Dockerfile b/Dockerfile index 4542bc9cf0bd2..0b9f74e08dc68 100644 --- a/Dockerfile +++ b/Dockerfile @@ -52,7 +52,7 @@ WORKDIR /workspace # after this step RUN --mount=type=cache,target=/root/.cache/pip \ if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \ - python3 -m pip install --index-url https://download.pytorch.org/whl/nightly/cu124 "torch==2.6.0.dev20241210+cu124" "torchvision==0.22.0.dev20241215"; \ + python3 -m pip install --index-url https://download.pytorch.org/whl/nightly/cu126 "torch==2.7.0.dev20250121+cu126" "torchvision==0.22.0.dev20250121"; \ fi COPY requirements-common.txt requirements-common.txt @@ -126,8 +126,8 @@ RUN --mount=type=cache,target=/root/.cache/ccache \ # Check the size of the wheel if RUN_WHEEL_CHECK is true COPY .buildkite/check-wheel-size.py check-wheel-size.py -# Default max size of the wheel is 250MB -ARG VLLM_MAX_SIZE_MB=250 +# sync the default value with .buildkite/check-wheel-size.py +ARG VLLM_MAX_SIZE_MB=300 ENV VLLM_MAX_SIZE_MB=$VLLM_MAX_SIZE_MB ARG RUN_WHEEL_CHECK=true RUN if [ "$RUN_WHEEL_CHECK" = "true" ]; then \ @@ -149,7 +149,8 @@ RUN --mount=type=cache,target=/root/.cache/pip \ #################### vLLM installation IMAGE #################### # image with vLLM installed -FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04 AS vllm-base +# TODO: Restore to base image after FlashInfer AOT wheel fixed +FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS vllm-base ARG CUDA_VERSION=12.4.1 ARG PYTHON_VERSION=3.12 WORKDIR /vllm-workspace @@ -194,12 +195,30 @@ RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist --mount=type=cache,target=/root/.cache/pip \ python3 -m pip install dist/*.whl --verbose +# How to build this FlashInfer wheel: +# $ export FLASHINFER_ENABLE_AOT=1 +# $ # Note we remove 7.0 from the arch list compared to the list below, since FlashInfer only supports sm75+ +# $ export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.6 8.9 9.0+PTX' +# $ git clone https://github.com/flashinfer-ai/flashinfer.git --recursive +# $ cd flashinfer +# $ git checkout 524304395bd1d8cd7d07db083859523fcaa246a4 +# $ python3 setup.py bdist_wheel --dist-dir=dist --verbose + RUN --mount=type=cache,target=/root/.cache/pip \ . /etc/environment && \ if [ "$TARGETPLATFORM" != "linux/arm64" ]; then \ - python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.6/flashinfer-0.1.6+cu121torch2.4-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl; \ + python3 -m pip install https://wheels.vllm.ai/flashinfer/524304395bd1d8cd7d07db083859523fcaa246a4/flashinfer_python-0.2.0.post1-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl; \ fi COPY examples examples + +# Although we build Flashinfer with AOT mode, there's still +# some issues w.r.t. JIT compilation. Therefore we need to +# install build dependencies for JIT compilation. +# TODO: Remove this once FlashInfer AOT wheel is fixed +COPY requirements-build.txt requirements-build.txt +RUN --mount=type=cache,target=/root/.cache/pip \ + python3 -m pip install -r requirements-build.txt + #################### vLLM installation IMAGE #################### #################### TEST IMAGE #################### diff --git a/Dockerfile.base b/Dockerfile.base index c064284897372..e33e73b303098 100644 --- a/Dockerfile.base +++ b/Dockerfile.base @@ -1,11 +1,12 @@ -ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:6.3-complete +ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:6.3.1-complete ARG HIPBLASLT_BRANCH="4d40e36" +ARG HIPBLAS_COMMON_BRANCH="7c1566b" ARG LEGACY_HIPBLASLT_OPTION= ARG RCCL_BRANCH="648a58d" ARG RCCL_REPO="https://github.com/ROCm/rccl" ARG TRITON_BRANCH="e5be006" ARG TRITON_REPO="https://github.com/triton-lang/triton.git" -ARG PYTORCH_BRANCH="8d4926e" +ARG PYTORCH_BRANCH="3a585126" ARG PYTORCH_VISION_BRANCH="v0.19.1" ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git" ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git" @@ -43,15 +44,24 @@ RUN pip install -U packaging cmake ninja wheel setuptools pybind11 Cython FROM base AS build_hipblaslt ARG HIPBLASLT_BRANCH +ARG HIPBLAS_COMMON_BRANCH # Set to "--legacy_hipblas_direct" for ROCm<=6.2 ARG LEGACY_HIPBLASLT_OPTION +RUN git clone https://github.com/ROCm/hipBLAS-common.git +RUN cd hipBLAS-common \ + && git checkout ${HIPBLAS_COMMON_BRANCH} \ + && mkdir build \ + && cd build \ + && cmake .. \ + && make package \ + && dpkg -i ./*.deb RUN git clone https://github.com/ROCm/hipBLASLt RUN cd hipBLASLt \ && git checkout ${HIPBLASLT_BRANCH} \ && ./install.sh -d --architecture ${PYTORCH_ROCM_ARCH} ${LEGACY_HIPBLASLT_OPTION} \ && cd build/release \ && make package -RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/install +RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/hipBLAS-common/build/*.deb /app/install FROM base AS build_rccl ARG RCCL_BRANCH @@ -133,6 +143,7 @@ ARG PYTORCH_VISION_REPO ARG FA_BRANCH ARG FA_REPO RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /app/versions.txt \ + && echo "HIPBLAS_COMMON_BRANCH: ${HIPBLAS_COMMON_BRANCH}" >> /app/versions.txt \ && echo "HIPBLASLT_BRANCH: ${HIPBLASLT_BRANCH}" >> /app/versions.txt \ && echo "LEGACY_HIPBLASLT_OPTION: ${LEGACY_HIPBLASLT_OPTION}" >> /app/versions.txt \ && echo "RCCL_BRANCH: ${RCCL_BRANCH}" >> /app/versions.txt \ @@ -144,4 +155,4 @@ RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /app/versions.txt \ && echo "PYTORCH_REPO: ${PYTORCH_REPO}" >> /app/versions.txt \ && echo "PYTORCH_VISION_REPO: ${PYTORCH_VISION_REPO}" >> /app/versions.txt \ && echo "FA_BRANCH: ${FA_BRANCH}" >> /app/versions.txt \ - && echo "FA_REPO: ${FA_REPO}" >> /app/versions.txt \ No newline at end of file + && echo "FA_REPO: ${FA_REPO}" >> /app/versions.txt diff --git a/Dockerfile.base_navi b/Dockerfile.base_navi index 8c571ae219794..389933840cd38 100644 --- a/Dockerfile.base_navi +++ b/Dockerfile.base_navi @@ -1,5 +1,6 @@ -ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:6.3-complete +ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:6.3.1-complete ARG HIPBLASLT_BRANCH="4d40e36" +ARG HIPBLAS_COMMON_BRANCH="7c1566b" ARG LEGACY_HIPBLASLT_OPTION= ARG RCCL_BRANCH="648a58d" ARG RCCL_REPO="https://github.com/ROCm/rccl" @@ -43,13 +44,21 @@ FROM base AS build_hipblaslt ARG HIPBLASLT_BRANCH # Set to "--legacy_hipblas_direct" for ROCm<=6.2 ARG LEGACY_HIPBLASLT_OPTION +RUN git clone https://github.com/ROCm/hipBLAS-common.git +RUN cd hipBLAS-common \ + && git checkout ${HIPBLAS_COMMON_BRANCH} \ + && mkdir build \ + && cd build \ + && cmake .. \ + && make package \ + && dpkg -i ./*.deb RUN git clone https://github.com/ROCm/hipBLASLt RUN cd hipBLASLt \ && git checkout ${HIPBLASLT_BRANCH} \ && ./install.sh -d --architecture ${PYTORCH_ROCM_ARCH} ${LEGACY_HIPBLASLT_OPTION} \ && cd build/release \ && make package -RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/install +RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/hipBLAS-common/build/*.deb /app/install FROM base AS build_rccl ARG RCCL_BRANCH @@ -121,6 +130,7 @@ ARG PYTORCH_VISION_BRANCH ARG PYTORCH_REPO ARG PYTORCH_VISION_REPO RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /app/versions.txt \ + && echo "HIPBLAS_COMMON_BRANCH: ${HIPBLAS_COMMON_BRANCH}" >> /app/versions.txt \ && echo "HIPBLASLT_BRANCH: ${HIPBLASLT_BRANCH}" >> /app/versions.txt \ && echo "LEGACY_HIPBLASLT_OPTION: ${LEGACY_HIPBLASLT_OPTION}" >> /app/versions.txt \ && echo "RCCL_BRANCH: ${RCCL_BRANCH}" >> /app/versions.txt \ diff --git a/Dockerfile.hpu b/Dockerfile.hpu index 87e0c1a6a934e..66cf68c32f2ca 100644 --- a/Dockerfile.hpu +++ b/Dockerfile.hpu @@ -1,4 +1,4 @@ -FROM vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest +FROM vault.habana.ai/gaudi-docker/1.19.1/ubuntu22.04/habanalabs/pytorch-installer-2.5.1:latest COPY ./ /workspace/vllm diff --git a/Dockerfile.rocm b/Dockerfile.rocm index 13b0a764cbd7e..009e929ebace1 100644 --- a/Dockerfile.rocm +++ b/Dockerfile.rocm @@ -14,7 +14,7 @@ ENV PYTORCH_ROCM_ARCH=${ARG_PYTORCH_ROCM_ARCH:-${PYTORCH_ROCM_ARCH}} RUN apt-get update -q -y && apt-get install -q -y \ sqlite3 libsqlite3-dev libfmt-dev libmsgpack-dev libsuitesparse-dev # Remove sccache -RUN python3 -m pip install --upgrade pip +RUN python3 -m pip install --upgrade pip && pip install setuptools_scm RUN apt-get purge -y sccache; python3 -m pip uninstall -y sccache; rm -f "$(which sccache)" ARG COMMON_WORKDIR WORKDIR ${COMMON_WORKDIR} @@ -45,13 +45,35 @@ RUN cd vllm \ FROM scratch AS export_vllm ARG COMMON_WORKDIR COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/dist/*.whl / -COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/rocm_patch /rocm_patch COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/requirements*.txt / COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/benchmarks /benchmarks COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/tests /tests COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/examples /examples COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/.buildkite /.buildkite +# ----------------------- +# Test vLLM image +FROM base AS test + +RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/* + +# Install vLLM +RUN --mount=type=bind,from=export_vllm,src=/,target=/install \ + cd /install \ + && pip install -U -r requirements-rocm.txt \ + && pip uninstall -y vllm \ + && pip install *.whl + +WORKDIR /vllm-workspace +ARG COMMON_WORKDIR +COPY --from=build_vllm ${COMMON_WORKDIR}/vllm /vllm-workspace + +# install development dependencies (for testing) +RUN cd /vllm-workspace \ + && rm -rf vllm \ + && python3 -m pip install -e tests/vllm_test_utils \ + && python3 -m pip install lm-eval[api]==0.4.4 \ + && python3 -m pip install pytest-shard # ----------------------- # Final vLLM image @@ -75,17 +97,9 @@ RUN if [ ${BUILD_RPD} -eq "1" ]; then \ && cd hipMarker && python3 setup.py install ; fi # Install vLLM -# Make sure punica kernels are built (for LoRA) -ENV VLLM_INSTALL_PUNICA_KERNELS=1 RUN --mount=type=bind,from=export_vllm,src=/,target=/install \ cd /install \ && pip install -U -r requirements-rocm.txt \ - && case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \ - *"rocm-6.0"*) \ - patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h rocm_patch/rocm_bf16.patch;; \ - *"rocm-6.1"*) \ - cp rocm_patch/libamdhip64.so.6 /opt/rocm/lib/libamdhip64.so.6;; \ - *) ;; esac \ && pip uninstall -y vllm \ && pip install *.whl @@ -93,10 +107,7 @@ ARG COMMON_WORKDIR # Copy over the benchmark scripts as well COPY --from=export_vllm /benchmarks ${COMMON_WORKDIR}/vllm/benchmarks -COPY --from=export_vllm /tests ${COMMON_WORKDIR}/vllm/tests COPY --from=export_vllm /examples ${COMMON_WORKDIR}/vllm/examples -COPY --from=export_vllm /.buildkite ${COMMON_WORKDIR}/vllm/.buildkite - ENV RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1 ENV TOKENIZERS_PARALLELISM=false @@ -104,9 +115,5 @@ ENV TOKENIZERS_PARALLELISM=false # Performance environment variable. ENV HIP_FORCE_DEV_KERNARG=1 -# install development dependencies (for testing) -RUN cd ${COMMON_WORKDIR}/vllm \ - && python3 -m pip install -e tests/vllm_test_utils - CMD ["/bin/bash"] diff --git a/Dockerfile.rocm_base b/Dockerfile.rocm_base new file mode 100644 index 0000000000000..5bbe98b0c2204 --- /dev/null +++ b/Dockerfile.rocm_base @@ -0,0 +1,158 @@ +ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:6.3.1-complete +ARG HIPBLASLT_BRANCH="4d40e36" +ARG HIPBLAS_COMMON_BRANCH="7c1566b" +ARG LEGACY_HIPBLASLT_OPTION= +ARG RCCL_BRANCH="648a58d" +ARG RCCL_REPO="https://github.com/ROCm/rccl" +ARG TRITON_BRANCH="e5be006" +ARG TRITON_REPO="https://github.com/triton-lang/triton.git" +ARG PYTORCH_BRANCH="8d4926e" +ARG PYTORCH_VISION_BRANCH="v0.19.1" +ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git" +ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git" +ARG FA_BRANCH="b7d29fb" +ARG FA_REPO="https://github.com/ROCm/flash-attention.git" + +FROM ${BASE_IMAGE} AS base + +ENV PATH=/opt/rocm/llvm/bin:$PATH +ENV ROCM_PATH=/opt/rocm +ENV LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib: +ARG PYTORCH_ROCM_ARCH=gfx90a;gfx942 +ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} + +ARG PYTHON_VERSION=3.12 + +RUN mkdir -p /app +WORKDIR /app +ENV DEBIAN_FRONTEND=noninteractive + +# Install Python and other dependencies +RUN apt-get update -y \ + && apt-get install -y software-properties-common git curl sudo vim less \ + && add-apt-repository ppa:deadsnakes/ppa \ + && apt-get update -y \ + && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \ + python${PYTHON_VERSION}-lib2to3 python-is-python3 \ + && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \ + && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \ + && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \ + && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \ + && python3 --version && python3 -m pip --version + +RUN pip install -U packaging cmake ninja wheel setuptools pybind11 Cython + +FROM base AS build_hipblaslt +ARG HIPBLASLT_BRANCH +ARG HIPBLAS_COMMON_BRANCH +# Set to "--legacy_hipblas_direct" for ROCm<=6.2 +ARG LEGACY_HIPBLASLT_OPTION +RUN git clone https://github.com/ROCm/hipBLAS-common.git +RUN cd hipBLAS-common \ + && git checkout ${HIPBLAS_COMMON_BRANCH} \ + && mkdir build \ + && cd build \ + && cmake .. \ + && make package \ + && dpkg -i ./*.deb +RUN git clone https://github.com/ROCm/hipBLASLt +RUN cd hipBLASLt \ + && git checkout ${HIPBLASLT_BRANCH} \ + && ./install.sh -d --architecture ${PYTORCH_ROCM_ARCH} ${LEGACY_HIPBLASLT_OPTION} \ + && cd build/release \ + && make package +RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/hipBLAS-common/build/*.deb /app/install + +FROM base AS build_rccl +ARG RCCL_BRANCH +ARG RCCL_REPO +RUN git clone ${RCCL_REPO} +RUN cd rccl \ + && git checkout ${RCCL_BRANCH} \ + && ./install.sh -p --amdgpu_targets ${PYTORCH_ROCM_ARCH} +RUN mkdir -p /app/install && cp /app/rccl/build/release/*.deb /app/install + +FROM base AS build_triton +ARG TRITON_BRANCH +ARG TRITON_REPO +RUN git clone ${TRITON_REPO} +RUN cd triton \ + && git checkout ${TRITON_BRANCH} \ + && cd python \ + && python3 setup.py bdist_wheel --dist-dir=dist +RUN mkdir -p /app/install && cp /app/triton/python/dist/*.whl /app/install + +FROM base AS build_amdsmi +RUN cd /opt/rocm/share/amd_smi \ + && pip wheel . --wheel-dir=dist +RUN mkdir -p /app/install && cp /opt/rocm/share/amd_smi/dist/*.whl /app/install + +FROM base AS build_pytorch +ARG PYTORCH_BRANCH +ARG PYTORCH_VISION_BRANCH +ARG PYTORCH_REPO +ARG PYTORCH_VISION_REPO +ARG FA_BRANCH +ARG FA_REPO +RUN git clone ${PYTORCH_REPO} pytorch +RUN cd pytorch && git checkout ${PYTORCH_BRANCH} && \ + pip install -r requirements.txt && git submodule update --init --recursive \ + && python3 tools/amd_build/build_amd.py \ + && CMAKE_PREFIX_PATH=$(python3 -c 'import sys; print(sys.prefix)') python3 setup.py bdist_wheel --dist-dir=dist \ + && pip install dist/*.whl +RUN git clone ${PYTORCH_VISION_REPO} vision +RUN cd vision && git checkout ${PYTORCH_VISION_BRANCH} \ + && python3 setup.py bdist_wheel --dist-dir=dist \ + && pip install dist/*.whl +RUN git clone ${FA_REPO} +RUN cd flash-attention \ + && git checkout ${FA_BRANCH} \ + && git submodule update --init \ + && MAX_JOBS=64 GPU_ARCHS=${PYTORCH_ROCM_ARCH} python3 setup.py bdist_wheel --dist-dir=dist +RUN mkdir -p /app/install && cp /app/pytorch/dist/*.whl /app/install \ + && cp /app/vision/dist/*.whl /app/install \ + && cp /app/flash-attention/dist/*.whl /app/install + +FROM base AS final +RUN --mount=type=bind,from=build_hipblaslt,src=/app/install/,target=/install \ + dpkg -i /install/*deb \ + && sed -i 's/, hipblaslt-dev \(.*\), hipcub-dev/, hipcub-dev/g' /var/lib/dpkg/status \ + && sed -i 's/, hipblaslt \(.*\), hipfft/, hipfft/g' /var/lib/dpkg/status +RUN --mount=type=bind,from=build_rccl,src=/app/install/,target=/install \ + dpkg -i /install/*deb \ + && sed -i 's/, rccl-dev \(.*\), rocalution/, rocalution/g' /var/lib/dpkg/status \ + && sed -i 's/, rccl \(.*\), rocalution/, rocalution/g' /var/lib/dpkg/status +RUN --mount=type=bind,from=build_triton,src=/app/install/,target=/install \ + pip install /install/*.whl +RUN --mount=type=bind,from=build_amdsmi,src=/app/install/,target=/install \ + pip install /install/*.whl +RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \ + pip install /install/*.whl + +ARG BASE_IMAGE +ARG HIPBLASLT_BRANCH +ARG LEGACY_HIPBLASLT_OPTION +ARG RCCL_BRANCH +ARG RCCL_REPO +ARG TRITON_BRANCH +ARG TRITON_REPO +ARG PYTORCH_BRANCH +ARG PYTORCH_VISION_BRANCH +ARG PYTORCH_REPO +ARG PYTORCH_VISION_REPO +ARG FA_BRANCH +ARG FA_REPO +RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /app/versions.txt \ + && echo "HIPBLAS_COMMON_BRANCH: ${HIPBLAS_COMMON_BRANCH}" >> /app/versions.txt \ + && echo "HIPBLASLT_BRANCH: ${HIPBLASLT_BRANCH}" >> /app/versions.txt \ + && echo "LEGACY_HIPBLASLT_OPTION: ${LEGACY_HIPBLASLT_OPTION}" >> /app/versions.txt \ + && echo "RCCL_BRANCH: ${RCCL_BRANCH}" >> /app/versions.txt \ + && echo "RCCL_REPO: ${RCCL_REPO}" >> /app/versions.txt \ + && echo "TRITON_BRANCH: ${TRITON_BRANCH}" >> /app/versions.txt \ + && echo "TRITON_REPO: ${TRITON_REPO}" >> /app/versions.txt \ + && echo "PYTORCH_BRANCH: ${PYTORCH_BRANCH}" >> /app/versions.txt \ + && echo "PYTORCH_VISION_BRANCH: ${PYTORCH_VISION_BRANCH}" >> /app/versions.txt \ + && echo "PYTORCH_REPO: ${PYTORCH_REPO}" >> /app/versions.txt \ + && echo "PYTORCH_VISION_REPO: ${PYTORCH_VISION_REPO}" >> /app/versions.txt \ + && echo "FA_BRANCH: ${FA_BRANCH}" >> /app/versions.txt \ + && echo "FA_REPO: ${FA_REPO}" >> /app/versions.txt diff --git a/Dockerfile.tpu b/Dockerfile.tpu index b617932a85b47..e268b39476665 100644 --- a/Dockerfile.tpu +++ b/Dockerfile.tpu @@ -1,4 +1,4 @@ -ARG NIGHTLY_DATE="20241017" +ARG NIGHTLY_DATE="20250124" ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE" FROM $BASE_IMAGE diff --git a/README.md b/README.md index 67c557bfe13a9..34fab8547d56f 100644 --- a/README.md +++ b/README.md @@ -20,6 +20,7 @@ The first vLLM meetup in 2025 is happening on January 22nd, Wednesday, with Goog --- *Latest News* 🔥 +- [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing). - [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone! - [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing). - [2024/10] We have just created a developer slack ([slack.vllm.ai](https://slack.vllm.ai)) focusing on coordinating contributions and discussing features. Please feel free to join us there! @@ -38,6 +39,8 @@ The first vLLM meetup in 2025 is happening on January 22nd, Wednesday, with Goog ## About vLLM is a fast and easy-to-use library for LLM inference and serving. +Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evloved into a community-driven project with contributions from both academia and industry. + vLLM is fast with: - State-of-the-art serving throughput @@ -72,16 +75,16 @@ Find the full list of supported models [here](https://docs.vllm.ai/en/latest/mod ## Getting Started -Install vLLM with `pip` or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source): +Install vLLM with `pip` or [from source](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/index.html#build-wheel-from-source): ```bash pip install vllm ``` -Visit our [documentation](https://vllm.readthedocs.io/en/latest/) to learn more. -- [Installation](https://vllm.readthedocs.io/en/latest/getting_started/installation.html) -- [Quickstart](https://vllm.readthedocs.io/en/latest/getting_started/quickstart.html) -- [List of Supported Models](https://vllm.readthedocs.io/en/latest/models/supported_models.html) +Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more. +- [Installation](https://docs.vllm.ai/en/latest/getting_started/installation/index.html) +- [Quickstart](https://docs.vllm.ai/en/latest/getting_started/quickstart.html) +- [List of Supported Models](https://docs.vllm.ai/en/latest/models/supported_models.html) ## Contributing diff --git a/SECURITY.md b/SECURITY.md index de0032d26c87b..47196a1f1221e 100644 --- a/SECURITY.md +++ b/SECURITY.md @@ -4,7 +4,7 @@ If you believe you have found a security vulnerability in vLLM, we encourage you to let us know right away. We will investigate all legitimate reports and do our best to quickly fix the problem. -Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new). Reports will then be triaged by the [vulnerability management team](https://docs.vllm.ai/contributing/vulnerability_management/). +Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new). Reports will then be triaged by the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html). --- diff --git a/benchmarks/backend_request_func.py b/benchmarks/backend_request_func.py index 9d71e4ecc4a37..0612e8778aca5 100644 --- a/benchmarks/backend_request_func.py +++ b/benchmarks/backend_request_func.py @@ -22,6 +22,7 @@ class RequestFuncInput: prompt_len: int output_len: int model: str + model_name: Optional[str] = None best_of: int = 1 logprobs: Optional[int] = None extra_body: Optional[dict] = None @@ -34,6 +35,7 @@ class RequestFuncOutput: generated_text: str = "" success: bool = False latency: float = 0.0 + output_tokens: int = 0 ttft: float = 0.0 # Time to first token itl: List[float] = field( default_factory=list) # List of inter-token latencies @@ -49,7 +51,8 @@ async def async_request_tgi( api_url = request_func_input.api_url assert api_url.endswith("generate_stream") - async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session: + async with aiohttp.ClientSession(trust_env=True, + timeout=AIOHTTP_TIMEOUT) as session: params = { "best_of": request_func_input.best_of, "max_new_tokens": request_func_input.output_len, @@ -78,7 +81,7 @@ async def async_request_tgi( continue chunk_bytes = chunk_bytes.decode("utf-8") - #NOTE: Sometimes TGI returns a ping response without + # NOTE: Sometimes TGI returns a ping response without # any data, we should skip it. if chunk_bytes.startswith(":"): continue @@ -121,7 +124,8 @@ async def async_request_trt_llm( api_url = request_func_input.api_url assert api_url.endswith("generate_stream") - async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session: + async with aiohttp.ClientSession(trust_env=True, + timeout=AIOHTTP_TIMEOUT) as session: assert request_func_input.best_of == 1 payload = { "accumulate_tokens": True, @@ -155,7 +159,7 @@ async def async_request_trt_llm( timestamp = time.perf_counter() # First token if ttft == 0.0: - ttft = time.perf_counter() - st + ttft = timestamp - st output.ttft = ttft # Decoding phase @@ -185,7 +189,8 @@ async def async_request_deepspeed_mii( request_func_input: RequestFuncInput, pbar: Optional[tqdm] = None, ) -> RequestFuncOutput: - async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session: + async with aiohttp.ClientSession(trust_env=True, + timeout=AIOHTTP_TIMEOUT) as session: assert request_func_input.best_of == 1 payload = { @@ -233,17 +238,23 @@ async def async_request_openai_completions( ("completions", "profile") ), "OpenAI Completions API URL must end with 'completions' or 'profile'." - async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session: + async with aiohttp.ClientSession(trust_env=True, + timeout=AIOHTTP_TIMEOUT) as session: payload = { - "model": request_func_input.model, + "model": request_func_input.model_name \ + if request_func_input.model_name else request_func_input.model, "prompt": request_func_input.prompt, "temperature": 0.0, "best_of": request_func_input.best_of, "max_tokens": request_func_input.output_len, "logprobs": request_func_input.logprobs, "stream": True, - "ignore_eos": request_func_input.ignore_eos, + "stream_options": { + "include_usage": True, + }, } + if request_func_input.ignore_eos: + payload["ignore_eos"] = request_func_input.ignore_eos if request_func_input.extra_body: payload.update(request_func_input.extra_body) headers = { @@ -254,7 +265,6 @@ async def async_request_openai_completions( output.prompt_len = request_func_input.prompt_len generated_text = "" - ttft = 0.0 st = time.perf_counter() most_recent_timestamp = st try: @@ -269,15 +279,16 @@ async def async_request_openai_completions( chunk = chunk_bytes.decode("utf-8").removeprefix( "data: ") - if chunk == "[DONE]": - latency = time.perf_counter() - st - else: + if chunk != "[DONE]": data = json.loads(chunk) # NOTE: Some completion API might have a last # usage summary response without a token so we # want to check a token was generated - if data["choices"][0]["text"]: + if choices := data.get("choices"): + # Note that text could be empty here + # e.g. for special tokens + text = choices[0].get("text") timestamp = time.perf_counter() # First token if not first_chunk_received: @@ -291,7 +302,10 @@ async def async_request_openai_completions( most_recent_timestamp) most_recent_timestamp = timestamp - generated_text += data["choices"][0]["text"] + generated_text += text or "" + elif usage := data.get("usage"): + output.output_tokens = usage.get( + "completion_tokens") if first_chunk_received: output.success = True else: @@ -300,7 +314,7 @@ async def async_request_openai_completions( "Never received a valid chunk to calculate TTFT." "This response will be marked as failed!") output.generated_text = generated_text - output.latency = latency + output.latency = most_recent_timestamp - st else: output.error = response.reason or "" output.success = False @@ -323,12 +337,14 @@ async def async_request_openai_chat_completions( "chat/completions" ), "OpenAI Chat Completions API URL must end with 'chat/completions'." - async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session: + async with aiohttp.ClientSession(trust_env=True, + timeout=AIOHTTP_TIMEOUT) as session: content = [{"type": "text", "text": request_func_input.prompt}] if request_func_input.multi_modal_content: content.append(request_func_input.multi_modal_content) payload = { - "model": request_func_input.model, + "model": request_func_input.model_name \ + if request_func_input.model_name else request_func_input.model, "messages": [ { "role": "user", @@ -338,8 +354,12 @@ async def async_request_openai_chat_completions( "temperature": 0.0, "max_completion_tokens": request_func_input.output_len, "stream": True, - "ignore_eos": request_func_input.ignore_eos, + "stream_options": { + "include_usage": True, + }, } + if request_func_input.ignore_eos: + payload["ignore_eos"] = request_func_input.ignore_eos if request_func_input.extra_body: payload.update(request_func_input.extra_body) headers = { @@ -365,17 +385,15 @@ async def async_request_openai_chat_completions( chunk = chunk_bytes.decode("utf-8").removeprefix( "data: ") - if chunk == "[DONE]": - latency = time.perf_counter() - st - else: + if chunk != "[DONE]": timestamp = time.perf_counter() data = json.loads(chunk) - delta = data["choices"][0]["delta"] - if delta.get("content", None): + if choices := data.get("choices"): + content = choices[0]["delta"].get("content") # First token if ttft == 0.0: - ttft = time.perf_counter() - st + ttft = timestamp - st output.ttft = ttft # Decoding phase @@ -383,13 +401,16 @@ async def async_request_openai_chat_completions( output.itl.append(timestamp - most_recent_timestamp) - generated_text += delta["content"] + generated_text += content or "" + elif usage := data.get("usage"): + output.output_tokens = usage.get( + "completion_tokens") most_recent_timestamp = timestamp output.generated_text = generated_text output.success = True - output.latency = latency + output.latency = most_recent_timestamp - st else: output.error = response.reason or "" output.success = False diff --git a/benchmarks/benchmark_serving.py b/benchmarks/benchmark_serving.py index 4eb0e1f8ac903..63d2c3f7c7dd9 100644 --- a/benchmarks/benchmark_serving.py +++ b/benchmarks/benchmark_serving.py @@ -25,6 +25,7 @@ import argparse import asyncio import base64 +import gc import io import json import os @@ -199,7 +200,7 @@ def sample_sonnet_requests( return sampled_requests -def sample_mmmu_pro_vision_requests( +def sample_vision_arena_requests( dataset, num_requests: int, tokenizer: PreTrainedTokenizerBase, @@ -211,13 +212,7 @@ def sample_mmmu_pro_vision_requests( if len(sampled_requests) == num_requests: break - # MMMU-Pro vision direct prompt - # Ref: https://github.com/MMMU-Benchmark/MMMU/blob/6ce42f4d8f70c1841c67867152648974415b5cac/mmmu-pro/prompts.yaml#L5 - prompt = ( - "Answer with the option letter from the given choices directly. " - "The last line of your response should be of the following " - "format: 'Answer: $LETTER' (without quotes) where LETTER is one of " - "options.") + prompt = data["turns"][0][0]['content'] prompt_token_ids = tokenizer(prompt).input_ids if fixed_output_len is None: @@ -229,10 +224,10 @@ def sample_mmmu_pro_vision_requests( output_len = fixed_output_len assert isinstance( - data["image"], + data["images"][0], Image), ("Input image format must be `PIL.Image.Image`, " f"given {type(data['image'])}.") - image: Image = data["image"] + image: Image = data["images"][0] image = image.convert("RGB") image_data = io.BytesIO() image.save(image_data, format='JPEG') @@ -251,7 +246,7 @@ def sample_mmmu_pro_vision_requests( def sample_hf_requests( dataset_path: str, - dataset_subset: str, + dataset_subset: Optional[str], dataset_split: str, num_requests: int, tokenizer: PreTrainedTokenizerBase, @@ -259,19 +254,17 @@ def sample_hf_requests( fixed_output_len: Optional[int] = None, ) -> List[Tuple[str, str, int, Optional[Dict[str, Collection[str]]]]]: - # Special case for MMMU-Pro vision dataset - if dataset_path == 'MMMU/MMMU_Pro' and dataset_subset == 'vision': - assert dataset_split == "test" + # Special case for vision_arena dataset + if dataset_path == 'lmarena-ai/vision-arena-bench-v0.1' \ + and dataset_subset is None: + assert dataset_split == "train" dataset = load_dataset(dataset_path, name=dataset_subset, split=dataset_split, streaming=True) - assert "image" in dataset.features, ( - "MMMU/MMMU_Pro vision dataset must have 'image' column.") - filter_func = lambda x: isinstance(x["image"], Image) - dataset = dataset.shuffle(seed=random_seed).filter(filter_func) - return sample_mmmu_pro_vision_requests(dataset, num_requests, - tokenizer, fixed_output_len) + dataset = dataset.shuffle(seed=random_seed) + return sample_vision_arena_requests(dataset, num_requests, tokenizer, + fixed_output_len) dataset = load_dataset(dataset_path, name=dataset_subset, @@ -423,7 +416,7 @@ def calculate_metrics( tokenizer: PreTrainedTokenizerBase, selected_percentile_metrics: List[str], selected_percentiles: List[float], - gootput_config_dict: Dict[str, float], + goodput_config_dict: Dict[str, float], ) -> Tuple[BenchmarkMetrics, List[int]]: actual_output_lens: List[int] = [] total_input = 0 @@ -436,19 +429,23 @@ def calculate_metrics( e2els: List[float] = [] for i in range(len(outputs)): if outputs[i].success: - # We use the tokenizer to count the number of output tokens for all - # serving backends instead of looking at len(outputs[i].itl) since - # multiple output tokens may be bundled together - # Note : this may inflate the output token count slightly - output_len = len( - tokenizer(outputs[i].generated_text, - add_special_tokens=False).input_ids) + output_len = outputs[i].output_tokens + + if output_len is None: + # We use the tokenizer to count the number of output tokens + # for some serving backends instead of looking at + # len(outputs[i].itl) since multiple output tokens may be + # bundled together + # Note : this may inflate the output token count slightly + output_len = len( + tokenizer(outputs[i].generated_text, + add_special_tokens=False).input_ids) actual_output_lens.append(output_len) total_input += input_requests[i][1] tpot = 0 if output_len > 1: - tpot = (outputs[i].latency - outputs[i].ttft) / (output_len - - 1) + latency_minus_ttft = outputs[i].latency - outputs[i].ttft + tpot = latency_minus_ttft / (output_len - 1) tpots.append(tpot) # Note: if output_len <= 1, we regard tpot as 0 for goodput all_tpots.append(tpot) @@ -459,21 +456,21 @@ def calculate_metrics( else: actual_output_lens.append(0) - if gootput_config_dict: + if goodput_config_dict: valid_metrics = [] slo_values = [] - if "ttft" in gootput_config_dict: + if "ttft" in goodput_config_dict: valid_metrics.append(ttfts) - slo_values.append(gootput_config_dict["ttft"] / + slo_values.append(goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION) - if "tpot" in gootput_config_dict: + if "tpot" in goodput_config_dict: valid_metrics.append(all_tpots) - slo_values.append(gootput_config_dict["tpot"] / + slo_values.append(goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION) - if "e2el" in gootput_config_dict: + if "e2el" in goodput_config_dict: valid_metrics.append(e2els) - slo_values.append(gootput_config_dict["e2el"] / + slo_values.append(goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION) for req_metric in zip(*valid_metrics): @@ -525,6 +522,7 @@ async def benchmark( api_url: str, base_url: str, model_id: str, + model_name: str, tokenizer: PreTrainedTokenizerBase, input_requests: List[Tuple[str, int, int]], logprobs: Optional[int], @@ -536,7 +534,7 @@ async def benchmark( selected_percentile_metrics: List[str], selected_percentiles: List[str], ignore_eos: bool, - gootput_config_dict: Dict[str, float], + goodput_config_dict: Dict[str, float], max_concurrency: Optional[int], ): if backend in ASYNC_REQUEST_FUNCS: @@ -553,6 +551,7 @@ async def benchmark( "Multi-modal content is only supported on 'openai-chat' backend.") test_input = RequestFuncInput( model=model_id, + model_name=model_name, prompt=test_prompt, api_url=api_url, prompt_len=test_prompt_len, @@ -573,6 +572,7 @@ async def benchmark( if profile: print("Starting profiler...") profile_input = RequestFuncInput(model=model_id, + model_name=model_name, prompt=test_prompt, api_url=base_url + "/start_profile", prompt_len=test_prompt_len, @@ -616,6 +616,7 @@ async def limited_request_func(request_func_input, pbar): async for request in get_request(input_requests, request_rate, burstiness): prompt, prompt_len, output_len, mm_content = request request_func_input = RequestFuncInput(model=model_id, + model_name=model_name, prompt=prompt, api_url=api_url, prompt_len=prompt_len, @@ -657,7 +658,7 @@ async def limited_request_func(request_func_input, pbar): tokenizer=tokenizer, selected_percentile_metrics=selected_percentile_metrics, selected_percentiles=selected_percentiles, - gootput_config_dict=gootput_config_dict, + goodput_config_dict=goodput_config_dict, ) print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='=')) @@ -669,7 +670,7 @@ async def limited_request_func(request_func_input, pbar): metrics.total_output)) print("{:<40} {:<10.2f}".format("Request throughput (req/s):", metrics.request_throughput)) - if gootput_config_dict: + if goodput_config_dict: print("{:<40} {:<10.2f}".format("Request goodput (req/s):", metrics.request_goodput)) print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):", @@ -684,7 +685,7 @@ async def limited_request_func(request_func_input, pbar): "total_output_tokens": metrics.total_output, "request_throughput": metrics.request_throughput, "request_goodput:": - metrics.request_goodput if gootput_config_dict else None, + metrics.request_goodput if goodput_config_dict else None, "output_throughput": metrics.output_throughput, "total_token_throughput": metrics.total_token_throughput, "input_lens": [output.prompt_len for output in outputs], @@ -740,11 +741,11 @@ def process_one_metric( def check_goodput_args(args): # Check and parse goodput arguments - gootput_config_dict = {} + goodput_config_dict = {} VALID_NAMES = ["ttft", "tpot", "e2el"] if args.goodput: - gootput_config_dict = parse_goodput(args.goodput) - for slo_name, slo_val in gootput_config_dict.items(): + goodput_config_dict = parse_goodput(args.goodput) + for slo_name, slo_val in goodput_config_dict.items(): if slo_name not in VALID_NAMES: raise ValueError( f"Invalid metric name found, {slo_name}: {slo_val}. " @@ -755,22 +756,22 @@ def check_goodput_args(args): f"Invalid value found, {slo_name}: {slo_val}. " "The service level objective value should be " "non-negative.") - return gootput_config_dict + return goodput_config_dict def parse_goodput(slo_pairs): - gootput_config_dict = {} + goodput_config_dict = {} try: for slo_pair in slo_pairs: slo_name, slo_val = slo_pair.split(":") - gootput_config_dict[slo_name] = float(slo_val) + goodput_config_dict[slo_name] = float(slo_val) except ValueError as err: raise argparse.ArgumentTypeError( "Invalid format found for service level objectives. " "Specify service level objectives for goodput as \"KEY:VALUE\" " "pairs, where the key is a metric name, and the value is a " "number in milliseconds.") from err - return gootput_config_dict + return goodput_config_dict def main(args: argparse.Namespace): @@ -780,6 +781,7 @@ def main(args: argparse.Namespace): backend = args.backend model_id = args.model + model_name = args.served_model_name tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model tokenizer_mode = args.tokenizer_mode @@ -869,7 +871,11 @@ def main(args: argparse.Namespace): else: raise ValueError(f"Unknown dataset: {args.dataset_name}") - gootput_config_dict = check_goodput_args(args) + goodput_config_dict = check_goodput_args(args) + + # Avoid GC processing "static" data - reduce pause times. + gc.collect() + gc.freeze() benchmark_result = asyncio.run( benchmark( @@ -877,6 +883,7 @@ def main(args: argparse.Namespace): api_url=api_url, base_url=base_url, model_id=model_id, + model_name=model_name, tokenizer=tokenizer, input_requests=input_requests, logprobs=args.logprobs, @@ -890,7 +897,7 @@ def main(args: argparse.Namespace): float(p) for p in args.metric_percentiles.split(",") ], ignore_eos=args.ignore_eos, - gootput_config_dict=gootput_config_dict, + goodput_config_dict=goodput_config_dict, max_concurrency=args.max_concurrency, )) @@ -1222,5 +1229,12 @@ def main(args: argparse.Namespace): 'always use the slow tokenizer. \n* ' '"mistral" will always use the `mistral_common` tokenizer.') + parser.add_argument("--served-model-name", + type=str, + default=None, + help="The model name used in the API. " + "If not specified, the model name will be the " + "same as the ``--model`` argument. ") + args = parser.parse_args() main(args) diff --git a/benchmarks/kernels/benchmark_lora.py b/benchmarks/kernels/benchmark_lora.py new file mode 100644 index 0000000000000..e1f613e1da509 --- /dev/null +++ b/benchmarks/kernels/benchmark_lora.py @@ -0,0 +1,1147 @@ +import argparse +import copy +import json +import pickle +import time +from dataclasses import dataclass +from enum import Enum, auto +from itertools import product +from pathlib import Path +from typing import Any, Callable, Dict, List, Optional, Tuple + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from utils import ArgPool, Bench, CudaGraphBenchParams +from weight_shapes import WEIGHT_SHAPES + +from vllm.lora.ops.triton_ops.bgmv_expand import bgmv_expand +from vllm.lora.ops.triton_ops.bgmv_expand_slice import bgmv_expand_slice +from vllm.lora.ops.triton_ops.bgmv_shrink import bgmv_shrink +from vllm.lora.ops.triton_ops.sgmv_expand import sgmv_expand +from vllm.lora.ops.triton_ops.sgmv_shrink import sgmv_shrink +from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) +DEFAULT_TP_SIZES = [1] +DEFAULT_BATCH_SIZES = [ + 1, 16, 32, 64, 128, 192, 256, 320, 384, 448, 512, 640, 768, 896, 1024, + 2048, 3072, 4096, 5120, 6144, 7168, 8192 +] +DEFAULT_HIDDEN_SIZES = [1024, 2048, 4096, 8192, 16384] +DEFAULT_LORA_RANKS = [16] +DEFAULT_NUM_LORAS = [1, 2, 3, 4] +DEFAULT_SORT_BY_LORA_IDS = [False, True] +DEFAULT_SEQ_LENGTHS = [1] +DEFAULT_EXPAND_FN_ADD_INPUTS = [True, False] + + +# Utilities +def dtype_to_str(dtype: torch.dtype): + if dtype == torch.float16: + return "f16" + if dtype == torch.bfloat16: + return "bf16" + if dtype == torch.float32: + return "f32" + raise ValueError(f"Unsupported dtype {dtype}") + + +def make_rand_lora_weight_tensor(k: int, + n: int, + num_loras: int, + dtype: torch.dtype, + device: str = "cuda") -> torch.Tensor: + + # LoRA weights column major + return torch.rand((num_loras, n, k), dtype=dtype).to(device) + + +def make_rand_tensors( + a_shape: Tuple[int], + b_shape: Tuple[int], + c_shape: Tuple[int], + a_dtype: torch.dtype, + b_dtype: torch.dtype, + c_dtype: torch.dtype, + num_slices: int, + device: str = "cuda", +) -> Tuple[torch.Tensor, List[torch.Tensor], torch.Tensor]: + """ + Make LoRA input/output matrices. + """ + A = torch.rand(a_shape, dtype=a_dtype).to(device) + + # LoRA weights column major + Bs = [ + torch.rand(b_shape, dtype=b_dtype).to(device) + for _ in range(num_slices) + ] + + C = torch.zeros(c_shape, dtype=c_dtype).to(device) + return A, Bs, C + + +def make_prompt_lora_mapping(num_prompts: int, num_active_loras: int, + sort_by_lora_id: bool, + device: str) -> torch.Tensor: + """ + All prompts are mapped to a Lora ID in range [0, num_active_loras). + where 0 refers to first lora, 1 refers to second lora and so on. + """ + assert num_active_loras > 0 + + if not sort_by_lora_id: + return torch.randint(0, + num_active_loras, (num_prompts, ), + dtype=torch.long) + + # Divide LoRAs equally and in order. + part_size = num_prompts // num_active_loras + part_size = max(part_size, 1) + + lora_id = 0 + prompt_lora_mapping = [] + while len(prompt_lora_mapping) < num_prompts: + prompt_lora_mapping.extend([lora_id] * part_size) + lora_id = lora_id + 1 if lora_id + 1 < num_active_loras else lora_id + return torch.tensor(prompt_lora_mapping[:num_prompts], + dtype=torch.long, + device=device) + + +def make_token_lora_mapping(num_tokens: int, num_prompts: int, + prompt_lora_mapping: torch.Tensor, + seq_len_tensor: torch.Tensor, device: str): + """ + Make token_lora_mapping from prompt_lora_mapping and seq_lens_tensor + """ + assert prompt_lora_mapping.shape[0] == num_prompts + + # token to lora index mapping + token_lora_mapping = [0] * num_tokens + current_offset = 0 + for b_id in range(num_prompts): + lora_index = prompt_lora_mapping[b_id].item() + s = current_offset + e = s + seq_len_tensor[b_id].item() + token_lora_mapping[s:e] = [lora_index] * (e - s) + current_offset += seq_len_tensor[b_id].item() + + return torch.tensor(token_lora_mapping, dtype=torch.long, device=device) + + +def ref_group_gemm(ref_out: torch.Tensor, input: torch.Tensor, + lora_weights: List[torch.Tensor], + seq_lens_cpu: torch.Tensor, + prompt_lora_mapping_cpu: torch.Tensor, scaling: float, + add_inputs: Optional[bool]): + """ + Torch group gemm reference implementation to test correctness of + benchmarking operations. + """ + batches = seq_lens_cpu.size(0) + out_list = [] + current_offset = 0 + for lora_index, b_length in zip(range(batches), seq_lens_cpu): + x = input[current_offset:b_length + current_offset, :] + current_offset += b_length + w = lora_weights[prompt_lora_mapping_cpu[lora_index]] + result = torch.nn.functional.linear(x, w) + result *= scaling + out_list.append(result) + torch.cat(out_list, dim=0) + + cat_result = torch.cat(out_list, dim=0) + + if add_inputs: + ref_out += cat_result + else: + ref_out.copy_(cat_result) + + +class OpType(Enum): + """ + LoRA Ops to benchmark and its properties. + """ + SGMV_SHRINK = auto() + BGMV_SHRINK = auto() + SGMV_EXPAND = auto() + BGMV_EXPAND = auto() + BGMV_EXPAND_SLICE = auto() + + @staticmethod + def from_str(s: str) -> "OpType": + if s.lower() == 'sgmv_shrink': + return OpType.SGMV_SHRINK + if s.lower() == 'sgmv_expand': + return OpType.SGMV_EXPAND + if s.lower() == 'bgmv_shrink': + return OpType.BGMV_SHRINK + if s.lower() == 'bgmv_expand': + return OpType.BGMV_EXPAND + if s.lower() == "bgmv_expand_slice": + return OpType.BGMV_EXPAND_SLICE + raise ValueError(f"Unrecognized str {s} to convert to OpType") + + def is_shrink_fn(self) -> bool: + return self in [OpType.SGMV_SHRINK, OpType.BGMV_SHRINK] + + def is_expand_fn(self) -> bool: + return self in [OpType.SGMV_EXPAND, OpType.BGMV_EXPAND] + + def is_prefill_op(self) -> bool: + return self in [OpType.SGMV_SHRINK, OpType.SGMV_EXPAND] + + def is_decode_op(self) -> bool: + return self in [ + OpType.BGMV_SHRINK, OpType.BGMV_EXPAND, OpType.BGMV_EXPAND_SLICE + ] + + def is_expand_slice_fn(self) -> bool: + return self in [OpType.BGMV_EXPAND_SLICE] + + def num_slices(self) -> List[int]: + if self in [OpType.SGMV_EXPAND, OpType.SGMV_SHRINK]: + # SGMV kernels supports slices + return [1, 2, 3] + if self in [OpType.BGMV_SHRINK, OpType.BGMV_EXPAND]: + return [1] + if self in [OpType.BGMV_EXPAND_SLICE]: + return [2, 3] + raise ValueError(f"Unrecognized OpType {self}") + + def mkn(self, batch_size: int, seq_length: int, hidden_size: int, + lora_rank: int) -> Tuple[int, int, int]: + num_tokens = batch_size * seq_length + if self.is_shrink_fn(): + m = num_tokens + k = hidden_size + n = lora_rank + else: + assert self.is_expand_fn() or self.is_expand_slice_fn() + m = num_tokens + k = lora_rank + n = hidden_size + return m, k, n + + def matmul_dtypes( + self, op_dtype: torch.dtype + ) -> Tuple[torch.dtype, torch.dtype, torch.dtype]: + """ + return a type, b type and c type for A x B = C + """ + if self.is_shrink_fn(): + return op_dtype, op_dtype, torch.float32 + else: + assert self.is_expand_fn() or self.is_expand_slice_fn() + return torch.float32, op_dtype, op_dtype + + def matmul_shapes( + self, batch_size: int, seq_length: int, hidden_size: int, + lora_rank: int, num_loras: int, + num_slices: int) -> Tuple[Tuple[int], Tuple[int], Tuple[int]]: + """ + Given num_slices, return the shapes of the A, B, and C matrices + in A x B = C, for the op_type + """ + m, k, n = self.mkn(batch_size, seq_length, hidden_size, lora_rank) + + b_shape = (num_loras, n, k) # col-major + if self == OpType.SGMV_SHRINK: + # SGMV shrink supports num_slices inherently in the kernel + return ((m, k), b_shape, (num_slices, m, n)) + if self == OpType.SGMV_EXPAND: + # SGMV expand supports num_slices inherently in the kernel + return ((num_slices, m, k), b_shape, (m, n * num_slices)) + if self == OpType.BGMV_SHRINK: + return ((m, k), b_shape, (m, n)) + if self == OpType.BGMV_EXPAND: + return ((m, k), b_shape, (m, n)) + if self == OpType.BGMV_EXPAND_SLICE: + return ((num_slices, m, k), b_shape, (m, n * num_slices)) + + raise ValueError(f"Unrecognized op_type {self}") + + def bench_fn(self) -> Callable: + + def emulate_bgmv_expand_slice(kwargs_list: List[Dict[str, Any]]): + for x in kwargs_list: + bgmv_expand_slice(**x) + + if self == OpType.SGMV_SHRINK: + return sgmv_shrink + if self == OpType.SGMV_EXPAND: + return sgmv_expand + if self == OpType.BGMV_SHRINK: + return bgmv_shrink + if self == OpType.BGMV_EXPAND: + return bgmv_expand + if self == OpType.BGMV_EXPAND_SLICE: + return emulate_bgmv_expand_slice + raise ValueError(f"Unrecognized optype {self}") + + def run_ref_group_gemm(self, output: torch.Tensor, input: torch.Tensor, + lora_weights: List[torch.Tensor], + **kwargs) -> Callable: + """Each benchmark operation expected the input, lora_weights and outputs + in a slightly different format. Refer to self.matmul_shapes(). + run_ref_group_gemm accounts for those differences in executing a + reference group gemm for correctness testing. + """ + w_dtype = lora_weights[0].dtype + num_slices = len(lora_weights) + if self == OpType.SGMV_SHRINK: + for slice_idx in range(num_slices): + ref_group_gemm(ref_out=output[slice_idx, :], + input=input, + lora_weights=lora_weights[slice_idx], + **kwargs) + if self == OpType.SGMV_EXPAND: + hidden_size = lora_weights[0].shape[1] + for slice_idx in range(num_slices): + slice_offset = slice_idx * hidden_size + ref_group_gemm( + ref_out=output[:, slice_offset:slice_offset + hidden_size], + input=input[slice_idx].clone().to(dtype=w_dtype), + lora_weights=lora_weights[slice_idx], + **kwargs) + if self == OpType.BGMV_SHRINK: + assert num_slices == 1 + ref_group_gemm(ref_out=output, + input=input, + lora_weights=lora_weights[0], + **kwargs) + if self == OpType.BGMV_EXPAND: + assert num_slices == 1 + ref_group_gemm(ref_out=output, + input=input.clone().to(dtype=w_dtype), + lora_weights=lora_weights[0], + **kwargs) + if self == OpType.BGMV_EXPAND_SLICE: + hidden_size = lora_weights[0].shape[1] + for slice_idx in range(num_slices): + slice_offset = slice_idx * hidden_size + ref_group_gemm( + ref_out=output[:, slice_offset:slice_offset + hidden_size], + input=input[slice_idx].clone().to(dtype=w_dtype), + lora_weights=lora_weights[slice_idx], + **kwargs) + raise ValueError(f"Unrecognized optype {self}") + + +@dataclass +class BenchmarkContext: + """ + LoRA benchmark context + """ + batch_size: int + hidden_size: int + num_loras: int + num_active_loras: int + lora_rank: int + sort_by_lora_id: bool + dtype: torch.dtype + seq_length: Optional[int] = None + num_slices: Optional[int] = None # num_slices for slice based ops + + def with_seq_length(self, seq_length: int) -> "BenchmarkContext": + ctx = copy.copy(self) + ctx.seq_length = seq_length + return ctx + + def with_num_slices(self, num_slices: int) -> "BenchmarkContext": + ctx = copy.copy(self) + ctx.num_slices = num_slices + return ctx + + def bench_label(self) -> str: + return f"lora-{self.dtype}" + + def bench_sublabel(self, op_type: OpType) -> str: + m, k, n = op_type.mkn(self.batch_size, self.seq_length, + self.hidden_size, self.lora_rank) + desc = { + 'bs': self.batch_size, + 'sl': self.seq_length, + 'm': m, + 'k': k, + 'n': n, + 'num_loras': self.num_loras, + 'sort_by_lora': self.sort_by_lora_id, + 'num_slices': self.num_slices, + } + return json.dumps(desc) + + +@dataclass +class BenchmarkTensors: + """ + Input/Output tensors used for benchmarks + """ + # matmul tensors + input: torch.Tensor + lora_weights_lst: List[torch.Tensor] + output: torch.Tensor + # metadata tensors + seq_lens: torch.Tensor + seq_start_loc: torch.Tensor + prompt_lora_mapping: torch.Tensor + token_lora_mapping: torch.Tensor + + def io_types(self) -> str: + return (f"{dtype_to_str(self.input.dtype)}x" + f"{dtype_to_str(self.lora_weights_lst[0].dtype)}=>" + f"{dtype_to_str(self.output.dtype)}") + + @staticmethod + def make(ctx: BenchmarkContext, + op_type: OpType, + device: str = "cuda") -> "BenchmarkTensors": + + # Make input / output matmul tensors. + a_shape, b_shape, c_shape = op_type.matmul_shapes( + ctx.batch_size, ctx.seq_length, ctx.hidden_size, ctx.lora_rank, + ctx.num_loras, ctx.num_slices) + a_type, b_type, c_type = op_type.matmul_dtypes(ctx.dtype) + input_tensor, lora_weights, output_tensor = \ + make_rand_tensors(a_shape, b_shape, c_shape, a_type, b_type, c_type, + num_slices = ctx.num_slices) + + # Make metadata tensors. + # Keep the metadata tensors in the CPU for further processing if needed. + # The tensors get moved to the GPU before benchmarking. + assert ctx.num_active_loras <= ctx.num_loras + total_tokens = ctx.batch_size * ctx.seq_length + + # Prepare seq lens tensor + seq_len_tensor = torch.randint(ctx.seq_length, ctx.seq_length + 1, + (ctx.batch_size, )) + # Prepare seq_start_loc tensor + seq_start_loc_tensor = torch.cumsum(torch.tensor( + [0] + seq_len_tensor[:-1].tolist(), dtype=torch.long), + dim=0) + assert total_tokens == seq_len_tensor.sum() + # Prepare prompt lora indices tensor + prompt_lora_indices_tensor = make_prompt_lora_mapping( + ctx.batch_size, ctx.num_active_loras, ctx.sort_by_lora_id, "cpu") + # Prepare token lora indices tensor + token_lora_indices_tensor = make_token_lora_mapping( + total_tokens, ctx.batch_size, prompt_lora_indices_tensor, + seq_len_tensor, "cpu") + + return BenchmarkTensors(input_tensor, lora_weights, output_tensor, + seq_len_tensor, seq_start_loc_tensor, + prompt_lora_indices_tensor, + token_lora_indices_tensor) + + def sanity_check(self) -> None: + """ + Fails asserts when non-conformality is detected. + """ + num_tokens = self.input.shape[-2] + # check metadata tensors + assert torch.sum(self.seq_lens) == num_tokens + num_seqs = self.seq_lens.shape[0] + assert self.seq_start_loc.shape[0] == num_seqs + assert self.prompt_lora_mapping.shape[0] == num_seqs + assert self.token_lora_mapping.shape[0] == num_tokens + + def to_device(self, device: str): + """ + Transfer tensors to device if the tensors aren't already on the device + """ + + def to_device(tensor: torch.Tensor): + if tensor.device != device: + tensor = tensor.to(device=device) + return tensor + + self.input = to_device(self.input) + self.output = to_device(self.output) + self.seq_lens = to_device(self.seq_lens) + self.seq_start_loc = to_device(self.seq_start_loc) + self.prompt_lora_mapping = to_device(self.prompt_lora_mapping) + self.token_lora_mapping = to_device(self.token_lora_mapping) + for i in range(len(self.lora_weights_lst)): + self.lora_weights_lst[i] = to_device(self.lora_weights_lst[i]) + + def metadata(self) -> Tuple[int, int, int]: + """ + Return num_seqs, num_tokens and max_seq_len + """ + num_seqs = self.seq_lens.shape[0] + num_tokens = self.token_lora_mapping.shape[0] + max_seq_len = torch.max(self.seq_lens).item() + num_slices = len(self.lora_weights_lst) + return num_seqs, num_tokens, max_seq_len, num_slices + + def convert_to_sgmv_benchmark_tensors(self): + """ + For sgmv punica kernels, when consecutive sequences have the + same LoRA ID, we just merge them together. + This happens in punica.py::compute_metadata + """ + + # Collapse seq_lens and seq_start_loc + _, seq_lens = torch.unique_consecutive(self.token_lora_mapping, + return_counts=True) + cum_result = torch.cumsum(seq_lens, dim=0) + seq_start_loc = torch.zeros_like(seq_lens) + seq_start_loc[1:].copy_(cum_result[:-1]) + + # Collapse prompt mapping + prompt_lora_mapping = torch.unique_consecutive( + self.prompt_lora_mapping) + + assert torch.sum(seq_lens) == torch.sum(self.seq_lens), \ + f"dont match - new {torch.sum(seq_lens)} vs {torch.sum(self.seq_lens)}" + + self.prompt_lora_mapping = prompt_lora_mapping.to( + dtype=self.prompt_lora_mapping.dtype) + self.seq_lens = seq_lens.to(dtype=self.seq_lens.dtype) + self.seq_start_loc = seq_start_loc.to(dtype=self.seq_start_loc.dtype) + + def as_sgmv_shrink_kwargs(self) -> Dict[str, Any]: + self.convert_to_sgmv_benchmark_tensors() + self.sanity_check() + self.to_device(self.input.device) + + num_seqs, num_tokens, max_seq_len, num_slices = self.metadata() + + # Sanity check matrix shapes. + i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[ + 0].shape, self.output.shape + # Expected input shape [num_tokens, hidden_size] + assert len(i_shape) == 2 + assert i_shape[0] == num_tokens + hidden_size = i_shape[1] + # Expected lora weight shape [num_loras, lora_rank, hidden_size] + assert len(lw_shape) == 3 + assert lw_shape[2] == hidden_size + lora_rank = lw_shape[1] + # Expected output shape [num_slices, num_tokens, lora_rank] + assert len(o_shape) == 3 + assert o_shape == (num_slices, num_tokens, lora_rank) + + return { + 'inputs': self.input, + 'lora_a_weights': self.lora_weights_lst, + 'output_tensor': self.output, + 'b_seq_start_loc': self.seq_start_loc, + 'seq_len_tensor': self.seq_lens, + 'lora_indices_tensor': self.prompt_lora_mapping, + 'batches': num_seqs, + 'max_seq_length': max_seq_len, + 'token_nums': num_tokens, + 'scaling': 1.0, + } + + def as_sgmv_expand_kwargs(self, add_inputs: bool) -> Dict[str, Any]: + + self.convert_to_sgmv_benchmark_tensors() + self.sanity_check() + self.to_device(self.input.device) + + num_seqs, num_tokens, max_seq_len, num_slices = self.metadata() + + # Sanity check matrix shapes. + i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[ + 0].shape, self.output.shape + # Expected input shape : [num_slices, num_tokens, lora_rank] + assert len(i_shape) == 3 + assert i_shape[0] == num_slices + assert i_shape[1] == num_tokens + lora_rank = i_shape[2] + # Expected lora weight shape : [num_lora, hidden_size, lora_rank] + assert len(lw_shape) == 3 + assert lw_shape[2] == lora_rank + hidden_size = lw_shape[1] + # Expected output shape : [num_tokens, hidden_size * num_slices] + assert len(o_shape) == 2 + assert o_shape == (num_tokens, hidden_size * num_slices) + + return { + 'inputs': self.input, + 'lora_b_weights': self.lora_weights_lst, + 'output_tensor': self.output, + 'b_seq_start_loc': self.seq_start_loc, + 'seq_len_tensor': self.seq_lens, + 'lora_indices_tensor': self.prompt_lora_mapping, + 'batches': num_seqs, + 'max_seq_length': max_seq_len, + 'token_nums': num_tokens, + 'offset_start': 0, + 'add_inputs': add_inputs, + } + + def as_bgmv_shrink_kwargs(self) -> Dict[str, Any]: + assert len(self.lora_weights_lst) == 1 + self.to_device(self.input.device) + + _, num_tokens, _, _ = self.metadata() + # Sanity check shapes + i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[ + 0].shape, self.output.shape + # Expected input shape [num_tokens, hidden_size] + assert len(i_shape) == 2 + assert i_shape[0] == num_tokens + hidden_size = i_shape[1] + # Expected lora weight shape [num_loras, lora_rank, hidden_size] + assert len(lw_shape) == 3 + assert lw_shape[2] == hidden_size + lora_rank = lw_shape[1] + # Expected output shape [num_tokens, lora_rank] + assert len(o_shape) == 2 + assert o_shape == (num_tokens, lora_rank) + + return { + 'inputs': self.input, + 'lora_a_weights': self.lora_weights_lst[0], + 'output_tensor': self.output, + 'lora_indices_tensor': self.token_lora_mapping, + 'scaling': 1.0 + } + + def as_bgmv_expand_kwargs(self, add_inputs: bool): + assert len(self.lora_weights_lst) == 1 + self.to_device(self.input.device) + + _, num_tokens, _, _ = self.metadata() + # Sanity check shapes + i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[ + 0].shape, self.output.shape + # Expected input shape [num_tokens, lora_rank] + assert len(i_shape) == 2 + assert i_shape[0] == num_tokens + lora_rank = i_shape[1] + # Expected lora weight shape [num_loras, hidden_size, lora_rank] + assert len(lw_shape) == 3 + assert lw_shape[2] == lora_rank + hidden_size = lw_shape[1] + # Expected output shape [num_tokens, hidden_size] + assert len(o_shape) == 2 + assert o_shape == (num_tokens, hidden_size) + + return { + 'inputs': self.input, + 'lora_b_weights': self.lora_weights_lst[0], + 'output_tensor': self.output, + 'lora_indices_tensor': self.token_lora_mapping, + 'add_inputs': add_inputs + } + + def as_bgmv_expand_slice_kwargs(self, add_inputs: bool) -> Dict[str, Any]: + + _, num_tokens, _, num_slices = self.metadata() + # Sanity check shapes + i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[ + 0].shape, self.output.shape + # Expected input shape [num_slices, num_tokens, lora_rank] + assert len(i_shape) == 3 + assert i_shape[0] == num_slices + assert i_shape[1] == num_tokens + lora_rank = i_shape[2] + # Expected lora weight shape [num_loras, hidden_size, lora_rank] + assert len(lw_shape) == 3 + assert lw_shape[2] == lora_rank + hidden_size = lw_shape[1] + # Expected output shape [num_tokens, hidden_size * num_slices] + assert len(o_shape) == 2 + assert o_shape == (num_tokens, hidden_size * num_slices) + + self.to_device(self.input.device) + + kwargs_list = [] + for i in range(num_slices): + kwargs_list.append({ + 'inputs': self.input[i], + 'lora_b_weights': self.lora_weights_lst[i], + 'output_tensor': self.output, + 'lora_indices_tensor': self.token_lora_mapping, + 'slice_offset': i * hidden_size, + 'slice_size': hidden_size, + 'add_inputs': add_inputs, + }) + return {'kwargs_list': kwargs_list} + + def bench_fn_kwargs(self, + op_type: OpType, + add_inputs: Optional[bool] = None) -> Dict[str, Any]: + if op_type.is_shrink_fn(): + assert add_inputs is None + else: + assert add_inputs is not None + + if op_type == OpType.SGMV_SHRINK: + return self.as_sgmv_shrink_kwargs() + if op_type == OpType.SGMV_EXPAND: + return self.as_sgmv_expand_kwargs(add_inputs) + if op_type == OpType.BGMV_SHRINK: + return self.as_bgmv_shrink_kwargs() + if op_type == OpType.BGMV_EXPAND: + return self.as_bgmv_expand_kwargs(add_inputs) + if op_type == OpType.BGMV_EXPAND_SLICE: + return self.as_bgmv_expand_slice_kwargs(add_inputs) + raise ValueError(f"Unrecognized optype {self}") + + def test_correctness(self, op_type: OpType, + expand_fn_add_inputs: Optional[bool]) -> bool: + """ + Test correctness of op_type implementation against a grouped gemm + reference implementation. + """ + seq_lens_cpu = self.seq_lens.to(device="cpu") + prompt_lora_mapping_cpu = self.prompt_lora_mapping.to(device="cpu") + ref_output = self.output.clone() + + self.output.zero_() + op_type.bench_fn()( + **self.bench_fn_kwargs(op_type, expand_fn_add_inputs)) + + op_type.run_ref_group_gemm( + ref_output, + self.input, + self.lora_weights_lst, + seq_lens_cpu=seq_lens_cpu, + prompt_lora_mapping_cpu=prompt_lora_mapping_cpu, + scaling=1.0, + add_inputs=expand_fn_add_inputs) + + rtol, atol = { + torch.float16: (6e-2, 6e-2), + torch.bfloat16: (6e-2, 6e-2), + torch.float32: (1e-2, 1e-2), + }[self.output.dtype] + + return torch.allclose(ref_output, self.output, rtol=rtol, atol=atol) + + +def bench_optype(ctx: BenchmarkContext, + arg_pool_size: int, + op_type: OpType, + cuda_graph_nops: Optional[int] = None, + expand_fn_add_inputs: Optional[bool] = None, + test_correctness: bool = False) -> TMeasurement: + + assert arg_pool_size >= 1 + if op_type.is_shrink_fn(): + assert expand_fn_add_inputs is None + else: + assert expand_fn_add_inputs is not None + + # BenchmarkContext -> BenchmarkTensors + bench_tensors : List[BenchmarkTensors] = \ + [BenchmarkTensors.make(ctx, op_type) for _ in range(arg_pool_size)] + for bt in bench_tensors: + bt.sanity_check() + + # Test correctness of our implementation. + if test_correctness: + assert all([ + bt.test_correctness(op_type, expand_fn_add_inputs) + for bt in bench_tensors + ]) + + # BenchmarkTensors -> Dict (kwargs) + kwargs_list = [ + bt.bench_fn_kwargs(op_type, add_inputs=expand_fn_add_inputs) + for bt in bench_tensors + ] + + # Clear LoRA optimization hash-maps. + _LORA_A_PTR_DICT.clear() + _LORA_B_PTR_DICT.clear() + # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are setup + for kwargs in kwargs_list: + op_type.bench_fn()(**kwargs) + torch.cuda.synchronize() + + # Merge into a single kwargs and qualify arguments as ArgPool + kwargs = {k: ArgPool([]) for k in kwargs_list[0]} + for _kwargs in kwargs_list: + for k, v in _kwargs.items(): + kwargs[k].values.append(v) + + describe_args = (f"add_inputs={expand_fn_add_inputs}" + if expand_fn_add_inputs is not None else "") + description = ( + f"{op_type.name}({describe_args}) ({bench_tensors[0].io_types()})") + + cuda_graph_params = None + if cuda_graph_nops: + cuda_graph_params = CudaGraphBenchParams(cuda_graph_nops) + timer = None + with Bench(cuda_graph_params, + ctx.bench_label(), ctx.bench_sublabel(op_type), description, + op_type.bench_fn(), **kwargs) as bench: + timer = bench.run() + return timer + + +def bench_torch_mm(ctx: BenchmarkContext, + arg_pool_size: int, + op_type: OpType, + cuda_graph_nops: Optional[int] = None) -> TMeasurement: + """ + Benchmark basic torch.mm as a roofline. + + When all the input tokens have the same LoRA ID, the LoRA kernels are just + a matmul. This torch.mm benchmark serves as a roofline for that case. + + input op_type is used in determining the m, k, n dimensions for the matmul. + """ + + batch_size, hidden_size, lora_rank, seq_length, dtype = (ctx.batch_size, + ctx.hidden_size, + ctx.lora_rank, + ctx.seq_length, + ctx.dtype) + + m, k, n = op_type.mkn(batch_size, seq_length, hidden_size, lora_rank) + # For a fairer comparison. + n = n * ctx.num_slices + + # Get matmul input and output tensors for A x B = C + As, Bs, Cs = [], [], [] + for _ in range(arg_pool_size): + As.append(torch.rand((m, k), dtype=dtype).to("cuda")) + Bs.append(torch.rand((n, k), dtype=dtype).to("cuda").t()) + Cs.append(torch.rand((m, n), dtype=dtype).to("cuda")) + + # Make torch.mm kwargs + mm_kwargs = {'input': ArgPool(As), 'mat2': ArgPool(Bs), 'out': ArgPool(Cs)} + + description = ( + f"single-lora roofline using torch.mm ({dtype_to_str(dtype)}" + f"x{dtype_to_str(dtype)}" + f"=>{dtype_to_str(dtype)})") + cuda_graph_params = None + if cuda_graph_nops: + cuda_graph_params = CudaGraphBenchParams(cuda_graph_nops) + with Bench(cuda_graph_params, ctx.bench_label(), + ctx.bench_sublabel(op_type), description, torch.mm, + **mm_kwargs) as bench: + return bench.run() + + +# runner +def use_cuda_graph_recommendation() -> str: + return """ + Triton kernels have a significant launch overhead with + launched directly via python. This overhead is more noticeable + for small the problem sizes. For these cases, it is recommended + to use the script with `--cuda-graph-nops N` to benchmark N + consecutive invocations of the benchmarking operations from + inside a CUDA Graph. Note that the returned measurement is for N + invocations of the operation. + """ + + +def print_timers(timers: List[TMeasurement], + args: Optional[argparse.Namespace] = None): + compare = TBenchmark.Compare(timers) + compare.print() + + if args and args.cuda_graph_nops: + print( + f"Note : The timings reported above is for {args.cuda_graph_nops} " + "consecutive invocations of the benchmarking functions. " + f"Please divide by {args.cuda_graph_nops} for single invocation " + "timings.") + + print("Note on Comparison with torch.mm : The torch.mm numbers are " + "benchmark numbers of a simple matmul emulating the single lora " + "case. It is provided as a roofline for comparing our LoRA Kernel " + "implementations. It is expected that the LoRA kernels will be " + "slower than torch.mm in cases where num_loras is big. But for " + "small num_loras the goal should be to match the torch.mm numbers.") + + +def run(args: argparse.Namespace, bench_ctxs: List[BenchmarkContext]): + + if args.cuda_graph_nops is not None: + assert args.cuda_graph_nops > 0 + print(f"Benchmarking {args.cuda_graph_nops} invocations inside a CUDA " + "Graph") + else: + print(f"CUDA Graphs not enabled.\n{use_cuda_graph_recommendation()}") + + timers = [] + for bench_ctx in bench_ctxs: + for seq_len in args.seq_lengths: + bench_ops: List[OpType] = [] + if seq_len == 1: + # bench all decode ops + bench_ops = [op for op in args.op_types if op.is_decode_op()] + else: + # bench all prefill ops + bench_ops = [op for op in args.op_types if op.is_prefill_op()] + + seq_len_timers = [] + for bench_op in bench_ops: + for num_slices in bench_op.num_slices(): + _ctx = bench_ctx.with_seq_length(seq_len).with_num_slices( + num_slices) + # Benchmark torch.mm as a roofline + seq_len_timers.append( + bench_torch_mm(_ctx, args.arg_pool_size, bench_op, + args.cuda_graph_nops)) + + # Benchmark bench_op + expand_fn_add_inputs = [ + None + ] if bench_op.is_shrink_fn() else args.expand_fn_add_inputs + for add_input_arg in expand_fn_add_inputs: + seq_len_timers.append( + bench_optype(_ctx, args.arg_pool_size, bench_op, + args.cuda_graph_nops, add_input_arg, + args.test_correctness)) + + print_timers(seq_len_timers) + timers.extend(seq_len_timers) + + # Result stdout dump + print("== All Results ====") + print_timers(timers, args) + + if args.output_directory: + # Result file dump + od = Path(args.output_directory) + if not od.exists(): + od.mkdir() + + timestamp = int(time.time()) + pkl_file = od / f"lora_bench-{timestamp}.pkl" + print(f"Writing benchmarks to {pkl_file}") + with open(pkl_file, "wb") as f: + pickle.dump(timers, f) + + +def as_benchmark_contexts(hidden_sizes: List[int], lora_ranks: List[int], + args: argparse.Namespace) -> List[BenchmarkContext]: + + ctxs: List[BenchmarkContext] = [] + for batch_size, hidden_size, lora_rank, num_loras, sort_by_lora_id in product( # noqa + args.batch_sizes, list(hidden_sizes), lora_ranks, args.num_loras, + args.sort_by_lora_id): + ctxs.append( + BenchmarkContext( + batch_size=batch_size, + hidden_size=hidden_size, + lora_rank=lora_rank, + num_loras=num_loras, + num_active_loras=args.num_active_loras + if args.num_active_loras else num_loras, + # To be filled based on the OpType to benchmark + seq_length=None, + sort_by_lora_id=sort_by_lora_id, + dtype=args.dtype, + # To be filled based on the OpType to benchmark + num_slices=None)) + + return ctxs + + +def run_list_bench(args: argparse.Namespace): + print(args) + + print("List bench :\n" + f" Hidden Sizes {args.hidden_sizes}" + f" LoRA Ranks {args.lora_ranks}") + + # Get all benchmarking contexts + bench_contexts: List[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=args.hidden_sizes, lora_ranks=args.lora_ranks, args=args) + + run(args, bench_contexts) + + +def run_range_bench(args: argparse.Namespace): + print(args) + + hidden_sizes = list( + range(args.hidden_sizes_start, args.hidden_sizes_end + 1, + args.hidden_sizes_increment)) + lora_ranks = list( + range(args.lora_ranks_start, args.lora_ranks_end + 1, + args.lora_ranks_increment)) + + print("Range bench :\n" + f" Hidden Sizes {hidden_sizes}" + f" LoRA Ranks {lora_ranks}") + + # Get all benchmarking contexts + bench_contexts: List[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=hidden_sizes, lora_ranks=lora_ranks, args=args) + + run(args, bench_contexts) + + +def run_model_bench(args: argparse.Namespace): + print(args) + + def hidden_sizes_from_model(model: str, tp_size: int) -> set[int]: + hidden_sizes = set() + for KN, tp_split_dim in WEIGHT_SHAPES[model]: + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + hidden_sizes.add(KN[1]) + return hidden_sizes + + # Get all hidden sizes + hidden_sizes: set[int] = set() + for model_name, tp_size in product(args.models, args.tp_sizes): + hidden_sizes = hidden_sizes.union( + hidden_sizes_from_model(model_name, tp_size)) + + print("Model bench :\n" + f" Hidden Sizes {hidden_sizes}" + f" LoRA Ranks {args.lora_ranks}") + + # Get all benchmarking contexts + bench_contexts: List[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=hidden_sizes, lora_ranks=args.lora_ranks, args=args) + + run(args, bench_contexts) + + +if __name__ == '__main__': + + def to_torch_dtype(dt): + if dt == "torch.float16": + return torch.float16 + if dt == "torch.bfloat16": + return torch.bfloat16 + raise ValueError("unsupported dtype") + + def get_bool(s: str) -> bool: + return s.lower() in ['true', '1'] + + def add_common_command_args(p: argparse.ArgumentParser): + p.add_argument( + "--dtype", + type=to_torch_dtype, + required=True, + help="Available options are ['torch.float16', 'torch.bfloat16']") + + p.add_argument( + "--arg-pool-size", + type=int, + default=32, + help="Run profiles with a pool of input/output/meta tensors instead" + "of simply reusing the same tensors for all runs. A bigger arg-pool" + "mitigates hardware caching effects during benchmarking.") + + p.add_argument( + "--cuda-graph-nops", + type=int, + help=("when set profiling is done using cudagraph, " + "with the given number of operations in a graph." + "Note that the measurement returned is the time " + "taken for N consecutive executions of the benchmarking " + "functions, where N is the value of this argument.")) + p.add_argument("--num-loras", + nargs="+", + type=int, + default=DEFAULT_NUM_LORAS) + p.add_argument("--num-active-loras", + type=int, + default=None, + help="Active LoRAs. When None, all LoRAs are active") + p.add_argument("--sort-by-lora-id", + nargs="+", + type=get_bool, + default=DEFAULT_SORT_BY_LORA_IDS) + p.add_argument("--op-types", + nargs="+", + type=OpType.from_str, + default=list(OpType)) + p.add_argument('--seq-lengths', + nargs="+", + type=int, + default=DEFAULT_SEQ_LENGTHS) + p.add_argument("--batch-sizes", + nargs="+", + type=int, + default=DEFAULT_BATCH_SIZES) + p.add_argument("--expand-fn-add-inputs", + nargs="+", + type=get_bool, + default=DEFAULT_EXPAND_FN_ADD_INPUTS) + p.add_argument( + '-o', + '--output-directory', + type=str, + help=("Output directory to store a the list of benchmarking" + "TMeasurement objects as a pickle file")) + + p.add_argument( + "--test-correctness", + action='store_true', + help=("When enabled, the benchmarking functions are tested" + "for correctness before the actual benchmarking")) + + parser = FlexibleArgumentParser( + description=f""" +Benchmark LoRA kernels: + {use_cuda_graph_recommendation()} + + list_bench example: + python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 + + model_bench example: + python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --lora-ranks 16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 + + range_bench example: + python3 benchmarks/kernels/benchmark_lora.py range_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter) + + subparsers = parser.add_subparsers(dest="cmd", required=True) + + list_parser = subparsers.add_parser("list_bench") + list_parser.add_argument("--hidden-sizes", + nargs="+", + type=int, + default=DEFAULT_HIDDEN_SIZES) + list_parser.add_argument("--lora-ranks", + nargs="+", + type=int, + default=DEFAULT_LORA_RANKS) + add_common_command_args(list_parser) + list_parser.set_defaults(func=run_list_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument("--hidden-sizes-start", type=int, required=True) + range_parser.add_argument("--hidden-sizes-end", type=int, required=True) + range_parser.add_argument("--hidden-sizes-increment", + type=int, + required=True) + range_parser.add_argument("--lora-ranks-start", type=int, required=True) + range_parser.add_argument("--lora-ranks-end", type=int, required=True) + range_parser.add_argument("--lora-ranks-increment", + type=int, + required=True) + add_common_command_args(range_parser) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument("--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys()) + model_parser.add_argument("--tp-sizes", + nargs="+", + type=int, + default=DEFAULT_TP_SIZES) + model_parser.add_argument("--lora-ranks", + nargs="+", + type=int, + default=DEFAULT_LORA_RANKS) + add_common_command_args(model_parser) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + args.func(args) diff --git a/benchmarks/kernels/benchmark_moe.py b/benchmarks/kernels/benchmark_moe.py index 8471541e14412..17658ba8f6482 100644 --- a/benchmarks/kernels/benchmark_moe.py +++ b/benchmarks/kernels/benchmark_moe.py @@ -12,10 +12,10 @@ from vllm.model_executor.layers.fused_moe.fused_moe import * from vllm.platforms import current_platform -from vllm.utils import FlexibleArgumentParser, is_navi +from vllm.utils import FlexibleArgumentParser FP8_DTYPE = torch.float8_e4m3fnuz if current_platform.is_rocm( -) and not is_navi() else torch.float8_e4m3fn +) else torch.float8_e4m3fn class BenchmarkConfig(TypedDict): @@ -41,12 +41,6 @@ def benchmark_config( ) -> float: init_dtype = torch.float16 if use_fp8_w8a8 else dtype x = torch.randn(num_tokens, hidden_size, dtype=dtype) - padding_size = 0 - if envs.VLLM_MOE_PADDING and not (use_fp8_w8a8 or use_int8_w8a16): - padding_size = 128 # fp16 padding size - if envs.VLLM_FP8_PADDING and use_fp8_w8a8: - padding_size = 256 # fp8 padding size. Ignoring int8 for now - if use_int8_w8a16: w1 = torch.randint(-127, 127, ( @@ -65,8 +59,8 @@ def benchmark_config( else: w1 = torch.randn(num_experts, shard_intermediate_size, - hidden_size + padding_size, - dtype=init_dtype)[..., :-padding_size] + hidden_size, + dtype=init_dtype) w2 = torch.randn(num_experts, hidden_size, shard_intermediate_size // 2, @@ -152,33 +146,29 @@ def run(): def get_rocm_tuning_space(use_fp16): - # small search space, no pruning required - # bypassLDS: block_n/num_warps=16 for perf - block_m_range = [16, 32, 64, 128, 256] - block_n_range = [128] if use_fp16 else [64] - block_k_range = [128] if use_fp16 else [256] - - num_warps_range = [8] if use_fp16 else [4] - group_m_range = [1] - # For now we see better perf with num_stages=0 for all gemm configs we care - # But keep this explicit so that we do not forget we may need to set it to - # other values in the future + block_mn_range = [16, 32, 64, 128, 256] + block_k_range = [16, 32, 64, 128, 256] + if not use_fp16: + block_k_range.remove(16) # BLOCK_K=16 not supported for fp8 + num_warps_range = [1, 2, 4, 8] + group_m_range = [1, 4, 8, 16, 32] num_stage_range = [2] waves_per_eu_range = [0] - matrix_instr_nonkdim_range = [16] - kpack_range = [2] + matrix_instr_nonkdim_range = [16, 32] if use_fp16 else [] + kpack_range = [1, 2] if use_fp16 else [] param_ranges = { - "BLOCK_SIZE_M": block_m_range, - "BLOCK_SIZE_N": block_n_range, + "BLOCK_SIZE_M": block_mn_range, + "BLOCK_SIZE_N": block_mn_range, "BLOCK_SIZE_K": block_k_range, "GROUP_SIZE_M": group_m_range, "num_warps": num_warps_range, "num_stages": num_stage_range, "waves_per_eu": waves_per_eu_range, - "matrix_instr_nonkdim": matrix_instr_nonkdim_range, - "kpack": kpack_range, } + if use_fp16: + param_ranges["matrix_instr_nonkdim"] = matrix_instr_nonkdim_range + param_ranges["kpack"] = kpack_range return param_ranges @@ -215,6 +205,112 @@ def get_configs_compute_bound(use_fp16) -> List[Dict[str, int]]: return configs +def prune_rocm_search_space(num_tokens, shard_intermediate_size, hidden_size, + search_space, is_fp16): + N1, K1 = shard_intermediate_size, hidden_size + N2, K2 = hidden_size, shard_intermediate_size // 2 + pruned_space_1 = prune_rocm_configs(num_tokens * 2, N1, K1, search_space, + is_fp16) + pruned_space_2 = prune_rocm_configs(num_tokens * 2, N2, K2, search_space, + is_fp16) + search_space = merge_unique_dicts(pruned_space_1, pruned_space_2) + return search_space + + +# The following code is inspired by ROCm/Triton GEMM tuning script: +# https://github.com/ROCm/triton/blob/triton-mlir/scripts/amd/gemm/tune_gemm.py#L89 +def prune_rocm_configs(M, N, K, configs, is_fp16=True): + pruned_configs = [] + elemBytes_a = 2 if is_fp16 else 1 + elemBytes_b = 2 if is_fp16 else 1 + + mfma = 16 if M < 32 or N < 32 else 32 + + # TODO (zhanglx): figure out the boundary between large and small gemms + large_gemm = False + if M >= 2048 and N >= 2048: + large_gemm = True + + for config in configs: + BLOCK_SIZE_M = config.get("BLOCK_SIZE_M") + BLOCK_SIZE_N = config.get("BLOCK_SIZE_N") + BLOCK_SIZE_K = config.get("BLOCK_SIZE_K") + num_warps = config.get("num_warps") + + if is_fp16: + matrix_instr_nonkdim = config.get("matrix_instr_nonkdim") + if matrix_instr_nonkdim > mfma: + continue + if mfma == 4 and BLOCK_SIZE_K < 64: + continue + # some layouts could not work properly in case + # number elements per thread is less 1 + if BLOCK_SIZE_M * BLOCK_SIZE_N < 64: + continue + SPLIT_K = config.get("SPLIT_K", 1) + GROUP_M = config.get("GROUP_SIZE_M") + if is_fp16: + if (matrix_instr_nonkdim > BLOCK_SIZE_M + or matrix_instr_nonkdim > BLOCK_SIZE_N): + continue + if (matrix_instr_nonkdim >= M + and matrix_instr_nonkdim != BLOCK_SIZE_M): + continue + if (matrix_instr_nonkdim >= N + and matrix_instr_nonkdim != BLOCK_SIZE_N): + continue + # Skip BLOCK_SIZE that is too large compare to M/N + # unless BLOCK_SIZE is already small enough + if M * 2 < BLOCK_SIZE_M and BLOCK_SIZE_M != 16: + continue + if N * 2 < BLOCK_SIZE_N and BLOCK_SIZE_N != 16: + continue + # skip large split_k when not necessary + if SPLIT_K != 1 and not need_split_k(M, N, K): + continue + # skip split_k that leads to EVEN_K = false + leap = SPLIT_K * BLOCK_SIZE_K + modv = K % leap + if modv != 0: + continue + # skip large GROUP_M + if GROUP_M * BLOCK_SIZE_M > M and GROUP_M != 1: + continue + # out of shared memory resource + # TODO (zhanglx): This does not consider the LDS usage in the epilogue + LDS = (BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a + + BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b) + if LDS > 65536: + continue + # Skip small block sizes and num_warps for large gemm + # For fp16 and f8, we want to only use BLOCK_SIZE >= 64 + if large_gemm: + if BLOCK_SIZE_M < 64 or BLOCK_SIZE_N < 64: + continue + if BLOCK_SIZE_K < 64: + continue + if num_warps < 4: + continue + + pruned_configs.append(config) + + return pruned_configs + + +def need_split_k(SIZE_M, SIZE_N, SIZE_K): + return (SIZE_M < 64 or SIZE_N < 64) and SIZE_K > 1024 + + +def merge_unique_dicts(list1, list2): + result = [] + combined_list = list1.copy() + combined_list.extend(list2) + for dictionary in combined_list: + if dictionary not in result: + result.append(dictionary) + return result + + @ray.remote(num_gpus=1) class BenchmarkWorker: @@ -222,8 +318,9 @@ def __init__(self, seed: int) -> None: torch.set_default_device("cuda") current_platform.seed_everything(seed) self.seed = seed - # get the device id to allocate the tensors and kernels explicitly - # on the respective GPU ID + # Get the device ID to allocate tensors and kernels + # on the respective GPU. This is required for Ray to work + # correctly with multi-GPU tuning on the ROCm platform. self.device_id = int(ray.get_gpu_ids()[0]) def benchmark( @@ -272,6 +369,12 @@ def tune( ) -> Dict[str, int]: best_config = None best_time = float("inf") + if current_platform.is_rocm(): + is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16) + search_space = prune_rocm_search_space(num_tokens, + shard_intermediate_size, + hidden_size, search_space, + is_fp16) with torch.cuda.device(self.device_id): for config in tqdm(search_space): @@ -373,7 +476,7 @@ def main(args: argparse.Namespace): if args.batch_size is None: batch_sizes = [ 1, 2, 4, 8, 16, 24, 32, 48, 64, 96, 128, 256, 512, 1024, 1536, - 2048, 3072, 4096, 8192, 16384, 18432, 20480 + 2048, 3072, 4096 ] else: batch_sizes = [args.batch_size] @@ -437,4 +540,4 @@ def _distribute(method: str, inputs: List[Any]) -> List[Any]: parser.add_argument("--tune", action="store_true") args = parser.parse_args() - main(args) + main(args) \ No newline at end of file diff --git a/benchmarks/kernels/benchmark_paged_attention.py b/benchmarks/kernels/benchmark_paged_attention.py index 483584dd804ef..f0478123c355c 100644 --- a/benchmarks/kernels/benchmark_paged_attention.py +++ b/benchmarks/kernels/benchmark_paged_attention.py @@ -9,8 +9,9 @@ from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser, create_kv_caches_with_random) -NUM_BLOCKS = 1024 * 1024 +NUM_BLOCKS = 128 * 1024 PARTITION_SIZE = 512 +PARTITION_SIZE_ROCM = 256 @torch.inference_mode() @@ -78,9 +79,12 @@ def main( # Prepare for the paged attention kernel. output = torch.empty_like(query) if version == "v2": - if current_platform.is_rocm() and not args.custom_paged_attn: + if current_platform.is_rocm(): global PARTITION_SIZE - PARTITION_SIZE = 1024 + if not args.custom_paged_attn: + PARTITION_SIZE = 1024 + else: + PARTITION_SIZE = PARTITION_SIZE_ROCM num_partitions = ((max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE) tmp_output = torch.empty( size=(num_seqs, num_query_heads, num_partitions, head_size), @@ -101,7 +105,9 @@ def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: start_time = time.perf_counter() # Using default kv_scale - k_scale = v_scale = 1.0 + k_scale = v_scale = torch.tensor(1.0, + dtype=torch.float32, + device=device) for _ in range(num_iters): if version == "v1": @@ -161,6 +167,8 @@ def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: kv_cache_dtype, k_scale, v_scale, + None, + PARTITION_SIZE, ) else: raise ValueError(f"Invalid version: {version}") @@ -174,13 +182,13 @@ def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: # Warmup. print("Warming up...") run_benchmark = run_cuda_benchmark - run_benchmark(num_iters=3, profile=False) + run_benchmark(num_iters=500, profile=False) # Benchmark. if do_profile: latency = run_benchmark(num_iters=1, profile=True) else: - latency = run_benchmark(num_iters=1000, profile=False) + latency = run_benchmark(num_iters=10000, profile=False) print(f"Kernel running time: {latency * 1000000:.3f} us") diff --git a/benchmarks/kernels/utils.py b/benchmarks/kernels/utils.py new file mode 100644 index 0000000000000..fee877b6f76fa --- /dev/null +++ b/benchmarks/kernels/utils.py @@ -0,0 +1,210 @@ +import dataclasses +from typing import Any, Callable, Iterable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement + + +@dataclasses.dataclass +class CudaGraphBenchParams: + num_ops_in_cuda_graph: int + + +@dataclasses.dataclass +class ArgPool: + """ + When some argument of the benchmarking function is annotated with this type, + the benchmarking class (BenchMM) will collapse the argument to a pick a + single value from the given list of values, during function invocation. + For every invocation during a benchmarking run, it will choose a + different value from the list. + """ + values: Iterable[Any] + + def __getitem__(self, index): + return self.values[index] + + +class Bench: + + class ArgsIterator: + + def __init__(self, args_list, kwargs_list): + assert len(args_list) == len(kwargs_list) + self.args_list = args_list + self.kwargs_list = kwargs_list + self.n = len(self.args_list) + self.idx = 0 + + def __next__(self): + while True: + yield (self.args_list[self.idx], self.kwargs_list[self.idx]) + self.idx += 1 + self.idx = self.idx % self.n + + def reset(self): + self.idx = 0 + + @property + def n_args(self): + return self.n + + def __init__(self, cuda_graph_params: Optional[CudaGraphBenchParams], + label: str, sub_label: str, description: str, fn: Callable, + *args, **kwargs): + + self.cuda_graph_params = cuda_graph_params + self.use_cuda_graph = self.cuda_graph_params is not None + self.label = label + self.sub_label = sub_label + self.description = description + self.fn = fn + + # Process args + self._args = args + self._kwargs = kwargs + self.args_list, self.kwargs_list = self.collapse_argpool( + *args, **kwargs) + self.args_iterator = self.ArgsIterator(self.args_list, + self.kwargs_list) + + # Cudagraph runner + self.g = None + if self.use_cuda_graph: + self.g = self.get_cuda_graph_runner() + + # benchmark run params + self.min_run_time = 1 + + def collapse_argpool(self, *args, **kwargs): + argpool_args = [arg for arg in args if isinstance(arg, ArgPool)] + [ + arg for arg in kwargs.values() if isinstance(arg, ArgPool) + ] + if len(argpool_args) == 0: + return [args], [kwargs] + + # Make sure all argpools are of the same size + argpool_size = len(argpool_args[0].values) + assert all([argpool_size == len(arg.values) for arg in argpool_args]) + + # create copies of the args + args_list = [] + kwargs_list = [] + for _ in range(argpool_size): + args_list.append(args) + kwargs_list.append(kwargs.copy()) + + for i in range(argpool_size): + # collapse args; Just pick the ith value + args_list[i] = tuple([ + arg[i] if isinstance(arg, ArgPool) else arg + for arg in args_list[i] + ]) + + # collapse kwargs + kwargs_i = kwargs_list[i] + arg_pool_keys = [ + k for k, v in kwargs_i.items() if isinstance(v, ArgPool) + ] + for k in arg_pool_keys: + # again just pick the ith value + kwargs_i[k] = kwargs_i[k][i] + kwargs_list[i] = kwargs_i + + return args_list, kwargs_list + + def get_cuda_graph_runner(self): + assert self.use_cuda_graph + assert self.args_iterator is not None + + num_graph_ops = self.cuda_graph_params.num_ops_in_cuda_graph + + # warmup + args_it = self.args_iterator.__next__() + for _ in range(2): + args, kwargs = next(args_it) + self.fn(*args, **kwargs) + + self.args_iterator.reset() + args_it = self.args_iterator.__next__() + stream = torch.cuda.Stream() + with torch.cuda.stream(stream): + g = torch.cuda.CUDAGraph() + with torch.cuda.graph(g): + for _ in range(num_graph_ops): + args, kwargs = next(args_it) + self.fn(*args, **kwargs) + return g + + def run_cudagrah(self) -> TMeasurement: + assert self.use_cuda_graph + globals = {'g': self.g} + + return TBenchmark.Timer( + stmt="g.replay()", + globals=globals, + label=( + f"{self.label}" + f" | cugraph {self.cuda_graph_params.num_ops_in_cuda_graph} ops" + ), + sub_label=self.sub_label, + description=self.description, + ).blocked_autorange(min_run_time=self.min_run_time) + + def run_eager(self) -> TMeasurement: + setup = None + stmt = None + globals = None + + has_arg_pool = self.args_iterator.n_args > 1 + if has_arg_pool: + setup = ''' + args_iterator.reset() + args_it = args_iterator.__next__() + ''' + stmt = ''' + args, kwargs = next(args_it) + fn(*args, **kwargs) + ''' + globals = {'fn': self.fn, 'args_iterator': self.args_iterator} + else: + # no arg pool. Just use the args and kwargs directly + self.args_iterator.reset() + args_it = self.args_iterator.__next__() + args, kwargs = next(args_it) + + setup = "" + stmt = ''' + fn(*args, **kwargs) + ''' + globals = {'fn': self.fn, 'args': args, 'kwargs': kwargs} + + return TBenchmark.Timer( + stmt=stmt, + setup=setup, + globals=globals, + label=self.label, + sub_label=self.sub_label, + description=self.description, + ).blocked_autorange(min_run_time=self.min_run_time) + + def run(self) -> TMeasurement: + timer = None + if self.use_cuda_graph: # noqa SIM108 + timer = self.run_cudagrah() + else: + timer = self.run_eager() + if not timer.meets_confidence() or timer.has_warnings: + print("Doesn't meet confidence - re-running bench ...") + return self.run() + return timer + + def __enter__(self): + return self + + def __exit__(self, exc_type, exc_value, traceback): + if exc_type: + print(f"exc type {exc_type}") + print(f"exc value {exc_value}") + print(f"exc traceback {traceback}") diff --git a/benchmarks/profiling/benchmark_throughput.py b/benchmarks/profiling/benchmark_throughput.py index 3bbdd7d4267ae..67e111f442d6c 100644 --- a/benchmarks/profiling/benchmark_throughput.py +++ b/benchmarks/profiling/benchmark_throughput.py @@ -272,7 +272,8 @@ def main(args: argparse.Namespace): args.tokenizer, trust_remote_code=args.trust_remote_code) if args.dataset is None: # Synthesize a prompt with the given input length. - prompt = "hi" * (args.input_len - 1) + prompt = { "prompt_token_ids" : [42] * (args.input_len - 1) } \ + if args.skip_tokenizer_init else "hi" * (args.input_len - 1) requests = [(prompt, args.input_len, args.output_len) for _ in range(args.num_prompts)] else: diff --git a/benchmarks/test_accuracy.py b/benchmarks/test_accuracy.py index 06c1150f9f20f..94b87d7e550cc 100644 --- a/benchmarks/test_accuracy.py +++ b/benchmarks/test_accuracy.py @@ -1,6 +1,7 @@ -from vllm import LLM, SamplingParams import time +from vllm import LLM, SamplingParams + def main(): llm = LLM( @@ -33,7 +34,7 @@ def main(): num_tokens = sum(out_lengths) print( - f"{num_tokens} tokens. {num_tokens / batch_size} on average. {num_tokens / elapsed_time:.2f} tokens/s. {elapsed_time} seconds" + f"{num_tokens} tokens. {num_tokens / batch_size} on average. {num_tokens / elapsed_time:.2f} tokens/s. {elapsed_time} seconds" # noqa: E501 ) for out in outs: print("===========") diff --git a/cmake/utils.cmake b/cmake/utils.cmake index 81ac34d55b03f..825fac8cd368b 100644 --- a/cmake/utils.cmake +++ b/cmake/utils.cmake @@ -58,8 +58,8 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS) # set(SRCS ${ORIG_SRCS}) set(CXX_SRCS ${ORIG_SRCS}) - list(FILTER SRCS EXCLUDE REGEX "\.(cc)|(cpp)$") - list(FILTER CXX_SRCS INCLUDE REGEX "\.(cc)|(cpp)$") + list(FILTER SRCS EXCLUDE REGEX "\.(cc)|(cpp)|(hip)$") + list(FILTER CXX_SRCS INCLUDE REGEX "\.(cc)|(cpp)|(hip)$") # # Generate ROCm/HIP source file names from CUDA file names. @@ -259,7 +259,7 @@ endmacro() # in `SRC_CUDA_ARCHS` that is less or equal to the version in `TGT_CUDA_ARCHS`. # We have special handling for 9.0a, if 9.0a is in `SRC_CUDA_ARCHS` and 9.0 is # in `TGT_CUDA_ARCHS` then we should remove 9.0a from `SRC_CUDA_ARCHS` and add -# 9.0a to the result. +# 9.0a to the result (and remove 9.0 from TGT_CUDA_ARCHS). # The result is stored in `OUT_CUDA_ARCHS`. # # Example: @@ -270,34 +270,47 @@ endmacro() # function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_ARCHS) list(REMOVE_DUPLICATES SRC_CUDA_ARCHS) + set(TGT_CUDA_ARCHS_ ${TGT_CUDA_ARCHS}) # if 9.0a is in SRC_CUDA_ARCHS and 9.0 is in CUDA_ARCHS then we should # remove 9.0a from SRC_CUDA_ARCHS and add 9.0a to _CUDA_ARCHS set(_CUDA_ARCHS) if ("9.0a" IN_LIST SRC_CUDA_ARCHS) list(REMOVE_ITEM SRC_CUDA_ARCHS "9.0a") - if ("9.0" IN_LIST TGT_CUDA_ARCHS) + if ("9.0" IN_LIST TGT_CUDA_ARCHS_) + list(REMOVE_ITEM TGT_CUDA_ARCHS_ "9.0") set(_CUDA_ARCHS "9.0a") endif() endif() list(SORT SRC_CUDA_ARCHS COMPARE NATURAL ORDER ASCENDING) - # for each ARCH in CUDA_ARCHS find the highest arch in SRC_CUDA_ARCHS that is - # less or eqault to ARCH - foreach(_ARCH ${CUDA_ARCHS}) - set(_TMP_ARCH) - foreach(_SRC_ARCH ${SRC_CUDA_ARCHS}) - if (_SRC_ARCH VERSION_LESS_EQUAL _ARCH) - set(_TMP_ARCH ${_SRC_ARCH}) - else() - break() + # for each ARCH in TGT_CUDA_ARCHS find the highest arch in SRC_CUDA_ARCHS that + # is less or equal to ARCH (but has the same major version since SASS binary + # compatibility is only forward compatible within the same major version). + foreach(_ARCH ${TGT_CUDA_ARCHS_}) + set(_TMP_ARCH) + # Extract the major version of the target arch + string(REGEX REPLACE "^([0-9]+)\\..*$" "\\1" TGT_ARCH_MAJOR "${_ARCH}") + foreach(_SRC_ARCH ${SRC_CUDA_ARCHS}) + # Extract the major version of the source arch + string(REGEX REPLACE "^([0-9]+)\\..*$" "\\1" SRC_ARCH_MAJOR "${_SRC_ARCH}") + # Check major-version match AND version-less-or-equal + if (_SRC_ARCH VERSION_LESS_EQUAL _ARCH) + if (SRC_ARCH_MAJOR STREQUAL TGT_ARCH_MAJOR) + set(_TMP_ARCH "${_SRC_ARCH}") + endif() + else() + # If we hit a version greater than the target, we can break + break() + endif() + endforeach() + + # If we found a matching _TMP_ARCH, append it to _CUDA_ARCHS + if (_TMP_ARCH) + list(APPEND _CUDA_ARCHS "${_TMP_ARCH}") endif() endforeach() - if (_TMP_ARCH) - list(APPEND _CUDA_ARCHS ${_TMP_ARCH}) - endif() - endforeach() list(REMOVE_DUPLICATES _CUDA_ARCHS) set(${OUT_CUDA_ARCHS} ${_CUDA_ARCHS} PARENT_SCOPE) diff --git a/csrc/activation_kernels.cu b/csrc/activation_kernels.cu index 8c3e74e502a55..41176c801fb04 100644 --- a/csrc/activation_kernels.cu +++ b/csrc/activation_kernels.cu @@ -13,8 +13,16 @@ namespace vllm { +template +__device__ __forceinline__ scalar_t compute(const scalar_t& x, + const scalar_t& y) { + return act_first ? ACT_FN(x) * y : x * ACT_FN(y); +} // Activation and gating kernel template. -template + +template __global__ void act_and_mul_kernel( scalar_t* __restrict__ out, // [..., d] const scalar_t* __restrict__ input, // [..., 2, d] @@ -23,7 +31,7 @@ __global__ void act_and_mul_kernel( for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) { const scalar_t x = VLLM_LDG(&input[token_idx * 2 * d + idx]); const scalar_t y = VLLM_LDG(&input[token_idx * 2 * d + d + idx]); - out[token_idx * d + idx] = ACT_FN(x) * y; + out[token_idx * d + idx] = compute(x, y); } } @@ -77,7 +85,9 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) { } // namespace vllm // Launch activation and gating kernel. -#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL) \ +// Use ACT_FIRST (bool) indicating whether to apply the activation function +// first. +#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, ACT_FIRST) \ int d = input.size(-1) / 2; \ int64_t num_tokens = input.numel() / input.size(-1); \ dim3 grid(num_tokens); \ @@ -86,7 +96,7 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) { const cudaStream_t stream = at::cuda::getCurrentCUDAStream(); \ VLLM_DISPATCH_FLOATING_TYPES( \ input.scalar_type(), "act_and_mul_kernel", [&] { \ - vllm::act_and_mul_kernel> \ + vllm::act_and_mul_kernel, ACT_FIRST> \ <<>>(out.data_ptr(), \ input.data_ptr(), d); \ }); @@ -113,7 +123,15 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) { void silu_and_mul(torch::Tensor& out, // [..., d] torch::Tensor& input) // [..., 2 * d] { - LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel); + LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, true); +} + +void mul_and_silu(torch::Tensor& out, // [..., d] + torch::Tensor& input) // [..., 2 * d] +{ + // The difference between mul_and_silu and silu_and_mul is that mul_and_silu + // applies the silu to the latter half of the input. + LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, false); } void scaled_silu_and_mul(torch::Tensor& out, // [..., d] @@ -127,13 +145,13 @@ void scaled_silu_and_mul(torch::Tensor& out, // [..., d] void gelu_and_mul(torch::Tensor& out, // [..., d] torch::Tensor& input) // [..., 2 * d] { - LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel); + LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel, true); } void gelu_tanh_and_mul(torch::Tensor& out, // [..., d] torch::Tensor& input) // [..., 2 * d] { - LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel); + LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel, true); } namespace vllm { diff --git a/csrc/attention/paged_attention_v1.cu b/csrc/attention/paged_attention_v1.cu index 3569b3c88abcd..87bea0f3b279c 100644 --- a/csrc/attention/paged_attention_v1.cu +++ b/csrc/attention/paged_attention_v1.cu @@ -100,6 +100,9 @@ void paged_attention_v1_launcher( // NOTE(woosuk): To reduce the compilation time, we only compile for the // head sizes that we use in the model. However, we can easily extend this // to support any head size which is a multiple of 16. + case 32: + LAUNCH_PAGED_ATTENTION_V1(32); + break; case 64: LAUNCH_PAGED_ATTENTION_V1(64); break; diff --git a/csrc/attention/paged_attention_v2.cu b/csrc/attention/paged_attention_v2.cu index bc543e713fe58..fc84a6774b8ea 100644 --- a/csrc/attention/paged_attention_v2.cu +++ b/csrc/attention/paged_attention_v2.cu @@ -114,6 +114,9 @@ void paged_attention_v2_launcher( // NOTE(woosuk): To reduce the compilation time, we only compile for the // head sizes that we use in the model. However, we can easily extend this // to support any head size which is a multiple of 16. + case 32: + LAUNCH_PAGED_ATTENTION_V2(32); + break; case 64: LAUNCH_PAGED_ATTENTION_V2(64); break; diff --git a/csrc/core/scalar_type.hpp b/csrc/core/scalar_type.hpp index 408e736d5bc0f..c2ae554c9f8e8 100644 --- a/csrc/core/scalar_type.hpp +++ b/csrc/core/scalar_type.hpp @@ -32,7 +32,7 @@ class ScalarType { signed_(signed_), bias(bias), finite_values_only(finite_values_only), - nan_repr(nan_repr){}; + nan_repr(nan_repr) {}; static constexpr ScalarType int_(uint8_t size_bits, int32_t bias = 0) { return ScalarType(0, size_bits - 1, true, bias); diff --git a/csrc/cpu/attention.cpp b/csrc/cpu/attention.cpp index ef5b14088c63b..b9764056e8a2d 100644 --- a/csrc/cpu/attention.cpp +++ b/csrc/cpu/attention.cpp @@ -460,11 +460,11 @@ void paged_attention_v1( torch::Tensor& value_cache, int64_t num_kv_heads, double scale, torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size, int64_t max_seq_len, const std::optional& alibi_slopes, - const std::string& kv_cache_dtype, double k_scale, double v_scale, - const int64_t tp_rank, const int64_t blocksparse_local_blocks, + const std::string& kv_cache_dtype, torch::Tensor& k_scale, + torch::Tensor& v_scale, const int64_t tp_rank, + const int64_t blocksparse_local_blocks, const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size, const int64_t blocksparse_head_sliding_step) { - TORCH_CHECK(k_scale == 1.0f && v_scale == 1.0f); TORCH_CHECK(blocksparse_vert_stride <= 1, "CPU backend does not support blocksparse attention yet."); VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v1_impl", @@ -782,11 +782,11 @@ void paged_attention_v2( torch::Tensor& value_cache, int64_t num_kv_heads, double scale, torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size, int64_t max_seq_len, const std::optional& alibi_slopes, - const std::string& kv_cache_dtype, double k_scale, double v_scale, - const int64_t tp_rank, const int64_t blocksparse_local_blocks, + const std::string& kv_cache_dtype, torch::Tensor& k_scale, + torch::Tensor& v_scale, const int64_t tp_rank, + const int64_t blocksparse_local_blocks, const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size, const int64_t blocksparse_head_sliding_step) { - TORCH_CHECK(k_scale == 1.0f && v_scale == 1.0f); TORCH_CHECK(blocksparse_vert_stride <= 1, "CPU backend does not support blocksparse attention yet."); VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v2_impl", diff --git a/csrc/cpu/cache.cpp b/csrc/cpu/cache.cpp index 31d454328b2c1..e3809acad7453 100644 --- a/csrc/cpu/cache.cpp +++ b/csrc/cpu/cache.cpp @@ -107,10 +107,8 @@ void copy_blocks(std::vector const& key_caches, void reshape_and_cache(torch::Tensor& key, torch::Tensor& value, torch::Tensor& key_cache, torch::Tensor& value_cache, torch::Tensor& slot_mapping, - const std::string& kv_cache_dtype, double k_scale, - double v_scale) { - TORCH_CHECK(k_scale == 1.0f && v_scale == 1.0f); - + const std::string& kv_cache_dtype, + torch::Tensor& k_scale, torch::Tensor& v_scale) { int num_tokens = key.size(0); int num_heads = key.size(1); int head_size = key.size(2); diff --git a/csrc/cpu/cpu_types.hpp b/csrc/cpu/cpu_types.hpp index 28db0479748bf..a71815106133a 100644 --- a/csrc/cpu/cpu_types.hpp +++ b/csrc/cpu/cpu_types.hpp @@ -2,13 +2,13 @@ #define CPU_TYPES_HPP #if defined(__x86_64__) - //x86 implementation + // x86 implementation #include "cpu_types_x86.hpp" #elif defined(__POWER9_VECTOR__) - //ppc implementation + // ppc implementation #include "cpu_types_vsx.hpp" #elif defined(__aarch64__) - //arm implementation + // arm implementation #include "cpu_types_arm.hpp" #else #warning "unsupported vLLM cpu implementation" diff --git a/csrc/cpu/cpu_types_arm.hpp b/csrc/cpu/cpu_types_arm.hpp index ae062a5b86892..990e99f2fc069 100644 --- a/csrc/cpu/cpu_types_arm.hpp +++ b/csrc/cpu/cpu_types_arm.hpp @@ -1,48 +1,50 @@ #include -#include +#include #include namespace vec_op { #ifdef ARM_BF16_SUPPORT - #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...) \ - AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \ - AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__) \ - AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) + #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...) \ + AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \ + AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__) \ + AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) #else - #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...) \ - AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \ + #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...) \ + AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \ AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__) #endif -#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \ +#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \ AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__)) #ifndef CPU_OP_GUARD -#define CPU_KERNEL_GUARD_IN(NAME) -#define CPU_KERNEL_GUARD_OUT(NAME) + #define CPU_KERNEL_GUARD_IN(NAME) + #define CPU_KERNEL_GUARD_OUT(NAME) #else -#define CPU_KERNEL_GUARD_IN(NAME) \ - std::cout << #NAME << " invoked." << std::endl; -#define CPU_KERNEL_GUARD_OUT(NAME) std::cout << #NAME << " exit." << std::endl; + #define CPU_KERNEL_GUARD_IN(NAME) \ + std::cout << #NAME << " invoked." << std::endl; + #define CPU_KERNEL_GUARD_OUT(NAME) \ + std::cout << #NAME << " exit." << std::endl; #endif #define FORCE_INLINE __attribute__((always_inline)) inline namespace { - template - constexpr void unroll_loop_item(std::integer_sequence, F &&f) { - (f(std::integral_constant{}), ...); - }; -}; +template +constexpr void unroll_loop_item(std::integer_sequence, F&& f) { + (f(std::integral_constant{}), ...); +}; +}; // namespace template >> -constexpr void unroll_loop(F &&f) { +constexpr void unroll_loop(F&& f) { unroll_loop_item(std::make_integer_sequence{}, std::forward(f)); } -template struct Vec { +template +struct Vec { constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; }; }; @@ -54,127 +56,124 @@ struct FP16Vec8 : public Vec { float16x8_t reg; - explicit FP16Vec8(const void *ptr) - : reg(vld1q_f16(static_cast(ptr))) {}; + explicit FP16Vec8(const void* ptr) + : reg(vld1q_f16(static_cast(ptr))) {}; - explicit FP16Vec8(const FP32Vec8 &); + explicit FP16Vec8(const FP32Vec8&); - void save(void *ptr) const { - vst1q_f16(static_cast<__fp16 *>(ptr), reg); - } + void save(void* ptr) const { vst1q_f16(static_cast<__fp16*>(ptr), reg); } }; struct FP16Vec16 : public Vec { - constexpr static int VEC_ELEM_NUM = 16; - - float16x8x2_t reg; - - explicit FP16Vec16(const void *ptr) { - reg.val[0] = vld1q_f16(reinterpret_cast(ptr)); - reg.val[1] = vld1q_f16(reinterpret_cast(ptr) + 8); - } - - explicit FP16Vec16(const FP32Vec16& vec); - - void save(void *ptr) const { - vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]); - vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]); + constexpr static int VEC_ELEM_NUM = 16; + + float16x8x2_t reg; + + explicit FP16Vec16(const void* ptr) { + reg.val[0] = vld1q_f16(reinterpret_cast(ptr)); + reg.val[1] = vld1q_f16(reinterpret_cast(ptr) + 8); + } + + explicit FP16Vec16(const FP32Vec16& vec); + + void save(void* ptr) const { + vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]); + vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]); + } + + void save(void* ptr, const int elem_num) const { + int full_blocks = elem_num / 8; + int remainder = elem_num % 8; + + if (full_blocks > 0) { + vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]); + if (full_blocks > 1) { + vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]); + } } - - void save(void *ptr, const int elem_num) const { - int full_blocks = elem_num / 8; - int remainder = elem_num % 8; - - if (full_blocks > 0) { - vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]); - if (full_blocks > 1) { - vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]); - } - } - - // Note: below is the unrolled version of the following code: - // - // for (int i = 0; i < remainder; ++i) { - // reinterpret_cast<__fp16*>(ptr)[full_blocks * 8 + i] = - // vgetq_lane_f16(temp, i); - // } - // - // For macOS build (Clang), the arm/neon intrinsics function - // `vgetq_lane_f16` needs the parameter `i` to be constant at compile - // time. - - if (remainder > 0) { - float16x8_t temp = reg.val[full_blocks]; - __fp16* fp16_ptr = reinterpret_cast<__fp16*>(ptr); - switch (remainder) - { - case 1: - fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); - break; - case 2: - fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); - fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); - break; - case 3: - fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); - fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); - fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2); - break; - case 4: - fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); - fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); - fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2); - fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3); - break; - case 5: - fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); - fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); - fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2); - fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3); - fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4); - break; - case 6: - fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); - fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); - fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2); - fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3); - fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4); - fp16_ptr[full_blocks * 8 + 5] = vgetq_lane_f16(temp, 5); - break; - case 7: - fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); - fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); - fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2); - fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3); - fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4); - fp16_ptr[full_blocks * 8 + 5] = vgetq_lane_f16(temp, 5); - fp16_ptr[full_blocks * 8 + 6] = vgetq_lane_f16(temp, 6); - break; - - default: - break; - } - } + + // Note: below is the unrolled version of the following code: + // + // for (int i = 0; i < remainder; ++i) { + // reinterpret_cast<__fp16*>(ptr)[full_blocks * 8 + i] = + // vgetq_lane_f16(temp, i); + // } + // + // For macOS build (Clang), the arm/neon intrinsics function + // `vgetq_lane_f16` needs the parameter `i` to be constant at compile + // time. + + if (remainder > 0) { + float16x8_t temp = reg.val[full_blocks]; + __fp16* fp16_ptr = reinterpret_cast<__fp16*>(ptr); + switch (remainder) { + case 1: + fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); + break; + case 2: + fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); + fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); + break; + case 3: + fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); + fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); + fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2); + break; + case 4: + fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); + fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); + fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2); + fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3); + break; + case 5: + fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); + fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); + fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2); + fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3); + fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4); + break; + case 6: + fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); + fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); + fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2); + fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3); + fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4); + fp16_ptr[full_blocks * 8 + 5] = vgetq_lane_f16(temp, 5); + break; + case 7: + fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0); + fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1); + fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2); + fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3); + fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4); + fp16_ptr[full_blocks * 8 + 5] = vgetq_lane_f16(temp, 5); + fp16_ptr[full_blocks * 8 + 6] = vgetq_lane_f16(temp, 6); + break; + + default: + break; + } } + } }; - #ifdef ARM_BF16_SUPPORT struct BF16Vec8 : public Vec { constexpr static int VEC_ELEM_NUM = 8; bfloat16x8_t reg; - explicit BF16Vec8(const void *ptr) - : reg(*reinterpret_cast(ptr)) {}; + explicit BF16Vec8(const void* ptr) + : reg(*reinterpret_cast(ptr)) {}; explicit BF16Vec8(bfloat16x8_t data) : reg(data) {}; - explicit BF16Vec8(const FP32Vec8 &); + explicit BF16Vec8(const FP32Vec8&); - explicit BF16Vec8(float32x4x2_t v) : reg(vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[0]), v.val[1])) {}; + explicit BF16Vec8(float32x4x2_t v) + : reg(vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[0]), v.val[1])) {}; - void save(void *ptr) const { *reinterpret_cast(ptr) = reg; } + void save(void* ptr) const { *reinterpret_cast(ptr) = reg; } }; struct BF16Vec16 : public Vec { @@ -182,19 +181,18 @@ struct BF16Vec16 : public Vec { bfloat16x8x2_t reg; - explicit BF16Vec16(const void *ptr) - : reg(*reinterpret_cast(ptr)) {}; + explicit BF16Vec16(const void* ptr) + : reg(*reinterpret_cast(ptr)) {}; explicit BF16Vec16(bfloat16x8x2_t data) : reg(data) {}; - explicit BF16Vec16(const FP32Vec16 &); + explicit BF16Vec16(const FP32Vec16&); - explicit BF16Vec16(float32x4x4_t v) : reg({ - vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[0]), v.val[1]), - vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[2]), v.val[3]) - }){}; + explicit BF16Vec16(float32x4x4_t v) + : reg({vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[0]), v.val[1]), + vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[2]), v.val[3])}) {}; - void save(void *ptr) const { *reinterpret_cast(ptr) = reg; }; + void save(void* ptr) const { *reinterpret_cast(ptr) = reg; }; }; struct BF16Vec32 : public Vec { @@ -202,19 +200,15 @@ struct BF16Vec32 : public Vec { bfloat16x8x4_t reg; - explicit BF16Vec32(const void *ptr) - : reg(*reinterpret_cast(ptr)) {}; + explicit BF16Vec32(const void* ptr) + : reg(*reinterpret_cast(ptr)) {}; explicit BF16Vec32(bfloat16x8x4_t data) : reg(data) {}; - explicit BF16Vec32(const BF16Vec8 &vec8_data) : reg({ - vec8_data.reg, - vec8_data.reg, - vec8_data.reg, - vec8_data.reg - }) {}; + explicit BF16Vec32(const BF16Vec8& vec8_data) + : reg({vec8_data.reg, vec8_data.reg, vec8_data.reg, vec8_data.reg}) {}; - void save(void *ptr) const { *reinterpret_cast(ptr) = reg; }; + void save(void* ptr) const { *reinterpret_cast(ptr) = reg; }; }; #endif @@ -232,11 +226,11 @@ struct FP32Vec4 : public Vec { explicit FP32Vec4() : reg(vdupq_n_f32(0.0f)) {}; - explicit FP32Vec4(const float *ptr) : reg(vld1q_f32(ptr)) {}; + explicit FP32Vec4(const float* ptr) : reg(vld1q_f32(ptr)) {}; explicit FP32Vec4(float32x4_t data) : reg(data) {}; - explicit FP32Vec4(const FP32Vec4 &data) : reg(data.reg) {}; + explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {}; }; struct FP32Vec8 : public Vec { @@ -252,32 +246,37 @@ struct FP32Vec8 : public Vec { explicit FP32Vec8() : reg({vmovq_n_f32(0.0), vmovq_n_f32(0.0)}) {}; - explicit FP32Vec8(const float *ptr) : reg({vld1q_f32(ptr), vld1q_f32(ptr + 4)}) {}; + explicit FP32Vec8(const float* ptr) + : reg({vld1q_f32(ptr), vld1q_f32(ptr + 4)}) {}; explicit FP32Vec8(float32x4x2_t data) : reg(data) {}; - explicit FP32Vec8(const FP32Vec8 &data) : reg(data.reg) {}; + explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {}; - explicit FP32Vec8(const FP16Vec8 &v) { - reg.val[0] = vcvt_f32_f16(vget_low_f16(v.reg)); - reg.val[1] = vcvt_f32_f16(vget_high_f16(v.reg)); - }; + explicit FP32Vec8(const FP16Vec8& v) { + reg.val[0] = vcvt_f32_f16(vget_low_f16(v.reg)); + reg.val[1] = vcvt_f32_f16(vget_high_f16(v.reg)); + }; - explicit FP32Vec8(float16x8_t v) : reg({vcvt_f32_f16(vget_low_f16(v)), vcvt_f32_f16(vget_high_f16(v))}) {}; + explicit FP32Vec8(float16x8_t v) + : reg({vcvt_f32_f16(vget_low_f16(v)), vcvt_f32_f16(vget_high_f16(v))}) {}; - #ifdef ARM_BF16_SUPPORT +#ifdef ARM_BF16_SUPPORT - explicit FP32Vec8(bfloat16x8_t v) : reg({vcvtq_low_f32_bf16(v), vcvtq_high_f32_bf16(v)}) {}; + explicit FP32Vec8(bfloat16x8_t v) + : reg({vcvtq_low_f32_bf16(v), vcvtq_high_f32_bf16(v)}) {}; - explicit FP32Vec8(const BF16Vec8 &v) : reg({vcvtq_low_f32_bf16(v.reg), vcvtq_high_f32_bf16(v.reg)}) {}; + explicit FP32Vec8(const BF16Vec8& v) + : reg({vcvtq_low_f32_bf16(v.reg), vcvtq_high_f32_bf16(v.reg)}) {}; - #endif +#endif float reduce_sum() const { AliasReg ar; ar.reg = reg; float answer = 0; - unroll_loop([&answer, &ar](int i) { answer += ar.values[i]; }); + unroll_loop( + [&answer, &ar](int i) { answer += ar.values[i]; }); return answer; } @@ -324,10 +323,14 @@ struct FP32Vec8 : public Vec { AliasReg ar; ar.reg = reg; - float32x2_t er_vec0 = {static_cast(erf(ar.values[0])), static_cast(erf(ar.values[1]))}; - float32x2_t er_vec1 = {static_cast(erf(ar.values[2])), static_cast(erf(ar.values[3]))}; - float32x2_t er_vec2 = {static_cast(erf(ar.values[4])), static_cast(erf(ar.values[5]))}; - float32x2_t er_vec3 = {static_cast(erf(ar.values[6])), static_cast(erf(ar.values[7]))}; + float32x2_t er_vec0 = {static_cast(erf(ar.values[0])), + static_cast(erf(ar.values[1]))}; + float32x2_t er_vec1 = {static_cast(erf(ar.values[2])), + static_cast(erf(ar.values[3]))}; + float32x2_t er_vec2 = {static_cast(erf(ar.values[4])), + static_cast(erf(ar.values[5]))}; + float32x2_t er_vec3 = {static_cast(erf(ar.values[6])), + static_cast(erf(ar.values[7]))}; float32x4_t result0 = vcombine_f32(er_vec0, er_vec1); float32x4_t result1 = vcombine_f32(er_vec2, er_vec3); @@ -337,25 +340,29 @@ struct FP32Vec8 : public Vec { result.val[1] = result1; return FP32Vec8(result); - } + } - FP32Vec8 operator*(const FP32Vec8 &b) const { - return FP32Vec8(float32x4x2_t({vmulq_f32(reg.val[0], b.reg.val[0]), vmulq_f32(reg.val[1], b.reg.val[1])})); + FP32Vec8 operator*(const FP32Vec8& b) const { + return FP32Vec8(float32x4x2_t({vmulq_f32(reg.val[0], b.reg.val[0]), + vmulq_f32(reg.val[1], b.reg.val[1])})); } - FP32Vec8 operator+(const FP32Vec8 &b) const { - return FP32Vec8(float32x4x2_t({vaddq_f32(reg.val[0], b.reg.val[0]), vaddq_f32(reg.val[1], b.reg.val[1])})); + FP32Vec8 operator+(const FP32Vec8& b) const { + return FP32Vec8(float32x4x2_t({vaddq_f32(reg.val[0], b.reg.val[0]), + vaddq_f32(reg.val[1], b.reg.val[1])})); } - FP32Vec8 operator-(const FP32Vec8 &b) const { - return FP32Vec8(float32x4x2_t({vsubq_f32(reg.val[0], b.reg.val[0]), vsubq_f32(reg.val[1], b.reg.val[1])})); + FP32Vec8 operator-(const FP32Vec8& b) const { + return FP32Vec8(float32x4x2_t({vsubq_f32(reg.val[0], b.reg.val[0]), + vsubq_f32(reg.val[1], b.reg.val[1])})); } - FP32Vec8 operator/(const FP32Vec8 &b) const { - return FP32Vec8(float32x4x2_t({vdivq_f32(reg.val[0], b.reg.val[0]), vdivq_f32(reg.val[1], b.reg.val[1])})); + FP32Vec8 operator/(const FP32Vec8& b) const { + return FP32Vec8(float32x4x2_t({vdivq_f32(reg.val[0], b.reg.val[0]), + vdivq_f32(reg.val[1], b.reg.val[1])})); } - void save(float *ptr) const { + void save(float* ptr) const { vst1q_f32(ptr, reg.val[0]); vst1q_f32(ptr + 4, reg.val[1]); } @@ -370,103 +377,100 @@ struct FP32Vec16 : public Vec { float32x4x4_t reg; - explicit FP32Vec16(float v) : reg({vmovq_n_f32(v), vmovq_n_f32(v), vmovq_n_f32(v), vmovq_n_f32(v)}) {} + explicit FP32Vec16(float v) + : reg({vmovq_n_f32(v), vmovq_n_f32(v), vmovq_n_f32(v), vmovq_n_f32(v)}) {} - explicit FP32Vec16() : reg({vmovq_n_f32(0.0), vmovq_n_f32(0.0), vmovq_n_f32(0.0), vmovq_n_f32(0.0)}) {} + explicit FP32Vec16() + : reg({vmovq_n_f32(0.0), vmovq_n_f32(0.0), vmovq_n_f32(0.0), + vmovq_n_f32(0.0)}) {} - explicit FP32Vec16(const float *ptr) : reg({vld1q_f32(ptr), vld1q_f32(ptr + 4), vld1q_f32(ptr + 8), vld1q_f32(ptr + 12)}) {} + explicit FP32Vec16(const float* ptr) + : reg({vld1q_f32(ptr), vld1q_f32(ptr + 4), vld1q_f32(ptr + 8), + vld1q_f32(ptr + 12)}) {} explicit FP32Vec16(float32x4x4_t data) : reg(data) {} - explicit FP32Vec16(const FP32Vec8 &data) { - reg.val[0] = data.reg.val[0]; - reg.val[1] = data.reg.val[1]; - reg.val[2] = data.reg.val[0]; - reg.val[3] = data.reg.val[1]; + explicit FP32Vec16(const FP32Vec8& data) { + reg.val[0] = data.reg.val[0]; + reg.val[1] = data.reg.val[1]; + reg.val[2] = data.reg.val[0]; + reg.val[3] = data.reg.val[1]; } - explicit FP32Vec16(const FP32Vec16 &data) : reg(data.reg) {} + explicit FP32Vec16(const FP32Vec16& data) : reg(data.reg) {} - explicit FP32Vec16(const FP16Vec8 &v) : FP32Vec16(FP32Vec8(v.reg)) {} + explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v.reg)) {} - #ifdef ARM_BF16_SUPPORT - explicit FP32Vec16(bfloat16x8x2_t v) : reg({ - vcvtq_low_f32_bf16(v.val[0]), - vcvtq_high_f32_bf16(v.val[0]), - vcvtq_low_f32_bf16(v.val[1]), - vcvtq_high_f32_bf16(v.val[1]) - }) {}; - #endif +#ifdef ARM_BF16_SUPPORT + explicit FP32Vec16(bfloat16x8x2_t v) + : reg({vcvtq_low_f32_bf16(v.val[0]), vcvtq_high_f32_bf16(v.val[0]), + vcvtq_low_f32_bf16(v.val[1]), vcvtq_high_f32_bf16(v.val[1])}) {}; +#endif - explicit FP32Vec16(const FP32Vec4 &data) { + explicit FP32Vec16(const FP32Vec4& data) { reg.val[0] = data.reg; reg.val[1] = data.reg; reg.val[2] = data.reg; reg.val[3] = data.reg; }; - #ifdef ARM_BF16_SUPPORT - explicit FP32Vec16(const BF16Vec16 &v) : reg({ - vcvtq_low_f32_bf16(v.reg.val[0]), - vcvtq_high_f32_bf16(v.reg.val[0]), - vcvtq_low_f32_bf16(v.reg.val[1]), - vcvtq_high_f32_bf16(v.reg.val[1]) - }) {}; - - explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}; - #endif - - explicit FP32Vec16(const FP16Vec16 &v) { - reg.val[0] = vcvt_f32_f16(vget_low_f16(v.reg.val[0])); - reg.val[1] = vcvt_f32_f16(vget_high_f16(v.reg.val[0])); - reg.val[2] = vcvt_f32_f16(vget_low_f16(v.reg.val[1])); - reg.val[3] = vcvt_f32_f16(vget_high_f16(v.reg.val[1])); +#ifdef ARM_BF16_SUPPORT + explicit FP32Vec16(const BF16Vec16& v) + : reg({vcvtq_low_f32_bf16(v.reg.val[0]), + vcvtq_high_f32_bf16(v.reg.val[0]), + vcvtq_low_f32_bf16(v.reg.val[1]), + vcvtq_high_f32_bf16(v.reg.val[1])}) {}; + + explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}; +#endif + + explicit FP32Vec16(const FP16Vec16& v) { + reg.val[0] = vcvt_f32_f16(vget_low_f16(v.reg.val[0])); + reg.val[1] = vcvt_f32_f16(vget_high_f16(v.reg.val[0])); + reg.val[2] = vcvt_f32_f16(vget_low_f16(v.reg.val[1])); + reg.val[3] = vcvt_f32_f16(vget_high_f16(v.reg.val[1])); }; - FP32Vec16 operator+(const FP32Vec16 &b) const { - return FP32Vec16(float32x4x4_t({ - vaddq_f32(reg.val[0], b.reg.val[0]), - vaddq_f32(reg.val[1], b.reg.val[1]), - vaddq_f32(reg.val[2], b.reg.val[2]), - vaddq_f32(reg.val[3], b.reg.val[3])})); + FP32Vec16 operator+(const FP32Vec16& b) const { + return FP32Vec16(float32x4x4_t({vaddq_f32(reg.val[0], b.reg.val[0]), + vaddq_f32(reg.val[1], b.reg.val[1]), + vaddq_f32(reg.val[2], b.reg.val[2]), + vaddq_f32(reg.val[3], b.reg.val[3])})); }; - FP32Vec16 operator*(const FP32Vec16 &b) const { - return FP32Vec16(float32x4x4_t({ - vmulq_f32(reg.val[0], b.reg.val[0]), - vmulq_f32(reg.val[1], b.reg.val[1]), - vmulq_f32(reg.val[2], b.reg.val[2]), - vmulq_f32(reg.val[3], b.reg.val[3])})); + FP32Vec16 operator*(const FP32Vec16& b) const { + return FP32Vec16(float32x4x4_t({vmulq_f32(reg.val[0], b.reg.val[0]), + vmulq_f32(reg.val[1], b.reg.val[1]), + vmulq_f32(reg.val[2], b.reg.val[2]), + vmulq_f32(reg.val[3], b.reg.val[3])})); }; - FP32Vec16 operator-(const FP32Vec16 &b) const { - return FP32Vec16(float32x4x4_t({ - vsubq_f32(reg.val[0], b.reg.val[0]), - vsubq_f32(reg.val[1], b.reg.val[1]), - vsubq_f32(reg.val[2], b.reg.val[2]), - vsubq_f32(reg.val[3], b.reg.val[3]) - })); + FP32Vec16 operator-(const FP32Vec16& b) const { + return FP32Vec16(float32x4x4_t({vsubq_f32(reg.val[0], b.reg.val[0]), + vsubq_f32(reg.val[1], b.reg.val[1]), + vsubq_f32(reg.val[2], b.reg.val[2]), + vsubq_f32(reg.val[3], b.reg.val[3])})); }; - FP32Vec16 operator/(const FP32Vec16 &b) const { - return FP32Vec16(float32x4x4_t({ - vdivq_f32(reg.val[0], b.reg.val[0]), - vdivq_f32(reg.val[1], b.reg.val[1]), - vdivq_f32(reg.val[2], b.reg.val[2]), - vdivq_f32(reg.val[3], b.reg.val[3]) - })); + FP32Vec16 operator/(const FP32Vec16& b) const { + return FP32Vec16(float32x4x4_t({vdivq_f32(reg.val[0], b.reg.val[0]), + vdivq_f32(reg.val[1], b.reg.val[1]), + vdivq_f32(reg.val[2], b.reg.val[2]), + vdivq_f32(reg.val[3], b.reg.val[3])})); }; float reduce_sum() const { AliasReg ar; ar.reg = reg; float answer = 0; - unroll_loop([&answer, &ar](int i) { answer += ar.values[i]; }); + unroll_loop( + [&answer, &ar](int i) { answer += ar.values[i]; }); return answer; }; - template float reduce_sub_sum(int idx) { + template + float reduce_sub_sum(int idx) { static_assert(VEC_ELEM_NUM % group_size == 0); AliasReg ar; @@ -479,7 +483,7 @@ struct FP32Vec16 : public Vec { return answer; }; - void save(float *ptr) const { + void save(float* ptr) const { vst1q_f32(ptr, reg.val[0]); vst1q_f32(ptr + 4, reg.val[1]); vst1q_f32(ptr + 8, reg.val[2]); @@ -487,43 +491,59 @@ struct FP32Vec16 : public Vec { }; }; -template struct VecType { using vec_type = void; }; +template +struct VecType { + using vec_type = void; +}; -template using vec_t = typename VecType::vec_type; +template +using vec_t = typename VecType::vec_type; -template <> struct VecType { using vec_type = FP32Vec8; }; +template <> +struct VecType { + using vec_type = FP32Vec8; +}; -template <> struct VecType { using vec_type = FP16Vec8; }; +template <> +struct VecType { + using vec_type = FP16Vec8; +}; #ifdef ARM_BF16_SUPPORT -template <> struct VecType { using vec_type = BF16Vec8; }; +template <> +struct VecType { + using vec_type = BF16Vec8; +}; #endif -template void storeFP32(float v, T *ptr) { *ptr = v; } +template +void storeFP32(float v, T* ptr) { + *ptr = v; +} -template <> inline void storeFP32(float v, c10::Half *ptr) { - *reinterpret_cast<__fp16 *>(ptr) = v; +template <> +inline void storeFP32(float v, c10::Half* ptr) { + *reinterpret_cast<__fp16*>(ptr) = v; } -inline FP16Vec16::FP16Vec16(const FP32Vec16 &v) { - float16x4_t low_0 = vcvt_f16_f32(v.reg.val[0]); - float16x4_t high_0 = vcvt_f16_f32(v.reg.val[1]); - float16x4_t low_1 = vcvt_f16_f32(v.reg.val[2]); - float16x4_t high_1 = vcvt_f16_f32(v.reg.val[3]); +inline FP16Vec16::FP16Vec16(const FP32Vec16& v) { + float16x4_t low_0 = vcvt_f16_f32(v.reg.val[0]); + float16x4_t high_0 = vcvt_f16_f32(v.reg.val[1]); + float16x4_t low_1 = vcvt_f16_f32(v.reg.val[2]); + float16x4_t high_1 = vcvt_f16_f32(v.reg.val[3]); - reg.val[0] = vcombine_f16(low_0, high_0); - reg.val[1] = vcombine_f16(low_1, high_1); + reg.val[0] = vcombine_f16(low_0, high_0); + reg.val[1] = vcombine_f16(low_1, high_1); }; -inline FP16Vec8 :: FP16Vec8(const FP32Vec8 &v) { - float16x4_t lower_half = vcvt_f16_f32(v.reg.val[0]); - float16x4_t upper_half = vcvt_f16_f32(v.reg.val[1]); +inline FP16Vec8 ::FP16Vec8(const FP32Vec8& v) { + float16x4_t lower_half = vcvt_f16_f32(v.reg.val[0]); + float16x4_t upper_half = vcvt_f16_f32(v.reg.val[1]); - reg = vcombine_f16(lower_half, upper_half); + reg = vcombine_f16(lower_half, upper_half); }; -inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) { - +inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) { acc.reg.val[0] = vfmaq_f32(acc.reg.val[0], a.reg.val[0], b.reg.val[0]); acc.reg.val[1] = vfmaq_f32(acc.reg.val[1], a.reg.val[1], b.reg.val[1]); acc.reg.val[2] = vfmaq_f32(acc.reg.val[2], a.reg.val[2], b.reg.val[2]); @@ -531,8 +551,7 @@ inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) { }; #ifdef ARM_BF16_SUPPORT -inline void fma(FP32Vec16 &acc, BF16Vec32 &a, BF16Vec32 &b) { - +inline void fma(FP32Vec16& acc, BF16Vec32& a, BF16Vec32& b) { float32x4_t a0_low = vcvt_f32_bf16(vget_low_bf16(a.reg.val[0])); float32x4_t a0_high = vcvt_f32_bf16(vget_high_bf16(a.reg.val[0])); float32x4_t a1_low = vcvt_f32_bf16(vget_low_bf16(a.reg.val[1])); @@ -551,22 +570,22 @@ inline void fma(FP32Vec16 &acc, BF16Vec32 &a, BF16Vec32 &b) { #endif #ifdef ARM_BF16_SUPPORT -inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) : reg(vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[0]), v.reg.val[1])) {}; - -inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) : reg({ - vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[0]), v.reg.val[1]), - vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[2]), v.reg.val[3]) - }){}; +inline BF16Vec8::BF16Vec8(const FP32Vec8& v) + : reg(vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[0]), v.reg.val[1])) { + }; + +inline BF16Vec16::BF16Vec16(const FP32Vec16& v) + : reg({vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[0]), v.reg.val[1]), + vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[2]), + v.reg.val[3])}) {}; #endif -inline void prefetch(const void *addr) { - __builtin_prefetch(addr, 0, 1); -}; +inline void prefetch(const void* addr) { __builtin_prefetch(addr, 0, 1); }; #ifdef ARM_BF16_SUPPORT template <> -inline void storeFP32(float v, c10::BFloat16 *ptr) { - *reinterpret_cast<__bf16 *>(ptr) = vcvth_bf16_f32(v); +inline void storeFP32(float v, c10::BFloat16* ptr) { + *reinterpret_cast<__bf16*>(ptr) = vcvth_bf16_f32(v); }; #endif -}; \ No newline at end of file +}; // namespace vec_op \ No newline at end of file diff --git a/csrc/cpu/cpu_types_vsx.hpp b/csrc/cpu/cpu_types_vsx.hpp index b50bdadc5713d..a8e1be37eb418 100644 --- a/csrc/cpu/cpu_types_vsx.hpp +++ b/csrc/cpu/cpu_types_vsx.hpp @@ -9,38 +9,40 @@ namespace vec_op { // FIXME: FP16 is not fully supported in Torch-CPU -#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...) \ - AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \ +#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...) \ + AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \ AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) -#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \ +#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \ AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__)) #ifndef CPU_OP_GUARD -#define CPU_KERNEL_GUARD_IN(NAME) -#define CPU_KERNEL_GUARD_OUT(NAME) + #define CPU_KERNEL_GUARD_IN(NAME) + #define CPU_KERNEL_GUARD_OUT(NAME) #else -#define CPU_KERNEL_GUARD_IN(NAME) \ - std::cout << #NAME << " invoked." << std::endl; -#define CPU_KERNEL_GUARD_OUT(NAME) std::cout << #NAME << " exit." << std::endl; + #define CPU_KERNEL_GUARD_IN(NAME) \ + std::cout << #NAME << " invoked." << std::endl; + #define CPU_KERNEL_GUARD_OUT(NAME) \ + std::cout << #NAME << " exit." << std::endl; #endif #define FORCE_INLINE __attribute__((always_inline)) inline namespace { template -constexpr void unroll_loop_item(std::integer_sequence, F &&f) { +constexpr void unroll_loop_item(std::integer_sequence, F&& f) { (f(std::integral_constant{}), ...); } -}; // namespace +}; // namespace template >> -constexpr void unroll_loop(F &&f) { +constexpr void unroll_loop(F&& f) { unroll_loop_item(std::make_integer_sequence{}, std::forward(f)); } -template struct Vec { +template +struct Vec { constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; } }; @@ -68,12 +70,14 @@ struct BF16Vec8 : public Vec { __vector signed short reg; - explicit BF16Vec8(const void *ptr) - : reg((__vector signed short)vec_xl(0, (__vector signed short *)ptr)) {} + explicit BF16Vec8(const void* ptr) + : reg((__vector signed short)vec_xl(0, (__vector signed short*)ptr)) {} - explicit BF16Vec8(const FP32Vec8 &); + explicit BF16Vec8(const FP32Vec8&); - void save(void *ptr) const { *reinterpret_cast<__vector signed short *>(ptr) = reg; } + void save(void* ptr) const { + *reinterpret_cast<__vector signed short*>(ptr) = reg; + } }; struct BF16Vec16 : public Vec { @@ -81,18 +85,18 @@ struct BF16Vec16 : public Vec { ss16x8x2_t reg; - explicit BF16Vec16(const void *ptr) { + explicit BF16Vec16(const void* ptr) { // Load 256 bits in two parts - reg.val[0] = (__vector signed short)vec_xl(0, (signed short *)ptr); - reg.val[1] = (__vector signed short)vec_xl(16, (signed short *)ptr); + reg.val[0] = (__vector signed short)vec_xl(0, (signed short*)ptr); + reg.val[1] = (__vector signed short)vec_xl(16, (signed short*)ptr); } - explicit BF16Vec16(const FP32Vec16 &); + explicit BF16Vec16(const FP32Vec16&); - void save(void *ptr) const { + void save(void* ptr) const { // Save 256 bits in two parts - vec_xst(reg.val[0], 0, (signed short *)ptr); - vec_xst(reg.val[1], 16, (signed short *)ptr); + vec_xst(reg.val[0], 0, (signed short*)ptr); + vec_xst(reg.val[1], 16, (signed short*)ptr); } }; @@ -102,19 +106,15 @@ struct BF16Vec32 : public Vec { constexpr static int VEC_ELEM_NUM = 32; ss16x8x4_t reg; - explicit BF16Vec32(const void *ptr) - : reg(*reinterpret_cast(ptr)) {} + explicit BF16Vec32(const void* ptr) + : reg(*reinterpret_cast(ptr)) {} explicit BF16Vec32(ss16x8x4_t data) : reg(data) {} - explicit BF16Vec32(const BF16Vec8 &vec8_data) : reg({ - vec8_data.reg, - vec8_data.reg, - vec8_data.reg, - vec8_data.reg - }) {} + explicit BF16Vec32(const BF16Vec8& vec8_data) + : reg({vec8_data.reg, vec8_data.reg, vec8_data.reg, vec8_data.reg}) {} - void save(void *ptr) const { *reinterpret_cast(ptr) = reg; } + void save(void* ptr) const { *reinterpret_cast(ptr) = reg; } }; struct FP32Vec4 : public Vec { @@ -130,11 +130,11 @@ struct FP32Vec4 : public Vec { explicit FP32Vec4() : reg(vec_splats(0.0f)) {} - explicit FP32Vec4(const float *ptr) : reg(vec_xl(0, ptr)) {} + explicit FP32Vec4(const float* ptr) : reg(vec_xl(0, ptr)) {} explicit FP32Vec4(__vector float data) : reg(data) {} - explicit FP32Vec4(const FP32Vec4 &data) : reg(data.reg) {} + explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {} }; struct FP32Vec8 : public Vec { @@ -156,19 +156,19 @@ struct FP32Vec8 : public Vec { reg.val[1] = vec_splats(0.0f); } - explicit FP32Vec8(const float *ptr) { + explicit FP32Vec8(const float* ptr) { reg.val[0] = vec_xl(0, ptr); reg.val[1] = vec_xl(16, ptr); } explicit FP32Vec8(f32x4x2_t data) : reg(data) {} - explicit FP32Vec8(const FP32Vec8 &data) { + explicit FP32Vec8(const FP32Vec8& data) { reg.val[0] = data.reg.val[0]; reg.val[1] = data.reg.val[1]; } - explicit FP32Vec8(const BF16Vec8 &v) { + explicit FP32Vec8(const BF16Vec8& v) { reg.val[0] = (__vector float)vec_mergeh(zero, v.reg); reg.val[1] = (__vector float)vec_mergel(zero, v.reg); } @@ -177,7 +177,8 @@ struct FP32Vec8 : public Vec { AliasReg ar; ar.reg = reg; float result = 0; - unroll_loop([&result, &ar](int i) { result += ar.values[i]; }); + unroll_loop( + [&result, &ar](int i) { result += ar.values[i]; }); return result; } @@ -230,23 +231,27 @@ struct FP32Vec8 : public Vec { return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]})); } - FP32Vec8 operator*(const FP32Vec8 &b) const { - return FP32Vec8({vec_mul(reg.val[0], b.reg.val[0]), vec_mul(reg.val[1], b.reg.val[1])}); + FP32Vec8 operator*(const FP32Vec8& b) const { + return FP32Vec8( + {vec_mul(reg.val[0], b.reg.val[0]), vec_mul(reg.val[1], b.reg.val[1])}); } - FP32Vec8 operator+(const FP32Vec8 &b) const { - return FP32Vec8({vec_add(reg.val[0], b.reg.val[0]), vec_add(reg.val[1], b.reg.val[1])}); + FP32Vec8 operator+(const FP32Vec8& b) const { + return FP32Vec8( + {vec_add(reg.val[0], b.reg.val[0]), vec_add(reg.val[1], b.reg.val[1])}); } - FP32Vec8 operator-(const FP32Vec8 &b) const { - return FP32Vec8({vec_sub(reg.val[0], b.reg.val[0]), vec_sub(reg.val[1], b.reg.val[1])}); + FP32Vec8 operator-(const FP32Vec8& b) const { + return FP32Vec8( + {vec_sub(reg.val[0], b.reg.val[0]), vec_sub(reg.val[1], b.reg.val[1])}); } - FP32Vec8 operator/(const FP32Vec8 &b) const { - return FP32Vec8({vec_div(reg.val[0], b.reg.val[0]), vec_div(reg.val[1], b.reg.val[1])}); + FP32Vec8 operator/(const FP32Vec8& b) const { + return FP32Vec8( + {vec_div(reg.val[0], b.reg.val[0]), vec_div(reg.val[1], b.reg.val[1])}); } - void save(float *ptr) const { + void save(float* ptr) const { vec_xst(reg.val[0], 0, ptr); vec_xst(reg.val[1], 16, ptr); } @@ -275,7 +280,7 @@ struct FP32Vec16 : public Vec { reg.val[3] = vec_splats(0.0f); } - explicit FP32Vec16(const float *ptr) { + explicit FP32Vec16(const float* ptr) { reg.val[0] = vec_xl(0, ptr); reg.val[1] = vec_xl(16, ptr); reg.val[2] = vec_xl(32, ptr); @@ -284,78 +289,76 @@ struct FP32Vec16 : public Vec { explicit FP32Vec16(f32x4x4_t data) : reg(data) {} - explicit FP32Vec16(const FP32Vec16 &data) { + explicit FP32Vec16(const FP32Vec16& data) { reg.val[0] = data.reg.val[0]; reg.val[1] = data.reg.val[1]; reg.val[2] = data.reg.val[2]; reg.val[3] = data.reg.val[3]; } - explicit FP32Vec16(const FP32Vec4 &data) { + explicit FP32Vec16(const FP32Vec4& data) { reg.val[0] = data.reg; reg.val[1] = data.reg; reg.val[2] = data.reg; reg.val[3] = data.reg; } - explicit FP32Vec16(const FP32Vec8 &data) { + explicit FP32Vec16(const FP32Vec8& data) { reg.val[0] = data.reg.val[0]; reg.val[1] = data.reg.val[1]; reg.val[2] = data.reg.val[0]; reg.val[3] = data.reg.val[1]; } - explicit FP32Vec16(const BF16Vec16 &v) { + explicit FP32Vec16(const BF16Vec16& v) { reg.val[0] = (__vector float)vec_mergeh(zero, v.reg.val[0]); reg.val[1] = (__vector float)vec_mergel(zero, v.reg.val[0]); reg.val[2] = (__vector float)vec_mergeh(zero, v.reg.val[1]); reg.val[3] = (__vector float)vec_mergel(zero, v.reg.val[1]); } - explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {} + explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {} - FP32Vec16 operator*(const FP32Vec16 &b) const { - return FP32Vec16(f32x4x4_t({ - vec_mul(reg.val[0], b.reg.val[0]), - vec_mul(reg.val[1], b.reg.val[1]), - vec_mul(reg.val[2], b.reg.val[2]), - vec_mul(reg.val[3], b.reg.val[3])})); + FP32Vec16 operator*(const FP32Vec16& b) const { + return FP32Vec16(f32x4x4_t({vec_mul(reg.val[0], b.reg.val[0]), + vec_mul(reg.val[1], b.reg.val[1]), + vec_mul(reg.val[2], b.reg.val[2]), + vec_mul(reg.val[3], b.reg.val[3])})); } - FP32Vec16 operator+(const FP32Vec16 &b) const { - return FP32Vec16(f32x4x4_t({ - vec_add(reg.val[0], b.reg.val[0]), - vec_add(reg.val[1], b.reg.val[1]), - vec_add(reg.val[2], b.reg.val[2]), - vec_add(reg.val[3], b.reg.val[3])})); + FP32Vec16 operator+(const FP32Vec16& b) const { + return FP32Vec16(f32x4x4_t({vec_add(reg.val[0], b.reg.val[0]), + vec_add(reg.val[1], b.reg.val[1]), + vec_add(reg.val[2], b.reg.val[2]), + vec_add(reg.val[3], b.reg.val[3])})); } - FP32Vec16 operator-(const FP32Vec16 &b) const { - return FP32Vec16(f32x4x4_t({ - vec_sub(reg.val[0], b.reg.val[0]), - vec_sub(reg.val[1], b.reg.val[1]), - vec_sub(reg.val[2], b.reg.val[2]), - vec_sub(reg.val[3], b.reg.val[3])})); + FP32Vec16 operator-(const FP32Vec16& b) const { + return FP32Vec16(f32x4x4_t({vec_sub(reg.val[0], b.reg.val[0]), + vec_sub(reg.val[1], b.reg.val[1]), + vec_sub(reg.val[2], b.reg.val[2]), + vec_sub(reg.val[3], b.reg.val[3])})); } - FP32Vec16 operator/(const FP32Vec16 &b) const { - return FP32Vec16(f32x4x4_t({ - vec_div(reg.val[0], b.reg.val[0]), - vec_div(reg.val[1], b.reg.val[1]), - vec_div(reg.val[2], b.reg.val[2]), - vec_div(reg.val[3], b.reg.val[3])})); + FP32Vec16 operator/(const FP32Vec16& b) const { + return FP32Vec16(f32x4x4_t({vec_div(reg.val[0], b.reg.val[0]), + vec_div(reg.val[1], b.reg.val[1]), + vec_div(reg.val[2], b.reg.val[2]), + vec_div(reg.val[3], b.reg.val[3])})); } float reduce_sum() const { AliasReg ar; ar.reg = reg; float result = 0; - unroll_loop([&result, &ar](int i) { result += ar.values[i]; }); + unroll_loop( + [&result, &ar](int i) { result += ar.values[i]; }); return result; } - template float reduce_sub_sum(int idx) { + template + float reduce_sub_sum(int idx) { static_assert(VEC_ELEM_NUM % group_size == 0); AliasReg ar; @@ -368,7 +371,7 @@ struct FP32Vec16 : public Vec { return result; } - void save(float *ptr) const { + void save(float* ptr) const { vec_xst(reg.val[0], 0, ptr); vec_xst(reg.val[1], 16, ptr); vec_xst(reg.val[2], 32, ptr); @@ -376,43 +379,62 @@ struct FP32Vec16 : public Vec { } }; -template struct VecType { using vec_type = void; }; +template +struct VecType { + using vec_type = void; +}; -template using vec_t = typename VecType::vec_type; +template +using vec_t = typename VecType::vec_type; -template <> struct VecType { using vec_type = FP32Vec8; }; +template <> +struct VecType { + using vec_type = FP32Vec8; +}; -template <> struct VecType { using vec_type = BF16Vec8; }; +template <> +struct VecType { + using vec_type = BF16Vec8; +}; -template void storeFP32(float v, T *ptr) { *ptr = v; } +template +void storeFP32(float v, T* ptr) { + *ptr = v; +} -inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) { +inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) { acc = acc + a * b; } -template <> inline void storeFP32(float v, c10::BFloat16 *ptr) { - c10::BFloat16 __attribute__((__may_alias__)) *v_ptr = - reinterpret_cast(&v); +template <> +inline void storeFP32(float v, c10::BFloat16* ptr) { + c10::BFloat16 __attribute__((__may_alias__))* v_ptr = + reinterpret_cast(&v); *ptr = *(v_ptr + 1); } #ifndef __VEC_CLASS_FP_NAN -#define __VEC_CLASS_FP_NAN (1 << 6) + #define __VEC_CLASS_FP_NAN (1 << 6) #endif -const static __vector unsigned char omask = { 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29 }; +const static __vector unsigned char omask = {0, 1, 4, 5, 8, 9, 12, 13, + 16, 17, 20, 21, 24, 25, 28, 29}; #ifndef _ARCH_PWR10 -const static __vector unsigned int bias = { 0x00007fff, 0x00007fff, 0x00007fff, 0x00007fff }; -const static __vector unsigned int nan = { 0x7fc00000, 0x7fc00000, 0x7fc00000, 0x7fc00000 }; -const static __vector unsigned int sh16 = { 16, 16, 16, 16 }; -const static __vector unsigned int one = { 1, 1, 1, 1 }; +const static __vector unsigned int bias = {0x00007fff, 0x00007fff, 0x00007fff, + 0x00007fff}; +const static __vector unsigned int nan = {0x7fc00000, 0x7fc00000, 0x7fc00000, + 0x7fc00000}; +const static __vector unsigned int sh16 = {16, 16, 16, 16}; +const static __vector unsigned int one = {1, 1, 1, 1}; #endif -inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) { +inline BF16Vec8::BF16Vec8(const FP32Vec8& v) { #ifdef _ARCH_PWR10 __vector signed short ret[2]; - ret[0] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[0]); - ret[1] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[1]); + ret[0] = (__vector signed short)__builtin_vsx_xvcvspbf16( + (__vector unsigned char)v.reg.val[0]); + ret[1] = (__vector signed short)__builtin_vsx_xvcvspbf16( + (__vector unsigned char)v.reg.val[1]); reg = vec_perm(ret[0], ret[1], omask); #elif defined(_ARCH_PWR9) __vector unsigned int inp0 = (__vector unsigned int)(v.reg.val[0]); @@ -425,8 +447,10 @@ inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) { __vector unsigned int rnd1 = vec_add(lsb1, bias); inp0 = vec_add(inp0, rnd0); inp1 = vec_add(inp1, rnd1); - __vector __bool int sel0 = vec_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN); - __vector __bool int sel1 = vec_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN); + __vector __bool int sel0 = + vec_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN); + __vector __bool int sel1 = + vec_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN); inp0 = vec_sel(inp0, nan, sel0); inp1 = vec_sel(inp1, nan, sel1); inp0 = vec_sr(inp0, sh16); @@ -435,13 +459,17 @@ inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) { #endif } -inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) { +inline BF16Vec16::BF16Vec16(const FP32Vec16& v) { #ifdef _ARCH_PWR10 __vector signed short ret[4]; - ret[0] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[0]); - ret[1] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[1]); - ret[2] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[2]); - ret[3] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[3]); + ret[0] = (__vector signed short)__builtin_vsx_xvcvspbf16( + (__vector unsigned char)v.reg.val[0]); + ret[1] = (__vector signed short)__builtin_vsx_xvcvspbf16( + (__vector unsigned char)v.reg.val[1]); + ret[2] = (__vector signed short)__builtin_vsx_xvcvspbf16( + (__vector unsigned char)v.reg.val[2]); + ret[3] = (__vector signed short)__builtin_vsx_xvcvspbf16( + (__vector unsigned char)v.reg.val[3]); reg.val[0] = vec_perm(ret[0], ret[1], omask); reg.val[1] = vec_perm(ret[2], ret[3], omask); #elif defined(_ARCH_PWR9) @@ -465,10 +493,14 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) { inp1 = vec_add(inp1, rnd1); inp2 = vec_add(inp2, rnd2); inp3 = vec_add(inp3, rnd3); - __vector __bool int sel0 = vec_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN); - __vector __bool int sel1 = vec_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN); - __vector __bool int sel2 = vec_test_data_class(v.reg.val[2], __VEC_CLASS_FP_NAN); - __vector __bool int sel3 = vec_test_data_class(v.reg.val[3], __VEC_CLASS_FP_NAN); + __vector __bool int sel0 = + vec_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN); + __vector __bool int sel1 = + vec_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN); + __vector __bool int sel2 = + vec_test_data_class(v.reg.val[2], __VEC_CLASS_FP_NAN); + __vector __bool int sel3 = + vec_test_data_class(v.reg.val[3], __VEC_CLASS_FP_NAN); inp0 = vec_sel(inp0, nan, sel0); inp1 = vec_sel(inp1, nan, sel1); inp2 = vec_sel(inp2, nan, sel2); @@ -482,10 +514,10 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) { #endif } -inline void prefetch(const void *addr) { +inline void prefetch(const void* addr) { __asm__ __volatile__("dcbt 0, %0" : : "r"(addr) : "memory"); } -}; // namespace vec_op +}; // namespace vec_op #endif diff --git a/csrc/cpu/cpu_types_x86.hpp b/csrc/cpu/cpu_types_x86.hpp index 4bb4eb0f491ac..a4ef2be2a58ca 100644 --- a/csrc/cpu/cpu_types_x86.hpp +++ b/csrc/cpu/cpu_types_x86.hpp @@ -11,39 +11,40 @@ static_assert(false, "AVX2 must be supported for the current implementation."); namespace vec_op { -#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...) \ - AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \ - AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \ +#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...) \ + AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \ + AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \ AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__) -#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \ +#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \ AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__)) #ifndef CPU_OP_GUARD -#define CPU_KERNEL_GUARD_IN(NAME) -#define CPU_KERNEL_GUARD_OUT(NAME) + #define CPU_KERNEL_GUARD_IN(NAME) + #define CPU_KERNEL_GUARD_OUT(NAME) #else -#define CPU_KERNEL_GUARD_IN(NAME) \ - RECORD_FUNCTION(#NAME, c10::ArrayRef({})); -#define CPU_KERNEL_GUARD_OUT(NAME) + #define CPU_KERNEL_GUARD_IN(NAME) \ + RECORD_FUNCTION(#NAME, c10::ArrayRef({})); + #define CPU_KERNEL_GUARD_OUT(NAME) #endif #define FORCE_INLINE __attribute__((always_inline)) inline namespace { template -constexpr void unroll_loop_item(std::integer_sequence, F &&f) { +constexpr void unroll_loop_item(std::integer_sequence, F&& f) { (f(std::integral_constant{}), ...); } -}; // namespace +}; // namespace template >> -constexpr void unroll_loop(F &&f) { +constexpr void unroll_loop(F&& f) { unroll_loop_item(std::make_integer_sequence{}, std::forward(f)); } -template struct Vec { +template +struct Vec { constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; } }; @@ -55,12 +56,12 @@ struct FP16Vec8 : public Vec { __m128i reg; - explicit FP16Vec8(const void *ptr) - : reg((__m128i)_mm_loadu_si128((__m128i *)ptr)) {} + explicit FP16Vec8(const void* ptr) + : reg((__m128i)_mm_loadu_si128((__m128i*)ptr)) {} - explicit FP16Vec8(const FP32Vec8 &); + explicit FP16Vec8(const FP32Vec8&); - void save(void *ptr) const { *reinterpret_cast<__m128i *>(ptr) = reg; } + void save(void* ptr) const { *reinterpret_cast<__m128i*>(ptr) = reg; } }; struct FP16Vec16 : public Vec { @@ -68,12 +69,12 @@ struct FP16Vec16 : public Vec { __m256i reg; - explicit FP16Vec16(const void *ptr) - : reg((__m256i)_mm256_loadu_si256((__m256i *)ptr)) {} + explicit FP16Vec16(const void* ptr) + : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {} - explicit FP16Vec16(const FP32Vec16 &); + explicit FP16Vec16(const FP32Vec16&); - void save(void *ptr) const { *reinterpret_cast<__m256i *>(ptr) = reg; } + void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; } void save(void* ptr, const int elem_num) const { constexpr uint32_t M = 0xFFFFFFFF; @@ -87,12 +88,12 @@ struct BF16Vec8 : public Vec { __m128i reg; - explicit BF16Vec8(const void *ptr) - : reg((__m128i)_mm_loadu_si128((__m128i *)ptr)) {} + explicit BF16Vec8(const void* ptr) + : reg((__m128i)_mm_loadu_si128((__m128i*)ptr)) {} - explicit BF16Vec8(const FP32Vec8 &); + explicit BF16Vec8(const FP32Vec8&); - void save(void *ptr) const { *reinterpret_cast<__m128i *>(ptr) = reg; } + void save(void* ptr) const { *reinterpret_cast<__m128i*>(ptr) = reg; } }; struct BF16Vec16 : public Vec { @@ -100,12 +101,12 @@ struct BF16Vec16 : public Vec { __m256i reg; - explicit BF16Vec16(const void *ptr) - : reg((__m256i)_mm256_loadu_si256((__m256i *)ptr)) {} + explicit BF16Vec16(const void* ptr) + : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {} - explicit BF16Vec16(const FP32Vec16 &); + explicit BF16Vec16(const FP32Vec16&); - void save(void *ptr) const { *reinterpret_cast<__m256i *>(ptr) = reg; } + void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; } void save(void* ptr, const int elem_num) const { constexpr uint32_t M = 0xFFFFFFFF; @@ -120,11 +121,11 @@ struct BF16Vec32 : public Vec { __m512i reg; - explicit BF16Vec32(const void *ptr) : reg((__m512i)_mm512_loadu_si512(ptr)) {} + explicit BF16Vec32(const void* ptr) : reg((__m512i)_mm512_loadu_si512(ptr)) {} explicit BF16Vec32(__m512i data) : reg(data) {} - explicit BF16Vec32(BF16Vec8 &vec8_data) + explicit BF16Vec32(BF16Vec8& vec8_data) : reg((__m512i)_mm512_inserti32x4( _mm512_inserti32x4(_mm512_inserti32x4(_mm512_castsi128_si512( (__m128i)vec8_data.reg), @@ -132,7 +133,7 @@ struct BF16Vec32 : public Vec { (__m128i)vec8_data.reg, 2), (__m128i)vec8_data.reg, 3)) {} - void save(void *ptr) const { *reinterpret_cast<__m512i *>(ptr) = reg; } + void save(void* ptr) const { *reinterpret_cast<__m512i*>(ptr) = reg; } }; #else struct BF16Vec32 : public Vec { @@ -141,24 +142,24 @@ struct BF16Vec32 : public Vec { __m256i reg_low; __m256i reg_high; - explicit BF16Vec32(const void *ptr) - : reg_low(_mm256_loadu_si256((__m256i const *)ptr)), - reg_high(_mm256_loadu_si256((__m256i const *)ptr + 1)) {} + explicit BF16Vec32(const void* ptr) + : reg_low(_mm256_loadu_si256((__m256i const*)ptr)), + reg_high(_mm256_loadu_si256((__m256i const*)ptr + 1)) {} - explicit BF16Vec32(__m256i low, __m256i high) : reg_low(low), - reg_high(high) {} + explicit BF16Vec32(__m256i low, __m256i high) + : reg_low(low), reg_high(high) {} - explicit BF16Vec32(BF16Vec8 &vec8_data) + explicit BF16Vec32(BF16Vec8& vec8_data) : reg_low((__m256i)_mm256_inserti32x4( - _mm256_castsi128_si256((__m128i)vec8_data.reg), - (__m128i)vec8_data.reg, 1)), + _mm256_castsi128_si256((__m128i)vec8_data.reg), + (__m128i)vec8_data.reg, 1)), reg_high((__m256i)_mm256_inserti32x4( - _mm256_castsi128_si256((__m128i)vec8_data.reg), - (__m128i)vec8_data.reg, 1)) {} + _mm256_castsi128_si256((__m128i)vec8_data.reg), + (__m128i)vec8_data.reg, 1)) {} - void save(void *ptr) const { - *reinterpret_cast<__m256i *>(ptr) = reg_low; - *reinterpret_cast<__m256i *>((__m256i *)ptr + 1) = reg_high; + void save(void* ptr) const { + *reinterpret_cast<__m256i*>(ptr) = reg_low; + *reinterpret_cast<__m256i*>((__m256i*)ptr + 1) = reg_high; } }; #endif @@ -176,11 +177,11 @@ struct FP32Vec4 : public Vec { explicit FP32Vec4() : reg(_mm_set1_ps(0.0)) {} - explicit FP32Vec4(const float *ptr) : reg(_mm_loadu_ps(ptr)) {} + explicit FP32Vec4(const float* ptr) : reg(_mm_loadu_ps(ptr)) {} explicit FP32Vec4(__m128 data) : reg(data) {} - explicit FP32Vec4(const FP32Vec4 &data) : reg(data.reg) {} + explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {} }; struct FP32Vec8 : public Vec { @@ -196,15 +197,15 @@ struct FP32Vec8 : public Vec { explicit FP32Vec8() : reg(_mm256_set1_ps(0.0)) {} - explicit FP32Vec8(const float *ptr) : reg(_mm256_loadu_ps(ptr)) {} + explicit FP32Vec8(const float* ptr) : reg(_mm256_loadu_ps(ptr)) {} explicit FP32Vec8(__m256 data) : reg(data) {} - explicit FP32Vec8(const FP32Vec8 &data) : reg(data.reg) {} + explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {} - explicit FP32Vec8(const FP16Vec8 &v) : reg(_mm256_cvtph_ps(v.reg)) {} + explicit FP32Vec8(const FP16Vec8& v) : reg(_mm256_cvtph_ps(v.reg)) {} - explicit FP32Vec8(const BF16Vec8 &v) + explicit FP32Vec8(const BF16Vec8& v) : reg(_mm256_castsi256_ps( _mm256_bslli_epi128(_mm256_cvtepu16_epi32(v.reg), 2))) {} @@ -212,7 +213,8 @@ struct FP32Vec8 : public Vec { AliasReg ar; ar.reg = reg; float result = 0; - unroll_loop([&result, &ar](int i) { result += ar.values[i]; }); + unroll_loop( + [&result, &ar](int i) { result += ar.values[i]; }); return result; } @@ -244,27 +246,27 @@ struct FP32Vec8 : public Vec { erf(ar.values[1]), erf(ar.values[0]))); } - FP32Vec8 operator*(const FP32Vec8 &b) const { + FP32Vec8 operator*(const FP32Vec8& b) const { return FP32Vec8(_mm256_mul_ps(reg, b.reg)); } - FP32Vec8 operator+(const FP32Vec8 &b) const { + FP32Vec8 operator+(const FP32Vec8& b) const { return FP32Vec8(_mm256_add_ps(reg, b.reg)); } - FP32Vec8 operator-(const FP32Vec8 &b) const { + FP32Vec8 operator-(const FP32Vec8& b) const { return FP32Vec8(_mm256_sub_ps(reg, b.reg)); } - FP32Vec8 operator/(const FP32Vec8 &b) const { + FP32Vec8 operator/(const FP32Vec8& b) const { return FP32Vec8(_mm256_div_ps(reg, b.reg)); } - void save(float *ptr) const { _mm256_storeu_ps(ptr, reg); } + void save(float* ptr) const { _mm256_storeu_ps(ptr, reg); } }; #ifdef __AVX512F__ -struct INT32Vec16: public Vec { +struct INT32Vec16 : public Vec { constexpr static int VEC_ELEM_NUM = 16; union AliasReg { __m512i reg; @@ -272,12 +274,11 @@ struct INT32Vec16: public Vec { }; __m512i reg; - - explicit INT32Vec16(const void* data_ptr) : reg(_mm512_loadu_epi32(data_ptr)) {} - void save(int32_t* ptr) const { - _mm512_storeu_epi32(ptr, reg); - } + explicit INT32Vec16(const void* data_ptr) + : reg(_mm512_loadu_epi32(data_ptr)) {} + + void save(int32_t* ptr) const { _mm512_storeu_epi32(ptr, reg); } void save(int32_t* ptr, const int elem_num) const { constexpr uint32_t M = 0xFFFFFFFF; @@ -301,11 +302,11 @@ struct FP32Vec16 : public Vec { explicit FP32Vec16() : reg(_mm512_set1_ps(0.0)) {} - explicit FP32Vec16(const float *ptr) : reg(_mm512_loadu_ps(ptr)) {} + explicit FP32Vec16(const float* ptr) : reg(_mm512_loadu_ps(ptr)) {} explicit FP32Vec16(__m512 data) : reg(data) {} - explicit FP32Vec16(const FP32Vec4 &data) + explicit FP32Vec16(const FP32Vec4& data) : reg((__m512)_mm512_inserti32x4( _mm512_inserti32x4( _mm512_inserti32x4(_mm512_castsi128_si512((__m128i)data.reg), @@ -313,36 +314,37 @@ struct FP32Vec16 : public Vec { (__m128i)data.reg, 2), (__m128i)data.reg, 3)) {} - explicit FP32Vec16(const FP32Vec8 &data) + explicit FP32Vec16(const FP32Vec8& data) : reg((__m512)_mm512_inserti32x8( _mm512_castsi256_si512((__m256i)data.reg), (__m256i)data.reg, 1)) {} - explicit FP32Vec16(const BF16Vec16 &v) + explicit FP32Vec16(const BF16Vec16& v) : reg(_mm512_castsi512_ps( _mm512_bslli_epi128(_mm512_cvtepu16_epi32(v.reg), 2))) {} - explicit FP32Vec16(const FP16Vec16 &v) : reg(_mm512_cvtph_ps(v.reg)) {} + explicit FP32Vec16(const FP16Vec16& v) : reg(_mm512_cvtph_ps(v.reg)) {} - explicit FP32Vec16(const FP16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {} + explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v)) {} - explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {} + explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {} - explicit FP32Vec16(const INT32Vec16 &v) - : reg(_mm512_cvt_roundepi32_ps(v.reg, _MM_FROUND_TO_NEAREST_INT |_MM_FROUND_NO_EXC)) {} + explicit FP32Vec16(const INT32Vec16& v) + : reg(_mm512_cvt_roundepi32_ps( + v.reg, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)) {} - FP32Vec16 operator*(const FP32Vec16 &b) const { + FP32Vec16 operator*(const FP32Vec16& b) const { return FP32Vec16(_mm512_mul_ps(reg, b.reg)); } - FP32Vec16 operator+(const FP32Vec16 &b) const { + FP32Vec16 operator+(const FP32Vec16& b) const { return FP32Vec16(_mm512_add_ps(reg, b.reg)); } - FP32Vec16 operator-(const FP32Vec16 &b) const { + FP32Vec16 operator-(const FP32Vec16& b) const { return FP32Vec16(_mm512_sub_ps(reg, b.reg)); } - FP32Vec16 operator/(const FP32Vec16 &b) const { + FP32Vec16 operator/(const FP32Vec16& b) const { return FP32Vec16(_mm512_div_ps(reg, b.reg)); } @@ -370,9 +372,7 @@ struct FP32Vec16 : public Vec { return FP32Vec16(_mm512_mask_min_ps(reg, mask, reg, b.reg)); } - FP32Vec16 abs() const { - return FP32Vec16(_mm512_abs_ps(reg)); - } + FP32Vec16 abs() const { return FP32Vec16(_mm512_abs_ps(reg)); } float reduce_sum() const { return _mm512_reduce_add_ps(reg); } @@ -380,14 +380,15 @@ struct FP32Vec16 : public Vec { float reduce_min() const { return _mm512_reduce_min_ps(reg); } - template float reduce_sub_sum(int idx) { + template + float reduce_sub_sum(int idx) { static_assert(VEC_ELEM_NUM % group_size == 0); constexpr uint32_t base_mask = (0xFFFF >> (16 - group_size)); __mmask16 mask = _cvtu32_mask16(base_mask << (idx * group_size)); return _mm512_mask_reduce_add_ps(mask, reg); } - void save(float *ptr) const { _mm512_storeu_ps(ptr, reg); } + void save(float* ptr) const { _mm512_storeu_ps(ptr, reg); } void save(float* ptr, const int elem_num) const { constexpr uint32_t M = 0xFFFFFFFF; @@ -407,32 +408,30 @@ struct FP32Vec16 : public Vec { __m256 reg_low; __m256 reg_high; - explicit FP32Vec16(float v) : reg_low(_mm256_set1_ps(v)), - reg_high(_mm256_set1_ps(v)) {} + explicit FP32Vec16(float v) + : reg_low(_mm256_set1_ps(v)), reg_high(_mm256_set1_ps(v)) {} - explicit FP32Vec16() : reg_low(_mm256_set1_ps(0.0)), - reg_high(_mm256_set1_ps(0.0)) {} + explicit FP32Vec16() + : reg_low(_mm256_set1_ps(0.0)), reg_high(_mm256_set1_ps(0.0)) {} - explicit FP32Vec16(const float *ptr) : reg_low(_mm256_loadu_ps(ptr)), - reg_high(_mm256_loadu_ps(ptr + 8)) {} + explicit FP32Vec16(const float* ptr) + : reg_low(_mm256_loadu_ps(ptr)), reg_high(_mm256_loadu_ps(ptr + 8)) {} explicit FP32Vec16(__m256 low, __m256 high) : reg_low(low), reg_high(high) {} - explicit FP32Vec16(const FP32Vec16 &data) : reg_low(data.reg_low), - reg_high(data.reg_high) {} + explicit FP32Vec16(const FP32Vec16& data) + : reg_low(data.reg_low), reg_high(data.reg_high) {} - explicit FP32Vec16(const FP32Vec4 &data) + explicit FP32Vec16(const FP32Vec4& data) : reg_low((__m256)_mm256_inserti128_si256( - _mm256_castsi128_si256((__m128i)data.reg), - (__m128i)data.reg, 1)), + _mm256_castsi128_si256((__m128i)data.reg), (__m128i)data.reg, 1)), reg_high((__m256)_mm256_inserti128_si256( - _mm256_castsi128_si256((__m128i)data.reg), - (__m128i)data.reg, 1)) {} + _mm256_castsi128_si256((__m128i)data.reg), (__m128i)data.reg, 1)) {} - explicit FP32Vec16(const FP32Vec8 &data) + explicit FP32Vec16(const FP32Vec8& data) : reg_low(data.reg), reg_high(data.reg) {} - explicit FP32Vec16(const FP16Vec16 &v) { + explicit FP32Vec16(const FP16Vec16& v) { __m128i low = _mm256_extractf128_si256(v.reg, 0); __m128i high = _mm256_extractf128_si256(v.reg, 1); @@ -440,9 +439,9 @@ struct FP32Vec16 : public Vec { reg_high = _mm256_cvtph_ps(high); } - explicit FP32Vec16(const FP16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {} + explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v)) {} - explicit FP32Vec16(const BF16Vec16 &v) { + explicit FP32Vec16(const BF16Vec16& v) { __m128i low = _mm256_extractf128_si256(v.reg, 0); __m128i high = _mm256_extractf128_si256(v.reg, 1); @@ -456,24 +455,24 @@ struct FP32Vec16 : public Vec { reg_high = _mm256_castsi256_ps(v_high_shifted); } - explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {} + explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {} - FP32Vec16 operator*(const FP32Vec16 &b) const { + FP32Vec16 operator*(const FP32Vec16& b) const { return FP32Vec16(_mm256_mul_ps(reg_low, b.reg_low), _mm256_mul_ps(reg_high, b.reg_high)); } - FP32Vec16 operator+(const FP32Vec16 &b) const { + FP32Vec16 operator+(const FP32Vec16& b) const { return FP32Vec16(_mm256_add_ps(reg_low, b.reg_low), _mm256_add_ps(reg_high, b.reg_high)); } - FP32Vec16 operator-(const FP32Vec16 &b) const { + FP32Vec16 operator-(const FP32Vec16& b) const { return FP32Vec16(_mm256_sub_ps(reg_low, b.reg_low), _mm256_sub_ps(reg_high, b.reg_high)); } - FP32Vec16 operator/(const FP32Vec16 &b) const { + FP32Vec16 operator/(const FP32Vec16& b) const { return FP32Vec16(_mm256_div_ps(reg_low, b.reg_low), _mm256_div_ps(reg_high, b.reg_high)); } @@ -484,7 +483,8 @@ struct FP32Vec16 : public Vec { return low.reduce_sum() + high.reduce_sum(); } - template float reduce_sub_sum(int idx) { + template + float reduce_sub_sum(int idx) { float sum = 0.0; static_assert(VEC_ELEM_NUM % group_size == 0); constexpr uint32_t base_mask = (0xFFFF >> (16 - group_size)); @@ -507,7 +507,7 @@ struct FP32Vec16 : public Vec { return sum; } - void save(float *ptr) const { + void save(float* ptr) const { _mm256_storeu_ps(ptr, reg_low); _mm256_storeu_ps(ptr + 8, reg_high); } @@ -515,7 +515,7 @@ struct FP32Vec16 : public Vec { #endif #ifdef __AVX512F__ -struct INT8Vec16: public Vec { +struct INT8Vec16 : public Vec { constexpr static int VEC_ELEM_NUM = 16; union AliasReg { __m128i reg; @@ -523,14 +523,12 @@ struct INT8Vec16: public Vec { }; __m128i reg; - - explicit INT8Vec16(const FP32Vec16& vec) : reg( - _mm512_cvtepi32_epi8(_mm512_cvt_roundps_epi32(vec.reg, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)) - ) {} - void save(int8_t* ptr) const { - _mm_storeu_epi8(ptr, reg); - } + explicit INT8Vec16(const FP32Vec16& vec) + : reg(_mm512_cvtepi32_epi8(_mm512_cvt_roundps_epi32( + vec.reg, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))) {} + + void save(int8_t* ptr) const { _mm_storeu_epi8(ptr, reg); } void save(int8_t* ptr, const int elem_num) const { constexpr uint32_t M = 0xFFFFFFFF; @@ -540,71 +538,92 @@ struct INT8Vec16: public Vec { }; #endif -template struct VecType { using vec_type = void; }; +template +struct VecType { + using vec_type = void; +}; -template using vec_t = typename VecType::vec_type; +template +using vec_t = typename VecType::vec_type; -template <> struct VecType { using vec_type = FP32Vec8; }; +template <> +struct VecType { + using vec_type = FP32Vec8; +}; -template <> struct VecType { using vec_type = FP16Vec8; }; +template <> +struct VecType { + using vec_type = FP16Vec8; +}; -template <> struct VecType { using vec_type = BF16Vec8; }; +template <> +struct VecType { + using vec_type = BF16Vec8; +}; -template void storeFP32(float v, T *ptr) { *ptr = v; } +template +void storeFP32(float v, T* ptr) { + *ptr = v; +} -inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) { +inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) { acc = acc + a * b; } -template <> inline void storeFP32(float v, c10::Half *ptr) { - *reinterpret_cast(ptr) = +template <> +inline void storeFP32(float v, c10::Half* ptr) { + *reinterpret_cast(ptr) = _cvtss_sh(v, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC); } -inline FP16Vec8::FP16Vec8(const FP32Vec8 &v) +inline FP16Vec8::FP16Vec8(const FP32Vec8& v) : reg(_mm256_cvtps_ph(v.reg, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)) {} #ifdef __AVX512F__ -inline FP16Vec16::FP16Vec16(const FP32Vec16 &v) +inline FP16Vec16::FP16Vec16(const FP32Vec16& v) : reg(_mm512_cvtps_ph(v.reg, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)) {} #else -inline FP16Vec16::FP16Vec16(const FP32Vec16 &v) - : reg(_mm256_insertf128_si256(_mm256_castsi128_si256(FP16Vec8(FP32Vec8(v.reg_low)).reg), FP16Vec8(FP32Vec8(v.reg_low)).reg, 1)) {} +inline FP16Vec16::FP16Vec16(const FP32Vec16& v) + : reg(_mm256_insertf128_si256( + _mm256_castsi128_si256(FP16Vec8(FP32Vec8(v.reg_low)).reg), + FP16Vec8(FP32Vec8(v.reg_low)).reg, 1)) {} #endif #ifdef __AVX512BF16__ -template <> inline void storeFP32(float v, c10::BFloat16 *ptr) { - *reinterpret_cast<__bfloat16 *>(ptr) = _mm_cvtness_sbh(v); +template <> +inline void storeFP32(float v, c10::BFloat16* ptr) { + *reinterpret_cast<__bfloat16*>(ptr) = _mm_cvtness_sbh(v); } -inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) +inline BF16Vec8::BF16Vec8(const FP32Vec8& v) : reg((__m128i)_mm256_cvtneps_pbh(v.reg)) {} -inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) +inline BF16Vec16::BF16Vec16(const FP32Vec16& v) : reg((__m256i)_mm512_cvtneps_pbh(v.reg)) {} -inline void fma(FP32Vec16 &acc, BF16Vec32 &a, BF16Vec32 &b) { +inline void fma(FP32Vec16& acc, BF16Vec32& a, BF16Vec32& b) { acc.reg = _mm512_dpbf16_ps(acc.reg, (__m512bh)a.reg, (__m512bh)b.reg); } #else -template <> inline void storeFP32(float v, c10::BFloat16 *ptr) { - c10::BFloat16 __attribute__((__may_alias__)) *v_ptr = - reinterpret_cast(&v); +template <> +inline void storeFP32(float v, c10::BFloat16* ptr) { + c10::BFloat16 __attribute__((__may_alias__))* v_ptr = + reinterpret_cast(&v); *ptr = *(v_ptr + 1); } -#ifdef __AVX512F__ -inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) + #ifdef __AVX512F__ +inline BF16Vec8::BF16Vec8(const FP32Vec8& v) : reg(_mm256_cvtepi32_epi16( _mm256_bsrli_epi128(_mm256_castps_si256(v.reg), 2))) {} -inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) +inline BF16Vec16::BF16Vec16(const FP32Vec16& v) : reg(_mm512_cvtepi32_epi16( _mm512_bsrli_epi128(_mm512_castps_si512(v.reg), 2))) {} -#else -namespace{ + #else +namespace { __m128i FP32Vec8_to_BF16Vec8_avx2(__m256 a) { __m256i ai = _mm256_castps_si256(a); ai = _mm256_srli_epi32(ai, 16); @@ -612,21 +631,21 @@ __m128i FP32Vec8_to_BF16Vec8_avx2(__m256 a) { ai = _mm256_permute4x64_epi64(ai, 0b00111001); return _mm256_extracti128_si256(ai, 0); } -} +} // namespace -inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) +inline BF16Vec8::BF16Vec8(const FP32Vec8& v) : reg(FP32Vec8_to_BF16Vec8_avx2(v.reg)) {} -inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) { +inline BF16Vec16::BF16Vec16(const FP32Vec16& v) { BF16Vec8 low = BF16Vec8(FP32Vec8(v.reg_low)); BF16Vec8 high = BF16Vec8(FP32Vec8(v.reg_high)); reg = _mm256_insertf128_si256(_mm256_castsi128_si256(low.reg), high.reg, 1); } -#endif // __AVX512F__ -#endif // __AVX512BF16__ + #endif // __AVX512F__ +#endif // __AVX512BF16__ -inline void prefetch(const void *addr) { _mm_prefetch(addr, _MM_HINT_T1); } +inline void prefetch(const void* addr) { _mm_prefetch(addr, _MM_HINT_T1); } -}; // namespace vec_op +}; // namespace vec_op #endif diff --git a/csrc/cpu/torch_bindings.cpp b/csrc/cpu/torch_bindings.cpp index 74e4d8189d403..5d1c5f4c83d3e 100644 --- a/csrc/cpu/torch_bindings.cpp +++ b/csrc/cpu/torch_bindings.cpp @@ -30,7 +30,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) { " Tensor value_cache, int num_kv_heads, float scale," " Tensor block_tables, Tensor seq_lens, int block_size," " int max_seq_len, Tensor? alibi_slopes," - " str kv_cache_dtype, float k_scale, float v_scale," + " str kv_cache_dtype, Tensor k_scale, Tensor v_scale," " int tp_rank, int blocksparse_local_blocks," " int blocksparse_vert_stride, int blocksparse_block_size," " int blocksparse_head_sliding_step) -> ()"); @@ -44,7 +44,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) { " Tensor value_cache, int num_kv_heads, float scale," " Tensor block_tables, Tensor seq_lens, int block_size," " int max_seq_len, Tensor? alibi_slopes," - " str kv_cache_dtype, float k_scale, float v_scale," + " str kv_cache_dtype, Tensor k_scale, Tensor v_scale," " int tp_rank, int blocksparse_local_blocks," " int blocksparse_vert_stride, int blocksparse_block_size," " int blocksparse_head_sliding_step) -> ()"); @@ -148,7 +148,7 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) { " Tensor! key_cache, Tensor! value_cache," " Tensor slot_mapping," " str kv_cache_dtype," - " float k_scale, float v_scale) -> ()"); + " Tensor k_scale, Tensor v_scale) -> ()"); cache_ops.impl("reshape_and_cache", torch::kCPU, &reshape_and_cache); } diff --git a/csrc/cumem_allocator.cpp b/csrc/cumem_allocator.cpp new file mode 100644 index 0000000000000..e8555d853b7ac --- /dev/null +++ b/csrc/cumem_allocator.cpp @@ -0,0 +1,310 @@ +// A CUDAPluggableAllocator based on cumem* APIs. +// Important: allocation size, CUdeviceptr and CUmemGenericAllocationHandle* +// need to be unsigned long long +#include + +extern "C" { + +#define PY_SSIZE_T_CLEAN +#include + +#include +#include +#include + +#define CUDA_CHECK(condition) \ + do { \ + CUresult error = condition; \ + if (error != 0) { \ + char* error_string; \ + cuGetErrorString(error, (const char**)&error_string); \ + std::cerr << "CUDA Error: " << error_string << " at " << __FILE__ << ":" \ + << __LINE__ << std::endl; \ + } \ + } while (0) + +// Global references to Python callables +// NOTE: this is borrowed reference, so we don't need to DECREF them. +// This brings the limitation that the allocator needs to be singleton. +static PyObject* g_python_malloc_callback = nullptr; +static PyObject* g_python_free_callback = nullptr; + +// --------------------------------------------------------------------------- +// Helper functions: + +void ensure_context(unsigned long long device) { + CUcontext pctx; + CUDA_CHECK(cuCtxGetCurrent(&pctx)); + if (!pctx) { + // Ensure device context. + CUDA_CHECK(cuDevicePrimaryCtxRetain(&pctx, device)); + CUDA_CHECK(cuCtxSetCurrent(pctx)); + } +} + +void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem, + CUmemGenericAllocationHandle* p_memHandle) { + ensure_context(device); + // Define memory allocation properties + CUmemAllocationProp prop = {}; + prop.type = CU_MEM_ALLOCATION_TYPE_PINNED; + prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE; + prop.location.id = device; + prop.allocFlags.compressionType = CU_MEM_ALLOCATION_COMP_NONE; + + // Allocate memory using cuMemCreate + CUDA_CHECK(cuMemCreate(p_memHandle, size, &prop, 0)); + CUDA_CHECK(cuMemMap(d_mem, size, 0, *p_memHandle, 0)); + + CUmemAccessDesc accessDesc = {}; + accessDesc.location.type = CU_MEM_LOCATION_TYPE_DEVICE; + accessDesc.location.id = device; + accessDesc.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE; + + CUDA_CHECK(cuMemSetAccess(d_mem, size, &accessDesc, 1)); + // std::cout << "create_and_map: device=" << device << ", size=" << size << ", + // d_mem=" << d_mem << ", p_memHandle=" << p_memHandle << std::endl; +} + +void unmap_and_release(unsigned long long device, ssize_t size, + CUdeviceptr d_mem, + CUmemGenericAllocationHandle* p_memHandle) { + // std::cout << "unmap_and_release: device=" << device << ", size=" << size << + // ", d_mem=" << d_mem << ", p_memHandle=" << p_memHandle << std::endl; + ensure_context(device); + CUDA_CHECK(cuMemUnmap(d_mem, size)); + CUDA_CHECK(cuMemRelease(*p_memHandle)); +} + +PyObject* create_tuple_from_c_integers(unsigned long long a, + unsigned long long b, + unsigned long long c, + unsigned long long d) { + // Create a new tuple of size 4 + PyObject* tuple = PyTuple_New(4); + if (!tuple) { + return NULL; // Return NULL on failure + } + + // Convert integers to Python objects and set them in the tuple + PyTuple_SetItem( + tuple, 0, + PyLong_FromUnsignedLongLong(a)); // Steals reference to the PyLong + PyTuple_SetItem(tuple, 1, PyLong_FromUnsignedLongLong(b)); + PyTuple_SetItem(tuple, 2, PyLong_FromUnsignedLongLong(c)); + PyTuple_SetItem(tuple, 3, PyLong_FromUnsignedLongLong(d)); + + // Note: PyTuple_SetItem "steals" a reference to each object, + // so we do not need to Py_DECREF the PyLong objects explicitly. + + return tuple; // Return the created tuple +} + +// --------------------------------------------------------------------------- +// Our exported C functions that call Python: + +// use CUstream instead of cudaStream_t, to avoid including cuda_runtime_api.h +void* my_malloc(ssize_t size, int device, CUstream stream) { + ensure_context(device); + + // first allocation, align the size, and reserve an address, and also allocate + // a CUmemGenericAllocationHandle + + // Define memory allocation properties + CUmemAllocationProp prop = {}; + prop.type = CU_MEM_ALLOCATION_TYPE_PINNED; + prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE; + prop.location.id = device; + prop.allocFlags.compressionType = CU_MEM_ALLOCATION_COMP_NONE; + + // Check if the allocation is supported + size_t granularity; + CUDA_CHECK(cuMemGetAllocationGranularity(&granularity, &prop, + CU_MEM_ALLOC_GRANULARITY_MINIMUM)); + + size_t alignedSize = ((size + granularity - 1) / granularity) * granularity; + + CUdeviceptr d_mem; + CUDA_CHECK(cuMemAddressReserve(&d_mem, alignedSize, 0, 0, 0)); + + // allocate the CUmemGenericAllocationHandle + CUmemGenericAllocationHandle* p_memHandle = + (CUmemGenericAllocationHandle*)malloc( + sizeof(CUmemGenericAllocationHandle)); + + if (!g_python_malloc_callback) { + std::cerr << "ERROR: g_python_malloc_callback not set.\n"; + return nullptr; + } + + // Acquire GIL (not in stable ABI officially, but often works) + PyGILState_STATE gstate = PyGILState_Ensure(); + + PyObject* arg_tuple = create_tuple_from_c_integers( + (unsigned long long)device, (unsigned long long)alignedSize, + (unsigned long long)d_mem, (unsigned long long)p_memHandle); + + // Call g_python_malloc_callback + PyObject* py_result = + PyObject_CallFunctionObjArgs(g_python_malloc_callback, arg_tuple, NULL); + Py_DECREF(arg_tuple); + + if (!py_result) { + PyErr_Print(); + PyGILState_Release(gstate); + return nullptr; + } + + PyGILState_Release(gstate); + + // do the final mapping + create_and_map(device, alignedSize, d_mem, p_memHandle); + + return (void*)d_mem; +} + +// use CUstream instead of cudaStream_t, to avoid including cuda_runtime_api.h +void my_free(void* ptr, ssize_t size, int device, CUstream stream) { + // get memory handle from the pointer + if (!g_python_free_callback) { + std::cerr << "ERROR: g_python_free_callback not set.\n"; + return; + } + + // Acquire GIL (not in stable ABI officially, but often works) + PyGILState_STATE gstate = PyGILState_Ensure(); + + PyObject* py_ptr = + PyLong_FromUnsignedLongLong(reinterpret_cast(ptr)); + + PyObject* py_result = + PyObject_CallFunctionObjArgs(g_python_free_callback, py_ptr, NULL); + + if (!py_result || !PyTuple_Check(py_result) || PyTuple_Size(py_result) != 4) { + PyErr_SetString(PyExc_TypeError, "Expected a tuple of size 4"); + return; + } + + unsigned long long recv_device, recv_size; + unsigned long long recv_d_mem, recv_p_memHandle; + // Unpack the tuple into four C integers + if (!PyArg_ParseTuple(py_result, "KKKK", &recv_device, &recv_size, + &recv_d_mem, &recv_p_memHandle)) { + // PyArg_ParseTuple sets an error if it fails + return; + } + + PyGILState_Release(gstate); + + // recv_size == size + // recv_device == device + + // Free memory + + CUdeviceptr d_mem = (CUdeviceptr)recv_d_mem; + CUmemGenericAllocationHandle* p_memHandle = + (CUmemGenericAllocationHandle*)recv_p_memHandle; + unmap_and_release(device, size, d_mem, p_memHandle); + + // free address and the handle + CUDA_CHECK(cuMemAddressFree(d_mem, size)); + free(p_memHandle); +} + +// --------------------------------------------------------------------------- +// Python extension boilerplate: + +// Python-exposed function: init_module(python_malloc, python_free) +static PyObject* py_init_module(PyObject* self, PyObject* args) { + PyObject* malloc_callback = nullptr; + PyObject* free_callback = nullptr; + + if (!PyArg_ParseTuple(args, "OO", &malloc_callback, &free_callback)) { + return nullptr; + } + + if (!PyCallable_Check(malloc_callback) || !PyCallable_Check(free_callback)) { + PyErr_SetString(PyExc_TypeError, "Both arguments must be callables"); + return nullptr; + } + + // Save the Python callables + // This module does not handle GC of these objects, so they must be kept alive + // outside of this module. + g_python_malloc_callback = malloc_callback; + g_python_free_callback = free_callback; + + Py_RETURN_NONE; +} + +static PyObject* python_unmap_and_release(PyObject* self, PyObject* args) { + if (!args || !PyTuple_Check(args) || PyTuple_Size(args) != 4) { + PyErr_SetString(PyExc_TypeError, "Expected a tuple of size 4"); + return nullptr; + } + + unsigned long long recv_device, recv_size; + unsigned long long recv_d_mem, recv_p_memHandle; + // Unpack the tuple into four C integers + if (!PyArg_ParseTuple(args, "KKKK", &recv_device, &recv_size, &recv_d_mem, + &recv_p_memHandle)) { + // PyArg_ParseTuple sets an error if it fails + return nullptr; + } + + CUdeviceptr d_mem_ptr = (CUdeviceptr)recv_d_mem; + CUmemGenericAllocationHandle* p_memHandle = + (CUmemGenericAllocationHandle*)recv_p_memHandle; + + unmap_and_release(recv_device, recv_size, d_mem_ptr, p_memHandle); + + Py_RETURN_NONE; +} + +static PyObject* python_create_and_map(PyObject* self, PyObject* args) { + if (!args || !PyTuple_Check(args) || PyTuple_Size(args) != 4) { + PyErr_SetString(PyExc_TypeError, "Expected a tuple of size 4"); + return nullptr; + } + + unsigned long long recv_device, recv_size; + unsigned long long recv_d_mem, recv_p_memHandle; + // Unpack the tuple into four C integers + if (!PyArg_ParseTuple(args, "KKKK", &recv_device, &recv_size, &recv_d_mem, + &recv_p_memHandle)) { + // PyArg_ParseTuple sets an error if it fails + return nullptr; + } + + CUdeviceptr d_mem_ptr = (CUdeviceptr)recv_d_mem; + CUmemGenericAllocationHandle* p_memHandle = + (CUmemGenericAllocationHandle*)recv_p_memHandle; + + create_and_map(recv_device, recv_size, d_mem_ptr, p_memHandle); + + Py_RETURN_NONE; +} + +static PyMethodDef module_methods[] = { + {"init_module", (PyCFunction)py_init_module, METH_VARARGS, + "Initialize module with python_malloc and python_free callables."}, + {"python_create_and_map", (PyCFunction)python_create_and_map, METH_VARARGS, + "Create and map memory on the device."}, + {"python_unmap_and_release", (PyCFunction)python_unmap_and_release, + METH_VARARGS, "Unmap and release memory on the device."}, + {NULL, NULL, 0, NULL} // sentinel +}; + +static struct PyModuleDef cumem_allocator_module = { + PyModuleDef_HEAD_INIT, "cumem_allocator", + "cumem-based allocator for CUDAPluggableAllocator", -1, module_methods}; + +PyMODINIT_FUNC PyInit_cumem_allocator(void) { + // Initialize the module + PyObject* module = PyModule_Create(&cumem_allocator_module); + if (!module) { + return NULL; + } + return module; +} +} // extern "C" diff --git a/csrc/cutlass_extensions/common.hpp b/csrc/cutlass_extensions/common.hpp index 85e359aa57113..07c9e46c27b06 100644 --- a/csrc/cutlass_extensions/common.hpp +++ b/csrc/cutlass_extensions/common.hpp @@ -27,8 +27,7 @@ inline int get_cuda_max_shared_memory_per_block_opt_in(int const device) { int max_shared_mem_per_block_opt_in = 0; cudaDeviceGetAttribute(&max_shared_mem_per_block_opt_in, - cudaDevAttrMaxSharedMemoryPerBlockOptin, - device); + cudaDevAttrMaxSharedMemoryPerBlockOptin, device); return max_shared_mem_per_block_opt_in; } diff --git a/csrc/moe/moe_align_sum_kernels.cu b/csrc/moe/moe_align_sum_kernels.cu index 16fccae403338..7e0a25afbfec4 100644 --- a/csrc/moe/moe_align_sum_kernels.cu +++ b/csrc/moe/moe_align_sum_kernels.cu @@ -21,7 +21,7 @@ __device__ __forceinline__ int32_t index(int32_t total_col, int32_t row, } } // namespace -template +template __global__ void moe_align_block_size_kernel(scalar_t* __restrict__ topk_ids, int32_t* sorted_token_ids, int32_t* expert_ids, @@ -32,12 +32,10 @@ __global__ void moe_align_block_size_kernel(scalar_t* __restrict__ topk_ids, const size_t start_idx = threadIdx.x * tokens_per_thread; extern __shared__ int32_t shared_mem[]; - - int32_t* tokens_cnts = - shared_mem; // 2d tensor with shape (blockDim.x + 1, num_experts) - int32_t* cumsum = - shared_mem + - (blockDim.x + 1) * num_experts; // 1d tensor with shape (num_experts + 1) + int32_t* cumsum = shared_mem; // 1d tensor with shape (num_experts + 1) + token_cnts_t* tokens_cnts = + (token_cnts_t*)(shared_mem + num_experts + + 1); // 2d tensor with shape (blockDim.x + 1, num_experts) for (int i = 0; i < num_experts; ++i) { tokens_cnts[index(num_experts, threadIdx.x + 1, i)] = 0; @@ -55,11 +53,11 @@ __global__ void moe_align_block_size_kernel(scalar_t* __restrict__ topk_ids, __syncthreads(); // For each expert we accumulate the token counts from the different threads. - for (int eid = threadIdx.x; eid < num_experts; eid += blockDim.x) { - tokens_cnts[index(num_experts, 0, eid)] = 0; + if (threadIdx.x < num_experts) { + tokens_cnts[index(num_experts, 0, threadIdx.x)] = 0; for (int i = 1; i <= blockDim.x; ++i) { - tokens_cnts[index(num_experts, i, eid)] += - tokens_cnts[index(num_experts, i - 1, eid)]; + tokens_cnts[index(num_experts, i, threadIdx.x)] += + tokens_cnts[index(num_experts, i - 1, threadIdx.x)]; } } @@ -74,7 +72,7 @@ __global__ void moe_align_block_size_kernel(scalar_t* __restrict__ topk_ids, block_size) * block_size; } - *total_tokens_post_pad = cumsum[num_experts]; + *total_tokens_post_pad = static_cast(cumsum[num_experts]); } __syncthreads(); @@ -83,9 +81,10 @@ __global__ void moe_align_block_size_kernel(scalar_t* __restrict__ topk_ids, * For each expert, each thread processes the tokens of the corresponding * blocks and stores the corresponding expert_id for each block. */ - for (int eid = threadIdx.x; eid < num_experts; eid += blockDim.x) { - for (int i = cumsum[eid]; i < cumsum[eid + 1]; i += block_size) { - expert_ids[i / block_size] = eid; + if (threadIdx.x < num_experts) { + for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1]; + i += block_size) { + expert_ids[i / block_size] = threadIdx.x; } } @@ -140,11 +139,11 @@ __global__ void moe_align_block_size_global_mem_kernel( __syncthreads(); // For each expert we accumulate the token counts from the different threads. - for (int eid = threadIdx.x; eid < num_experts; eid += blockDim.x) { - tokens_cnts[index(num_experts, 0, eid)] = 0; + if (threadIdx.x < num_experts) { + tokens_cnts[index(num_experts, 0, threadIdx.x)] = 0; for (int i = 1; i <= blockDim.x; ++i) { - tokens_cnts[index(num_experts, i, eid)] += - tokens_cnts[index(num_experts, i - 1, eid)]; + tokens_cnts[index(num_experts, i, threadIdx.x)] += + tokens_cnts[index(num_experts, i - 1, threadIdx.x)]; } } @@ -168,9 +167,10 @@ __global__ void moe_align_block_size_global_mem_kernel( * For each expert, each thread processes the tokens of the corresponding * blocks and stores the corresponding expert_id for each block. */ - for (int eid = threadIdx.x; eid < num_experts; eid += blockDim.x) { - for (int i = cumsum[eid]; i < cumsum[eid + 1]; i += block_size) { - expert_ids[i / block_size] = eid; + if (threadIdx.x < num_experts) { + for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1]; + i += block_size) { + expert_ids[i / block_size] = threadIdx.x; } } @@ -221,25 +221,88 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts, torch::Tensor experts_ids, torch::Tensor num_tokens_post_pad) { const cudaStream_t stream = at::cuda::getCurrentCUDAStream(); - VLLM_DISPATCH_INTEGRAL_TYPES( - topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] { - // calc needed amount of shared mem for `tokens_cnts` and `cumsum` - // tensors - const int32_t num_thread = WARP_SIZE; - const int32_t shared_mem = - ((num_thread + 1) * num_experts + (num_experts + 1)) * - sizeof(int32_t); - - // set dynamic shared mem - auto kernel = vllm::moe::moe_align_block_size_kernel; - AT_CUDA_CHECK(VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize( - (void*)kernel, shared_mem)); - kernel<<<1, num_thread, shared_mem, stream>>>( - topk_ids.data_ptr(), sorted_token_ids.data_ptr(), - experts_ids.data_ptr(), - num_tokens_post_pad.data_ptr(), num_experts, block_size, - topk_ids.numel()); - }); + + int device_max_shared_mem; + auto dev = topk_ids.get_device(); + cudaDeviceGetAttribute(&device_max_shared_mem, + cudaDevAttrMaxSharedMemoryPerBlockOptin, dev); + + const int32_t num_thread = max((int32_t)num_experts, WARP_SIZE); + const int32_t shared_mem_i32 = + ((num_thread + 1) * num_experts + (num_experts + 1)) * sizeof(int32_t); + const int32_t shared_mem_i16 = + ((num_thread + 1) * num_experts) * sizeof(uint16_t) + + (num_experts + 1) * sizeof(int32_t); + + bool use_global_memory = false; + bool use_i16 = false; // Use uint16_t for shared memory token counts + if (shared_mem_i32 < device_max_shared_mem) { + // Do nothing in this case. We're all set to use int32_t token counts + } else if (shared_mem_i16 < device_max_shared_mem && + topk_ids.numel() <= 65535) { + // when nelements of topk_ids is smaller than 65535 (max value of uint16), + // element value of token_cnts would also smaller than 65535, + // so we can use uint16 as dtype of token_cnts + use_i16 = true; + } else { + use_global_memory = true; + } + + if (use_global_memory) { + VLLM_DISPATCH_INTEGRAL_TYPES( + topk_ids.scalar_type(), "moe_align_block_size_global_mem_kernel", [&] { + // calc needed amount of shared mem for `tokens_cnts` and `cumsum` + // tensors + const int32_t num_thread = max((int32_t)num_experts, WARP_SIZE); + + auto options_int = torch::TensorOptions() + .dtype(torch::kInt) + .device(topk_ids.device()); + torch::Tensor token_cnts_buffer = + torch::empty({(num_experts + 1) * num_experts}, options_int); + torch::Tensor cumsum_buffer = + torch::empty({num_experts + 1}, options_int); + + auto kernel = + vllm::moe::moe_align_block_size_global_mem_kernel; + kernel<<<1, num_thread, 0, stream>>>( + topk_ids.data_ptr(), + sorted_token_ids.data_ptr(), + experts_ids.data_ptr(), + num_tokens_post_pad.data_ptr(), num_experts, block_size, + topk_ids.numel(), token_cnts_buffer.data_ptr(), + cumsum_buffer.data_ptr()); + }); + } else if (use_i16) { + VLLM_DISPATCH_INTEGRAL_TYPES( + topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] { + // set dynamic shared mem + auto kernel = + vllm::moe::moe_align_block_size_kernel; + AT_CUDA_CHECK(VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize( + (void*)kernel, shared_mem_i16)); + kernel<<<1, num_thread, shared_mem_i16, stream>>>( + topk_ids.data_ptr(), + sorted_token_ids.data_ptr(), + experts_ids.data_ptr(), + num_tokens_post_pad.data_ptr(), num_experts, block_size, + topk_ids.numel()); + }); + } else { + VLLM_DISPATCH_INTEGRAL_TYPES( + topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] { + auto kernel = + vllm::moe::moe_align_block_size_kernel; + AT_CUDA_CHECK(VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize( + (void*)kernel, shared_mem_i32)); + kernel<<<1, num_thread, shared_mem_i32, stream>>>( + topk_ids.data_ptr(), + sorted_token_ids.data_ptr(), + experts_ids.data_ptr(), + num_tokens_post_pad.data_ptr(), num_experts, block_size, + topk_ids.numel()); + }); + } } void moe_sum(torch::Tensor& input, // [num_tokens, topk, hidden_size] @@ -283,4 +346,4 @@ void moe_sum(torch::Tensor& input, // [num_tokens, topk, hidden_size] at::sum_out(output, input, 1); break; } -} +} \ No newline at end of file diff --git a/csrc/ops.h b/csrc/ops.h index e9cc8d2e215e2..830424a4a9d20 100644 --- a/csrc/ops.h +++ b/csrc/ops.h @@ -88,6 +88,8 @@ void batched_rotary_embedding(torch::Tensor& positions, torch::Tensor& query, void silu_and_mul(torch::Tensor& out, torch::Tensor& input); +void mul_and_silu(torch::Tensor& out, torch::Tensor& input); + void scaled_silu_and_mul(torch::Tensor& out, torch::Tensor& input, torch::Tensor& scale); diff --git a/csrc/prepare_inputs/advance_step.cu b/csrc/prepare_inputs/advance_step.cu index bd184ee22682e..c3902f4c2a163 100644 --- a/csrc/prepare_inputs/advance_step.cu +++ b/csrc/prepare_inputs/advance_step.cu @@ -95,6 +95,16 @@ __global__ void advance_step_flashinfer_kernel( long* input_positions_ptr, int* seq_lens_ptr, long* slot_mapping_ptr, int const* block_tables_ptr, int64_t const block_tables_stride, int* paged_kv_last_page_len_ptr, int* block_table_bound_ptr) { + int const n_pad = num_seqs - num_queries; + if (n_pad && blockIdx.x == 0) { + // Handle cuda graph padding + int const offset = num_queries; + for (int i = threadIdx.x; i < n_pad; i += blockDim.x) { + input_tokens_ptr[offset + i] = 0; + input_positions_ptr[offset + i] = 0; + slot_mapping_ptr[offset + i] = -1; + } + } int num_query_blocks = div_ceil(num_queries, num_threads); if (blockIdx.x < num_query_blocks) { diff --git a/csrc/rocm/attention.cu b/csrc/rocm/attention.cu index ab8edd6d0f57b..828d324ceb333 100644 --- a/csrc/rocm/attention.cu +++ b/csrc/rocm/attention.cu @@ -51,6 +51,9 @@ using floatx4 = __attribute__((__vector_size__(4 * sizeof(float)))) float; using float16x4 = __attribute__((__vector_size__(4 * sizeof(_Float16)))) _Float16; typedef float16x4 _Half4; +using float16x2 = + __attribute__((__vector_size__(2 * sizeof(_Float16)))) _Float16; +typedef float16x2 _Half2; typedef struct _Half8 { _Half4 xy[2]; } _Half8; @@ -63,24 +66,34 @@ typedef struct _B16x8 { } _B16x8; using _B8x8 = uint2; +using _B8x4 = int32_t; // used in builtins using bit8_t = uint8_t; -////// Non temporal load stores /////// +typedef struct _B8x16 { + _B8x8 xy[2]; +} _B8x16; +////// Non temporal loads /////// template -__device__ __forceinline__ T load(T* addr) { - return addr[0]; +__device__ __forceinline__ T loadnt(T* addr) { + return __builtin_nontemporal_load(addr); } -template -__device__ __forceinline__ void store(T value, T* addr) { - addr[0] = value; +__device__ __forceinline__ _B16x8 load_ntmprl_16Byte(const _B16x8* addr) { + auto addr_alias = reinterpret_cast(addr); + auto dat0 = loadnt(addr_alias); + auto dat1 = loadnt(addr_alias + 1); + auto dat2 = loadnt(addr_alias + 2); + auto dat3 = loadnt(addr_alias + 3); + auto res = make_float4(dat0, dat1, dat2, dat3); + return *reinterpret_cast<_B16x8*>(&res); } +/////////////////////////////////// template -__device__ __forceinline__ floatx4 gcn_mfma_instr(const _B16x4& inpA, - const _B16x4& inpB, - const floatx4& inpC) { +__device__ __forceinline__ floatx4 gcn_mfma4x4x4_instr(const _B16x4& inpA, + const _B16x4& inpB, + const floatx4& inpC) { if constexpr (std::is_same::value) { return __builtin_amdgcn_mfma_f32_4x4x4f16(inpA, inpB, inpC, absz, cbid, blgp); @@ -92,6 +105,21 @@ __device__ __forceinline__ floatx4 gcn_mfma_instr(const _B16x4& inpA, } } +template +__device__ __forceinline__ floatx4 gcn_mfma16x16x16_instr(const _B16x4& inpA, + const _B16x4& inpB, + const floatx4& inpC) { + if constexpr (std::is_same::value) { + return __builtin_amdgcn_mfma_f32_16x16x16f16(inpA, inpB, inpC, absz, cbid, + blgp); + } else if constexpr (std::is_same::value) { + return __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(inpA, inpB, inpC, absz, + cbid, blgp); + } else { + static_assert(false, "unsupported 16b dtype"); + } +} + template __device__ __forceinline__ float to_float(const T& inp) { if constexpr (std::is_same::value) { @@ -140,17 +168,22 @@ __device__ __forceinline__ _B16x4 from_floatx4(const floatx4& inp) { } t16; _B16x4 ret; if constexpr (std::is_same::value) { - #pragma unroll - for (int i = 0; i < 4; i++) { - t16.f = (_Float16)inp[i]; - ret[i] = t16.u; - } - return ret; + union h2cvt { + __half2 h2[2]; + _B16x4 b16x4; + } u; + u.h2[0] = __float22half2_rn(make_float2(inp[0], inp[1])); + u.h2[1] = __float22half2_rn(make_float2(inp[2], inp[3])); + return u.b16x4; } else if constexpr (std::is_same::value) { - #pragma unroll for (int i = 0; i < 4; i++) { - t16.b = __float2bfloat16(inp[i]); - ret[i] = t16.u; + union fcvt { + uint32_t u32; + float f32; + } u; + u.f32 = inp[i]; + u.u32 += 0x7fff + ((u.u32 >> 16) & 1); // BF16 RNE with no nan/inf check + ret[i] = uint16_t(u.u32 >> 16); } return ret; } else { @@ -168,21 +201,25 @@ __device__ __forceinline__ _B16x4 addx4(const _B16x4& inp1, } t1, t2, res; _B16x4 ret; if constexpr (std::is_same::value) { - #pragma unroll - for (int i = 0; i < 4; i++) { - t1.u = inp1[i]; - t2.u = inp2[i]; - res.f = t1.f + t2.f; - ret[i] = res.u; - } - return ret; + union h2cvt { + _B16x4 b16x4; + __half2 h2[2]; + } u1, u2, s; + u1.b16x4 = inp1; + u2.b16x4 = inp2; + s.h2[0] = u1.h2[0] + u2.h2[0]; + s.h2[1] = u1.h2[1] + u2.h2[1]; + return s.b16x4; } else if constexpr (std::is_same::value) { - #pragma unroll for (int i = 0; i < 4; i++) { - t1.u = inp1[i]; - t2.u = inp2[i]; - res.b = t1.b + t2.b; - ret[i] = res.u; + union fcvt { + float f32; + uint32_t i32; + } u1, u2, s; + u1.i32 = uint32_t(inp1[i]) << 16; + u2.i32 = uint32_t(inp2[i]) << 16; + s.f32 = u1.f32 + u2.f32; + ret[i] = uint16_t(s.i32 >> 16); } return ret; } else { @@ -210,15 +247,600 @@ __device__ __forceinline__ _B16x8 scaled_convert_b8x8(const _B8x8 input, } } +template +__device__ __forceinline__ _B16x8 +scaled_convert_b8x8_custom(const _B8x8 input, const float scale) { + union { + floatx4 f32x4[2]; + vllm::Float8_ f32x8; + } tmpf8; + tmpf8.f32x8 = vllm::fp8::vec_conversion( + *reinterpret_cast(&input)); + + tmpf8.f32x4[0] *= scale; + tmpf8.f32x4[1] *= scale; + + _B16x8 ret; + ret.xy[0] = from_floatx4(tmpf8.f32x4[0]); + ret.xy[1] = from_floatx4(tmpf8.f32x4[1]); + return ret; +} + +__device__ __forceinline__ floatx4 to_float_fp8x4(const _B8x4& inp) { + #if defined(__gfx90a__) + float4 f32x4 = vllm::fp8::vec_conversion( + *reinterpret_cast(&inp)); + return *reinterpret_cast(&f32x4); + #else // MI3xx+ optimized builtins + const auto f0 = __builtin_amdgcn_cvt_pk_f32_fp8(inp, false); + const auto f1 = __builtin_amdgcn_cvt_pk_f32_fp8(inp, true); + floatx4 ret; + ret[0] = f0[0]; + ret[1] = f0[1]; + ret[2] = f1[0]; + ret[3] = f1[1]; + return ret; + #endif +} + +template +__device__ __forceinline__ _B16x4 from_floatx4_rtz(const floatx4& inp) { + _B16x4 ret; + if constexpr (std::is_same::value) { + union h2cvt { + _Half2 h2[2]; + _B16x4 b16x4; + } u; + u.h2[0] = __builtin_amdgcn_cvt_pkrtz(inp[0], inp[1]); + u.h2[1] = __builtin_amdgcn_cvt_pkrtz(inp[2], inp[3]); + return u.b16x4; + } else if constexpr (std::is_same::value) { + for (int i = 0; i < 4; i++) { + union fcvt { + uint32_t i32; + float f32; + } u; + u.f32 = inp[i]; + ret[i] = uint16_t(u.i32 >> 16); + } + return ret; + } else { + static_assert(false, "unsupported 16b dtype"); + } +} + +template +__device__ __forceinline__ _B16x8 convert_b8x8_custom(const _B8x8 input) { + union { + _B8x8 b8x8; + _B8x4 b8x4[2]; + } tmp; + tmp.b8x8 = input; + _B16x8 ret; + for (int i = 0; i < 2; i++) { + ret.xy[i] = from_floatx4_rtz(to_float_fp8x4(tmp.b8x4[i])); + } + return ret; +} + /////////////////////////////////////// +// grid (num_seqs, num_partitions,num_kv_heads) +// block (256) +template +__global__ +__launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel( + const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size] + const cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads, + // head_size/x, block_size, x] + const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads, + // head_size, block_size] + const int num_kv_heads, const float scale, + const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq] + const int* __restrict__ context_lens, // [num_seqs] + const int max_num_blocks_per_seq, + const float* __restrict__ alibi_slopes, // [num_heads] + const int q_stride, const int kv_block_stride, const int kv_head_stride, + float* __restrict__ exp_sums, // [num_seqs, num_heads, max_num_partitions] + float* __restrict__ max_logits, // [num_seqs, num_heads, + // max_num_partitions] + scalar_t* __restrict__ out, // [num_seqs, num_heads, max_num_partitions, + // head_size] + OUTT* __restrict__ final_out, // [num_seqs, num_heads, head_size] + int max_ctx_blocks, const float* k_scale_ptr, const float* v_scale_ptr, + const float* __restrict__ fp8_out_scale_ptr) { + constexpr int NWARPS = NUM_THREADS / WARP_SIZE; + const int warpid = threadIdx.x / WARP_SIZE; + const int laneid = threadIdx.x % WARP_SIZE; + const int lane4id = laneid % 4; + const int lane16id = laneid % 16; + const int rowid = laneid / 16; + + const int seq_idx = blockIdx.x; + const int partition_idx = blockIdx.y; + + constexpr int T_PAR_SIZE = 256; // token partition size set to 256 + + const int max_num_partitions = gridDim.y; + + const int context_len = context_lens[seq_idx]; + + const int partition_start_token_idx = + partition_idx * T_PAR_SIZE; // partition_size; + // exit if partition is out of context for seq + if (partition_start_token_idx >= context_len) { + return; + } + + constexpr int GQA_RATIO4 = DIVIDE_ROUND_UP(GQA_RATIO, 4); + + __shared__ float shared_qk_max[NWARPS][16 + 1]; + __shared__ float shared_exp_sum[NWARPS][16 + 1]; + // shared_logits is used for multiple purposes + __shared__ _B16x4 shared_logits[NWARPS][4][16][4]; + + // for QK mfma16x16, layout is QHead/Tokenx16 across every 16 lanes, 16 Bytes + // HeadElements in each lane, 4x16B HeadElements across 4 rows of warp + constexpr int ROWS_PER_WARP = + WARP_SIZE / 16; // rows refers to 16 lanes; refer dpp terminology + constexpr int CONTIGUOUS_KV_ELEMS_16B_LOAD = + 16 / sizeof(cache_t); // 8 for 16 bit cache type, 16 for 8 bit types + constexpr int QKHE_PER_FETCH = + CONTIGUOUS_KV_ELEMS_16B_LOAD * + ROWS_PER_WARP; // each fetch across a warp fetches these many elements + constexpr int QK_SIZE_RATIO = + sizeof(scalar_t) / + sizeof(cache_t); // 1 for 16bit types, 2 for 8bit types + constexpr int QKHELOOP = HEAD_SIZE / QKHE_PER_FETCH; // 4xQKHE_16B across + // warp + + _B16x8 Qlocal[QKHELOOP] + [QK_SIZE_RATIO]; // note that 16 contiguous elements of Q should + // be fetched per lane for 8 bit cache types : + // QK_SIZE_RATIO changes for this + + constexpr int CONTIGUOUS_SCALAR_ELEMS_16B = 16 / sizeof(scalar_t); + + constexpr int TOKENS_PER_WARP = + T_PAR_SIZE / + NWARPS; // sub partition of tokens per warp for qk calculation + constexpr int TLOOP = + TOKENS_PER_WARP / + 16; // each mfma16x16x16 instruction processes 16 tokens + + _B16x8 Klocal[TLOOP][QKHELOOP]; // can be interpreted as B8x16 for 8 bit + // types + + const int wg_start_head_idx = blockIdx.z * GQA_RATIO; + const int wg_start_kv_head_idx = blockIdx.z; + const int total_num_heads = gridDim.z * GQA_RATIO; + + // for QK mfma, tokens in multiples of TOKENS_PER_WARP are spread across warps + // each mfma takes QH16xT16x16HE across warp + // repeat mfmas across QKHELOOP dimension + // output layout from QKmfma : QH16xT4x4 16 qheads across 16 lanes, 16 tokens + // across 4 rows x 4 tokens per lane + + const int num_context_blocks = DIVIDE_ROUND_UP(context_len, BLOCK_SIZE); + const int last_ctx_block = num_context_blocks - 1; + + const int* block_table_seq = block_tables + seq_idx * max_num_blocks_per_seq; + + int kphysical_block_number[TLOOP]; + + // fetch k physical block numbers + for (int token_depth = 0; token_depth < TLOOP; token_depth++) { + const int klocal_token_idx = + TOKENS_PER_WARP * warpid + token_depth * 16 + lane16id; + const int kglobal_token_idx = partition_start_token_idx + klocal_token_idx; + const int kblock_idx = (kglobal_token_idx < context_len) + ? kglobal_token_idx / BLOCK_SIZE + : last_ctx_block; + kphysical_block_number[token_depth] = block_table_seq[kblock_idx]; + } + + // fetch Q in shared across warps and then write to registers + const int local_qhead_idx = 4 * warpid + rowid; + const int global_qhead_idx = wg_start_head_idx + local_qhead_idx; + const int64_t seq_idx64 = static_cast(seq_idx); + const scalar_t* q_ptr = + q + seq_idx64 * q_stride + global_qhead_idx * HEAD_SIZE; + + const int qhead_element = lane16id * CONTIGUOUS_SCALAR_ELEMS_16B; + if ((local_qhead_idx < GQA_RATIO) && (qhead_element < HEAD_SIZE)) { + const scalar_t* q_fetch_ptr = q_ptr + qhead_element; + const _B16x8* q_fetch_ptr_16B = + reinterpret_cast(q_fetch_ptr); + _B16x8 tmp = *q_fetch_ptr_16B; + if constexpr (KV_DTYPE == vllm::Fp8KVCacheDataType::kAuto) { + const int offset1 = + lane16id / + 4; // 16 contiguous chunks of head elems are spread across 4x4lanes + shared_logits[offset1][lane4id][local_qhead_idx][0] = tmp.xy[0]; + shared_logits[offset1][lane4id][local_qhead_idx][1] = tmp.xy[1]; + } else { + for (int i = 0; i < 2; i++) { + const int head_elem = lane16id * 2 + i; // element id in _B16x4 terms + const int offset3 = head_elem % 4; + const int offset2 = (head_elem / 4) % 4; + const int offset1 = head_elem / 4 / 4; + shared_logits[offset1][offset2][local_qhead_idx][offset3] = tmp.xy[i]; + } + } + } + __syncthreads(); + for (int qkhe_depth = 0; qkhe_depth < QKHELOOP; qkhe_depth++) { + for (int qkratio = 0; qkratio < QK_SIZE_RATIO; qkratio++) { + for (int i = 0; i < 2; i++) { + Qlocal[qkhe_depth][qkratio].xy[i] = + shared_logits[qkhe_depth][rowid][lane16id % GQA_RATIO] + [2 * qkratio + i]; + } + } + } + + // set to true to enable non temporal kv loads: has some benefit in very high + // batch size cases + constexpr bool NT_KV_LOAD = false; + + constexpr int KX = + 16 / sizeof(cache_t); // vLLM defines x as 16 Bytes of kv cache elements + const cache_t* k_ptr = k_cache + wg_start_kv_head_idx * kv_head_stride; + + const int row_head_elem = rowid * CONTIGUOUS_KV_ELEMS_16B_LOAD; + // fetch K values + for (int token_depth = 0; token_depth < TLOOP; token_depth++) { + const int64_t kblock_number = + static_cast(kphysical_block_number[token_depth]); + const cache_t* k_ptr2 = k_ptr + kblock_number * kv_block_stride; + const int klocal_token_idx = + TOKENS_PER_WARP * warpid + token_depth * 16 + lane16id; + const int kglobal_token_idx = partition_start_token_idx + klocal_token_idx; + const int kphysical_block_offset = klocal_token_idx % BLOCK_SIZE; + const cache_t* k_ptr3 = k_ptr2 + kphysical_block_offset * KX; + + for (int qkhe_depth = 0; qkhe_depth < QKHELOOP; qkhe_depth++) { + const int head_elem = row_head_elem + qkhe_depth * QKHE_PER_FETCH; + const int offset1 = head_elem / KX; + const int offset2 = head_elem % KX; + const cache_t* k_fetch_ptr = k_ptr3 + offset1 * BLOCK_SIZE * KX + offset2; + const _B16x8* k_fetch_ptr_16B = + reinterpret_cast(k_fetch_ptr); + if constexpr (NT_KV_LOAD) { + Klocal[token_depth][qkhe_depth] = load_ntmprl_16Byte(k_fetch_ptr_16B); + } else { + Klocal[token_depth][qkhe_depth] = *k_fetch_ptr_16B; + } + } + } + + float alibi_slope; + if constexpr (ALIBI_ENABLED) { + const int alibi_head_idx = wg_start_head_idx + lane16id; + alibi_slope = (lane16id < GQA_RATIO) ? alibi_slopes[alibi_head_idx] : 0.f; + } + + constexpr int VTOKENS_PER_LANE = + TOKENS_PER_WARP / ROWS_PER_WARP; // 64/4 = 16 contiguous vtokens per lane + constexpr int VBLOCKS_PER_LANE = + 1; // assumes block size >=16, each lane can correspond to 1 block only + constexpr int VTLOOP = NWARPS; // corresponds to tokens across warps + constexpr int VTLANELOOP = DIVIDE_ROUND_UP( + VTOKENS_PER_LANE, + CONTIGUOUS_KV_ELEMS_16B_LOAD); // optimized for 16B fetches; assumes + // minimum block size is 16 + constexpr int VHELOOP = + HEAD_SIZE / 16 / NWARPS; // head_size distributed across warps; each mfma + // instr works on 16 head elements + + int vphysical_block_number[VTLOOP][VBLOCKS_PER_LANE]; + + // fetch v physical block numbers + for (int vtoken_depth = 0; vtoken_depth < VTLOOP; vtoken_depth++) { + for (int vblock_depth = 0; vblock_depth < VBLOCKS_PER_LANE; + vblock_depth++) { + const int vlocal_token_idx = + vtoken_depth * VTOKENS_PER_LANE * ROWS_PER_WARP + + rowid * VTOKENS_PER_LANE + vblock_depth * BLOCK_SIZE; + const int vglobal_token_idx = + partition_start_token_idx + vlocal_token_idx; + const int vblock_idx = (vglobal_token_idx < context_len) + ? vglobal_token_idx / BLOCK_SIZE + : last_ctx_block; + vphysical_block_number[vtoken_depth][vblock_depth] = + block_table_seq[vblock_idx]; + } + } + + _B16x8 Vlocal[VTLOOP][VHELOOP] + [VTLANELOOP]; // this can be interpreted as B8x16 too + + const cache_t* v_ptr = v_cache + wg_start_kv_head_idx * kv_head_stride + + ((rowid * VTOKENS_PER_LANE) % BLOCK_SIZE); + + // v fetches are 16head elems across lanes x 16 tokens per lane + for (int vhe_depth = 0; vhe_depth < VHELOOP; vhe_depth++) { + const int vhead_elem = vhe_depth * NWARPS * 16 + warpid * 16 + lane16id; + const cache_t* v_ptr2 = v_ptr + vhead_elem * BLOCK_SIZE; + + for (int vtoken_depth = 0; vtoken_depth < VTLOOP; vtoken_depth++) { + for (int vfetch_depth = 0; vfetch_depth < VTLANELOOP; vfetch_depth++) { + const int vblock_depth = 0; + const int64_t vblock_number = static_cast( + vphysical_block_number[vtoken_depth][vblock_depth]); + const cache_t* v_ptr3 = v_ptr2 + (vblock_number * kv_block_stride); + + const cache_t* v_fetch_ptr = + v_ptr3 + vfetch_depth * CONTIGUOUS_KV_ELEMS_16B_LOAD; + const _B16x8* v_fetch_ptr_16B = + reinterpret_cast(v_fetch_ptr); + if constexpr (NT_KV_LOAD) { + Vlocal[vtoken_depth][vhe_depth][vfetch_depth] = + load_ntmprl_16Byte(v_fetch_ptr_16B); + } else { + Vlocal[vtoken_depth][vhe_depth][vfetch_depth] = *v_fetch_ptr_16B; + } + } + } + } + + // calculate post qk mfma scale + float scale2 = scale; + if constexpr (KV_DTYPE != vllm::Fp8KVCacheDataType::kAuto) { + // multiply by k_scale if fp8 kv cache + scale2 *= *k_scale_ptr; + } + + floatx4 dout[TLOOP]; + // qk mfma + for (int token_depth = 0; token_depth < TLOOP; token_depth++) { + dout[token_depth] = {0}; + for (int qkhe_depth = 0; qkhe_depth < QKHELOOP; qkhe_depth++) { + if constexpr (KV_DTYPE == vllm::Fp8KVCacheDataType::kAuto) { + for (int qkratio = 0; qkratio < QK_SIZE_RATIO; qkratio++) { + for (int i = 0; i < 2; i++) { + dout[token_depth] = gcn_mfma16x16x16_instr( + Klocal[token_depth][qkhe_depth].xy[i], + Qlocal[qkhe_depth][qkratio].xy[i], dout[token_depth]); + } + } + } else { // kv cache dtype fp8 + auto Ktmp = Klocal[token_depth][qkhe_depth]; + _B8x16 Ktmp8x16 = *reinterpret_cast<_B8x16*>(&Ktmp); + for (int qkratio = 0; qkratio < QK_SIZE_RATIO; qkratio++) { + _B8x8 Ktmp8x8 = Ktmp8x16.xy[qkratio]; + _B16x8 Klocaltmp = convert_b8x8_custom(Ktmp8x8); + for (int i = 0; i < 2; i++) { + dout[token_depth] = gcn_mfma16x16x16_instr( + Klocaltmp.xy[i], Qlocal[qkhe_depth][qkratio].xy[i], + dout[token_depth]); + } + } + } + } + dout[token_depth] *= scale2; + } + + const int qkout_token_idx = + partition_start_token_idx + TOKENS_PER_WARP * warpid + rowid * 4; + + // apply alibi + if constexpr (ALIBI_ENABLED) { + for (int token_depth = 0; token_depth < TLOOP; token_depth++) { + const int local_token_idx = qkout_token_idx + token_depth * 16; + const int alibi_offset = local_token_idx - context_len + 1; + for (int i = 0; i < 4; i++) { + dout[token_depth][i] += alibi_slope * (alibi_offset + i); + } + } + } + + // calculate qk_max and exp_sum per warp and write to shared memory + float qk_max = -FLT_MAX; + float exp_sum = 0.0f; + + for (int token_depth = 0; token_depth < TLOOP; token_depth++) { + const int local_token_idx = qkout_token_idx + token_depth * 16; + for (int i = 0; i < 4; i++) { + const float tmp = + (local_token_idx + i < context_len) ? dout[token_depth][i] : -FLT_MAX; + qk_max = fmaxf(qk_max, tmp); + } + } + + for (int mask = WARP_SIZE / 2; mask >= 16; mask /= 2) { + qk_max = fmaxf(qk_max, __shfl_xor(qk_max, mask)); + } + + for (int token_depth = 0; token_depth < TLOOP; token_depth++) { + const int local_token_idx = qkout_token_idx + token_depth * 16; + for (int i = 0; i < 4; i++) { + const float tmp = (local_token_idx + i < context_len) + ? __expf(dout[token_depth][i] - qk_max) + : 0.0f; + dout[token_depth][i] = tmp; + exp_sum += tmp; + } + } + + for (int mask = WARP_SIZE / 2; mask >= 16; mask /= 2) { + exp_sum += __shfl_xor(exp_sum, mask); + } + + __syncthreads(); // sync before writing to shared mem + + float* shared_mem = reinterpret_cast(shared_logits); + if (laneid < 16) { + const int qk_max_offset = warpid * 16 + lane16id; + shared_mem[qk_max_offset] = qk_max; + const int exp_sum_offset = NWARPS * 16 + qk_max_offset; + shared_mem[exp_sum_offset] = exp_sum; + } + + __syncthreads(); + + // calculate partition qk_max and exp_sum + float partition_qk_max = -FLT_MAX; + float warp_qk_max_exp[NWARPS]; + float partition_exp_sum = 0.0f; + + for (int w = 0; w < NWARPS; w++) { + warp_qk_max_exp[w] = shared_mem[w * 16 + lane16id]; + partition_qk_max = fmaxf(partition_qk_max, warp_qk_max_exp[w]); + } + + for (int w = 0; w < NWARPS; w++) { + warp_qk_max_exp[w] = __expf(warp_qk_max_exp[w] - partition_qk_max); + partition_exp_sum += + shared_mem[NWARPS * 16 + w * 16 + lane16id] * warp_qk_max_exp[w]; + } + + const float inv_sum_scale = + __fdividef(1.f, partition_exp_sum + 1e-6f) * warp_qk_max_exp[warpid]; + + __syncthreads(); + + // disable rtz conversion due to its impact on accuracy. + constexpr bool LOGITS_RTZ_CONVERSION = false; + + // write logits to shared mem + for (int token_depth = 0; token_depth < TLOOP; token_depth++) { + dout[token_depth] *= inv_sum_scale; + if constexpr (LOGITS_RTZ_CONVERSION) { + // use rtz conversion for better performance, with negligible impact on + // accuracy. + shared_logits[warpid][token_depth][lane16id][rowid] = + from_floatx4_rtz(dout[token_depth]); + } else { + shared_logits[warpid][token_depth][lane16id][rowid] = + from_floatx4(dout[token_depth]); + } + } + // write out partition max_logits and exp_sum + if (threadIdx.x < GQA_RATIO) { + const int qhead_idx = lane16id; + const int offset = seq_idx * total_num_heads * max_num_partitions + + (wg_start_head_idx + qhead_idx) * max_num_partitions + + partition_idx; + max_logits[offset] = partition_qk_max; + exp_sums[offset] = partition_exp_sum; + } + + __syncthreads(); + + constexpr int ELEMS8_ELEMS4_RATIO = 8 / 4; + constexpr int ELEMS16_ELEMS8_RATIO = 16 / 8; + + _B16x4 outelems[VHELOOP]; + // Softmax V mfma + // v layout: 16he across lanes x 16 tokens per lane + for (int vhe_depth = 0; vhe_depth < VHELOOP; vhe_depth++) { + floatx4 tmp_out = {0}; + + for (int vtoken_depth = 0; vtoken_depth < VTLOOP; vtoken_depth++) { + if constexpr (KV_DTYPE == vllm::Fp8KVCacheDataType::kAuto) { + for (int vfetch_depth = 0; vfetch_depth < VTLANELOOP; vfetch_depth++) { + for (int i = 0; i < ELEMS8_ELEMS4_RATIO; i++) { + const int offset = rowid * VTLANELOOP * ELEMS8_ELEMS4_RATIO + + vfetch_depth * ELEMS8_ELEMS4_RATIO + i; + const int offset1 = offset % ROWS_PER_WARP; + const int offset2 = offset / ROWS_PER_WARP; + // output format is 16 qheads across 16 lanes, 16 head elems spread + // across 4 rows + tmp_out = gcn_mfma16x16x16_instr( + Vlocal[vtoken_depth][vhe_depth][vfetch_depth].xy[i], + shared_logits[vtoken_depth][offset2][lane16id][offset1], + tmp_out); + } + } + // KV cache fp8 + } else { + for (int vfetch_depth = 0; vfetch_depth < VTLANELOOP; vfetch_depth++) { + _B16x8 Vtmp = Vlocal[vtoken_depth][vhe_depth][vfetch_depth]; + // reinterpret V format as 16 elements of 8bits + _B8x16 Vtmp8x16 = *reinterpret_cast<_B8x16*>(&Vtmp); + for (int j = 0; j < ELEMS16_ELEMS8_RATIO; j++) { + _B8x8 Vtmp8x8 = Vtmp8x16.xy[j]; + _B16x8 Vlocaltmp = convert_b8x8_custom(Vtmp8x8); + for (int i = 0; i < ELEMS8_ELEMS4_RATIO; i++) { + const int offset = + rowid * ELEMS16_ELEMS8_RATIO * ELEMS8_ELEMS4_RATIO + + j * ELEMS8_ELEMS4_RATIO + i; + const int offset1 = offset % ROWS_PER_WARP; + const int offset2 = offset / ROWS_PER_WARP; + // output format is 16 qheads across 16 lanes, 16 head elems + // spread across 4 rows + tmp_out = gcn_mfma16x16x16_instr( + Vlocaltmp.xy[i], + shared_logits[vtoken_depth][offset2][lane16id][offset1], + tmp_out); + } + } + } + } + } + // apply post Softmax V mfma v_scale + if constexpr (KV_DTYPE != vllm::Fp8KVCacheDataType::kAuto) { + tmp_out *= *v_scale_ptr; + } + outelems[vhe_depth] = from_floatx4(tmp_out); + } + + __syncthreads(); + + // store Softmax-V mfma output to shared mem + for (int vhe_depth = 0; vhe_depth < VHELOOP; vhe_depth++) { + // lane16 id head dimension; rowid head element dimension + shared_logits[warpid][vhe_depth][lane16id][rowid] = outelems[vhe_depth]; + } + + __syncthreads(); -// grid (num_seqs, num_partitions,num_heads/gqa_ratio) -// block (partition size) + // write to tmp_out with coalesced writes after reading from shared mem + if (warpid == 0) { + _B16x8 vout[GQA_RATIO4]; + // each lane writes out 16Bytes of tmp_out along head elem dimension + const int head_elem_idx = lane16id * 8; + if (head_elem_idx < HEAD_SIZE) { + for (int h = 0; h < GQA_RATIO4; h++) { + const int local_head_idx = 4 * h + rowid; + const int offset1 = (head_elem_idx / 16) % 4; + const int offset2 = head_elem_idx / 16 / NWARPS; + const int offset3 = (head_elem_idx / 4) % 4; + for (int i = 0; i < 2; i++) { + vout[h].xy[i] = + shared_logits[offset1][offset2][local_head_idx][offset3 + i]; + } + } + + const int hsz_maxp_mult = HEAD_SIZE * max_num_partitions; + scalar_t* out_ptr = out + seq_idx * total_num_heads * hsz_maxp_mult + + partition_idx * HEAD_SIZE; + for (int h = 0; h < GQA_RATIO4; h++) { + const int local_head_idx = 4 * h + rowid; + if (local_head_idx < GQA_RATIO) { + const int out_head_idx = wg_start_head_idx + local_head_idx; + scalar_t* out_ptr2 = out_ptr + out_head_idx * hsz_maxp_mult; + scalar_t* out_ptr3 = out_ptr2 + head_elem_idx; + _B16x8* out_ptr_B16x8 = reinterpret_cast<_B16x8*>(out_ptr3); + *out_ptr_B16x8 = vout[h]; + } + } + } + } +} + +///////////////////////////////////////////////////////////// +// grid (num_seqs, num_partitions, num_kv_heads) +// block (256 : partition size) +// each WG handles 1 partition per sequence template -__global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( +__global__ +__launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel( const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size] const cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads, // head_size/x, block_size, x] @@ -254,27 +876,35 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( if (partition_start_token_idx >= context_len) { return; } - constexpr int QHLOOP = - DIVIDE_ROUND_UP(GQA_RATIO, 4); // each 4 lanes fetch 4 different qheads, - // total qheads =8, so qhloop is 2 + // every 4 lanes fetch 4 different qheads + // qhloop = num loops over qhead dimension + constexpr int QHLOOP = DIVIDE_ROUND_UP(GQA_RATIO, 4); constexpr int GQA_RATIO4 = 4 * QHLOOP; __shared__ float shared_qk_max[NWARPS][GQA_RATIO4 + 1]; __shared__ float shared_exp_sum[NWARPS][GQA_RATIO4 + 1]; _B16x8 Qlocal[QHLOOP]; constexpr int x = 16 / sizeof(scalar_t); + // kheloop = num loops over head_size for 16Bytes of Q/dequantized K elements constexpr int KHELOOP = HEAD_SIZE / x; _B16x8 Klocal[KHELOOP]; _B8x8 Klocalb8[KHELOOP]; - constexpr int VHELOOP = - HEAD_SIZE / - WARP_SIZE; // v head_size dimension is distributed across lanes - constexpr int VTLOOP = 8; // 16 separate 4xtokens across warp -> 16/2 - // 8xtokens + // for SoftMax-V Gemm, V head_size dimension is distributed across warp + // vheloop = num loops to cover v head size dimension + constexpr int VHELOOP = HEAD_SIZE / WARP_SIZE; + // softmax out has warp_size tokens across warp + // vtloop = num loops to cover warp_size(64) tokens with 16Bytes of + // dequantized V elements + constexpr int VTLOOP = WARP_SIZE / 8; + // num vblocks to cover warp_size(64) v elements + constexpr int VBLOCKS = 8 * VTLOOP / BLOCK_SIZE; + int vphysical_blocks[VBLOCKS]; _B16x8 Vlocal[VHELOOP][VTLOOP]; _B8x8 Vlocalb8[VHELOOP][VTLOOP]; floatx4 dout[QHLOOP]; float qk_max[QHLOOP]; - #pragma unroll + + __shared__ _B16x4 vout_shared[QHLOOP][VHELOOP][WARP_SIZE][NWARPS + 1]; + for (int h = 0; h < QHLOOP; h++) { dout[h] = {0}; qk_max[h] = -FLT_MAX; @@ -286,52 +916,51 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( const int warp_start_token_idx = partition_start_token_idx + warpid * WARP_SIZE; - if (warp_start_token_idx >= context_len) { // warp out of context + // entire warp out of context + if (warp_start_token_idx >= context_len) { #pragma unroll for (int h = 0; h < GQA_RATIO4; h++) { shared_qk_max[warpid][h] = -FLT_MAX; shared_exp_sum[warpid][h] = 0.0f; } - } else { // warp within context - + // warp within context + } else { const int num_context_blocks = DIVIDE_ROUND_UP(context_len, BLOCK_SIZE); const int last_ctx_block = num_context_blocks - 1; const int* block_table = block_tables + seq_idx * max_num_blocks_per_seq; - + // token id within partition const int local_token_idx = threadIdx.x; + // token id within sequence const int global_token_idx = partition_start_token_idx + local_token_idx; + // fetch block number for k const int block_idx = (global_token_idx < context_len) ? global_token_idx / BLOCK_SIZE : last_ctx_block; - // fetch block number for q and k - // int32 physical_block_number leads to overflow when multiplied with - // kv_block_stride + + // fetch k physical block number + // int32 physical_block_number leads to overflow when multiplied with + // kv_block_stride const int64_t physical_block_number = static_cast(block_table[block_idx]); // fetch vphysical block numbers up front - constexpr int VBLOCKS = 8 * VTLOOP / BLOCK_SIZE; - int vphysical_blocks[VBLOCKS]; - const int warp_start_block_idx = warp_start_token_idx / BLOCK_SIZE; - if constexpr (GQA_RATIO < 12) { - #pragma unroll - for (int b = 0; b < VBLOCKS; b++) { - const int vblock_idx = warp_start_block_idx + b; - const int vblock_idx_ctx = - (vblock_idx <= last_ctx_block) ? vblock_idx : last_ctx_block; - vphysical_blocks[b] = block_table[vblock_idx_ctx]; - } + for (int b = 0; b < VBLOCKS; b++) { + const int vblock_idx = warp_start_block_idx + b; + const int vblock_idx_ctx = + (vblock_idx <= last_ctx_block) ? vblock_idx : last_ctx_block; + vphysical_blocks[b] = block_table[vblock_idx_ctx]; } - // each 4 lanes fetch 8 helems, so warp fetches 8*16 = 128 helems + // fetch q elements + // every 4 lanes fetch 8 elems, so warp fetches 8*16 = 128 elems const scalar_t* q_ptr = q + seq_idx * q_stride + wg_start_head_idx * HEAD_SIZE; const _B16x8* q_ptrh8 = reinterpret_cast(q_ptr); const int qhead_elemh8 = laneid / 4; - #pragma unroll + for (int h = 0; h < QHLOOP - 1; h++) { const int qhead_idx = h * 4 + lane4id; Qlocal[h] = q_ptrh8[qhead_idx * HEAD_SIZE / 8 + qhead_elemh8]; @@ -345,22 +974,24 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( Qlocal[QHLOOP - 1].xy[1] = {0}; } + // fetch k elements const cache_t* k_ptr = k_cache + physical_block_number * kv_block_stride + wg_start_kv_head_idx * kv_head_stride; - const int physical_block_offset = - local_token_idx % BLOCK_SIZE; // since x=half8, physical_block_offset - // is already cast as _H8 + // physical_block_offset is already cast in terms of _B16x8 + const int physical_block_offset = local_token_idx % BLOCK_SIZE; + + // each K fetch is for 8 elements of cache_t which are later dequantized to + // scalar_t for fp8 if constexpr (KV_DTYPE == vllm::Fp8KVCacheDataType::kAuto) { const _B16x8* k_ptrh8 = reinterpret_cast(k_ptr); - #pragma unroll for (int d = 0; d < KHELOOP; d++) { Klocal[d] = k_ptrh8[d * BLOCK_SIZE + physical_block_offset]; } } else { + // vllm defines X as 16 Bytes of elements of cache_t constexpr int X = 16 / sizeof(cache_t); const cache_t* k_ptr2 = k_ptr + physical_block_offset * X; - #pragma unroll for (int d = 0; d < KHELOOP; d++) { const int head_elem = d * 8; const int offset1 = head_elem / X; @@ -370,9 +1001,9 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( } } + // optional alibi fetch float alibi_slope[QHLOOP]; - if (alibi_slopes != nullptr) { - #pragma unroll + if constexpr (ALIBI_ENABLED) { for (int h = 0; h < QHLOOP; h++) { const int qhead_idx = h * 4 + lane4id; alibi_slope[h] = (qhead_idx < GQA_RATIO) @@ -381,22 +1012,11 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( } } - // fetch vphysical block numbers up front - if constexpr (GQA_RATIO >= 12) { - #pragma unroll - for (int b = 0; b < VBLOCKS; b++) { - const int vblock_idx = warp_start_block_idx + b; - const int vblock_idx_ctx = - (vblock_idx <= last_ctx_block) ? vblock_idx : last_ctx_block; - vphysical_blocks[b] = block_table[vblock_idx_ctx]; - } - } - const cache_t* v_ptr = v_cache + wg_start_kv_head_idx * kv_head_stride; + // fetch vcache in kv cache auto case if constexpr (KV_DTYPE == vllm::Fp8KVCacheDataType::kAuto) { const _B16x8* v_ptrh8 = reinterpret_cast(v_ptr); // iterate over each v block - #pragma unroll for (int b = 0; b < VBLOCKS; b++) { // int32 physical_block_number leads to overflow when multiplied with // kv_block_stride @@ -405,21 +1025,20 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( const _B16x8* v_ptrh8b = v_ptrh8 + (vphysical_block_number * kv_block_stride) / 8; // iterate over each head elem (within head_size) - #pragma unroll for (int h = 0; h < VHELOOP; h++) { const int head_size_elem = h * WARP_SIZE + laneid; const _B16x8* v_ptrh8be = v_ptrh8b + head_size_elem * BLOCK_SIZE / 8; // iterate over all velems within block - #pragma unroll for (int d = 0; d < BLOCK_SIZE / 8; d++) { Vlocal[h][b * BLOCK_SIZE / 8 + d] = v_ptrh8be[d]; } } } - } else { + } // if constexpr (KV_DTYPE == vllm::Fp8KVCacheDataType::kAuto) + // fetch vcache in fp8 case + else { // if constexpr (KV_DTYPE != vllm::Fp8KVCacheDataType::kAuto) const _B8x8* v_ptrh8 = reinterpret_cast(v_ptr); // iterate over each v block - #pragma unroll for (int b = 0; b < VBLOCKS; b++) { // int32 physical_block_number leads to overflow when multiplied with // kv_block_stride @@ -428,164 +1047,153 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( const _B8x8* v_ptrh8b = v_ptrh8 + (vphysical_block_number * kv_block_stride) / 8; // iterate over each head elem (within head_size) - #pragma unroll for (int h = 0; h < VHELOOP; h++) { const int head_size_elem = h * WARP_SIZE + laneid; const _B8x8* v_ptrh8be = v_ptrh8b + head_size_elem * BLOCK_SIZE / 8; // iterate over all velems within block - #pragma unroll for (int d = 0; d < BLOCK_SIZE / 8; d++) { - // Vlocalb8[h][b * BLOCK_SIZE / 8 + d] = v_ptrh8be[d]; - const _B8x8 Vlocalb8 = v_ptrh8be[d]; - Vlocal[h][b * BLOCK_SIZE / 8 + d] = - scaled_convert_b8x8(Vlocalb8, *v_scale_ptr); + Vlocalb8[h][b * BLOCK_SIZE / 8 + d] = v_ptrh8be[d]; } } } } + #define QK_mfma(x) \ + if constexpr (KV_DTYPE != vllm::Fp8KVCacheDataType::kAuto) { \ + Klocal[x] = convert_b8x8_custom(Klocalb8[x]); \ + } \ + for (int h = 0; h < QHLOOP; h++) { \ + dout[h] = gcn_mfma4x4x4_instr( \ + Qlocal[h].xy[0], Klocal[x].xy[0], dout[h]); \ + dout[h] = gcn_mfma4x4x4_instr( \ + Qlocal[h].xy[1], Klocal[x].xy[1], dout[h]); \ + } + // QK mfma with Q mfma block broadcast + // Q values across head_size dimension stored across lanes + // K values across head_size dimension are stored depthwise within lane + // Q broadcast with absz, cbid of mfma instruction + QK_mfma(0); + QK_mfma(1); + QK_mfma(2); + QK_mfma(3); + QK_mfma(4); + QK_mfma(5); + QK_mfma(6); + QK_mfma(7); + // below only needed for head size 128 + if constexpr (KHELOOP > 8) { + QK_mfma(8); + QK_mfma(9); + QK_mfma(10); + QK_mfma(11); + QK_mfma(12); + QK_mfma(13); + QK_mfma(14); + QK_mfma(15); + } + #undef QK_mfma + + float scale2 = scale; if constexpr (KV_DTYPE != vllm::Fp8KVCacheDataType::kAuto) { - #pragma unroll - for (int d = 0; d < KHELOOP; d++) { - Klocal[d] = - scaled_convert_b8x8(Klocalb8[d], *k_scale_ptr); - } + // post mfma scaling for fp8 + scale2 *= *k_scale_ptr; } - #pragma unroll for (int h = 0; h < QHLOOP; h++) { - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[0].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[0].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[1].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[1].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[2].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[2].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[3].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[3].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[4].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[4].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[5].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[5].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[6].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[6].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[7].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[7].xy[1], dout[h]); - if constexpr (KHELOOP > 8) { - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[8].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[8].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[9].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[9].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[10].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[10].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[11].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[11].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[12].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[12].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[13].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[13].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[14].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[14].xy[1], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[0], - Klocal[15].xy[0], dout[h]); - dout[h] = gcn_mfma_instr(Qlocal[h].xy[1], - Klocal[15].xy[1], dout[h]); - } // KHELOOP>8 - dout[h] *= scale; + dout[h] *= scale2; } - // transpose dout so that 4 token ids are in each lane, and 4 heads are across - // 4 lanes - #pragma unroll + + // transpose dout so that 4 token ids are in each lane, and 4 heads are + // across 4 lanes for (int h = 0; h < QHLOOP; h++) { floatx4 tmp = {0}; - #pragma unroll for (int i = 0; i < 4; i++) { const float B = (lane4id == i) ? 1.0f : 0.0f; - // const float A = (global_token_idx < context_len) ? dout[h][i] : 0.0f; tmp = __builtin_amdgcn_mfma_f32_4x4x1f32(dout[h][i], B, tmp, 0, 0, 0); - // tmp = __builtin_amdgcn_mfma_f32_4x4x1f32(A, B, tmp, 0, 0, 0); } dout[h] = tmp; } const int lane4_token_idx = 4 * (global_token_idx >> 2); - const int alibi_offset = lane4_token_idx - context_len + 1; - if (alibi_slopes != nullptr) { - #pragma unroll + + if constexpr (ALIBI_ENABLED) { + const int alibi_offset = lane4_token_idx - context_len + 1; for (int h = 0; h < QHLOOP; h++) { - #pragma unroll for (int i = 0; i < 4; i++) { dout[h][i] += alibi_slope[h] * (alibi_offset + i); } } } - #pragma unroll + const int bpermute_mask = 4 * (16 * ((laneid >> 2) % 4) + lane4id); + for (int h = 0; h < QHLOOP; h++) { qk_max[h] = -FLT_MAX; - #pragma unroll for (int i = 0; i < 4; i++) { qk_max[h] = (lane4_token_idx + i < context_len) ? fmaxf(qk_max[h], dout[h][i]) : qk_max[h]; } - #pragma unroll - for (int mask = WARP_SIZE / 2; mask >= 4; mask /= 2) { - qk_max[h] = fmaxf(qk_max[h], __shfl_xor(qk_max[h], mask)); - } + + // for (int mask = WARP_SIZE / 2; mask >= 4; mask /= 2) { + // qk_max[h] = fmaxf(qk_max[h], __shfl_xor(qk_max[h], mask)); + // } + // faster version of above code with dpp + asm("v_nop\n v_nop\n v_max_f32_dpp %0, %1, %2 row_ror:4" + : "=v"(qk_max[h]) + : "v"(qk_max[h]), "v"(qk_max[h])); + asm("v_nop\n v_nop\n v_max_f32_dpp %0, %1, %2 row_ror:8" + : "=v"(qk_max[h]) + : "v"(qk_max[h]), "v"(qk_max[h])); + + auto tmp = __builtin_amdgcn_ds_bpermute( + bpermute_mask, *reinterpret_cast(&qk_max[h])); + qk_max[h] = *reinterpret_cast(&tmp); + asm("v_nop\n v_nop\n v_max_f32_dpp %0, %1, %2 row_ror:4" + : "=v"(qk_max[h]) + : "v"(qk_max[h]), "v"(qk_max[h])); + asm("v_nop\n v_nop\n v_max_f32_dpp %0, %1, %2 row_ror:8" + : "=v"(qk_max[h]) + : "v"(qk_max[h]), "v"(qk_max[h])); } float exp_sum[QHLOOP]; - #pragma unroll for (int h = 0; h < QHLOOP; h++) { exp_sum[h] = 0.0f; - #pragma unroll for (int i = 0; i < 4; i++) { dout[h][i] = (lane4_token_idx + i < context_len) ? __expf(dout[h][i] - qk_max[h]) : 0.0f; exp_sum[h] += dout[h][i]; } - #pragma unroll - for (int mask = WARP_SIZE / 2; mask >= 4; mask /= 2) { - exp_sum[h] += __shfl_xor(exp_sum[h], mask); - } + // for (int mask = WARP_SIZE / 2; mask >= 4; mask /= 2) { + // exp_sum[h] += __shfl_xor(exp_sum[h], mask); + // } + // faster version of above code with dpp + asm("v_nop\n v_nop\n v_add_f32_dpp %0, %1, %2 row_ror:4" + : "=v"(exp_sum[h]) + : "v"(exp_sum[h]), "v"(exp_sum[h])); + asm("v_nop\n v_nop\n v_add_f32_dpp %0, %1, %2 row_ror:8" + : "=v"(exp_sum[h]) + : "v"(exp_sum[h]), "v"(exp_sum[h])); + + auto tmp = __builtin_amdgcn_ds_bpermute( + bpermute_mask, *reinterpret_cast(&exp_sum[h])); + exp_sum[h] = *reinterpret_cast(&tmp); + asm("v_nop\n v_nop\n v_add_f32_dpp %0, %1, %2 row_ror:4" + : "=v"(exp_sum[h]) + : "v"(exp_sum[h]), "v"(exp_sum[h])); + asm("v_nop\n v_nop\n v_add_f32_dpp %0, %1, %2 row_ror:8" + : "=v"(exp_sum[h]) + : "v"(exp_sum[h]), "v"(exp_sum[h])); } - #pragma unroll - for (int h = 0; h < QHLOOP; h++) { - const int head_idx = 4 * h + lane4id; - shared_qk_max[warpid][head_idx] = qk_max[h]; - shared_exp_sum[warpid][head_idx] = exp_sum[h]; + if (laneid < 4) { + for (int h = 0; h < QHLOOP; h++) { + const int head_idx = 4 * h + lane4id; + shared_qk_max[warpid][head_idx] = qk_max[h]; + shared_exp_sum[warpid][head_idx] = exp_sum[h]; + } } } // warp within context @@ -596,18 +1204,16 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( max_logits + seq_idx * num_heads * max_num_partitions + partition_idx; float* exp_sums_ptr = exp_sums + seq_idx * num_heads * max_num_partitions + partition_idx; - #pragma unroll + // calculate qk_max and exp_sums for partition for (int h = 0; h < QHLOOP; h++) { float global_qk_max = -FLT_MAX; float warp_qk_max[NWARPS]; const int head_idx = 4 * h + lane4id; - #pragma unroll for (int w = 0; w < NWARPS; w++) { warp_qk_max[w] = shared_qk_max[w][head_idx]; global_qk_max = fmaxf(global_qk_max, warp_qk_max[w]); } float global_exp_sum = 0.0f; - #pragma unroll for (int w = 0; w < NWARPS; w++) { global_exp_sum += shared_exp_sum[w][head_idx] * __expf(warp_qk_max[w] - global_qk_max); @@ -622,86 +1228,79 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( __expf(qk_max[h] - global_qk_max); dout[h] *= global_inv_sum_scale; } + constexpr bool LOGITS_RTZ_CONVERSION = false; // logits[h] -> every 4 lanes hold 4 heads, each lane holds 4 tokens, there // are 4x16 tokens across warp _B16x4 logits[QHLOOP]; - #pragma unroll for (int h = 0; h < QHLOOP; h++) { - logits[h] = from_floatx4(dout[h]); + if constexpr (LOGITS_RTZ_CONVERSION) { + // use rtz for faster performance with no perceivable accuracy loss + logits[h] = from_floatx4_rtz(dout[h]); + } else { + logits[h] = from_floatx4(dout[h]); + } } - __shared__ _B16x4 vout_shared[QHLOOP][VHELOOP][WARP_SIZE][NWARPS + 1]; - if (warp_start_token_idx >= context_len) { // warp out of context - #pragma unroll for (int qh = 0; qh < QHLOOP; qh++) { - #pragma unroll for (int vh = 0; vh < VHELOOP; vh++) { vout_shared[qh][vh][laneid][warpid] = {0}; } } } else { // warp in context - // iterate across heads - #pragma unroll - for (int qh = 0; qh < QHLOOP; qh++) { - // iterate over each v head elem (within head_size) - #pragma unroll - for (int vh = 0; vh < VHELOOP; vh++) { - floatx4 acc = {0}; - // iterate over tokens - acc = gcn_mfma_instr(logits[qh], Vlocal[vh][0].xy[0], - acc); - acc = gcn_mfma_instr(logits[qh], Vlocal[vh][0].xy[1], - acc); - acc = gcn_mfma_instr(logits[qh], Vlocal[vh][1].xy[0], - acc); - acc = gcn_mfma_instr(logits[qh], Vlocal[vh][1].xy[1], - acc); - acc = gcn_mfma_instr(logits[qh], Vlocal[vh][2].xy[0], - acc); - acc = gcn_mfma_instr(logits[qh], Vlocal[vh][2].xy[1], - acc); - acc = gcn_mfma_instr(logits[qh], Vlocal[vh][3].xy[0], - acc); - acc = gcn_mfma_instr(logits[qh], Vlocal[vh][3].xy[1], - acc); - acc = gcn_mfma_instr(logits[qh], Vlocal[vh][4].xy[0], - acc); - acc = gcn_mfma_instr(logits[qh], Vlocal[vh][4].xy[1], - acc); - acc = gcn_mfma_instr(logits[qh], - Vlocal[vh][5].xy[0], acc); - acc = gcn_mfma_instr(logits[qh], - Vlocal[vh][5].xy[1], acc); - acc = gcn_mfma_instr(logits[qh], - Vlocal[vh][6].xy[0], acc); - acc = gcn_mfma_instr(logits[qh], - Vlocal[vh][6].xy[1], acc); - acc = gcn_mfma_instr(logits[qh], - Vlocal[vh][7].xy[0], acc); - acc = gcn_mfma_instr(logits[qh], - Vlocal[vh][7].xy[1], acc); - vout_shared[qh][vh][laneid][warpid] = from_floatx4(acc); + #define SV_mfma(x) \ + if constexpr (KV_DTYPE != vllm::Fp8KVCacheDataType::kAuto) { \ + Vlocal[vh][x] = convert_b8x8_custom(Vlocalb8[vh][x]); \ + } \ + for (int qh = 0; qh < QHLOOP; qh++) { \ + acc[qh] = gcn_mfma4x4x4_instr( \ + logits[qh], Vlocal[vh][x].xy[0], acc[qh]); \ + acc[qh] = gcn_mfma4x4x4_instr( \ + logits[qh], Vlocal[vh][x].xy[1], acc[qh]); \ + } + + for (int vh = 0; vh < VHELOOP; vh++) { + floatx4 acc[QHLOOP]; + for (int qh = 0; qh < QHLOOP; qh++) { + acc[qh] = {0}; + } + // SoftMax-V calculation + // logits -> token dimension is distributed across lanes + // Vlocal -> token dimension is depthwise within lane + // uses mfma instruction block broadcast for logits + SV_mfma(0); + SV_mfma(1); + SV_mfma(2); + SV_mfma(3); + SV_mfma(4); + SV_mfma(5); + SV_mfma(6); + SV_mfma(7); + + for (int qh = 0; qh < QHLOOP; qh++) { + if constexpr (KV_DTYPE != vllm::Fp8KVCacheDataType::kAuto) { + // post mfma v scale for fp8 + acc[qh] *= *v_scale_ptr; + } + vout_shared[qh][vh][laneid][warpid] = from_floatx4(acc[qh]); } } + + #undef SV_mfma } // warp in context __syncthreads(); + // final write to tmp_out after vout accumulation if (warpid == 0) { - // const float out_scale = (fp8_out_scale_ptr != nullptr) ? - // __fdividef(1.0f,(*fp8_out_scale_ptr)) : 1.0f; const float out_scale = (fp8_out_scale_ptr != nullptr) ? 1.0f / (*fp8_out_scale_ptr) : 1.0f; _B16x4 vout[QHLOOP][VHELOOP]; // iterate across heads - #pragma unroll for (int qh = 0; qh < QHLOOP; qh++) { - // iterate over each v head elem (within head_size) - #pragma unroll + // iterate over each v head elem (within head_size) for (int vh = 0; vh < VHELOOP; vh++) { vout[qh][vh] = {0}; - #pragma unroll for (int w = 0; w < NWARPS; w++) { vout[qh][vh] = addx4(vout[qh][vh], vout_shared[qh][vh][laneid][w]); @@ -709,58 +1308,20 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( } } - if (context_len > partition_size) { - scalar_t* out_ptr = out + - seq_idx * num_heads * max_num_partitions * HEAD_SIZE + - partition_idx * HEAD_SIZE; - const int out_num_partitions = max_num_partitions; - bit16_t* out_ptr_b16 = reinterpret_cast(out_ptr); - #pragma unroll - for (int qh = 0; qh < QHLOOP; qh++) { - #pragma unroll - for (int vh = 0; vh < VHELOOP; vh++) { - const int head_size_elem = vh * WARP_SIZE + laneid; - #pragma unroll - for (int i = 0; i < 4; i++) { - const int head_idx = 4 * qh + i; - if (head_idx < GQA_RATIO) { - out_ptr_b16[(wg_start_head_idx + head_idx) * out_num_partitions * - HEAD_SIZE + - head_size_elem] = vout[qh][vh][i]; - } - } - } - } - } // context_len > partition_size - else { - bit8_t* final_out_ptr_b8; - bit16_t* final_out_ptr_b16; - if constexpr (std::is_same::value) { - final_out_ptr_b8 = final_out + seq_idx * num_heads * HEAD_SIZE; - } else { - OUTT* out_ptr = final_out + seq_idx * num_heads * HEAD_SIZE; - final_out_ptr_b16 = reinterpret_cast(out_ptr); - } - #pragma unroll - for (int qh = 0; qh < QHLOOP; qh++) { - #pragma unroll - for (int vh = 0; vh < VHELOOP; vh++) { - const int head_size_elem = vh * WARP_SIZE + laneid; - #pragma unroll - for (int i = 0; i < 4; i++) { - const int head_idx = 4 * qh + i; - if (head_idx < GQA_RATIO) { - if constexpr (std::is_same::value) { - const float tmpf = - out_scale * to_float_b16(vout[qh][vh][i]); - const OUTT tmp = hip_fp8(tmpf).data; - final_out_ptr_b8[(wg_start_head_idx + head_idx) * HEAD_SIZE + - head_size_elem] = tmp; - } else { - final_out_ptr_b16[(wg_start_head_idx + head_idx) * HEAD_SIZE + - head_size_elem] = vout[qh][vh][i]; - } - } + scalar_t* out_ptr = out + + seq_idx * num_heads * max_num_partitions * HEAD_SIZE + + partition_idx * HEAD_SIZE; + const int out_num_partitions = max_num_partitions; + bit16_t* out_ptr_b16 = reinterpret_cast(out_ptr); + for (int qh = 0; qh < QHLOOP; qh++) { + for (int vh = 0; vh < VHELOOP; vh++) { + const int head_size_elem = vh * WARP_SIZE + laneid; + for (int i = 0; i < 4; i++) { + const int head_idx = 4 * qh + i; + if (head_idx < GQA_RATIO) { + out_ptr_b16[(wg_start_head_idx + head_idx) * out_num_partitions * + HEAD_SIZE + + head_size_elem] = vout[qh][vh][i]; } } } @@ -787,12 +1348,6 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel( const int seq_idx = blockIdx.y; const int context_len = context_lens[seq_idx]; const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE); - if (num_partitions == 1) { - // if num_partitions==1, main kernel will write to out directly, no work in - // reduction kernel - return; - } - constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE; const int warpid = threadIdx.x / WARP_SIZE; const int laneid = threadIdx.x % WARP_SIZE; @@ -957,8 +1512,6 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel( const float inv_global_exp_sum = __fdividef(1.0f, shared_global_exp_sum + 1e-6f); - // const float out_scale = (fp8_out_scale_ptr != nullptr) ? - // __fdividef(1.0f,(*fp8_out_scale_ptr)) : 1.0f; const float out_scale = (fp8_out_scale_ptr != nullptr) ? 1.0f / (*fp8_out_scale_ptr) : 1.0f; acc *= inv_global_exp_sum; @@ -975,9 +1528,10 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel( template -__global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( +__global__ +__launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma16_kernel( const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size] const cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads, // head_size/x, block_size, x] @@ -1000,6 +1554,34 @@ __global__ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_kernel( UNREACHABLE_CODE } +template +__global__ +__launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel( + const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size] + const cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads, + // head_size/x, block_size, x] + const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads, + // head_size, block_size] + const int num_kv_heads, const float scale, + const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq] + const int* __restrict__ context_lens, // [num_seqs] + const int max_num_blocks_per_seq, + const float* __restrict__ alibi_slopes, // [num_heads] + const int q_stride, const int kv_block_stride, const int kv_head_stride, + float* __restrict__ exp_sums, // [num_seqs, num_heads, max_num_partitions] + float* __restrict__ max_logits, // [num_seqs, num_heads, + // max_num_partitions] + scalar_t* __restrict__ out, // [num_seqs, num_heads, max_num_partitions, + // head_size] + OUTT* __restrict__ final_out, // [num_seqs, num_heads, head_size] + int max_ctx_blocks, const float* k_scale, const float* v_scale, + const float* __restrict__ fp8_out_scale_ptr) { + UNREACHABLE_CODE +} + // Grid: (num_heads, num_seqs). template @@ -1018,9 +1600,21 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel( #endif // defined(__HIP__MI300_MI250__) TODO: Add NAVI support -#define LAUNCH_CUSTOM_ATTENTION(GQA_RATIO) \ - paged_attention_ll4mi_QKV_kernel \ +#define LAUNCH_CUSTOM_ATTENTION_MFMA16(GQA_RATIO) \ + paged_attention_ll4mi_QKV_mfma16_kernel \ + <<>>( \ + query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale, \ + block_tables_ptr, context_lens_ptr, max_num_blocks_per_seq, \ + alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, \ + exp_sums_ptr, max_logits_ptr, tmp_out_ptr, out_ptr, max_ctx_blocks, \ + k_scale_ptr, v_scale_ptr, fp8_out_scale_ptr); + +#define LAUNCH_CUSTOM_ATTENTION_MFMA4(GQA_RATIO) \ + paged_attention_ll4mi_QKV_mfma4_kernel \ <<>>( \ query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale, \ block_tables_ptr, context_lens_ptr, max_num_blocks_per_seq, \ @@ -1036,7 +1630,8 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel( context_lens_ptr, max_num_partitions, fp8_out_scale_ptr); template + int BLOCK_SIZE, int HEAD_SIZE, typename OUTT, int PARTITION_SIZE_OLD, + bool ALIBI_ENABLED> void paged_attention_custom_launcher( torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits, torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache, @@ -1070,7 +1665,6 @@ void paged_attention_custom_launcher( const float* k_scale_ptr = reinterpret_cast(k_scale.data_ptr()); const float* v_scale_ptr = reinterpret_cast(v_scale.data_ptr()); - // NOTE: fp8_out_scale is optional. const float* fp8_out_scale_ptr = fp8_out_scale @@ -1079,129 +1673,137 @@ void paged_attention_custom_launcher( OUTT* out_ptr = reinterpret_cast(out.data_ptr()); const int max_ctx_blocks = DIVIDE_ROUND_UP(max_context_len, BLOCK_SIZE); + + // partition size is fixed at 256 since both mfma4 and mfma16 kernels support + // it mfma4 kernel also supports partition size 512 + constexpr int PARTITION_SIZE = 256; const int max_num_partitions = DIVIDE_ROUND_UP(max_context_len, PARTITION_SIZE); const int gqa_ratio = num_heads / num_kv_heads; assert(num_heads % num_kv_heads == 0); assert(head_size == HEAD_SIZE); - constexpr int NTHR = PARTITION_SIZE; + constexpr int NTHR = 256; dim3 grid(num_seqs, max_num_partitions, num_kv_heads); dim3 block(NTHR); const at::cuda::OptionalCUDAGuard device_guard(device_of(query)); const cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + // mfma4 kernel is faster than mfma16 for gqa_ratio <= 4 switch (gqa_ratio) { case 1: - LAUNCH_CUSTOM_ATTENTION(1); + LAUNCH_CUSTOM_ATTENTION_MFMA4(1); break; case 2: - LAUNCH_CUSTOM_ATTENTION(2); + LAUNCH_CUSTOM_ATTENTION_MFMA4(2); break; case 3: - LAUNCH_CUSTOM_ATTENTION(3); + LAUNCH_CUSTOM_ATTENTION_MFMA4(3); break; case 4: - LAUNCH_CUSTOM_ATTENTION(4); + LAUNCH_CUSTOM_ATTENTION_MFMA4(4); break; case 5: - LAUNCH_CUSTOM_ATTENTION(5); + LAUNCH_CUSTOM_ATTENTION_MFMA16(5); break; case 6: - LAUNCH_CUSTOM_ATTENTION(6); + LAUNCH_CUSTOM_ATTENTION_MFMA16(6); break; case 7: - LAUNCH_CUSTOM_ATTENTION(7); + LAUNCH_CUSTOM_ATTENTION_MFMA16(7); break; case 8: - LAUNCH_CUSTOM_ATTENTION(8); + LAUNCH_CUSTOM_ATTENTION_MFMA16(8); break; case 9: - LAUNCH_CUSTOM_ATTENTION(9); + LAUNCH_CUSTOM_ATTENTION_MFMA16(9); break; case 10: - LAUNCH_CUSTOM_ATTENTION(10); + LAUNCH_CUSTOM_ATTENTION_MFMA16(10); break; case 11: - LAUNCH_CUSTOM_ATTENTION(11); + LAUNCH_CUSTOM_ATTENTION_MFMA16(11); break; case 12: - LAUNCH_CUSTOM_ATTENTION(12); + LAUNCH_CUSTOM_ATTENTION_MFMA16(12); break; case 13: - LAUNCH_CUSTOM_ATTENTION(13); + LAUNCH_CUSTOM_ATTENTION_MFMA16(13); break; case 14: - LAUNCH_CUSTOM_ATTENTION(14); + LAUNCH_CUSTOM_ATTENTION_MFMA16(14); break; case 15: - LAUNCH_CUSTOM_ATTENTION(15); + LAUNCH_CUSTOM_ATTENTION_MFMA16(15); break; case 16: - LAUNCH_CUSTOM_ATTENTION(16); + LAUNCH_CUSTOM_ATTENTION_MFMA16(16); break; default: TORCH_CHECK(false, "Unsupported gqa ratio: ", gqa_ratio); break; } - // reduction kernel is only required if max_context_len > partition size, - // otherwise main kernel writes directly to final output - // note there are cases with graphing where max_context_len is the max - // supported by graphing, not the actual max among all the sequences: in that - // case reduction kernel will still run but return immediately - if (max_context_len > PARTITION_SIZE) { - dim3 reduce_grid(num_heads, num_seqs); - dim3 reduce_block(head_size); - const int npar_loops = DIVIDE_ROUND_UP(max_num_partitions, WARP_SIZE); - // support upto 8*64*256=128K context length - switch (npar_loops) { - case 1: - LAUNCH_CUSTOM_REDUCTION(1); - break; - case 2: - LAUNCH_CUSTOM_REDUCTION(2); - break; - case 3: - LAUNCH_CUSTOM_REDUCTION(3); - break; - case 4: - LAUNCH_CUSTOM_REDUCTION(4); - break; - case 5: - LAUNCH_CUSTOM_REDUCTION(5); - break; - case 6: - LAUNCH_CUSTOM_REDUCTION(6); - break; - case 7: - LAUNCH_CUSTOM_REDUCTION(7); - break; - case 8: - LAUNCH_CUSTOM_REDUCTION(8); - break; - default: - TORCH_CHECK(false, "Unsupported npar_loops: ", npar_loops); - break; - } + dim3 reduce_grid(num_heads, num_seqs); + dim3 reduce_block(head_size); + const int npar_loops = DIVIDE_ROUND_UP(max_num_partitions, WARP_SIZE); + // reduction kernel supports upto 8 NPAR_loops * 64 (warp_size) * 256 + // (partition size) = 128K context length + switch (npar_loops) { + case 1: + LAUNCH_CUSTOM_REDUCTION(1); + break; + case 2: + LAUNCH_CUSTOM_REDUCTION(2); + break; + case 3: + LAUNCH_CUSTOM_REDUCTION(3); + break; + case 4: + LAUNCH_CUSTOM_REDUCTION(4); + break; + case 5: + LAUNCH_CUSTOM_REDUCTION(5); + break; + case 6: + LAUNCH_CUSTOM_REDUCTION(6); + break; + case 7: + LAUNCH_CUSTOM_REDUCTION(7); + break; + case 8: + LAUNCH_CUSTOM_REDUCTION(8); + break; + default: + TORCH_CHECK(false, "Unsupported npar_loops: ", npar_loops); + break; } } #define CALL_CUSTOM_LAUNCHER(T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE, OUTT, \ - PSIZE) \ + PSIZE, ALIBI_ENABLED) \ paged_attention_custom_launcher( \ + PSIZE, ALIBI_ENABLED>( \ out, exp_sums, max_logits, tmp_out, query, key_cache, value_cache, \ num_kv_heads, scale, block_tables, context_lens, max_context_len, \ alibi_slopes, k_scale, v_scale, fp8_out_scale); +#define CALL_CUSTOM_LAUNCHER_ALIBI(T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE, \ + OUTT, PSIZE) \ + if (alibi_slopes) { \ + CALL_CUSTOM_LAUNCHER(T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE, OUTT, PSIZE, \ + true); \ + } else { \ + CALL_CUSTOM_LAUNCHER(T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE, OUTT, PSIZE, \ + false); \ + } + #define CALL_CUSTOM_LAUNCHER_PSIZE(T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE, \ OUTT) \ switch (partition_size) { \ case 256: \ - CALL_CUSTOM_LAUNCHER(T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE, OUTT, 256); \ - break; \ - case 512: \ - CALL_CUSTOM_LAUNCHER(T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE, OUTT, 512); \ + CALL_CUSTOM_LAUNCHER_ALIBI(T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE, OUTT, \ + 256); \ break; \ default: \ TORCH_CHECK(false, "Unsupported partition size: ", partition_size); \ @@ -1249,7 +1851,6 @@ void paged_attention_custom_launcher( TORCH_CHECK(false, "Unsupported head size: ", head_size); \ break; \ } - void paged_attention( torch::Tensor& out, // [num_seqs, num_heads, head_size] torch::Tensor& exp_sums, // [num_seqs, num_heads, max_num_partitions] @@ -1298,4 +1899,4 @@ void paged_attention( #undef WARP_SIZE #undef MAX #undef MIN -#undef DIVIDE_ROUND_UP +#undef DIVIDE_ROUND_UP \ No newline at end of file diff --git a/csrc/torch_bindings.cpp b/csrc/torch_bindings.cpp index f3b1169060bb9..8cebcf9b86ac1 100644 --- a/csrc/torch_bindings.cpp +++ b/csrc/torch_bindings.cpp @@ -61,6 +61,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) { ops.def("scaled_silu_and_mul(Tensor! out, Tensor input, Tensor scale) -> ()"); ops.impl("scaled_silu_and_mul", torch::kCUDA, &scaled_silu_and_mul); + ops.def("mul_and_silu(Tensor! out, Tensor input) -> ()"); + ops.impl("mul_and_silu", torch::kCUDA, &mul_and_silu); + // Activation function used in GeGLU with `none` approximation. ops.def("gelu_and_mul(Tensor! out, Tensor input) -> ()"); ops.impl("gelu_and_mul", torch::kCUDA, &gelu_and_mul); diff --git a/docs/dev-docker/README.md b/docs/dev-docker/README.md index c3496358c15d9..c1af8eb645fd5 100644 --- a/docs/dev-docker/README.md +++ b/docs/dev-docker/README.md @@ -1,319 +1,313 @@ -# vllm FP8 Latency and Throughput benchmarks on AMD MI300x +# vllm FP8 Latency and Throughput benchmarks with vLLM on the AMD Instinct™ MI300X accelerator -Documentation for vLLM Inferencing on AMD Instinct platforms. +Documentation for Inferencing with vLLM on AMD Instinct™ MI300X platforms. ## Overview vLLM is a toolkit and library for large language model (LLM) inference and serving. It deploys the PagedAttention algorithm, which reduces memory consumption and increases throughput by leveraging dynamic key and value allocation in GPU memory. vLLM also incorporates many recent LLM acceleration and quantization algorithms, such as fp8 GeMM, fp8 KV cache, continuous batching, flash attention, hip graph, tensor parallel, GPTQ, AWQ, and token speculation. In addition, AMD implements high-performance custom kernels and modules in vLLM to enhance performance further. -This documentation shows some reference performance numbers and the steps to reproduce it for the popular Llama 3.1 series models from Meta with a pre-built AMD vLLM docker optimized for an AMD Instinct™ MI300X accelerator. +This documentation includes information for running the popular Llama 3.1 series models from Meta using a pre-built AMD vLLM docker image optimized for an AMD Instinct™ MI300X or MI325X accelerator. The container is publicly available at [AMD Infinity Hub](https://www.amd.com/en/developer/resources/infinity-hub.html) -It includes: +The pre-built image includes: -- ROCm™ 6.3 -- vLLM 0.6.3 +- ROCm™ 6.3.1 +- vLLM 0.6.6 - PyTorch 2.6dev (nightly) -## System configuration +## Pull latest Docker Image -The performance data below was measured on a server with MI300X accelerators with the following system configuration. The performance might vary with different system configurations. +Pull the most recent validated docker image with `docker pull rocm/vllm-dev:main` -| System | MI300X with 8 GPUs | -|---|---| -| BKC | 24.13 | -| ROCm | version ROCm 6.3 | -| amdgpu | build 2009461 | -| OS | Ubuntu 22.04 | -| Linux Kernel | 5.15.0-117-generic | -| BMCVersion | C2789.BC.0809.00 | -| BiosVersion | C2789.5.BS.1C11.AG.1 | -| CpldVersion | 02.02.00 | -| DCSCMCpldVersion | 02.02.00 | -| CX7 | FW 28.40.1000 | -| RAM | 1 TB | -| Host CPU | Intel(R) Xeon(R) Platinum 8480C | -| Cores | 224 | -| VRAM | 192 GB | -| Power cap | 750 W | -| SCLK/MCLK | 2100 Mhz / 1300 Mhz | +## What is New -## Pull latest +20250124: +- Fix accuracy issue with 405B FP8 Triton FA +- Fixed accuracy issue with TP8 +20250117: +- [Experimental DeepSeek-V3 and DeepSeek-R1 support](#running-deepseek-v3-and-deepseek-r1) -You can pull the image with `docker pull rocm/vllm-dev:main` +## Performance Results -### What is New +The data in the following tables is a reference point to help users validate observed performance. It should not be considered as the peak performance that can be delivered by AMD Instinct™ MI300X accelerator with vLLM. See the MLPerf section in this document for information about MLPerf 4.1 inference results. The performance numbers above were collected using the steps below. -- ROCm 6.3 support -- Potential bug with Tunable Ops not saving due to a PyTorch issue +### Throughput Measurements -Gemms are tuned using PyTorch's Tunable Ops feature (https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/cuda/tunable/README.md) -The gemms are automatically enabled in the docker image, and all stored gemm configs are kept in /app/_gemm_csv in the same image +The table below shows performance data where a local inference client is fed requests at an infinite rate and shows the throughput client-server scenario under maximum load. -### Reproducing benchmark results +| Model | Precision | TP Size | Input | Output | Num Prompts | Max Num Seqs | Throughput (tokens/s) | +|-------|-----------|---------|-------|--------|-------------|--------------|-----------------------| +| Llama 3.1 70B (amd/Llama-3.1-70B-Instruct-FP8-KV) | FP8 | 8 | 128 | 2048 | 3200 | 3200 | 15105 | +| | | | 128 | 4096 | 1500 | 1500 | 10505 | +| | | | 500 | 2000 | 2000 | 2000 | 12664 | +| | | | 2048 | 2048 | 1500 | 1500 | 8239 | +| Llama 3.1 405B (amd/Llama-3.1-405B-Instruct-FP8-KV) | FP8 | 8 | 128 | 2048 | 1500 | 1500 | 4065 | +| | | | 128 | 4096 | 1500 | 1500 | 3171 | +| | | | 500 | 2000 | 2000 | 2000 | 2985 | +| | | | 2048 | 2048 | 500 | 500 | 1999 | -### Use pre-quantized models - -To make it easier to run fp8 Llama 3.1 models on MI300X, the quantized checkpoints are available on AMD Huggingface space as follows - -- -- -- -- +*TP stands for Tensor Parallelism.* -Currently these models are private. Please join to access. +## Latency Measurements -Download the model you want to run. +The table below shows latency measurement, which typically involves assessing the time from when the system receives an input to when the model produces a result. -These FP8 quantized checkpoints were generated with AMD’s Quark Quantizer. For more information about Quark, please refer to +| Model | Precision | TP Size | Batch Size | Input | Output | MI300X Latency (ms) | +|-------|-----------|----------|------------|--------|---------|-------------------| +| Llama 3.1 70B (amd/Llama-3.1-70B-Instruct-FP8-KV) | FP8 | 8 | 1 | 128 | 2048 | 19088.59 | +| | | | 2 | 128 | 2048 | 19610.46 | +| | | | 4 | 128 | 2048 | 19911.30 | +| | | | 8 | 128 | 2048 | 21858.80 | +| | | | 16 | 128 | 2048 | 23537.59 | +| | | | 32 | 128 | 2048 | 25342.94 | +| | | | 64 | 128 | 2048 | 32548.19 | +| | | | 128 | 128 | 2048 | 45216.37 | +| | | | 1 | 2048 | 2048 | 19154.43 | +| | | | 2 | 2048 | 2048 | 19670.60 | +| | | | 4 | 2048 | 2048 | 19976.32 | +| | | | 8 | 2048 | 2048 | 22485.63 | +| | | | 16 | 2048 | 2048 | 25246.27 | +| | | | 32 | 2048 | 2048 | 28967.08 | +| | | | 64 | 2048 | 2048 | 39920.41 | +| | | | 128 | 2048 | 2048 | 59514.25 | +| Llama 3.1 405B (amd/Llama-3.1-70B-Instruct-FP8-KV) | FP8 | 8 | 1 | 128 | 2048 | 51739.70 | +| | | | 2 | 128 | 2048 | 52769.15 | +| | | | 4 | 128 | 2048 | 54557.07 | +| | | | 8 | 128 | 2048 | 56901.86 | +| | | | 16 | 128 | 2048 | 60432.12 | +| | | | 32 | 128 | 2048 | 67353.01 | +| | | | 64 | 128 | 2048 | 81085.33 | +| | | | 128 | 128 | 2048 | 116138.51 | +| | | | 1 | 2048 | 2048 | 52217.76 | +| | | | 2 | 2048 | 2048 | 53227.47 | +| | | | 4 | 2048 | 2048 | 55512.44 | +| | | | 8 | 2048 | 2048 | 59931.41 | +| | | | 16 | 2048 | 2048 | 66890.14 | +| | | | 32 | 2048 | 2048 | 80687.64 | +| | | | 64 | 2048 | 2048 | 108503.12 | +| | | | 128 | 2048 | 2048 | 168845.50 | -### Quantize your own models +*TP stands for Tensor Parallelism.* -This step is optional for you to use quantized models on your own. Take Llama 3.1 405B as an example. +## Reproducing Benchmarked Results -Download the Model View the Llama-3.1-405B model at . Ensure that you have been granted access, and apply for it if you do not have access. +### Preparation - Obtaining access to models -If you do not already have a HuggingFace token, open your user profile (https://huggingface.co/settings/profile), select "Access Tokens", press "+ Create New Token", and create a new Read token. +The vllm-dev docker image should work with any model supported by vLLM. When running with FP8, AMD has quantized models available for a variety of popular models, or you can quantize models yourself using Quark. If needed, the vLLM benchmark scripts will automatically download models and then store them in a Hugging Face cache directory for reuse in future tests. Alternatively, you can choose to download the model to the cache (or to another directory on the system) in advance. -Install the `huggingface-cli` (if not already available on your system) and log in with the token you created earlier and download the model. The instructions in this document assume that the model will be stored under `/data/llama-3.1`. You can store the model in a different location, but then you'll need to update other commands accordingly. The model is quite large and will take some time to download; it is recommended to use tmux or screen to keep your session running without getting disconnected. +Many HuggingFace models, including Llama-3.1, have gated access. You will need to set up an account at (https://huggingface.co), search for the model of interest, and request access if necessary. You will also need to create a token for accessing these models from vLLM: open your user profile (https://huggingface.co/settings/profile), select "Access Tokens", press "+ Create New Token", and create a new Read token. - sudo pip install -U "huggingface_hub[cli]" - - huggingface-cli login +### System optimization -Enter the token you created earlier; you do NOT need to save it as a git credential +Before running performance tests you should ensure the system is optimized according to the [ROCm Documentation](https://rocm.docs.amd.com/en/latest/how-to/system-optimization/mi300x.html). In particular, it is important to ensure that NUMA auto-balancing is disabled. -Create the directory for Llama 3.1 models (if it doesn't already exist) +*Note: Check that NUMA balancing is properly set by inspecting the output of the command below, which should have a value of 0, with, `cat /proc/sys/kernel/numa_balancing`* - sudo mkdir -p /data/llama-3.1 - - sudo chmod -R a+w /data/llama-3.1 +### Launch AMD vLLM Docker -Download the model +Download and launch the docker. The HF_TOKEN is required to be set (either here or after launching the container) if you want to allow vLLM to download gated models automatically; use your HuggingFace token in place of `` in the command below: - huggingface-cli download meta-llama/Llama-3.1-405B-Instruct --exclude "original/*" --local-dir /data/llama-3.1/Llama-3.1-405B-Instruct +```bash +docker run -it --rm --ipc=host --network=host --group-add render \ + --privileged --security-opt seccomp=unconfined \ + --cap-add=CAP_SYS_ADMIN --cap-add=SYS_PTRACE \ + --device=/dev/kfd --device=/dev/dri --device=/dev/mem \ + -e HF_HOME=/data \ + -e HF_TOKEN= \ + -v /data:/data \ + rocm/vllm-dev:main +``` -Similarly, you can download Llama-3.1-70B and Llama-3.1-8B. +Note: The instructions in this document use `/data` to store the models. If you choose a different directory, you will also need to make that change to the host volume mount when launching the docker container. For example, `-v /home/username/models:/data` in place of `-v /data:/data` would store the models in /home/username/models on the host. Some models can be quite large; please ensure that you have sufficient disk space prior to downloading the model. Since the model download may take a long time, you can use `tmux` or `screen` to avoid getting disconnected. -[Download and install Quark](https://quark.docs.amd.com/latest/install.html) +### Downloading models with huggingface-cli -Run the quantization script in the example folder using the following command line: -export MODEL_DIR = [local model checkpoint folder] or meta-llama/Llama-3.1-405B-Instruct - -#### single GPU - - python3 quantize_quark.py \ - --model_dir $MODEL_DIR \ - --output_dir Llama-3.1-405B-Instruct-FP8-KV \ - --quant_scheme w_fp8_a_fp8 \ - --kv_cache_dtype fp8 \ - --num_calib_data 128 \ - --model_export quark_safetensors \ - --no_weight_matrix_merge - -#### If model size is too large for single GPU, please use multi GPU instead - - python3 quantize_quark.py \ - --model_dir $MODEL_DIR \ - --output_dir Llama-3.1-405B-Instruct-FP8-KV \ - --quant_scheme w_fp8_a_fp8 \ - --kv_cache_dtype fp8 \ - --num_calib_data 128 \ - --model_export quark_safetensors \ - --no_weight_matrix_merge \ - --multi_gpu +If you would like want to download models directly (instead of allowing vLLM to download them automatically), you can use the huggingface-cli inside the running docker container. (remove an extra white space) Login using the token that you created earlier. (Note, it is not necessary to save it as a git credential.) -### Launch AMD vLLM Docker +```bash +huggingface-cli login +``` -Download and launch the docker, +You can download a model to the huggingface-cache directory using a command similar to the following (substituting the name of the model you wish to download): - docker run -it --rm --ipc=host --network=host --group-add render \ - --privileged --security-opt seccomp=unconfined \ - --cap-add=CAP_SYS_ADMIN --cap-add=SYS_PTRACE \ - --device=/dev/kfd --device=/dev/dri --device=/dev/mem \ - -v /data/llama-3.1:/data/llm \ - rocm/vllm-dev:main +```bash +sudo mkdir -p /data/huggingface-cache +sudo chmod -R a+w /data/huggingface-cache +HF_HOME=/data/huggingface-cache huggingface-cli download meta-llama/Llama-3.1-405B-Instruct --exclude "original/*" +``` -### Benchmark with AMD vLLM Docker +Alternatively, you may wish to download the model to a specific directory, e.g. so you can quantize the model with Quark: -There are some system settings to be configured for optimum performance on MI300X. +```bash +sudo mkdir -p /data/llama-3.1 +sudo chmod -R a+w /data/llama-3.1 +huggingface-cli download meta-llama/Llama-3.1-405B-Instruct --exclude "original/*" --local-dir /data/llama-3.1/Llama-3.1-405B-Instruct +``` -#### NUMA balancing setting +In the benchmark commands provided later in this document, replace the model name (e.g. `amd/Llama-3.1-405B-Instruct-FP8-KV`) with the path to the model (e.g. `/data/llama-3.1/Llama-3.1-405B-Instruct`) -To optimize performance, disable automatic NUMA balancing. Otherwise, the GPU might hang until the periodic balancing is finalized. For further details, refer to the AMD Instinct MI300X system optimization guide. +### Use pre-quantized models -Disable automatic NUMA balancing +AMD has provided [FP8-quantized versions](https://huggingface.co/collections/amd/quark-quantized-ocp-fp8-models-66db7936d18fcbaf95d4405c) of several models in order to make them easier to run on MI300X / MI325X, including: - sh -c 'echo 0 > /proc/sys/kernel/numa_balancing' +- +- +- -Check if NUMA balancing is disabled (returns 0 if disabled) +Some models may be private to those who are members of . - cat /proc/sys/kernel/numa_balancing - 0 +These FP8 quantized checkpoints were generated with AMD’s Quark Quantizer. For more information about Quark, please refer to -#### LLM performance settings +### Quantize your own models -Some environment variables enhance the performance of the vLLM kernels and PyTorch's tunableOp on the MI300X accelerator. The settings below are already preconfigured in the Docker image. See the AMD Instinct MI300X workload optimization guide for more information. +This is an optional step if you would like to quantize your own model instead of using AMD's pre-quantized models. These instructions use Llama-3.1-405B as an example, but the commands are similar for other models. -##### vLLM performance environment variables +First download the model from to the /data/llama-3.1 directory as described above. - export VLLM_USE_TRITON_FLASH_ATTN=0 - export NCCL_MIN_NCHANNELS=112 - export VLLM_FP8_PADDING=1 +[Download and install Quark](https://quark.docs.amd.com/latest/install.html) -You can set both PYTORCH_TUNABLEOP_ENABLED and PYTORCH_TUNABLEOP_TUNING to 1 to performance GEMM tuning for the 1st benchmark run. -It will take some time to complete the tuning during the benchmark. After tuning, it will generate several csv files as the performance lookup database. For the subsequent benchmark runs, you can keep +Run the quantization script in the example folder using the following command line: -PYTORCH_TUNABLEOP_ENABLED as 1 and set -PYTORCH_TUNABLEOP_TUNING to 0 to use the selected kernels. +```bash +# path to quark quantization script +export QUARK_DIR=/data/quark-0.6.0+dba9ca364/examples/torch/language_modeling/llm_ptq/quantize_quark.py +# path to Model +export MODEL_DIR=/data/llama-3.1/Llama-3.1-405B-Instruct +python3 $QUARK_DIR \ +--model_dir $MODEL_DIR \ +--output_dir Llama-3.1-405B-Instruct-FP8-KV \ +--kv_cache_dtype fp8 \ +--quant_scheme w_fp8_a_fp8 \ +--num_calib_data 128 \ +--model_export quark_safetensors \ +--no_weight_matrix_merge \ +--multi_gpu +``` -##### vLLM engine performance settings +Note: the `--multi_gpu` parameter can be omitted for small models that fit on a single GPU. -vLLM provides a number of engine options which can be changed to improve performance. -Refer for the complete list of vLLM engine options. -Below is a list of options which are useful: -- **--max-model-len** : Maximum context length supported by the model instance. Can be set to a lower value than model configuration value to improve performance and gpu memory utilization. -- **--max-num-batched-tokens** : The maximum prefill size, i.e., how many prompt tokens can be packed together in a single prefill. Set to a higher value to improve prefill performance at the cost of higher gpu memory utilization. 65536 works well for LLama models. -- **--max-num-seqs** : The maximum decode batch size. Set to a value higher than the default(256) to improve decode throughput. Higher values will also utilize more KV cache memory. Too high values can cause KV cache space to run out which will lead to decode preemption. 512/1024 works well for LLama models. -- **--max-seq-len-to-capture** : Maximum sequence length for which Hip-graphs are captured and utilized. It's recommended to use Hip-graphs for the best decode performance. The default value of this parameter is 8K, which is lower than the large context lengths supported by recent models such as LLama. Set this parameter to max-model-len or maximum context length supported by the model for best performance. -- **--gpu-memory-utilization** : The ratio of GPU memory reserved by a vLLM instance. Default value is 0.9. It's recommended to set this to 0.99 to increase KV cache space. +## Performance testing with AMD vLLM Docker -Note: vLLM's server creation command line (vllm serve) supports the above parameters as command line arguments. - -##### Online Gemm Tuning +### Performance environment variables -Online Gemm tuning for small decode batch sizes can improve performance in some cases. e.g. Llama 70B upto Batch size 8 +Some environment variables enhance the performance of the vLLM kernels on the MI300X / MI325X accelerator. See the AMD Instinct MI300X workload optimization guide for more information. -If you want to do limited online tuning use --enforce-eager and tune for particular batch sizes. See example below. +```bash +export VLLM_USE_TRITON_FLASH_ATTN=0 +``` - export PYTORCH_TUNABLEOP_TUNING=1 - export PYTORCH_TUNABLEOP_ENABLED=1 - export PYTORCH_TUNABLEOP_MAX_TUNING_DURATION_MS=100 - export PYTORCH_TUNABLEOP_MAX_WARMUP_DURATION_MS=10 - export PYTORCH_TUNABLEOP_ROTATING_BUFFER_SIZE=1024 - export PYTORCH_TUNABLEOP_FILENAME=/app/tuned_gemm_csv/bench_latency_tune_device_%d_full.csv +### vLLM engine performance settings - Run the following command for BS=1/2/4/8: +vLLM provides a number of engine options which can be changed to improve performance. Refer to the [vLLM Engine Args](https://docs.vllm.ai/en/stable/usage/engine_args.html) documentation for the complete list of vLLM engine options. - python /app/vllm/benchmarks/benchmark_latency.py \ - --model \ - --quantization fp8 \ - --kv-cache-dtype fp8 \ - --dtype float16 \ - --max-model-len 8192 \ - --num-iters-warmup 5 \ - --num-iters 5 \ - --tensor-parallel-size 8 \ - --input-len 4096 \ - --output-len 512 \ - --batch-size \ - --num-scheduler-steps 10 \ - --enforce-eager +Below is a list of a few of the key vLLM engine arguments for performance; these can be passed to the vLLM benchmark scripts: +- **--max-model-len** : Maximum context length supported by the model instance. Can be set to a lower value than model configuration value to improve performance and gpu memory utilization. +- **--max-num-batched-tokens** : The maximum prefill size, i.e., how many prompt tokens can be packed together in a single prefill. Set to a higher value to improve prefill performance at the cost of higher gpu memory utilization. 65536 works well for LLama models. +- **--max-num-seqs** : The maximum decode batch size (default 256). Using larger values will allow more prompts to be processed concurrently, resulting in increased throughput (possibly at the expense of higher latency). If the value is too large, there may not be enough GPU memory for the KV cache, resulting in requests getting preempted. The optimal value will depend on the GPU memory, model size, and maximum context length. +- **--max-seq-len-to-capture** : Maximum sequence length for which Hip-graphs are captured and utilized. It's recommended to use Hip-graphs for the best decode performance. The default value of this parameter is 8K, which is lower than the large context lengths supported by recent models such as LLama. Set this parameter to max-model-len or maximum context length supported by the model for best performance. +- **--gpu-memory-utilization** : The ratio of GPU memory reserved by a vLLM instance. Default value is 0.9. Increasing the value (potentially as high as 0.99) will increase the amount of memory available for KV cache. When running in graph mode (i.e. not using `--enforce-eager`), it may be necessary to use a slightly smaller value of 0.92 - 0.95 to ensure adequate memory is available for the HIP graph. -The tuned file will be generated for device 0 only at /app/tuned_gemm_csv/bench_latency_tune_device_0_full.csv. Copy this file to /app/tuned_gemm_csv/bench_latency_tune_device__full.csv for D=1 through 7. +### Latency Benchmark -After the above steps, retain the environment variables set earlier, but set export PYTORCH_TUNABLEOP_TUNING=0 to disable online tuning, and use the tuned solutions. +vLLM's benchmark_latency.py script measures end-to-end latency for a specified model, input/output length, and batch size. -##### Latency Benchmark +You can run latency tests for FP8 models with: -Benchmark Meta-Llama-3.1-405B FP8 with input 128 tokens, output 128 tokens, batch size 32 and tensor parallelism 8 as an example, +```bash +export VLLM_USE_TRITON_FLASH_ATTN=0 +MODEL=amd/Llama-3.1-405B-Instruct-FP8-KV +BS=1 +IN=128 +OUT=2048 +TP=8 - python /app/vllm/benchmarks/benchmark_latency.py \ - --model /data/llm/Meta-Llama-3.1-405B-Instruct-FP8-KV \ +python3 /app/vllm/benchmarks/benchmark_latency.py \ + --distributed-executor-backend mp \ --quantization fp8 \ --kv-cache-dtype fp8 \ - --dtype half \ - --gpu-memory-utilization 0.99 \ - --distributed-executor-backend mp \ - --tensor-parallel-size 8 \ - --batch size 32 \ - --input-len 128 \ - --output-len 128 + --dtype float16 \ + --gpu-memory-utilization 0.9 \ + --trust-remote-code \ + --model $MODEL \ + --batch-size $BS \ + --input-len $IN \ + --output-len $OUT \ + --tensor-parallel-size $TP \ + --num-iters-warmup 3 \ + --num-iters 5 \ + --output-json output.json +``` -If you want to run Meta-Llama-3.1-405B FP16, please run +For FP16 models, remove `--quantization fp8 --kv-cache-dtype fp8`. - python /app/vllm/benchmarks/benchmark_latency.py \ - --model /data/llm/Meta-Llama-3.1-405B-Instruct \ - --dtype float16 \ - --gpu-memory-utilization 0.99 \ - --distributed-executor-backend mp \ - --tensor-parallel-size 8 \ - --batch size 32 \ - --input-len 128 \ - --output-len 128 +When measuring models with long context lengths, performance may improve by setting `--max-model-len` to a smaller value. It is important, however, to ensure that the `--max-model-len` is at least as large as the IN + OUT token counts. -You can change various input-len, output-len, batch size and run the benchmark as well. When output-len is 1, it measures prefill latency (TTFT). -Decoding latency (TPOT) can be calculated based on the measured latency. +To estimate Time To First Token (TTFT) with the benchmark_latency.py script, set the OUT to 1 token. It is also recommended to use `--enforce-eager` to get a more accurate measurement of the time that it actually takes to generate the first token. (For a more comprehensive measurement of TTFT, use the Online Serving Benchmark.) -For more information about the parameters, please run +For additional information about the available parameters run: - /app/vllm/benchmarks/benchmark_latency.py -h +```bash +/app/vllm/benchmarks/benchmark_latency.py -h +``` -##### Throughput Benchmark +### Throughput Benchmark -Benchmark Meta-Llama-3.1-405B FP8 with input 128 tokens, output 128 tokens and tensor parallelism 8 as an example, +vLLM's benchmark_throughput.py script measures offline throughput. It can either use an input dataset or random prompts with fixed input/output lengths. - python /app/vllm/benchmarks/benchmark_throughput.py \ - --model /data/llm/Meta-Llama-3.1-405B-Instruct-FP8-KV \ - --quantization fp8 \ - --kv-cache-dtype fp8 \ - --dtype half \ - --gpu-memory-utilization 0.99 \ - --num-prompts 2000 \ - --distributed-executor-backend mp \ - --num-scheduler-steps 10 \ - --tensor-parallel-size 8 \ - --input-len 128 \ - --output-len 128 +You can run latency tests for FP8 models with: -If you want to run Meta-Llama-3.1-405B FP16, please run +```bash +export VLLM_USE_TRITON_FLASH_ATTN=0 +MODEL=amd/Llama-3.1-405B-Instruct-FP8-KV +IN=128 +OUT=2048 +TP=8 +PROMPTS=1500 +MAX_NUM_SEQS=1500 - python /app/vllm/benchmarks/benchmark_throughput.py \ - --model /data/llm/Meta-Llama-3.1-405B-Instruct \ +python3 /app/vllm/benchmarks/benchmark_throughput.py \ + --distributed-executor-backend mp \ + --quantization fp8 \ + --kv-cache-dtype fp8 \ --dtype float16 \ --gpu-memory-utilization 0.9 \ - --num-prompts 2000 \ - --distributed-executor-backend mp \ + --trust-remote-code \ --num-scheduler-steps 10 \ - --tensor-parallel-size 8 \ - --input-len 128 \ - --output-len 128 \ - --swap-space 16 \ + --enable-chunked-prefill False \ + --model $MODEL \ --max-model-len 8192 \ - --max-num-batched-tokens 65536 \ - --swap-space - --max-model-len - --gpu-memory-utilization 0.99 - -For fp8 quantized Llama3.18B/70B models: - - Recommend TP:1 for Llama3.1-8B, 8 for Llama3.1-70B - Recommend NSCHED: 10 for Llama3.1-8B, 8 for Llama3.1-70B + --max-num-batched-tokens 131072 \ + --max-seq-len-to-capture 131072 \ + --input-len $IN \ + --output-len $OUT \ + --tensor-parallel-size $TP \ + --num-prompts $PROMPTS \ + --max-num-seqs $MAX_NUM_SEQS \ + --output-json output.json +``` -You can change various input-len, output-len, num-prompts and run the benchmark as well. -Please note num-scheduler-step is a new feature added in vLLM 0.6.0. It can improve the decoding latency and throughput, however, it may increase the prefill latency. +For FP16 models, remove `--quantization fp8 --kv-cache-dtype fp8`. -For more information about the parameters, please run +When measuring models with long context lengths, performance may improve by setting `--max-model-len` to a smaller value (8192 in this example). It is important, however, to ensure that the `--max-model-len` is at least as large as the IN + OUT token counts. - /app/vllm/benchmarks/benchmark_throughput.py -h +It is important to tune vLLM’s --max-num-seqs value to an appropriate value depending on the model and input/output lengths. Larger values will allow vLLM to leverage more of the GPU memory for KV Cache and process more prompts concurrently. But if the value is too large, the KV cache will reach its capacity and vLLM will have to cancel and re-process some prompts. Suggested values for various models and configurations are listed below. -Tensor parallelism (TP) parameters depends on the model size. For Llama 3.1 70B and 8B model, TP 1 can be used as well for MI300X. In general, TP 8 and 1 is recommended to achieve the optimum performance. +For models that fit on a single GPU, it is usually best to run with `--tensor-parallel-size 1`. Requests can be distributed across multiple copies of vLLM running on different GPUs. This will be more efficient than running a single copy of the model with `--tensor-parallel-size 8`. (Note: the benchmark_throughput.py script does not include direct support for using multiple copies of vLLM) -##### Online Server Benchmark +For optimal performance, the PROMPTS value should be a multiple of the MAX_NUM_SEQS value -- for example, if MAX_NUM_SEQS=1500 then the PROMPTS value could be 1500, 3000, etc. If PROMPTS is smaller than MAX_NUM_SEQS then there won’t be enough prompts for vLLM to maximize concurrency. -Make the following changes if required +For additional information about the available parameters run: -/app/vllm/benchmarks/backend_request_func.py +```bash +python3 /app/vllm/benchmarks/benchmark_throughput.py -h +``` -line 242 + "ignore_eos": True, +### Online Serving Benchmark -/app/vllm/benchmarks/benchmark_serving.py -line 245 - interval = np.random.exponential(1.0 / request_rate) -line 245 + ## interval = np.random.exponential(1.0 / request_rate) -line 246 + interval = 1.0 / request_rate +Benchmark Llama-3.1-70B with input 4096 tokens, output 512 tokens and tensor parallelism 8 as an example, -Benchmark Meta-Llama-3.1-70B with input 4096 tokens, output 512 tokens and tensor parallelism 8 as an example, - - vllm serve /data/llm/Meta-Llama-3.1-70B-Instruct-FP8-KV \ +```bash +export VLLM_USE_TRITON_FLASH_ATTN=0 +vllm serve amd/Llama-3.1-70B-Instruct-FP8-KV \ --swap-space 16 \ --disable-log-requests \ --quantization fp8 \ @@ -324,40 +318,87 @@ Benchmark Meta-Llama-3.1-70B with input 4096 tokens, output 512 tokens and tenso --max-num-batched-tokens 65536 \ --gpu-memory-utilization 0.99 \ --num_scheduler-steps 10 +``` Change port (for example --port 8005) if port=8000 is currently being used by other processes. -run client in a separate terminal. Use port_id from previous step else port-id=8000. +Run client in a separate terminal. Use port_id from previous step else port-id=8000. - python /app/vllm/benchmarks/benchmark_serving.py \ +```bash +python /app/vllm/benchmarks/benchmark_serving.py \ --port 8000 \ - --model /data/llm/Meta-Llama-3.1-70B-Instruct-FP8-KV \ + --model amd/Llama-3.1-70B-Instruct-FP8-KV \ --dataset-name random \ --random-input-len 4096 \ --random-output-len 512 \ --request-rate 1 \ + --ignore-eos \ --num-prompts 500 \ --percentile-metrics ttft,tpot,itl,e2el +``` Once all prompts are processed, terminate the server gracefully (ctrl+c). -##### CPX mode +### Running DeepSeek-V3 and DeepSeek-R1 + +We have experimental support for running both DeepSeek-V3 and DeepSeek-R1 models. +*Note there are currently limitations and `--max-model-len` cannot be greater than 32768* + +```bash +docker run -it --rm --ipc=host --network=host --group-add render \ + --privileged --security-opt seccomp=unconfined \ + --cap-add=CAP_SYS_ADMIN --cap-add=SYS_PTRACE \ + --device=/dev/kfd --device=/dev/dri --device=/dev/mem \ + -e VLLM_USE_TRITON_FLASH_ATTN=0 \ + -e VLLM_FP8_PADDING=0 \ + rocm/vllm-dev:main +# Online serving +vllm serve deepseek-ai/DeepSeek-V3 \ + --disable-log-requests \ + --tensor-parallel-size 8 \ + --trust-remote-code \ + --max-model-len 32768 + +python3 /app/vllm/benchmarks/benchmark_serving.py \ + --backend vllm \ + --model deepseek-ai/DeepSeek-V3 \ + --max-concurrency 256\ + --dataset-name random \ + --random-input-len 128 \ + --random-output-len 128 \ + --num-prompts 1000 + +# Offline throughput +python3 /app/vllm/benchmarks/benchmark_throughput.py --model deepseek-ai/DeepSeek-V3 \ + --input-len <> --output-len <> --tensor-parallel-size 8 \ + --quantization fp8 --kv-cache-dtype fp8 --dtype float16 \ + --max-model-len 32768 --trust-remote-code +# Offline Latency +python benchmarks/benchmark_latency.py --model deepseek-ai/DeepSeek-V3 \ +--tensor-parallel-size 8 --trust-remote-code --max-model-len 32768 \ +--batch-size <> --input-len <> --output-len <> +``` + +### CPX mode Currently only CPX-NPS1 mode is supported. So ONLY tp=1 is supported in CPX mode. But multiple instances can be started simultaneously (if needed) in CPX-NPS1 mode. -Set GPUs in CPX mode +Set GPUs in CPX mode with: - rocm-smi --setcomputepartition cpx +```bash +rocm-smi --setcomputepartition cpx +``` Example of running Llama3.1-8B on 1 CPX-NPS1 GPU with input 4096 and output 512. As mentioned above, tp=1. - HIP_VISIBLE_DEVICES=0 \ - python3 /app/vllm/benchmarks/benchmark_throughput.py \ +```bash +HIP_VISIBLE_DEVICES=0 \ +python3 /app/vllm/benchmarks/benchmark_throughput.py \ --max-model-len 4608 \ --num-scheduler-steps 10 \ --num-prompts 100 \ - --model /data/llm/Meta-Llama-3.1-70B-Instruct-FP8-KV \ + --model amd/Llama-3.1-8B-Instruct-FP8-KV \ --input-len 4096 \ --output-len 512 \ --dtype float16 \ @@ -365,10 +406,13 @@ Example of running Llama3.1-8B on 1 CPX-NPS1 GPU with input 4096 and output 512. --output-json \ --quantization fp8 \ --gpu-memory-utilization 0.99 +``` Set GPU to SPX mode. - rocm-smi --setcomputepartition spx +```bash +rocm-smi --setcomputepartition spx +``` ### Speculative Decoding @@ -376,27 +420,33 @@ Speculative decoding is one of the key features in vLLM. It has been supported o Without Speculative Decoding - - python benchmark_latency.py --model /models/models--amd--Meta-Llama-3.1-405B-Instruct-FP8-KV/ --max-model-len 26720 -tp 8 --batch-size 1 --use-v2-block-manager --input-len 1024 --output-len 128 +```bash +export VLLM_USE_TRITON_FLASH_ATTN=0 +python /app/vllm/benchmarks/benchmark_latency.py --model amd/Llama-3.1-405B-Instruct-FP8-KV --max-model-len 26720 -tp 8 --batch-size 1 --input-len 1024 --output-len 128 +``` With Speculative Decoding - - python benchmark_latency.py --model /models/models--amd--Meta-Llama-3.1-405B-Instruct-FP8-KV/ --max-model-len 26720 -tp 8 --batch-size 1 --use-v2-block-manager --input-len 1024 --output-len 128 --speculative-model /models/models--amd--Meta-Llama-3.1-8B-Instruct-FP8-KV/ --num-speculative-tokens 5 +```bash +export VLLM_USE_TRITON_FLASH_ATTN=0 +python /app/vllm/benchmarks/benchmark_latency.py --model amd/Llama-3.1-405B-Instruct-FP8-KV --max-model-len 26720 -tp 8 --batch-size 1 --input-len 1024 --output-len 128 --speculative-model amd/Llama-3.1-8B-Instruct-FP8-KV --num-speculative-tokens 5 +``` You should see some performance improvement about the e2e latency. -### MMLU_PRO_Biology Accuracy Eval +## MMLU_PRO_Biology Accuracy Evaluation -### fp16 +### FP16 -vllm (pretrained=models--meta-llama--Meta-Llama-3.1-405B-Instruct/snapshots/069992c75aed59df00ec06c17177e76c63296a26,dtype=float16,tensor_parallel_size=8), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 64 +vllm (pretrained=models--meta-llama--Llama-3.1-405B-Instruct/snapshots/069992c75aed59df00ec06c17177e76c63296a26,dtype=float16,tensor_parallel_size=8), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 64 | Tasks |Version| Filter |n-shot| Metric | |Value | |Stderr| |-------|------:|--------------|-----:|-----------|---|-----:|---|-----:| |biology| 0|custom-extract| 5|exact_match|↑ |0.8466|± |0.0135| -### fp8 +### FP8 -vllm (pretrained=models--meta-llama--Meta-Llama-3.1-405B-Instruct/snapshots/069992c75aed59df00ec06c17177e76c63296a26,dtype=float16,quantization=fp8,quantized_weights_path=/llama.safetensors,tensor_parallel_size=8), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 32 +vllm (pretrained=models--meta-llama--Llama-3.1-405B-Instruct/snapshots/069992c75aed59df00ec06c17177e76c63296a26,dtype=float16,quantization=fp8,quantized_weights_path=/llama.safetensors,tensor_parallel_size=8), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 32 | Tasks |Version| Filter |n-shot| Metric | |Value| |Stderr| |-------|------:|--------------|-----:|-----------|---|----:|---|-----:| @@ -404,25 +454,19 @@ vllm (pretrained=models--meta-llama--Meta-Llama-3.1-405B-Instruct/snapshots/0699 ## Performance -### LLaMA2/3 *MLPerf* 70B - -Please refer to the MLPerf instructions for recreating the MLPerf numbers. - -## Version - -### Release Notes - -20240906a: Legacy quantization formats required `--quantization fp8_rocm` as a flag instead of `--quantization fp8` +### MLPerf Performance Results -Updated: +#### LLama-2-70B -vLLM: +Please refer to the [Benchmarking Machine Learning using ROCm and AMD GPUs: Reproducing Our MLPerf Inference Submission — ROCm Blogs](https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-4-1/README.html) for information on reproducing MLPerf 4.1 Inference results. Note that due to changes in vLLM, it is not possible to use these instructions with the current rocm/vllm-dev docker image. Due to recent changes in vLLM, the instructions for MLPerf 4.1 submission do not apply to the current rocm/vllm-dev docker image. -### Docker Manifest +## Docker Manifest To reproduce the release docker: +```bash git clone https://github.com/ROCm/vllm.git cd vllm - git checkout 2c60adc83981ada77a77b2adda78ef109d2e2e2b + git checkout 8e87b08c2a284c1a20eb3d8e0fbdc84918bf27dc docker build -f Dockerfile.rocm -t --build-arg BUILD_HIPBLASLT=1 --build-arg USE_CYTHON=1 . +``` diff --git a/docs/source/api/multimodal/inputs.md b/docs/source/api/multimodal/inputs.md index 76b2fb95a5009..21bd938be9e89 100644 --- a/docs/source/api/multimodal/inputs.md +++ b/docs/source/api/multimodal/inputs.md @@ -43,7 +43,7 @@ ``` ```{eval-rst} -.. autoclass:: vllm.multimodal.inputs.MultiModalInputsV2 +.. autoclass:: vllm.multimodal.inputs.MultiModalInputs :members: :show-inheritance: ``` diff --git a/docs/source/community/blog.md b/docs/source/community/blog.md new file mode 100644 index 0000000000000..e8030edfa02ee --- /dev/null +++ b/docs/source/community/blog.md @@ -0,0 +1,3 @@ +# vLLM Blog + +vLLM blog posts are published [here](https://blog.vllm.ai/). diff --git a/docs/source/community/meetups.md b/docs/source/community/meetups.md index 43fa9ee616096..ab5ea147f4c6a 100644 --- a/docs/source/community/meetups.md +++ b/docs/source/community/meetups.md @@ -4,6 +4,7 @@ We host regular meetups in San Francisco Bay Area every 2 months. We will share the project updates from the vLLM team and have guest speakers from the industry to share their experience and insights. Please find the materials of our previous meetups below: +- [The eighth vLLM meetup](https://lu.ma/zep56hui), with Google Cloud, January 22nd 2025. [[Slides]](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing) - [The seventh vLLM meetup](https://lu.ma/h0qvrajz), with Snowflake, November 14th 2024. [[Slides]](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing) - [The sixth vLLM meetup](https://lu.ma/87q3nvnh), with NVIDIA, September 9th 2024. [[Slides]](https://docs.google.com/presentation/d/1wrLGwytQfaOTd5wCGSPNhoaW3nq0E-9wqyP7ny93xRs/edit?usp=sharing) - [The fifth vLLM meetup](https://lu.ma/lp0gyjqr), with AWS, July 24th 2024. [[Slides]](https://docs.google.com/presentation/d/1RgUD8aCfcHocghoP3zmXzck9vX3RCI9yfUAB2Bbcl4Y/edit?usp=sharing) diff --git a/docs/source/contributing/overview.md b/docs/source/contributing/overview.md index e92104399342d..36cf8e7440eca 100644 --- a/docs/source/contributing/overview.md +++ b/docs/source/contributing/overview.md @@ -25,10 +25,12 @@ Check out the [building from source](#build-from-source) documentation for detai ```bash pip install -r requirements-dev.txt -# linting and formatting -bash format.sh -# Static type checking -mypy +# Linting, formatting and static type checking +pre-commit install + +# You can manually run pre-commit with +pre-commit run --all-files + # Unit tests pytest tests/ ``` @@ -88,7 +90,8 @@ If the PR spans more than one category, please include all relevant prefixes. The PR needs to meet the following code quality standards: - We adhere to [Google Python style guide](https://google.github.io/styleguide/pyguide.html) and [Google C++ style guide](https://google.github.io/styleguide/cppguide.html). -- Pass all linter checks. Please use to format your code. +- Pass all linter checks. Please use `pre-commit` to format your code. See + if `pre-commit` is new to you. - The code needs to be well-documented to ensure future contributors can easily understand the code. - Include sufficient tests to ensure the project stays correct and robust. This diff --git a/docs/source/contributing/vulnerability_management.md b/docs/source/contributing/vulnerability_management.md index 422dc13e6a644..a9bbfde2af770 100644 --- a/docs/source/contributing/vulnerability_management.md +++ b/docs/source/contributing/vulnerability_management.md @@ -41,3 +41,20 @@ You may use the `#security` channel in the [VLLM Slack](https://slack.vllm.ai) to discuss security-related topics. However, please do not disclose any vulnerabilities in this channel. If you need to report a vulnerability, please use the GitHub security advisory system or contact a VMT member privately. + +## Vulnerability Disclosure + +The process for disclosing vulnerabilities is the following: + +- The VMT will work with the project maintainers to develop a fix for the + vulnerability. +- The VMT will coordinate with the reporter and project maintainers to prepare a + security advisory that adequately describes the vulnerability and its impact. +- The VMT will coordinate with the project maintainers to publish a fix and + release an update that includes that fix. +- The VMT will publish the security advisory on GitHub. Release notes will be + updated to include a reference to the security advisory. + +The VMT and project maintainers will work to minimize the amount of time in +between disclosing any public information about the vulnerability and making a +release and advisory available. diff --git a/docs/source/deployment/docker.md b/docs/source/deployment/docker.md index 9e301483ef7f9..438be47316f3b 100644 --- a/docs/source/deployment/docker.md +++ b/docs/source/deployment/docker.md @@ -19,6 +19,8 @@ $ docker run --runtime nvidia --gpus all \ --model mistralai/Mistral-7B-v0.1 ``` +You can add any other you need after the image tag (`vllm/vllm-openai:latest`). + ```{note} You can either use the `ipc=host` flag or `--shm-size` flag to allow the container to access the host's shared memory. vLLM uses PyTorch, which uses shared @@ -40,6 +42,9 @@ DOCKER_BUILDKIT=1 docker build . --target vllm-openai --tag vllm/vllm-openai By default vLLM will build for all GPU types for widest distribution. If you are just building for the current GPU type the machine is running on, you can add the argument `--build-arg torch_cuda_arch_list=""` for vLLM to find the current GPU type and build for that. + +If you are using Podman instead of Docker, you might need to disable SELinux labeling by +adding `--security-opt label=disable` when running `podman build` command to avoid certain [existing issues](https://github.com/containers/buildah/discussions/4184). ``` ## Building for Arm64/aarch64 diff --git a/docs/source/features/compatibility_matrix.md b/docs/source/features/compatibility_matrix.md index 86a82eb36df33..47ab616b30686 100644 --- a/docs/source/features/compatibility_matrix.md +++ b/docs/source/features/compatibility_matrix.md @@ -307,7 +307,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar - ✅ - ? - ? - - ✅ + - [✗](gh-issue:11484) - ✅ - ✗ - ? diff --git a/docs/source/features/quantization/fp8.md b/docs/source/features/quantization/fp8.md index da49cd2747228..1398e8a324201 100644 --- a/docs/source/features/quantization/fp8.md +++ b/docs/source/features/quantization/fp8.md @@ -54,16 +54,15 @@ The quantization process involves three main steps: ### 1. Loading the Model -Use `SparseAutoModelForCausalLM`, which wraps `AutoModelForCausalLM`, for saving and loading quantized models: +Load your model and tokenizer using the standard `transformers` AutoModel classes: ```python -from llmcompressor.transformers import SparseAutoModelForCausalLM -from transformers import AutoTokenizer +from transformers import AutoTokenizer, AutoModelForCausalLM MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct" - -model = SparseAutoModelForCausalLM.from_pretrained( - MODEL_ID, device_map="auto", torch_dtype="auto") +model = AutoModelForCausalLM.from_pretrained( + MODEL_ID, device_map="auto", torch_dtype="auto", +) tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) ``` diff --git a/docs/source/features/quantization/fp8_e5m2_kvcache.md b/docs/source/features/quantization/fp8_e5m2_kvcache.md deleted file mode 100644 index 3a81ab17f332f..0000000000000 --- a/docs/source/features/quantization/fp8_e5m2_kvcache.md +++ /dev/null @@ -1,31 +0,0 @@ -(fp8-kv-cache)= - -# FP8 E5M2 KV Cache - -The int8/int4 quantization scheme requires additional scale GPU memory storage, which reduces the expected GPU memory benefits. -The FP8 data format retains 2~3 mantissa bits and can convert float/fp16/bfloat16 and fp8 to each other. - -Here is an example of how to enable this feature: - -```python -from vllm import LLM, SamplingParams -# Sample prompts. -prompts = [ - "Hello, my name is", - "The president of the United States is", - "The capital of France is", - "The future of AI is", -] -# Create a sampling params object. -sampling_params = SamplingParams(temperature=0.8, top_p=0.95) -# Create an LLM. -llm = LLM(model="facebook/opt-125m", kv_cache_dtype="fp8") -# Generate texts from the prompts. The output is a list of RequestOutput objects -# that contain the prompt, generated text, and other information. -outputs = llm.generate(prompts, sampling_params) -# Print the outputs. -for output in outputs: - prompt = output.prompt - generated_text = output.outputs[0].text - print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") -``` diff --git a/docs/source/features/quantization/index.md b/docs/source/features/quantization/index.md index 861cb165c11c2..56ccdb5f00c34 100644 --- a/docs/source/features/quantization/index.md +++ b/docs/source/features/quantization/index.md @@ -14,6 +14,5 @@ bnb gguf int8 fp8 -fp8_e5m2_kvcache -fp8_e4m3_kvcache +quantized_kvcache ``` diff --git a/docs/source/features/quantization/int8.md b/docs/source/features/quantization/int8.md index 82a15d76d352f..592a60d3988b2 100644 --- a/docs/source/features/quantization/int8.md +++ b/docs/source/features/quantization/int8.md @@ -30,14 +30,13 @@ The quantization process involves four main steps: ### 1. Loading the Model -Use `SparseAutoModelForCausalLM`, which wraps `AutoModelForCausalLM`, for saving and loading quantized models: +Load your model and tokenizer using the standard `transformers` AutoModel classes: ```python -from llmcompressor.transformers import SparseAutoModelForCausalLM -from transformers import AutoTokenizer +from transformers import AutoTokenizer, AutoModelForCausalLM MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct" -model = SparseAutoModelForCausalLM.from_pretrained( +model = AutoModelForCausalLM.from_pretrained( MODEL_ID, device_map="auto", torch_dtype="auto", ) tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) diff --git a/docs/source/features/quantization/quantized_kvcache.md b/docs/source/features/quantization/quantized_kvcache.md new file mode 100644 index 0000000000000..9f36c2949e0dd --- /dev/null +++ b/docs/source/features/quantization/quantized_kvcache.md @@ -0,0 +1,147 @@ +(quantized-kvcache)= + +# Quantized KV Cache + +## FP8 KV Cache + +Quantizing the KV cache to FP8 reduces its memory footprint. This increases the number of tokens that can be stored in the cache, improving throughput. + +### FP8 Formats + +[OCP (Open Compute Project)](https://www.opencompute.org) specifies two common 8-bit floating point data formats: + +- E5M2 (5 exponent bits and 2 mantissa bits) +- E4M3FN (4 exponent bits and 3 mantissa bits, often shortened as E4M3) + +The E4M3 format offers higher precision compared to E5M2. However, due to its small dynamic range (±240.0), E4M3 typically requires a higher-precision (FP32) scaling factor alongside each quantized tensor. + +### Current Limitations + +For now, only per-tensor (scalar) scaling factors are supported. Development is ongoing to support scaling factors of a finer granularity (e.g. per-channel). + +### Performance Impact + +The current FP8 KV cache implementation primarily benefits throughput by allowing approximately double the amount of space for KV cache allocation. This enables either: + +- Processing longer context lengths for individual requests, or +- Handling more concurrent request batches + +However, there are currently no latency improvements as the implementation does not yet include fused dequantization and attention operations. Future releases will support quantized attention with hardware acceleration, which should provide additional performance benefits. While the most recent silicon offerings (e.g. AMD MI300, NVIDIA Hopper or later) support native hardware conversion between FP8 and other formats (fp32, fp16, bf16), this benefit is not yet fully realized. + +Studies have shown that FP8 E4M3 quantization typically only minimally degrades inference accuracy, making it a practical choice for throughput optimization. + +## Usage Example + +Here is an example of how to enable FP8 quantization: + +```python +# To calculate kv cache scales on the fly enable the calculate_kv_scales +# parameter + +from vllm import LLM, SamplingParams + +sampling_params = SamplingParams(temperature=0.7, top_p=0.8) +llm = LLM(model="meta-llama/Llama-2-7b-chat-hf", + kv_cache_dtype="fp8", + calculate_kv_scales=True) +prompt = "London is the capital of" +out = llm.generate(prompt, sampling_params)[0].outputs[0].text +print(out) +``` + +The `kv_cache_dtype` argument specifies the data type for KV cache storage: +- `"auto"`: Uses the model's default "unquantized" data type +- `"fp8"` or `"fp8_e4m3"`: Supported on CUDA 11.8+ and ROCm (AMD GPU) +- `"fp8_e5m2"`: Supported on CUDA 11.8+ + +## Calibrated Scales for Better Accuracy + +For optimal model quality when using FP8 KV Cache, we recommend using calibrated scales tuned to representative inference data. [LLM Compressor](https://github.com/vllm-project/llm-compressor/) is the recommended tool for this process. + +### Installation + +First, install the required dependencies: + +```console +pip install llmcompressor +``` + +### Example Usage + +Here's a complete example using `meta-llama/Llama-3.1-8B-Instruct` (most models can use this same pattern): + +```python +from datasets import load_dataset +from transformers import AutoModelForCausalLM, AutoTokenizer +from llmcompressor.transformers import oneshot + +# Select model and load it +MODEL_ID = "meta-llama/Llama-3.1-8B-Instruct" +model = AutoModelForCausalLM.from_pretrained(MODEL_ID, device_map="auto", torch_dtype="auto") +tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) + +# Select calibration dataset +DATASET_ID = "HuggingFaceH4/ultrachat_200k" +DATASET_SPLIT = "train_sft" + +# Configure calibration parameters +NUM_CALIBRATION_SAMPLES = 512 # 512 samples is a good starting point +MAX_SEQUENCE_LENGTH = 2048 + +# Load and preprocess dataset +ds = load_dataset(DATASET_ID, split=DATASET_SPLIT) +ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES)) + +def process_and_tokenize(example): + text = tokenizer.apply_chat_template(example["messages"], tokenize=False) + return tokenizer( + text, + padding=False, + max_length=MAX_SEQUENCE_LENGTH, + truncation=True, + add_special_tokens=False, + ) + +ds = ds.map(process_and_tokenize, remove_columns=ds.column_names) + +# Configure quantization settings +recipe = """ +quant_stage: + quant_modifiers: + QuantizationModifier: + kv_cache_scheme: + num_bits: 8 + type: float + strategy: tensor + dynamic: false + symmetric: true +""" + +# Apply quantization +oneshot( + model=model, + dataset=ds, + recipe=recipe, + max_seq_length=MAX_SEQUENCE_LENGTH, + num_calibration_samples=NUM_CALIBRATION_SAMPLES, +) + +# Save quantized model +SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-KV" +model.save_pretrained(SAVE_DIR, save_compressed=True) +tokenizer.save_pretrained(SAVE_DIR) +``` + +The above script will create a folder in your current directory containing your quantized model (e.g., `Llama-3.1-8B-Instruct-FP8-KV`) with calibrated scales. + +When running the model you must specify `kv_cache_dtype="fp8"` in order to enable the kv cache quantization and use the scales. + +```python +from vllm import LLM, SamplingParams + +sampling_params = SamplingParams(temperature=0.7, top_p=0.8) +llm = LLM(model="Llama-3.1-8B-Instruct-FP8-KV", kv_cache_dtype="fp8") +prompt = "London is the capital of" +out = llm.generate(prompt, sampling_params)[0].outputs[0].text +print(out) +``` diff --git a/docs/source/features/quantization/supported_hardware.md b/docs/source/features/quantization/supported_hardware.md index 988288a82d9bc..f5c0a95ea426e 100644 --- a/docs/source/features/quantization/supported_hardware.md +++ b/docs/source/features/quantization/supported_hardware.md @@ -113,7 +113,7 @@ The table below shows the compatibility of various quantization implementations - ✅︎ - ✅︎ - ✅︎ - - ✗ + - ✅︎ - ✗ - ✗ - ✗ diff --git a/docs/source/getting_started/installation/ai_accelerator/hpu-gaudi.inc.md b/docs/source/getting_started/installation/ai_accelerator/hpu-gaudi.inc.md index b4695d504b601..ae42dd0c0d08f 100644 --- a/docs/source/getting_started/installation/ai_accelerator/hpu-gaudi.inc.md +++ b/docs/source/getting_started/installation/ai_accelerator/hpu-gaudi.inc.md @@ -59,6 +59,7 @@ To build and install vLLM from source, run: ```console git clone https://github.com/vllm-project/vllm.git cd vllm +pip install -r requirements-hpu.txt python setup.py develop ``` @@ -68,6 +69,7 @@ Currently, the latest features and performance optimizations are developed in Ga git clone https://github.com/HabanaAI/vllm-fork.git cd vllm-fork git checkout habana_main +pip install -r requirements-hpu.txt python setup.py develop ``` diff --git a/docs/source/getting_started/installation/gpu/rocm.inc.md b/docs/source/getting_started/installation/gpu/rocm.inc.md index f6f9d3c303f89..3e3e01ffe3ef8 100644 --- a/docs/source/getting_started/installation/gpu/rocm.inc.md +++ b/docs/source/getting_started/installation/gpu/rocm.inc.md @@ -13,6 +13,14 @@ vLLM supports AMD GPUs with ROCm 6.2. Currently, there are no pre-built ROCm wheels. +However, the [AMD Infinity hub for vLLM](https://hub.docker.com/r/rocm/vllm/tags) offers a prebuilt, optimized +docker image designed for validating inference performance on the AMD Instinct™ MI300X accelerator. + +```{tip} +Please check [LLM inference performance validation on AMD Instinct MI300X](https://rocm.docs.amd.com/en/latest/how-to/performance-validation/mi300x/vllm-benchmark.html) +for instructions on how to use this prebuilt docker image. +``` + ### Build wheel from source 0. Install prerequisites (skip if you are already in an environment/docker with the following installed): @@ -70,7 +78,7 @@ Currently, there are no pre-built ROCm wheels. # Install PyTorch $ pip uninstall torch -y - $ pip install --no-cache-dir --pre torch==2.6.0.dev20241024 --index-url https://download.pytorch.org/whl/nightly/rocm6.2 + $ pip install --no-cache-dir --pre torch --index-url https://download.pytorch.org/whl/rocm6.2 # Build & install AMD SMI $ pip install /opt/rocm/share/amd_smi @@ -123,11 +131,14 @@ It is important that the user kicks off the docker build using buildkit. Either uses ROCm 6.2 by default, but also supports ROCm 5.7, 6.0 and 6.1 in older vLLM branches. It provides flexibility to customize the build of docker image using the following arguments: -- `BASE_IMAGE`: specifies the base image used when running `docker build`, specifically the PyTorch on ROCm base image. +- `BASE_IMAGE`: specifies the base image used when running `docker build`. The default value `rocm/vllm-dev:base` is an image published and maintained by AMD. It is being built using . - `BUILD_FA`: specifies whether to build CK flash-attention. The default is 1. For [Radeon RX 7900 series (gfx1100)](https://rocm.docs.amd.com/projects/radeon/en/latest/index.html), this should be set to 0 before flash-attention supports this target. - `FX_GFX_ARCHS`: specifies the GFX architecture that is used to build CK flash-attention, for example, `gfx90a;gfx942` for MI200 and MI300. The default is `gfx90a;gfx942` - `FA_BRANCH`: specifies the branch used to build the CK flash-attention in [ROCm's flash-attention repo](https://github.com/ROCmSoftwarePlatform/flash-attention). The default is `ae7928c` - `BUILD_TRITON`: specifies whether to build triton flash-attention. The default value is 1. +- `USE_CYTHON`: An option to run cython compilation on a subset of python files upon docker build +- `BUILD_RPD`: Include RocmProfileData profiling tool in the image +- `ARG_PYTORCH_ROCM_ARCH`: Allows to override the gfx architecture values from the base docker image Their values can be passed in when running `docker build` with `--build-arg` options. @@ -137,10 +148,10 @@ To build vllm on ROCm 6.2 for MI200 and MI300 series, you can use the default: DOCKER_BUILDKIT=1 docker build -f Dockerfile.rocm -t vllm-rocm . ``` -To build vllm on ROCm 6.2 for Radeon RX7900 series (gfx1100), you should specify `BUILD_FA` as below: +To build vllm on ROCm 6.2 for Radeon RX7900 series (gfx1100), you should pick the alternative base image and specify `BUILD_FA` as below: ```console -DOCKER_BUILDKIT=1 docker build --build-arg BUILD_FA="0" -f Dockerfile.rocm -t vllm-rocm . +DOCKER_BUILDKIT=1 docker build --build-arg BASE_IMAGE="rocm/vllm-dev:navi_base" BUILD_FA="0" -f Dockerfile.rocm -t vllm-rocm. ``` To run the above docker image `vllm-rocm`, use the below command: diff --git a/docs/source/getting_started/troubleshooting.md b/docs/source/getting_started/troubleshooting.md index 1e290d2b4c0bd..7bfe9b4036adf 100644 --- a/docs/source/getting_started/troubleshooting.md +++ b/docs/source/getting_started/troubleshooting.md @@ -22,9 +22,9 @@ It'd be better to store the model in a local disk. Additionally, have a look at To isolate the model downloading and loading issue, you can use the `--load-format dummy` argument to skip loading the model weights. This way, you can check if the model downloading and loading is the bottleneck. ``` -## Model is too large +## Out of memory -If the model is too large to fit in a single GPU, you might want to [consider tensor parallelism](#distributed-serving) to split the model across multiple GPUs. In that case, every process will read the whole model and split it into chunks, which makes the disk reading time even longer (proportional to the size of tensor parallelism). You can convert the model checkpoint to a sharded checkpoint using . The conversion process might take some time, but later you can load the sharded checkpoint much faster. The model loading time should remain constant regardless of the size of tensor parallelism. +If the model is too large to fit in a single GPU, you will get an out-of-memory (OOM) error. Consider [using tensor parallelism](#distributed-serving) to split the model across multiple GPUs. In that case, every process will read the whole model and split it into chunks, which makes the disk reading time even longer (proportional to the size of tensor parallelism). You can convert the model checkpoint to a sharded checkpoint using . The conversion process might take some time, but later you can load the sharded checkpoint much faster. The model loading time should remain constant regardless of the size of tensor parallelism. ## Enable more logging @@ -197,6 +197,63 @@ if __name__ == '__main__': llm = vllm.LLM(...) ``` +## `torch.compile` Error + +vLLM heavily depends on `torch.compile` to optimize the model for better performance, which introduces the dependency on the `torch.compile` functionality and the `triton` library. By default, we use `torch.compile` to [optimize some functions](https://github.com/vllm-project/vllm/pull/10406) in the model. Before running vLLM, you can check if `torch.compile` is working as expected by running the following script: + +```python +import torch + +@torch.compile +def f(x): + # a simple function to test torch.compile + x = x + 1 + x = x * 2 + x = x.sin() + return x + +x = torch.randn(4, 4).cuda() +print(f(x)) +``` + +If it raises errors from `torch/_inductor` directory, usually it means you have a custom `triton` library that is not compatible with the version of PyTorch you are using. See [this issue](https://github.com/vllm-project/vllm/issues/12219) for example. + +## Model failed to be inspected + +If you see an error like: + +```text + File "vllm/model_executor/models/registry.py", line xxx, in _raise_for_unsupported + raise ValueError( +ValueError: Model architectures [''] failed to be inspected. Please check the logs for more details. +``` + +It means that vLLM failed to import the model file. +Usually, it is related to missing dependencies or outdated binaries in the vLLM build. +Please read the logs carefully to determine the root cause of the error. + +## Model not supported + +If you see an error like: + +```text +Traceback (most recent call last): +... + File "vllm/model_executor/models/registry.py", line xxx, in inspect_model_cls + for arch in architectures: +TypeError: 'NoneType' object is not iterable +``` + +or: + +```text + File "vllm/model_executor/models/registry.py", line xxx, in _raise_for_unsupported + raise ValueError( +ValueError: Model architectures [''] are not supported for now. Supported architectures: [...] +``` + +But you are sure that the model is in the [list of supported models](#supported-models), there may be some issue with vLLM's model resolution. In that case, please follow [these steps](#model-resolution) to explicitly specify the vLLM implementation for the model. + ## Known Issues - In `v0.5.2`, `v0.5.3`, and `v0.5.3.post1`, there is a bug caused by [zmq](https://github.com/zeromq/pyzmq/issues/2000) , which can occasionally cause vLLM to hang depending on the machine configuration. The solution is to upgrade to the latest version of `vllm` to include the [fix](gh-pr:6759). diff --git a/docs/source/index.md b/docs/source/index.md index 8f9493d77186e..2c302d3f3e863 100644 --- a/docs/source/index.md +++ b/docs/source/index.md @@ -23,6 +23,8 @@ vLLM is a fast and easy-to-use library for LLM inference and serving. +Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evloved into a community-driven project with contributions from both academia and industry. + vLLM is fast with: - State-of-the-art serving throughput @@ -182,6 +184,7 @@ api/model/index :caption: Community :maxdepth: 1 +community/blog community/meetups community/sponsors ``` diff --git a/docs/source/models/supported_models.md b/docs/source/models/supported_models.md index 642ef3c9655b8..8cdc663a0320f 100644 --- a/docs/source/models/supported_models.md +++ b/docs/source/models/supported_models.md @@ -216,6 +216,11 @@ See [this page](#generative-models) for more information on how to use generativ - `internlm/internlm2-7b`, `internlm/internlm2-chat-7b`, etc. - ✅︎ - ✅︎ +* - `InternLM3ForCausalLM` + - InternLM3 + - `internlm/internlm3-8b-instruct`, etc. + - ✅︎ + - ✅︎ * - `JAISLMHeadModel` - Jais - `inceptionai/jais-13b`, `inceptionai/jais-13b-chat`, `inceptionai/jais-30b-v3`, `inceptionai/jais-30b-chat-v3`, etc. @@ -297,8 +302,8 @@ See [this page](#generative-models) for more information on how to use generativ - ✅︎ - ✅︎ * - `Phi3ForCausalLM` - - Phi-3 - - `microsoft/Phi-3-mini-4k-instruct`, `microsoft/Phi-3-mini-128k-instruct`, `microsoft/Phi-3-medium-128k-instruct`, etc. + - Phi-4, Phi-3 + - `microsoft/Phi-4`, `microsoft/Phi-3-mini-4k-instruct`, `microsoft/Phi-3-mini-128k-instruct`, `microsoft/Phi-3-medium-128k-instruct`, etc. - ✅︎ - ✅︎ * - `Phi3SmallForCausalLM` @@ -465,6 +470,11 @@ of the whole prompt are extracted from the normalized hidden state corresponding - `Qwen/Qwen2.5-Math-RM-72B`, etc. - ✅︎ - ✅︎ +* - `Qwen2ForProcessRewardModel` + - Qwen2-based + - `Qwen/Qwen2.5-Math-PRM-7B`, `Qwen/Qwen2.5-Math-PRM-72B`, etc. + - ✅︎ + - ✅︎ ``` If your model is not in the above list, we will try to automatically convert the model using @@ -613,7 +623,7 @@ See [this page](#generative-models) for more information on how to use generativ * - `DeepseekVLV2ForCausalLM` - DeepSeek-VL2 - T + I+ - - `deepseek-ai/deepseek-vl2-tiny`(WIP), `deepseek-ai/deepseek-vl2-small`, `deepseek-ai/deepseek-vl2` etc. (see note) + - `deepseek-ai/deepseek-vl2-tiny`, `deepseek-ai/deepseek-vl2-small`, `deepseek-ai/deepseek-vl2` etc. (see note) - - ✅︎ - ✅︎ @@ -749,7 +759,7 @@ See [this page](#generative-models) for more information on how to use generativ - `Qwen/QVQ-72B-Preview`, `Qwen/Qwen2-VL-7B-Instruct`, `Qwen/Qwen2-VL-72B-Instruct`, etc. - ✅︎ - ✅︎ - - + - ✅︎ * - `UltravoxModel` - Ultravox - T + AE+ @@ -762,17 +772,10 @@ See [this page](#generative-models) for more information on how to use generativ E Pre-computed embeddings can be inputted for this modality. + Multiple items can be inputted per text prompt for this modality. -````{note} -The `deepseek-ai/deepseek-vl2-tiny` is not supported yet. - -To use `DeepSeek-VL2` series models, you need to install a fork version `deepseek_vl2` package: -```shell -pip install git+https://github.com/Isotr0py/DeepSeek-VL2.git +```{note} +To use `DeepSeek-VL2` series models, you have to pass `--hf_overrides '{"architectures": ["DeepseekVLV2ForCausalLM"]}'` when running vLLM. ``` -Besides, to run `DeepSeek-VL2` series models, you have to pass `--hf_overrides '{"architectures": ["DeepseekVLV2ForCausalLM"]}'` when running vLLM. -```` - ```{note} To use `TIGER-Lab/Mantis-8B-siglip-llama3`, you have to pass `--hf_overrides '{"architectures": ["MantisForConditionalGeneration"]}'` when running vLLM. ``` diff --git a/docs/source/serving/offline_inference.md b/docs/source/serving/offline_inference.md index 94703a1c32ade..8a18598665a70 100644 --- a/docs/source/serving/offline_inference.md +++ b/docs/source/serving/offline_inference.md @@ -31,6 +31,30 @@ Please refer to the above pages for more details about each API. This section lists the most common options for running the vLLM engine. For a full list, refer to the [Engine Arguments](#engine-args) page. +(model-resolution)= + +### Model resolution + +vLLM loads HuggingFace-compatible models by inspecting the `architectures` field in `config.json` of the model repository +and finding the corresponding implementation that is registered to vLLM. +Nevertheless, our model resolution may fail for the following reasons: + +- The `config.json` of the model repository lacks the `architectures` field. +- Unofficial repositories refer to a model using alternative names which are not recorded in vLLM. +- The same architecture name is used for multiple models, creating ambiguity as to which model should be loaded. + +To fix this, explicitly specify the model architecture by passing `config.json` overrides to the `hf_overrides` option. +For example: + +```python +model = LLM( + model="cerebras/Cerebras-GPT-1.3B", + hf_overrides={"architectures": ["GPT2LMHeadModel"]}, # GPT-2 +) +``` + +Our [list of supported models](#supported-models) shows the model architectures that are recognized by vLLM. + ### Reducing memory usage Large models might cause your machine to run out of memory (OOM). Here are some options that help alleviate this problem. diff --git a/docs/source/serving/openai_compatible_server.md b/docs/source/serving/openai_compatible_server.md index e49bbb06695f8..8bc234545befd 100644 --- a/docs/source/serving/openai_compatible_server.md +++ b/docs/source/serving/openai_compatible_server.md @@ -50,6 +50,11 @@ In addition, we have the following custom APIs: - Applicable to all [pooling models](../models/pooling_models.md). - [Score API](#score-api) (`/score`) - Only applicable to [cross-encoder models](../models/pooling_models.md) (`--task score`). +- [Re-rank API](#rerank-api) (`/rerank`, `/v1/rerank`, `/v2/rerank`) + - Implements [Jina AI's v1 re-rank API](https://jina.ai/reranker/) + - Also compatible with [Cohere's v1 & v2 re-rank APIs](https://docs.cohere.com/v2/reference/rerank) + - Jina and Cohere's APIs are very similar; Jina's includes extra information in the rerank endpoint's response. + - Only applicable to [cross-encoder models](../models/pooling_models.md) (`--task score`). (chat-template)= @@ -473,3 +478,90 @@ The following extra parameters are supported: :start-after: begin-score-extra-params :end-before: end-score-extra-params ``` + +(rerank-api)= + +### Re-rank API + +Our Re-rank API applies a cross-encoder model to predict relevant scores between a single query, and +each of a list of documents. Usually, the score for a sentence pair refers to the similarity between two sentences, on +a scale of 0 to 1. + +You can find the documentation for these kind of models at [sbert.net](https://www.sbert.net/docs/package_reference/cross_encoder/cross_encoder.html). + +The rerank endpoints support popular re-rank models such as `BAAI/bge-reranker-base` and other models supporting the +`score` task. Additionally, `/rerank`, `/v1/rerank`, and `/v2/rerank` +endpoints are compatible with both [Jina AI's re-rank API interface](https://jina.ai/reranker/) and +[Cohere's re-rank API interface](https://docs.cohere.com/v2/reference/rerank) to ensure compatibility with +popular open-source tools. + +Code example: + +#### Example Request + +Note that the `top_n` request parameter is optional and will default to the length of the `documents` field. +Result documents will be sorted by relevance, and the `index` property can be used to determine original order. + +Request: + +```bash +curl -X 'POST' \ + 'http://127.0.0.1:8000/v1/rerank' \ + -H 'accept: application/json' \ + -H 'Content-Type: application/json' \ + -d '{ + "model": "BAAI/bge-reranker-base", + "query": "What is the capital of France?", + "documents": [ + "The capital of Brazil is Brasilia.", + "The capital of France is Paris.", + "Horses and cows are both animals" + ] +}' +``` + +Response: + +```bash +{ + "id": "rerank-fae51b2b664d4ed38f5969b612edff77", + "model": "BAAI/bge-reranker-base", + "usage": { + "total_tokens": 56 + }, + "results": [ + { + "index": 1, + "document": { + "text": "The capital of France is Paris." + }, + "relevance_score": 0.99853515625 + }, + { + "index": 0, + "document": { + "text": "The capital of Brazil is Brasilia." + }, + "relevance_score": 0.0005860328674316406 + } + ] +} +``` + +#### Extra parameters + +The following [pooling parameters](#pooling-params) are supported. + +```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py +:language: python +:start-after: begin-rerank-pooling-params +:end-before: end-rerank-pooling-params +``` + +The following extra parameters are supported: + +```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py +:language: python +:start-after: begin-rerank-extra-params +:end-before: end-rerank-extra-params +``` diff --git a/examples/offline_inference/openai/openai_batch.md b/examples/offline_inference/openai/openai_batch.md index a4774e57cd9a5..953e6ef130f18 100644 --- a/examples/offline_inference/openai/openai_batch.md +++ b/examples/offline_inference/openai/openai_batch.md @@ -13,7 +13,7 @@ The OpenAI batch file format consists of a series of json objects on new lines. Each line represents a separate request. See the [OpenAI package reference](https://platform.openai.com/docs/api-reference/batch/requestInput) for more details. ```{note} -We currently only support `/v1/chat/completions` and `/v1/embeddings` endpoints (completions coming soon). +We currently support `/v1/chat/completions`, `/v1/embeddings`, and `/v1/score` endpoints (completions coming soon). ``` ## Pre-requisites @@ -203,3 +203,34 @@ $ cat results.jsonl {"id":"vllm-db0f71f7dec244e6bce530e0b4ef908b","custom_id":"request-1","response":{"status_code":200,"request_id":"vllm-batch-3580bf4d4ae54d52b67eee266a6eab20","body":{"id":"embd-33ac2efa7996430184461f2e38529746","object":"list","created":444647,"model":"intfloat/e5-mistral-7b-instruct","data":[{"index":0,"object":"embedding","embedding":[0.016204833984375,0.0092010498046875,0.0018358230590820312,-0.0028228759765625,0.001422882080078125,-0.0031147003173828125,...]}],"usage":{"prompt_tokens":8,"total_tokens":8,"completion_tokens":0}}},"error":null} ... ``` + +## Example 5: Using score endpoint + +### Additional prerequisites + +* Ensure you are using `vllm >= 0.7.0`. + +### Step 1: Create your batch file + +Add score requests to your batch file. The following is an example: + +``` +{"custom_id": "request-1", "method": "POST", "url": "/v1/score", "body": {"model": "BAAI/bge-reranker-v2-m3", "text_1": "What is the capital of France?", "text_2": ["The capital of Brazil is Brasilia.", "The capital of France is Paris."]}} +{"custom_id": "request-2", "method": "POST", "url": "/v1/score", "body": {"model": "BAAI/bge-reranker-v2-m3", "text_1": "What is the capital of France?", "text_2": ["The capital of Brazil is Brasilia.", "The capital of France is Paris."]}} +``` + +You can mix chat completion, embedding, and score requests in the batch file, as long as the model you are using supports them all (note that all requests must use the same model). + +### Step 2: Run the batch + +You can run the batch using the same command as in earlier examples. + +### Step 3: Check your results + +You can check your results by running `cat results.jsonl` + +``` +$ cat results.jsonl +{"id":"vllm-f87c5c4539184f618e555744a2965987","custom_id":"request-1","response":{"status_code":200,"request_id":"vllm-batch-806ab64512e44071b37d3f7ccd291413","body":{"id":"score-4ee45236897b4d29907d49b01298cdb1","object":"list","created":1737847944,"model":"BAAI/bge-reranker-v2-m3","data":[{"index":0,"object":"score","score":0.0010900497436523438},{"index":1,"object":"score","score":1.0}],"usage":{"prompt_tokens":37,"total_tokens":37,"completion_tokens":0,"prompt_tokens_details":null}}},"error":null} +{"id":"vllm-41990c51a26d4fac8419077f12871099","custom_id":"request-2","response":{"status_code":200,"request_id":"vllm-batch-73ce66379026482699f81974e14e1e99","body":{"id":"score-13f2ffe6ba40460fbf9f7f00ad667d75","object":"list","created":1737847944,"model":"BAAI/bge-reranker-v2-m3","data":[{"index":0,"object":"score","score":0.001094818115234375},{"index":1,"object":"score","score":1.0}],"usage":{"prompt_tokens":37,"total_tokens":37,"completion_tokens":0,"prompt_tokens_details":null}}},"error":null} +``` diff --git a/examples/offline_inference/rlhf.py b/examples/offline_inference/rlhf.py new file mode 100644 index 0000000000000..5c4918008dcb3 --- /dev/null +++ b/examples/offline_inference/rlhf.py @@ -0,0 +1,186 @@ +""" +a simple demonstration of RLHF with vLLM, inspired by +the OpenRLHF framework https://github.com/OpenRLHF/OpenRLHF . +It follows the design that, training processes and inference processes +are different, and they live on different GPUs. +Training processes send prompts to inference processes to generate data, +and also synchronize the weights of the model by broadcasting the weights +from the training process to the inference process. +Note that this is a simple demonstration of one training instance and one +inference instance. In practice, there could be multiple training instances +and multiple inference instances. For the full implementation, please refer +to the OpenRLHF framework. +""" +import os + +import ray +import torch +from ray.util.placement_group import placement_group +from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy +from transformers import AutoModelForCausalLM + +from vllm import LLM, SamplingParams +from vllm.utils import get_ip, get_open_port +from vllm.worker.worker import Worker + + +def stateless_init_process_group(master_address, master_port, rank, world_size, + device): + """ + vLLM provides `StatelessProcessGroup` to create a process group + without considering the global process group in torch.distributed. + It is recommended to create `StatelessProcessGroup`, and then initialize + the data-plane communication (NCCL) between external (train processes) + and vLLM workers. + """ + from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator + from vllm.distributed.utils import StatelessProcessGroup + pg = StatelessProcessGroup.create(host=master_address, + port=master_port, + rank=rank, + world_size=world_size) + pynccl = PyNcclCommunicator(pg, device=device) + return pynccl + + +class MyWorker(Worker): + """ + The `MyWorker` class inherits from `Worker` to provide custom functions. + For simplicity, we define the `MyWorker` class in this self-contained + script. Normally, we should define the `MyWorker` class in a separate + file and pass the qualified name of the class to the `worker_cls` + parameter. + """ + + def init_weight_update_group(self, master_address, master_port, + rank_offset, world_size): + from vllm.distributed.parallel_state import get_world_group + rank = get_world_group().rank + rank_offset + self.model_update_group = stateless_init_process_group( + master_address, + master_port, + rank, + world_size, + self.device, + ) + + def update_weight(self, name, dtype, shape): + weight = torch.empty(shape, dtype=dtype, device="cuda") + self.model_update_group.broadcast(weight, + src=0, + stream=torch.cuda.current_stream()) + + self.model_runner.model.load_weights(weights=[(name, weight)]) + + del weight + + def check_weights_changed(self): + """ + Check if the weights are updated to 0. + """ + weights_updated = True + for name, p in self.model_runner.model.named_parameters(): + weights_updated = weights_updated and torch.allclose( + p, torch.zeros_like(p)) + return weights_updated + + +class MyLLM(LLM): + + def __init__(self, *args, **kwargs): + # a hack to make the script work. + # stop ray from manipulating CUDA_VISIBLE_DEVICES + # at the top-level + del os.environ["CUDA_VISIBLE_DEVICES"] + super().__init__(*args, **kwargs) + + +""" +Start the training process, here we use huggingface transformers +as an example to hold a model on GPU 0. +""" + +train_model = AutoModelForCausalLM.from_pretrained("facebook/opt-125m") +train_model.to("cuda:0") +""" +Start the inference process, here we use vLLM to hold a model on GPU 1 and +GPU 2. For the details on how to use ray, please refer to the ray +documentation https://docs.ray.io/en/latest/ . +""" +os.environ["CUDA_VISIBLE_DEVICES"] = "1,2" +ray.init() + +pg_inference = placement_group([{"GPU": 1, "CPU": 0}] * 2) +ray.get(pg_inference.ready()) +scheduling_inference = PlacementGroupSchedulingStrategy( + placement_group=pg_inference, + placement_group_capture_child_tasks=True, + placement_group_bundle_index=0, +) +""" +launch the vLLM inference engine. +here we use `enforce_eager` to reduce the start time. +""" +llm = ray.remote( + num_cpus=0, + num_gpus=0, + scheduling_strategy=scheduling_inference, +)(MyLLM).remote( + model="facebook/opt-125m", + enforce_eager=True, + worker_cls=MyWorker, + tensor_parallel_size=2, + distributed_executor_backend="ray", +) + +# Generate texts from the prompts. +prompts = [ + "Hello, my name is", + "The president of the United States is", + "The capital of France is", + "The future of AI is", +] + +sampling_params = SamplingParams(temperature=0) + +outputs = ray.get(llm.generate.remote(prompts, sampling_params)) + +for output in outputs: + prompt = output.prompt + generated_text = output.outputs[0].text + print(f"Prompt: {prompt!r}, " + f"Generated text: {generated_text!r}") + +# set up the communication between the training process +# and the inference engine. +master_address = get_ip() +master_port = get_open_port() + +handle = llm.collective_rpc.remote("init_weight_update_group", + args=(master_address, master_port, 1, 3)) +model_update_group = stateless_init_process_group(master_address, master_port, + 0, 3, torch.device("cuda:0")) +ray.get(handle) + +# simulate training, modify the weights of the model. +for name, p in train_model.named_parameters(): + p.data.zero_() + +# sync weight from the training process to the inference engine. +for name, p in train_model.named_parameters(): + handle = llm.collective_rpc.remote("update_weight", + args=(name, p.dtype, p.shape)) + model_update_group.broadcast(p, src=0, stream=torch.cuda.current_stream()) + ray.get(handle) + +# check if the weights are updated. +assert all(ray.get(llm.collective_rpc.remote("check_weights_changed"))) + +# use the updated model to generate texts, they will be nonsense +# because the weights are all zeros. +outputs_updated = ray.get(llm.generate.remote(prompts, sampling_params)) +for output in outputs_updated: + prompt = output.prompt + generated_text = output.outputs[0].text + print(f"Prompt: {prompt!r}, " + f"Generated text: {generated_text!r}") diff --git a/examples/offline_inference/torchrun_example.py b/examples/offline_inference/torchrun_example.py new file mode 100644 index 0000000000000..b6de73eb7266e --- /dev/null +++ b/examples/offline_inference/torchrun_example.py @@ -0,0 +1,64 @@ +""" +experimental support for tensor-parallel inference with torchrun, +see https://github.com/vllm-project/vllm/issues/11400 for +the motivation and use case for this example. +run the script with `torchrun --nproc-per-node=2 torchrun_example.py`, +the argument 2 should match the `tensor_parallel_size` below. +see `tests/distributed/test_torchrun_example.py` for the unit test. +""" + +from vllm import LLM, SamplingParams + +# Create prompts, the same across all ranks +prompts = [ + "Hello, my name is", + "The president of the United States is", + "The capital of France is", + "The future of AI is", +] + +# Create sampling parameters, the same across all ranks +sampling_params = SamplingParams(temperature=0.8, top_p=0.95) + +# Use `distributed_executor_backend="external_launcher"` so that +# this llm engine/instance only creates one worker. +llm = LLM( + model="facebook/opt-125m", + tensor_parallel_size=2, + distributed_executor_backend="external_launcher", +) + +outputs = llm.generate(prompts, sampling_params) + +# all ranks will have the same outputs +for output in outputs: + prompt = output.prompt + generated_text = output.outputs[0].text + print(f"Prompt: {prompt!r}, " + f"Generated text: {generated_text!r}") +""" +Further tips: + +1. to communicate control messages across all ranks, use the cpu group, +a PyTorch ProcessGroup with GLOO backend. + +```python +from vllm.distributed.parallel_state import get_world_group +cpu_group = get_world_group().cpu_group +torch_rank = dist.get_rank(group=cpu_group) +if torch_rank == 0: + # do something for rank 0, e.g. saving the results to disk. +``` + +2. to communicate data across all ranks, use the model's device group, +a PyTorch ProcessGroup with NCCL backend. +```python +from vllm.distributed.parallel_state import get_world_group +device_group = get_world_group().device_group +``` + +3. to access the model directly in every rank, use the following code: +```python +llm.llm_engine.model_executor.driver_worker.worker.model_runner.model +``` +""" diff --git a/examples/offline_inference/vision_language.py b/examples/offline_inference/vision_language.py index ad32b9fe242e9..8bc715a50e0db 100644 --- a/examples/offline_inference/vision_language.py +++ b/examples/offline_inference/vision_language.py @@ -70,7 +70,7 @@ def run_chameleon(question: str, modality: str): def run_deepseek_vl2(question: str, modality: str): assert modality == "image" - model_name = "deepseek-ai/deepseek-vl2-small" + model_name = "deepseek-ai/deepseek-vl2-tiny" llm = LLM(model=model_name, max_model_len=4096, diff --git a/examples/offline_inference/vision_language_multi_image.py b/examples/offline_inference/vision_language_multi_image.py index c6cf3f30c31cb..fd289c22d4148 100644 --- a/examples/offline_inference/vision_language_multi_image.py +++ b/examples/offline_inference/vision_language_multi_image.py @@ -55,7 +55,7 @@ def load_aria(question, image_urls: List[str]) -> ModelRequestData: def load_deepseek_vl2(question: str, image_urls: List[str]): - model_name = "deepseek-ai/deepseek-vl2-small" + model_name = "deepseek-ai/deepseek-vl2-tiny" llm = LLM(model=model_name, max_model_len=4096, @@ -394,7 +394,7 @@ def load_qwen2_vl(question, image_urls: List[str]) -> ModelRequestData: model_example_map = { "aria": load_aria, - "deepseek_vl2": load_deepseek_vl2, + "deepseek_vl_v2": load_deepseek_vl2, "h2ovl_chat": load_h2onvl, "idefics3": load_idefics3, "internvl_chat": load_internvl, diff --git a/examples/online_serving/cohere_rerank_client.py b/examples/online_serving/cohere_rerank_client.py new file mode 100644 index 0000000000000..a07affe3351ce --- /dev/null +++ b/examples/online_serving/cohere_rerank_client.py @@ -0,0 +1,32 @@ +""" +Example of using the OpenAI entrypoint's rerank API which is compatible with +the Cohere SDK: https://github.com/cohere-ai/cohere-python + +run: vllm serve BAAI/bge-reranker-base +""" +import cohere + +# cohere v1 client +co = cohere.Client(base_url="http://localhost:8000", api_key="sk-fake-key") +rerank_v1_result = co.rerank( + model="BAAI/bge-reranker-base", + query="What is the capital of France?", + documents=[ + "The capital of France is Paris", "Reranking is fun!", + "vLLM is an open-source framework for fast AI serving" + ]) + +print(rerank_v1_result) + +# or the v2 +co2 = cohere.ClientV2("sk-fake-key", base_url="http://localhost:8000") + +v2_rerank_result = co2.rerank( + model="BAAI/bge-reranker-base", + query="What is the capital of France?", + documents=[ + "The capital of France is Paris", "Reranking is fun!", + "vLLM is an open-source framework for fast AI serving" + ]) + +print(v2_rerank_result) diff --git a/examples/online_serving/disaggregated_prefill.sh b/examples/online_serving/disaggregated_prefill.sh index 87155273a81d1..2bb2824c6c86f 100644 --- a/examples/online_serving/disaggregated_prefill.sh +++ b/examples/online_serving/disaggregated_prefill.sh @@ -3,6 +3,8 @@ # We will launch 2 vllm instances (1 for prefill and 1 for decode), # and then transfer the KV cache between them. +set -xe + echo "🚧🚧 Warning: The usage of disaggregated prefill is experimental and subject to change 🚧🚧" sleep 1 @@ -69,7 +71,7 @@ wait_for_server 8200 # instance # NOTE: the usage of this API is subject to change --- in the future we will # introduce "vllm connect" to connect between prefill and decode instances -python3 ../benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py & +python3 ../../benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py & sleep 1 # serve two example requests diff --git a/examples/online_serving/jinaai_rerank_client.py b/examples/online_serving/jinaai_rerank_client.py new file mode 100644 index 0000000000000..bf4de76ddf362 --- /dev/null +++ b/examples/online_serving/jinaai_rerank_client.py @@ -0,0 +1,33 @@ +""" +Example of using the OpenAI entrypoint's rerank API which is compatible with +Jina and Cohere https://jina.ai/reranker + +run: vllm serve BAAI/bge-reranker-base +""" +import json + +import requests + +url = "http://127.0.0.1:8000/rerank" + +headers = {"accept": "application/json", "Content-Type": "application/json"} + +data = { + "model": + "BAAI/bge-reranker-base", + "query": + "What is the capital of France?", + "documents": [ + "The capital of Brazil is Brasilia.", + "The capital of France is Paris.", "Horses and cows are both animals" + ] +} +response = requests.post(url, headers=headers, json=data) + +# Check the response +if response.status_code == 200: + print("Request successful!") + print(json.dumps(response.json(), indent=2)) +else: + print(f"Request failed with status code: {response.status_code}") + print(response.text) diff --git a/examples/template_deepseek_vl2.jinja b/examples/template_deepseek_vl2.jinja new file mode 100644 index 0000000000000..fbf3d320094d5 --- /dev/null +++ b/examples/template_deepseek_vl2.jinja @@ -0,0 +1,23 @@ +{%- if messages[0]['role'] == 'system' -%} + {%- set system_message = messages[0]['content'] -%} + {%- set messages = messages[1:] -%} +{%- else -%} + {% set system_message = '' -%} +{%- endif -%} + +{{ bos_token + system_message }} +{%- for message in messages -%} + {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%} + {{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }} + {%- endif -%} + + {%- if message['role'] == 'user' -%} + {{ '<|User|>: ' + message['content'] + '\n' }} + {%- elif message['role'] == 'assistant' -%} + {{ '<|Assistant|>: ' + message['content'] + eos_token + '\n' }} + {%- endif -%} +{%- endfor -%} + +{%- if add_generation_prompt -%} + {{ '<|Assistant|>: ' }} +{% endif %} diff --git a/format.sh b/format.sh index 2277eef93c745..4bcd0be0c96e5 100755 --- a/format.sh +++ b/format.sh @@ -1,321 +1,5 @@ -#!/usr/bin/env bash -# YAPF formatter, adapted from ray and skypilot. -# -# Usage: -# # Do work and commit your work. +#!/bin/bash -# # Format files that differ from origin/main. -# bash format.sh - -# # Commit changed files with message 'Run yapf and ruff' -# -# -# YAPF + Clang formatter (if installed). This script formats all changed files from the last mergebase. -# You are encouraged to run this locally before pushing changes for review. - -# Cause the script to exit if a single command fails -set -eo pipefail - -# this stops git rev-parse from failing if we run this from the .git directory -builtin cd "$(dirname "${BASH_SOURCE:-$0}")" -ROOT="$(git rev-parse --show-toplevel)" -builtin cd "$ROOT" || exit 1 - -check_command() { - if ! command -v "$1" &> /dev/null; then - echo "❓❓$1 is not installed, please run \`pip install -r requirements-lint.txt\`" - exit 1 - fi -} - -check_command yapf -check_command ruff -check_command mypy -check_command codespell -check_command isort -check_command clang-format - -YAPF_VERSION=$(yapf --version | awk '{print $2}') -RUFF_VERSION=$(ruff --version | awk '{print $2}') -MYPY_VERSION=$(mypy --version | awk '{print $2}') -CODESPELL_VERSION=$(codespell --version) -ISORT_VERSION=$(isort --vn) -CLANGFORMAT_VERSION=$(clang-format --version | awk '{print $3}') -PYMARKDOWNLNT_VERSION=$(pymarkdownlnt version | awk '{print $1}') - -# # params: tool name, tool version, required version -tool_version_check() { - expected=$(grep "$1" requirements-lint.txt | cut -d'=' -f3) - if [[ "$2" != "$expected" ]]; then - echo "❓❓Wrong $1 version installed: $expected is required, not $2." - exit 1 - fi -} - -tool_version_check "yapf" "$YAPF_VERSION" -tool_version_check "ruff" "$RUFF_VERSION" -tool_version_check "mypy" "$MYPY_VERSION" -tool_version_check "isort" "$ISORT_VERSION" -tool_version_check "codespell" "$CODESPELL_VERSION" -tool_version_check "clang-format" "$CLANGFORMAT_VERSION" -tool_version_check "pymarkdownlnt" "$PYMARKDOWNLNT_VERSION" - -YAPF_FLAGS=( - '--recursive' - '--parallel' -) - -YAPF_EXCLUDES=( - '--exclude' 'build/**' -) - -# Format specified files -format() { - yapf --in-place "${YAPF_FLAGS[@]}" "$@" -} - -# Format files that differ from main branch. Ignores dirs that are not slated -# for autoformat yet. -format_changed() { - # The `if` guard ensures that the list of filenames is not empty, which - # could cause yapf to receive 0 positional arguments, making it hang - # waiting for STDIN. - # - # `diff-filter=ACM` and $MERGEBASE is to ensure we only format files that - # exist on both branches. - MERGEBASE="$(git merge-base origin/main HEAD)" - - if ! git diff --diff-filter=ACM --quiet --exit-code "$MERGEBASE" -- '*.py' '*.pyi' &>/dev/null; then - git diff --name-only --diff-filter=ACM "$MERGEBASE" -- '*.py' '*.pyi' | xargs -P 5 \ - yapf --in-place "${YAPF_EXCLUDES[@]}" "${YAPF_FLAGS[@]}" - fi - -} - -# Format all files -format_all() { - yapf --in-place "${YAPF_FLAGS[@]}" "${YAPF_EXCLUDES[@]}" . -} - -## This flag formats individual files. --files *must* be the first command line -## arg to use this option. -if [[ "$1" == '--files' ]]; then - format "${@:2}" - # If `--all` is passed, then any further arguments are ignored and the - # entire python directory is formatted. -elif [[ "$1" == '--all' ]]; then - format_all -else - # Format only the files that changed in last commit. - format_changed -fi -echo 'vLLM yapf: Done' - -# Run mypy -echo 'vLLM mypy:' -tools/mypy.sh -echo 'vLLM mypy: Done' - - -# If git diff returns a file that is in the skip list, the file may be checked anyway: -# https://github.com/codespell-project/codespell/issues/1915 -# Avoiding the "./" prefix and using "/**" globs for directories appears to solve the problem -CODESPELL_EXCLUDES=( - '--skip' 'tests/prompts/**,./benchmarks/sonnet.txt,*tests/lora/data/**,build/**' -) - -# check spelling of specified files -spell_check() { - codespell "$@" -} - -spell_check_all(){ - codespell --toml pyproject.toml "${CODESPELL_EXCLUDES[@]}" -} - -# Spelling check of files that differ from main branch. -spell_check_changed() { - # The `if` guard ensures that the list of filenames is not empty, which - # could cause ruff to receive 0 positional arguments, making it hang - # waiting for STDIN. - # - # `diff-filter=ACM` and $MERGEBASE is to ensure we only lint files that - # exist on both branches. - MERGEBASE="$(git merge-base origin/main HEAD)" - if ! git diff --diff-filter=ACM --quiet --exit-code "$MERGEBASE" -- '*.py' '*.pyi' &>/dev/null; then - git diff --name-only --diff-filter=ACM "$MERGEBASE" -- '*.py' '*.pyi' | xargs \ - codespell "${CODESPELL_EXCLUDES[@]}" - fi -} - -# Run Codespell -## This flag runs spell check of individual files. --files *must* be the first command line -## arg to use this option. -if [[ "$1" == '--files' ]]; then - spell_check "${@:2}" - # If `--all` is passed, then any further arguments are ignored and the - # entire python directory is linted. -elif [[ "$1" == '--all' ]]; then - spell_check_all -else - # Check spelling only of the files that changed in last commit. - spell_check_changed -fi -echo 'vLLM codespell: Done' - - -# Lint specified files -lint() { - ruff check "$@" -} - -# Lint files that differ from main branch. Ignores dirs that are not slated -# for autolint yet. -lint_changed() { - # The `if` guard ensures that the list of filenames is not empty, which - # could cause ruff to receive 0 positional arguments, making it hang - # waiting for STDIN. - # - # `diff-filter=ACM` and $MERGEBASE is to ensure we only lint files that - # exist on both branches. - MERGEBASE="$(git merge-base origin/main HEAD)" - - if ! git diff --diff-filter=ACM --quiet --exit-code "$MERGEBASE" -- '*.py' '*.pyi' &>/dev/null; then - git diff --name-only --diff-filter=ACM "$MERGEBASE" -- '*.py' '*.pyi' | xargs \ - ruff check - fi - -} - -# Run Ruff -### This flag lints individual files. --files *must* be the first command line -### arg to use this option. -if [[ "$1" == '--files' ]]; then - lint "${@:2}" - # If `--all` is passed, then any further arguments are ignored and the - # entire python directory is linted. -elif [[ "$1" == '--all' ]]; then - lint vllm tests -else - # Format only the files that changed in last commit. - lint_changed -fi -echo 'vLLM ruff: Done' - -# check spelling of specified files -isort_check() { - isort "$@" -} - -isort_check_all(){ - isort . -} - -# Spelling check of files that differ from main branch. -isort_check_changed() { - # The `if` guard ensures that the list of filenames is not empty, which - # could cause ruff to receive 0 positional arguments, making it hang - # waiting for STDIN. - # - # `diff-filter=ACM` and $MERGEBASE is to ensure we only lint files that - # exist on both branches. - MERGEBASE="$(git merge-base origin/main HEAD)" - - if ! git diff --diff-filter=ACM --quiet --exit-code "$MERGEBASE" -- '*.py' '*.pyi' &>/dev/null; then - git diff --name-only --diff-filter=ACM "$MERGEBASE" -- '*.py' '*.pyi' | xargs \ - isort - fi -} - -# Run Isort -# This flag runs spell check of individual files. --files *must* be the first command line -# arg to use this option. -if [[ "$1" == '--files' ]]; then - isort_check "${@:2}" - # If `--all` is passed, then any further arguments are ignored and the - # entire python directory is linted. -elif [[ "$1" == '--all' ]]; then - isort_check_all -else - # Check spelling only of the files that changed in last commit. - isort_check_changed -fi -echo 'vLLM isort: Done' - -# Clang-format section -# Exclude some files for formatting because they are vendored -# NOTE: Keep up to date with .github/workflows/clang-format.yml -CLANG_FORMAT_EXCLUDES=( - 'csrc/moe/topk_softmax_kernels.cu' - 'csrc/quantization/gguf/ggml-common.h' - 'csrc/quantization/gguf/dequantize.cuh' - 'csrc/quantization/gguf/vecdotq.cuh' - 'csrc/quantization/gguf/mmq.cuh' - 'csrc/quantization/gguf/mmvq.cuh' -) - -# Format specified files with clang-format -clang_format() { - clang-format -i "$@" -} - -# Format files that differ from main branch with clang-format. -clang_format_changed() { - # The `if` guard ensures that the list of filenames is not empty, which - # could cause clang-format to receive 0 positional arguments, making it hang - # waiting for STDIN. - # - # `diff-filter=ACM` and $MERGEBASE is to ensure we only format files that - # exist on both branches. - MERGEBASE="$(git merge-base origin/main HEAD)" - - # Get the list of changed files, excluding the specified ones - changed_files=$(git diff --name-only --diff-filter=ACM "$MERGEBASE" -- '*.h' '*.cpp' '*.cu' '*.cuh' | (grep -vFf <(printf "%s\n" "${CLANG_FORMAT_EXCLUDES[@]}") || echo -e)) - if [ -n "$changed_files" ]; then - echo "$changed_files" | xargs -P 5 clang-format -i - fi -} - -# Format all files with clang-format -clang_format_all() { - find csrc/ \( -name '*.h' -o -name '*.cpp' -o -name '*.cu' -o -name '*.cuh' \) -print \ - | grep -vFf <(printf "%s\n" "${CLANG_FORMAT_EXCLUDES[@]}") \ - | xargs clang-format -i -} - -# Run clang-format -if [[ "$1" == '--files' ]]; then - clang_format "${@:2}" -elif [[ "$1" == '--all' ]]; then - clang_format_all -else - clang_format_changed -fi -echo 'vLLM clang-format: Done' - -echo 'vLLM actionlint:' -tools/actionlint.sh -color -echo 'vLLM actionlint: Done' - -echo 'vLLM shellcheck:' -tools/shellcheck.sh -echo 'vLLM shellcheck: Done' - -echo 'excalidraw png check:' -tools/png-lint.sh -echo 'excalidraw png check: Done' - -if ! git diff --quiet &>/dev/null; then - echo - echo "🔍🔍There are files changed by the format checker or by you that are not added and committed:" - git --no-pager diff --name-only - echo "🔍🔍Format checker passed, but please add, commit and push all the files above to include changes made by the format checker." - - exit 1 -else - echo "✨🎉 Format check passed! Congratulations! 🎉✨" -fi - -echo 'vLLM doc-lint:' -tools/doc-lint.sh -echo 'vLLM doc-lint: Done' +echo "vLLM linting system has been moved from format.sh to pre-commit hook." +echo "Please run 'pip install -r requirements-lint.txt' and 'pre-commit install' to install the pre-commit hook." +echo "Then linters will run automatically before each commit." diff --git a/pyproject.toml b/pyproject.toml index 2a777e6638d61..ab3083d649ec8 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -15,6 +15,11 @@ build-backend = "setuptools.build_meta" [tool.setuptools_scm] # version_file = "vllm/_version.py" # currently handled by `setup.py:get_version()` +[tool.yapfignore] +ignore_patterns = [ + "build/**", +] + [tool.ruff] # Allow lines to be as long as 80. line-length = 80 @@ -52,6 +57,9 @@ ignore = [ "B007", # f-string format "UP032", + # Python 3.8 typing + "UP006", "UP035", + ] [tool.mypy] diff --git a/requirements-common.txt b/requirements-common.txt index 6c390bcfd18e6..7051ca8cb50cd 100644 --- a/requirements-common.txt +++ b/requirements-common.txt @@ -19,7 +19,7 @@ pillow # Required for image processing prometheus-fastapi-instrumentator >= 7.0.0 tiktoken >= 0.6.0 # Required for DBRX tokenizer lm-format-enforcer >= 0.10.9, < 0.11 -outlines == 0.1.11 # Requires pytorch +outlines == 0.1.11 lark == 1.2.2 xgrammar >= 0.1.6; platform_machine == "x86_64" typing_extensions >= 4.10 @@ -34,6 +34,6 @@ pyyaml six>=1.16.0; python_version > '3.11' # transitive dependency of pandas that needs to be the latest version for python 3.12 setuptools>=74.1.1; python_version > '3.11' # Setuptools is used by triton, we need to ensure a modern version is installed for 3.12+ so that it does not try to import distutils, which was removed in 3.12 einops # Required for Qwen2-VL. -compressed-tensors == 0.8.1 # required for compressed-tensors, requires pytorch +compressed-tensors == 0.9.0 # required for compressed-tensors depyf==0.18.0 # required for profiling and debugging with compilation config cloudpickle # allows pickling lambda functions in model_executor/models/registry.py diff --git a/requirements-hpu.txt b/requirements-hpu.txt index f4fb89ef42834..63a5f8b18f6b8 100644 --- a/requirements-hpu.txt +++ b/requirements-hpu.txt @@ -3,7 +3,7 @@ # Dependencies for HPU code ray -triton +triton==3.1.0 pandas tabulate setuptools>=61 diff --git a/requirements-lint.txt b/requirements-lint.txt index ffc73f90a0d48..62446f94048df 100644 --- a/requirements-lint.txt +++ b/requirements-lint.txt @@ -1,15 +1,2 @@ # formatting -yapf==0.32.0 -toml==0.10.2 -tomli==2.0.2 -ruff==0.6.5 -codespell==2.3.0 -isort==5.13.2 -clang-format==18.1.5 -pymarkdownlnt==0.9.26 - -# type checking -mypy==1.11.1 -types-PyYAML -types-requests -types-setuptools +pre-commit==4.0.1 diff --git a/requirements-test.in b/requirements-test.in index 4b4dc376d1fa5..bc76a91ad5356 100644 --- a/requirements-test.in +++ b/requirements-test.in @@ -29,4 +29,7 @@ lm-eval[api]==0.4.4 # required for model evaluation test bitsandbytes>=0.45.0 buildkite-test-collector==0.1.9 +genai_perf==0.0.8 +tritonclient==2.51.0 + numpy < 2.0.0 diff --git a/requirements-test.txt b/requirements-test.txt index f576e42afcbbf..09e009c2e21f4 100644 --- a/requirements-test.txt +++ b/requirements-test.txt @@ -37,7 +37,7 @@ audioread==3.0.1 # via librosa awscli==1.35.23 # via -r requirements-test.in -bitsandbytes>=0.45.0 +bitsandbytes==0.45.0 # via -r requirements-test.in black==24.10.0 # via datamodel-code-generator @@ -75,6 +75,8 @@ colorama==0.4.6 # tqdm-multiprocess contourpy==1.3.0 # via matplotlib +cramjam==2.9.0 + # via fastparquet cupy-cuda12x==13.3.0 # via ray cycler==0.12.1 @@ -109,6 +111,8 @@ email-validator==2.2.0 # via pydantic evaluate==0.4.3 # via lm-eval +fastparquet==2024.11.0 + # via genai-perf fastrlock==0.8.2 # via cupy-cuda12x filelock==3.16.1 @@ -130,8 +134,11 @@ fsspec[http]==2024.9.0 # via # datasets # evaluate + # fastparquet # huggingface-hub # torch +genai-perf==0.0.8 + # via -r requirements-test.in genson==1.3.0 # via datamodel-code-generator h11==0.14.0 @@ -186,6 +193,8 @@ jsonschema==4.23.0 # ray jsonschema-specifications==2024.10.1 # via jsonschema +kaleido==0.2.1 + # via genai-perf kiwisolver==1.4.7 # via matplotlib lazy-loader==0.4 @@ -200,6 +209,8 @@ lm-eval[api]==0.4.4 # via -r requirements-test.in lxml==5.3.0 # via sacrebleu +markdown-it-py==3.0.0 + # via rich markupsafe==3.0.2 # via jinja2 matplotlib==3.9.2 @@ -209,6 +220,8 @@ mbstrdecoder==1.1.3 # dataproperty # pytablewriter # typepy +mdurl==0.1.2 + # via markdown-it-py mistral-common[opencv]==1.5.1 # via # -r requirements-test.in @@ -249,6 +262,8 @@ numpy==1.26.4 # datasets # decord # evaluate + # fastparquet + # genai-perf # librosa # matplotlib # mistral-common @@ -256,15 +271,18 @@ numpy==1.26.4 # numexpr # opencv-python-headless # pandas + # patsy # peft # rouge-score # sacrebleu # scikit-learn # scipy # soxr + # statsmodels # tensorizer # torchvision # transformers + # tritonclient nvidia-cublas-cu12==12.4.5.8 # via # nvidia-cudnn-cu12 @@ -306,30 +324,39 @@ packaging==24.1 # datamodel-code-generator # datasets # evaluate + # fastparquet # huggingface-hub # lazy-loader # matplotlib # peft + # plotly # pooch # pytest # pytest-rerunfailures # ray + # statsmodels # transformers # typepy pandas==2.2.3 # via # datasets # evaluate + # fastparquet + # genai-perf + # statsmodels pathspec==0.12.1 # via black pathvalidate==3.2.1 # via pytablewriter +patsy==1.0.1 + # via statsmodels peft==0.13.2 # via # -r requirements-test.in # lm-eval pillow==10.4.0 # via + # genai-perf # matplotlib # mistral-common # sentence-transformers @@ -338,6 +365,8 @@ platformdirs==4.3.6 # via # black # pooch +plotly==5.24.1 + # via genai-perf pluggy==1.5.0 # via pytest pooch==1.8.2 @@ -360,7 +389,9 @@ psutil==6.1.0 py==1.11.0 # via pytest-forked pyarrow==18.0.0 - # via datasets + # via + # datasets + # genai-perf pyasn1==0.6.1 # via rsa pybind11==2.13.6 @@ -373,6 +404,8 @@ pydantic[email]==2.9.2 # mistral-common pydantic-core==2.23.4 # via pydantic +pygments==2.18.0 + # via rich pyparsing==3.2.0 # via matplotlib pytablewriter==1.2.0 @@ -381,14 +414,18 @@ pytest==8.3.3 # via # -r requirements-test.in # buildkite-test-collector + # genai-perf # pytest-asyncio # pytest-forked + # pytest-mock # pytest-rerunfailures # pytest-shard pytest-asyncio==0.24.0 # via -r requirements-test.in pytest-forked==1.6.0 # via -r requirements-test.in +pytest-mock==3.14.0 + # via genai-perf pytest-rerunfailures==14.0 # via -r requirements-test.in pytest-shard==0.1.2 @@ -399,6 +436,8 @@ python-dateutil==2.9.0.post0 # matplotlib # pandas # typepy +python-rapidjson==1.20 + # via tritonclient pytz==2024.2 # via # pandas @@ -409,9 +448,11 @@ pyyaml==6.0.2 # awscli # datamodel-code-generator # datasets + # genai-perf # huggingface-hub # peft # ray + # responses # timm # transformers ray[adag]==2.40.0 @@ -438,8 +479,13 @@ requests==2.32.3 # mistral-common # pooch # ray + # responses # tiktoken # transformers +responses==0.25.3 + # via genai-perf +rich==13.9.4 + # via genai-perf rouge-score==0.1.2 # via lm-eval rpds-py==0.20.1 @@ -470,6 +516,7 @@ scipy==1.13.1 # librosa # scikit-learn # sentence-transformers + # statsmodels sentence-transformers==3.2.1 # via -r requirements-test.in sentencepiece==0.2.0 @@ -490,6 +537,8 @@ soxr==0.5.0.post1 # via librosa sqlitedict==2.1.0 # via lm-eval +statsmodels==0.14.4 + # via genai-perf sympy==1.13.1 # via torch tabledata==1.3.3 @@ -499,7 +548,9 @@ tabulate==0.9.0 tcolorpy==0.1.6 # via pytablewriter tenacity==9.0.0 - # via lm-eval + # via + # lm-eval + # plotly tensorizer==2.9.0 # via -r requirements-test.in threadpoolctl==3.5.0 @@ -540,6 +591,7 @@ tqdm-multiprocess==0.0.11 # via lm-eval transformers==4.47.0 # via + # genai-perf # lm-eval # peft # sentence-transformers @@ -548,6 +600,10 @@ transformers-stream-generator==0.0.5 # via -r requirements-test.in triton==3.1.0 # via torch +tritonclient==2.51.0 + # via + # -r requirements-test.in + # genai-perf typepy[datetime]==1.3.2 # via # dataproperty @@ -555,6 +611,7 @@ typepy[datetime]==1.3.2 # tabledata typing-extensions==4.12.2 # via + # bitsandbytes # huggingface-hub # librosa # mistral-common @@ -563,10 +620,12 @@ typing-extensions==4.12.2 # torch tzdata==2024.2 # via pandas -urllib3==1.26.20 +urllib3==2.2.3 # via # botocore # requests + # responses + # tritonclient word2number==1.1 # via lm-eval xxhash==3.5.0 diff --git a/requirements-tpu.txt b/requirements-tpu.txt index 8ab18b3770ae8..51a0c65eac5aa 100644 --- a/requirements-tpu.txt +++ b/requirements-tpu.txt @@ -10,16 +10,17 @@ wheel jinja2 ray[default] -# Install torch_xla ---pre ---extra-index-url https://download.pytorch.org/whl/nightly/cpu +# Install torch, torch_xla --find-links https://storage.googleapis.com/libtpu-releases/index.html --find-links https://storage.googleapis.com/jax-releases/jax_nightly_releases.html --find-links https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html -torch==2.6.0.dev20241126+cpu -torchvision==0.20.0.dev20241126+cpu -torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241126-cp39-cp39-linux_x86_64.whl ; python_version == "3.9" -torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241126-cp310-cp310-linux_x86_64.whl ; python_version == "3.10" -torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241126-cp311-cp311-linux_x86_64.whl ; python_version == "3.11" -jaxlib==0.4.36.dev20241122 -jax==0.4.36.dev20241122 +# Note: This torch whl can be slightly different from the official torch nightly whl +# since they are not built on the same commit (but on the same day). This difference may cause C++ undefined symbol issue +# if some change between the 2 commits introduce some C++ API change. +# Here we install the exact torch whl from which torch_xla is built from, to avoid potential C++ undefined symbol issue. +torch @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-2.7.0.dev20250124-cp39-cp39-linux_x86_64.whl ; python_version == "3.9" +torch @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-2.7.0.dev20250124-cp310-cp310-linux_x86_64.whl ; python_version == "3.10" +torch @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-2.7.0.dev20250124-cp311-cp311-linux_x86_64.whl ; python_version == "3.11" +torch_xla[pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.7.0.dev20250124-cp39-cp39-linux_x86_64.whl ; python_version == "3.9" +torch_xla[pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.7.0.dev20250124-cp310-cp310-linux_x86_64.whl ; python_version == "3.10" +torch_xla[pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.7.0.dev20250124-cp311-cp311-linux_x86_64.whl ; python_version == "3.11" diff --git a/setup.py b/setup.py old mode 100644 new mode 100755 index b7bc64fd950e7..e50e91a0f94ce --- a/setup.py +++ b/setup.py @@ -228,8 +228,11 @@ def target_name(s: str) -> str: # CMake appends the extension prefix to the install path, # and outdir already contains that prefix, so we need to remove it. + # We assume only the final component of extension prefix is added by + # CMake, this is currently true for current extensions but may not + # always be the case. prefix = outdir - for i in range(ext.name.count('.')): + if '.' in ext.name: prefix = prefix.parent # prefix here should actually be the same for all components @@ -298,9 +301,11 @@ def run(self) -> None: files_to_copy = [ "vllm/_C.abi3.so", "vllm/_moe_C.abi3.so", - "vllm/vllm_flash_attn/vllm_flash_attn_c.abi3.so", + "vllm/vllm_flash_attn/_vllm_fa2_C.abi3.so", + "vllm/vllm_flash_attn/_vllm_fa3_C.abi3.so", "vllm/vllm_flash_attn/flash_attn_interface.py", "vllm/vllm_flash_attn/__init__.py", + "vllm/cumem_allocator.abi3.so", # "vllm/_version.py", # not available in nightly wheels yet ] file_members = filter(lambda x: x.filename in files_to_copy, @@ -324,21 +329,26 @@ def run(self) -> None: def _is_hpu() -> bool: - is_hpu_available = True + # if VLLM_TARGET_DEVICE env var was set explicitly, skip HPU autodetection + if os.getenv("VLLM_TARGET_DEVICE", None) == VLLM_TARGET_DEVICE: + return VLLM_TARGET_DEVICE == "hpu" + + # if VLLM_TARGET_DEVICE was not set explicitly, check if hl-smi succeeds, + # and if it doesn't, check if habanalabs driver is loaded + is_hpu_available = False try: - subprocess.run(["hl-smi"], capture_output=True, check=True) + out = subprocess.run(["hl-smi"], capture_output=True, check=True) + is_hpu_available = out.returncode == 0 except (FileNotFoundError, PermissionError, subprocess.CalledProcessError): - if not os.path.exists('/dev/accel/accel0') and not os.path.exists( - '/dev/accel/accel_controlD0'): - # last resort... + if sys.platform.startswith("linux"): try: output = subprocess.check_output( 'lsmod | grep habanalabs | wc -l', shell=True) is_hpu_available = int(output) > 0 except (ValueError, FileNotFoundError, PermissionError, subprocess.CalledProcessError): - is_hpu_available = False - return is_hpu_available or VLLM_TARGET_DEVICE == "hpu" + pass + return is_hpu_available def _no_device() -> bool: @@ -548,7 +558,7 @@ def _read_requirements(filename: str) -> List[str]: return resolved_requirements if _no_device(): - requirements = _read_requirements("requirements-cuda.txt") + requirements = _read_requirements("requirements-cpu.txt") elif _is_cuda(): requirements = _read_requirements("requirements-cuda.txt") cuda_major, cuda_minor = torch.version.cuda.split(".") @@ -592,8 +602,12 @@ def _read_requirements(filename: str) -> List[str]: ext_modules.append(CMakeExtension(name="vllm._gradlib_C")) if _is_cuda(): - ext_modules.append( - CMakeExtension(name="vllm.vllm_flash_attn.vllm_flash_attn_c")) + ext_modules.append(CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa2_C")) + if envs.VLLM_USE_PRECOMPILED or get_nvcc_cuda_version() >= Version("12.0"): + # FA3 requires CUDA 12.0 or later + ext_modules.append( + CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa3_C")) + ext_modules.append(CMakeExtension(name="vllm.cumem_allocator")) if _build_custom_ops(): ext_modules.append(CMakeExtension(name="vllm._C")) diff --git a/tests/async_engine/test_api_server.py b/tests/async_engine/test_api_server.py index 83c71b5cf6eb7..91ac35dd67bbf 100644 --- a/tests/async_engine/test_api_server.py +++ b/tests/async_engine/test_api_server.py @@ -25,27 +25,32 @@ def _query_server_long(prompt: str) -> dict: @pytest.fixture -def api_server(tokenizer_pool_size: int, worker_use_ray: bool): +def api_server(tokenizer_pool_size: int, distributed_executor_backend: str): script_path = Path(__file__).parent.joinpath( "api_server_async_engine.py").absolute() commands = [ - sys.executable, "-u", - str(script_path), "--model", "facebook/opt-125m", "--host", - "127.0.0.1", "--tokenizer-pool-size", - str(tokenizer_pool_size) + sys.executable, + "-u", + str(script_path), + "--model", + "facebook/opt-125m", + "--host", + "127.0.0.1", + "--tokenizer-pool-size", + str(tokenizer_pool_size), + "--distributed-executor-backend", + distributed_executor_backend, ] - if worker_use_ray: - commands.append("--worker-use-ray") uvicorn_process = subprocess.Popen(commands) yield uvicorn_process.terminate() @pytest.mark.parametrize("tokenizer_pool_size", [0, 2]) -@pytest.mark.parametrize("worker_use_ray", [False, True]) +@pytest.mark.parametrize("distributed_executor_backend", ["mp", "ray"]) def test_api_server(api_server, tokenizer_pool_size: int, - worker_use_ray: bool): + distributed_executor_backend: str): """ Run the API server and test it. diff --git a/tests/basic_correctness/test_basic_correctness.py b/tests/basic_correctness/test_basic_correctness.py index 31a101e48e026..23285040642a8 100644 --- a/tests/basic_correctness/test_basic_correctness.py +++ b/tests/basic_correctness/test_basic_correctness.py @@ -61,9 +61,10 @@ def test_models( if backend == "FLASHINFER" and current_platform.is_rocm(): pytest.skip("Flashinfer does not support ROCm/HIP.") - if backend == "XFORMERS" and model == "google/gemma-2-2b-it": + if backend in ("XFORMERS", + "FLASHINFER") and model == "google/gemma-2-2b-it": pytest.skip( - "XFORMERS does not support gemma2 with full context length.") + f"{backend} does not support gemma2 with full context length.") os.environ["VLLM_ATTENTION_BACKEND"] = backend diff --git a/tests/basic_correctness/test_cumem.py b/tests/basic_correctness/test_cumem.py new file mode 100644 index 0000000000000..53f4ef08f36a2 --- /dev/null +++ b/tests/basic_correctness/test_cumem.py @@ -0,0 +1,112 @@ +import torch + +from vllm import LLM, SamplingParams +from vllm.device_allocator.cumem import CuMemAllocator +from vllm.utils import GiB_bytes + +from ..utils import fork_new_process_for_each_test + + +@fork_new_process_for_each_test +def test_basic_cumem(): + # some tensors from default memory pool + shape = (1024, 1024) + x = torch.empty(shape, device='cuda') + x.zero_() + + # some tensors from custom memory pool + allocator = CuMemAllocator.get_instance() + with allocator.use_memory_pool(): + # custom memory pool + y = torch.empty(shape, device='cuda') + y.zero_() + y += 1 + z = torch.empty(shape, device='cuda') + z.zero_() + z += 2 + + # they can be used together + output = x + y + z + assert torch.allclose(output, torch.ones_like(output) * 3) + + free_bytes = torch.cuda.mem_get_info()[0] + allocator.sleep() + free_bytes_after_sleep = torch.cuda.mem_get_info()[0] + assert free_bytes_after_sleep > free_bytes + allocator.wake_up() + + # they can be used together + output = x + y + z + assert torch.allclose(output, torch.ones_like(output) * 3) + + +@fork_new_process_for_each_test +def test_cumem_with_cudagraph(): + allocator = CuMemAllocator.get_instance() + with allocator.use_memory_pool(): + weight = torch.eye(1024, device='cuda') + with allocator.use_memory_pool(tag="discard"): + cache = torch.empty(1024, 1024, device='cuda') + + def model(x): + out = x @ weight + cache[:out.size(0)].copy_(out) + return out + 1 + + x = torch.empty(128, 1024, device='cuda') + + # warmup + model(x) + + # capture cudagraph + model_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(model_graph): + y = model(x) + + free_bytes = torch.cuda.mem_get_info()[0] + allocator.sleep() + free_bytes_after_sleep = torch.cuda.mem_get_info()[0] + assert free_bytes_after_sleep > free_bytes + allocator.wake_up() + + # after waking up, the content in the weight tensor + # should be restored, but the content in the cache tensor + # should be discarded + + # this operation is also compatible with cudagraph + + x.random_() + model_graph.replay() + + # cache content is as expected + assert torch.allclose(x, cache[:x.size(0)]) + + # output content is as expected + assert torch.allclose(y, x + 1) + + +@fork_new_process_for_each_test +def test_end_to_end(): + free, total = torch.cuda.mem_get_info() + used_bytes_baseline = total - free # in case other process is running + llm = LLM("meta-llama/Llama-3.2-1B", enable_sleep_mode=True) + prompt = "How are you?" + sampling_params = SamplingParams(temperature=0, max_tokens=10) + output = llm.generate(prompt, sampling_params) + + # the benefit of `llm.sleep(level=2)` is mainly CPU memory usage, + # which is difficult to measure in the test. therefore, we only + # test sleep level 1 here. + llm.sleep(level=1) + + free_gpu_bytes_after_sleep, total = torch.cuda.mem_get_info() + used_bytes = total - free_gpu_bytes_after_sleep - used_bytes_baseline + # now the memory usage is mostly cudagraph memory pool, + # and it should be less than the model weights (1B model, 2GiB weights) + assert used_bytes < 2 * GiB_bytes + + llm.wake_up() + output2 = llm.generate(prompt, sampling_params) + + # cmp output + assert output[0].outputs[0].text == output2[0].outputs[0].text diff --git a/tests/basic_correctness/test_preemption.py b/tests/basic_correctness/test_preemption.py index 4e502cfb5f4f8..4b27dcbc8609f 100644 --- a/tests/basic_correctness/test_preemption.py +++ b/tests/basic_correctness/test_preemption.py @@ -29,10 +29,10 @@ def check_settings(): @pytest.fixture -def worker_use_ray() -> bool: - # When SPMD worker is used, use ray_use_worker=True +def distributed_executor_backend() -> str: + # When SPMD worker is used, use distributed_executor_backend="ray" # to test delta input optimization works with preemption. - return envs.VLLM_USE_RAY_SPMD_WORKER + return "ray" if envs.VLLM_USE_RAY_SPMD_WORKER else "mp" @pytest.mark.parametrize("model", MODELS) @@ -47,7 +47,7 @@ def test_chunked_prefill_recompute( dtype: str, max_tokens: int, chunked_prefill_token_size: int, - worker_use_ray: bool, + distributed_executor_backend: str, ) -> None: """Ensure that chunked prefill works with preemption.""" max_num_seqs = min(chunked_prefill_token_size, 256) @@ -66,7 +66,7 @@ def test_chunked_prefill_recompute( max_num_batched_tokens=max_num_batched_tokens, enable_chunked_prefill=enable_chunked_prefill, max_num_seqs=max_num_seqs, - worker_use_ray=worker_use_ray, + distributed_executor_backend=distributed_executor_backend, disable_log_stats=False, ) as vllm_model: vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens) @@ -93,7 +93,7 @@ def test_preemption( model: str, dtype: str, max_tokens: int, - worker_use_ray: bool, + distributed_executor_backend: str, ) -> None: """By default, recompute preemption is enabled""" @@ -104,7 +104,7 @@ def test_preemption( model, dtype=dtype, disable_log_stats=False, - worker_use_ray=worker_use_ray, + distributed_executor_backend=distributed_executor_backend, ) as vllm_model: vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens) assert (vllm_model.model.llm_engine.scheduler[0].artificial_preempt_cnt @@ -144,7 +144,7 @@ def test_preemption_infeasible( model: str, dtype: str, max_tokens: int, - worker_use_ray: bool, + distributed_executor_backend: str, ) -> None: """Verify infeasible preemption request will be ignored.""" BLOCK_SIZE = 16 @@ -159,7 +159,7 @@ def test_preemption_infeasible( # ignored instead of hanging forever. num_gpu_blocks_override=prefill_blocks + decode_blocks // 2, max_model_len=((prefill_blocks + decode_blocks // 2) * BLOCK_SIZE), - worker_use_ray=worker_use_ray, + distributed_executor_backend=distributed_executor_backend, ) as vllm_model: sampling_params = SamplingParams(max_tokens=max_tokens, ignore_eos=True) diff --git a/tests/compile/test_basic_correctness.py b/tests/compile/test_basic_correctness.py index 87d5aefea6cb4..1945479fc3031 100644 --- a/tests/compile/test_basic_correctness.py +++ b/tests/compile/test_basic_correctness.py @@ -58,7 +58,7 @@ class TestSetting: model_args=["--task", "embed"], pp_size=1, tp_size=1, - attn_backend="FLASHINFER", + attn_backend="FLASH_ATTN", method="encode", fullgraph=True, ), diff --git a/tests/conftest.py b/tests/conftest.py index 95af4ac1eb17b..279c1bf9a3776 100644 --- a/tests/conftest.py +++ b/tests/conftest.py @@ -244,6 +244,7 @@ def video_assets() -> _VideoAssets: _T = TypeVar("_T", nn.Module, torch.Tensor, BatchEncoding, BatchFeature, dict) +_R = TypeVar("_R") class HfRunner: @@ -930,6 +931,10 @@ def score( req_outputs = self.model.score(text_1, text_2) return [req_output.outputs.score for req_output in req_outputs] + def apply_model(self, func: Callable[[nn.Module], _R]) -> list[_R]: + executor = self.model.llm_engine.model_executor + return executor.apply_model(func) + def __enter__(self): return self diff --git a/tests/core/block/test_prefix_caching_block.py b/tests/core/block/test_prefix_caching_block.py index 29ac3a3c86cb4..6642174c17d8b 100644 --- a/tests/core/block/test_prefix_caching_block.py +++ b/tests/core/block/test_prefix_caching_block.py @@ -796,6 +796,44 @@ def test_find_cached_blocks_prefix(): block_hashes=block_hashes_seq1) assert len(cached_blocks) == len(blocks_seq1) - num_evicted_blocks + # Test reset prefix cache + @staticmethod + @pytest.mark.parametrize("num_blocks", [10]) + @pytest.mark.parametrize("block_size", [16]) + def test_reset_prefix_cache(num_blocks: int, block_size: int): + """This test case simulates the case of resetting the prefix cache.""" + + allocator = PrefixCachingBlockAllocator(num_blocks=num_blocks, + block_size=block_size) + token_ids = list(range(3 * block_size)) + + first_chain = TestPrefixCachingBlockAllocator.create_immutable_chain( + block_size=block_size, + token_ids=token_ids, + allocator=allocator, + ) + second_chain = TestPrefixCachingBlockAllocator.create_immutable_chain( + block_size=block_size, + token_ids=token_ids, + allocator=allocator, + ) + + # Free each block in the first chain. + for block in first_chain: + allocator.free(block) + + # Failed to reset prefix cache because some blocks are not freed yet. + assert not allocator.reset_prefix_cache() + assert allocator.get_prefix_cache_hit_rate() > 0.0 + + # Free each block in the second chain. + for block in second_chain: + allocator.free(block) + + # Reset prefix cache. + assert allocator.reset_prefix_cache() + assert allocator.get_prefix_cache_hit_rate() == 0.0 + @staticmethod def create_immutable_chain( block_size: int, diff --git a/tests/distributed/test_torchrun_example.py b/tests/distributed/test_torchrun_example.py new file mode 100644 index 0000000000000..7aa03d7f0402a --- /dev/null +++ b/tests/distributed/test_torchrun_example.py @@ -0,0 +1,56 @@ +# unit test for `examples/offline_inference/torchrun_example.py` + +import random + +import torch.distributed as dist + +from vllm import LLM, SamplingParams +from vllm.distributed.parallel_state import get_world_group + +# Create prompts +prompts = [ + "Hello, my name is", + "The president of the United States is", + "The capital of France is", + "The future of AI is", +] + +sampling_params = SamplingParams(temperature=0.8, top_p=0.95) + +# set different `gpu_memory_utilization` and `swap_space` for different ranks, +# to test if all ranks agree on the same kv cache configuration. +llm = LLM(model="facebook/opt-125m", + tensor_parallel_size=2, + distributed_executor_backend="external_launcher", + gpu_memory_utilization=random.uniform(0.7, 0.9), + swap_space=random.randint(1, 4)) + +outputs = llm.generate(prompts, sampling_params) + +cpu_group = get_world_group().cpu_group + +torch_rank = dist.get_rank(group=cpu_group) + + +def test_consistent_across_ranks(obj): + if torch_rank == 0: + dist.broadcast_object_list([obj], src=0, group=cpu_group) + else: + container = [None] + dist.broadcast_object_list(container, src=0, group=cpu_group) + assert container[0] == obj + + +test_consistent_across_ranks( + llm.llm_engine.vllm_config.cache_config.num_cpu_blocks) +test_consistent_across_ranks( + llm.llm_engine.vllm_config.cache_config.num_gpu_blocks) + +# all ranks should have the same outputs +for output in outputs: + prompt = output.prompt + generated_text = output.outputs[0].text + test_consistent_across_ranks(prompt) + test_consistent_across_ranks(generated_text) + print(f"Rank {torch_rank}, Prompt: {prompt!r}, " + f"Generated text: {generated_text!r}") diff --git a/tests/engine/test_custom_executor.py b/tests/engine/test_custom_executor.py index bbabb936e92ba..0e33f3662da82 100644 --- a/tests/engine/test_custom_executor.py +++ b/tests/engine/test_custom_executor.py @@ -1,12 +1,13 @@ import asyncio import os +from typing import Any, Callable, Dict, List, Optional, Tuple, Union import pytest from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs from vllm.engine.async_llm_engine import AsyncLLMEngine from vllm.engine.llm_engine import LLMEngine -from vllm.executor.gpu_executor import GPUExecutor, GPUExecutorAsync +from vllm.executor.uniproc_executor import UniProcExecutor from vllm.sampling_params import SamplingParams @@ -14,21 +15,20 @@ class Mock: ... -class CustomGPUExecutor(GPUExecutor): +class CustomUniExecutor(UniProcExecutor): - def execute_model(self, *args, **kwargs): + def collective_rpc(self, + method: Union[str, Callable], + timeout: Optional[float] = None, + args: Tuple = (), + kwargs: Optional[Dict] = None) -> List[Any]: # Drop marker to show that this was ran with open(".marker", "w"): ... - return super().execute_model(*args, **kwargs) + return super().collective_rpc(method, timeout, args, kwargs) -class CustomGPUExecutorAsync(GPUExecutorAsync): - - async def execute_model_async(self, *args, **kwargs): - with open(".marker", "w"): - ... - return await super().execute_model_async(*args, **kwargs) +CustomUniExecutorAsync = CustomUniExecutor @pytest.mark.parametrize("model", ["facebook/opt-125m"]) @@ -41,10 +41,6 @@ def test_custom_executor_type_checking(model): engine_args = AsyncEngineArgs(model=model, distributed_executor_backend=Mock) AsyncLLMEngine.from_engine_args(engine_args) - with pytest.raises(TypeError): - engine_args = AsyncEngineArgs( - model=model, distributed_executor_backend=CustomGPUExecutor) - AsyncLLMEngine.from_engine_args(engine_args) @pytest.mark.parametrize("model", ["facebook/opt-125m"]) @@ -55,7 +51,9 @@ def test_custom_executor(model, tmp_path): assert not os.path.exists(".marker") engine_args = EngineArgs( - model=model, distributed_executor_backend=CustomGPUExecutor) + model=model, + distributed_executor_backend=CustomUniExecutor, + ) engine = LLMEngine.from_engine_args(engine_args) sampling_params = SamplingParams(max_tokens=1) @@ -75,7 +73,7 @@ def test_custom_executor_async(model, tmp_path): assert not os.path.exists(".marker") engine_args = AsyncEngineArgs( - model=model, distributed_executor_backend=CustomGPUExecutorAsync) + model=model, distributed_executor_backend=CustomUniExecutorAsync) engine = AsyncLLMEngine.from_engine_args(engine_args) sampling_params = SamplingParams(max_tokens=1) diff --git a/tests/engine/test_multiproc_workers.py b/tests/engine/test_multiproc_workers.py index e07dd6deef5bf..04505fcaae24b 100644 --- a/tests/engine/test_multiproc_workers.py +++ b/tests/engine/test_multiproc_workers.py @@ -6,16 +6,15 @@ import pytest +from vllm.config import VllmConfig from vllm.executor.multiproc_worker_utils import (ProcessWorkerWrapper, ResultHandler, WorkerMonitor) +from vllm.worker.worker_base import WorkerWrapperBase -class DummyWorker: +class DummyWorkerWrapper(WorkerWrapperBase): """Dummy version of vllm.worker.worker.Worker""" - def __init__(self, rank: int): - self.rank = rank - def worker_method(self, worker_input: Any) -> Tuple[int, Any]: sleep(0.05) @@ -23,14 +22,15 @@ def worker_method(self, worker_input: Any) -> Tuple[int, Any]: # simulate error case raise worker_input - return self.rank, input + return self.rpc_rank, input def _start_workers() -> Tuple[List[ProcessWorkerWrapper], WorkerMonitor]: result_handler = ResultHandler() + vllm_config = VllmConfig() workers = [ - ProcessWorkerWrapper(result_handler, partial(DummyWorker, rank=rank)) - for rank in range(8) + ProcessWorkerWrapper(result_handler, DummyWorkerWrapper, vllm_config, + rank) for rank in range(8) ] worker_monitor = WorkerMonitor(workers, result_handler) diff --git a/tests/entrypoints/llm/test_collective_rpc.py b/tests/entrypoints/llm/test_collective_rpc.py new file mode 100644 index 0000000000000..22473ce275295 --- /dev/null +++ b/tests/entrypoints/llm/test_collective_rpc.py @@ -0,0 +1,36 @@ +import pytest + +from vllm import LLM + +from ...utils import fork_new_process_for_each_test + + +@pytest.mark.parametrize("tp_size", [1, 2]) +@pytest.mark.parametrize("backend", ["mp", "ray"]) +@fork_new_process_for_each_test +def test_collective_rpc(tp_size, backend): + if tp_size == 1 and backend == "ray": + pytest.skip("Skip duplicate test case") + if tp_size == 1: + backend = None + + # intentionally define the method and class in the test function, + # to test if they can be serialized and sent to the workers + def echo_rank(self): + return self.rank + + from vllm.worker.worker import Worker + + class MyWorker(Worker): + + def echo_rank(self): + return self.rank + + llm = LLM(model="meta-llama/Llama-3.2-1B-Instruct", + enforce_eager=True, + load_format="dummy", + tensor_parallel_size=tp_size, + distributed_executor_backend=backend, + worker_cls=MyWorker) + for method in ["echo_rank", echo_rank]: + assert llm.collective_rpc(method) == list(range(tp_size)) diff --git a/tests/entrypoints/llm/test_encode.py b/tests/entrypoints/llm/test_encode.py index 41163809237e9..3906ad766e0b6 100644 --- a/tests/entrypoints/llm/test_encode.py +++ b/tests/entrypoints/llm/test_encode.py @@ -105,3 +105,10 @@ def test_multiple_pooling_params(llm: LLM): # pooling_params is None, default params should be applied outputs = llm.encode(PROMPTS, pooling_params=None) assert len(PROMPTS) == len(outputs) + + +@pytest.mark.skip_global_cleanup +def test_right_side_truncation(llm: LLM): + # Embeddings models should truncate the end of the prompt + tokenizer = llm.get_tokenizer() + assert tokenizer.truncation_side == "right" diff --git a/tests/entrypoints/openai/test_lora_adapters.py b/tests/entrypoints/openai/test_lora_adapters.py index 46a064f6d9e68..6ff99f6faa143 100644 --- a/tests/entrypoints/openai/test_lora_adapters.py +++ b/tests/entrypoints/openai/test_lora_adapters.py @@ -17,6 +17,33 @@ # generation quality here LORA_NAME = "typeof/zephyr-7b-beta-lora" +BADREQUEST_CASES = [ + ( + "test_rank", + { + "r": 1024 + }, + "is greater than max_lora_rank", + ), + ( + "test_bias", + { + "bias": "all" + }, + "Adapter bias cannot be used without bias_enabled", + ), + ("test_dora", { + "use_dora": True + }, "does not yet support DoRA"), + ( + "test_modules_to_save", + { + "modules_to_save": ["lm_head"] + }, + "only supports modules_to_save being None", + ), +] + @pytest.fixture(scope="module") def zephyr_lora_files(): @@ -138,32 +165,36 @@ async def test_dynamic_lora_invalid_files(client: openai.AsyncOpenAI, @pytest.mark.asyncio -async def test_dynamic_lora_invalid_lora_rank(client: openai.AsyncOpenAI, - tmp_path, zephyr_lora_files): - invalid_rank = tmp_path / "invalid_rank" - - # Copy adapter from zephyr_lora_files to invalid_rank - shutil.copytree(zephyr_lora_files, invalid_rank) - - with open(invalid_rank / "adapter_config.json") as f: +@pytest.mark.parametrize("test_name,config_change,expected_error", + BADREQUEST_CASES) +async def test_dynamic_lora_badrequests(client: openai.AsyncOpenAI, tmp_path, + zephyr_lora_files, test_name: str, + config_change: dict, + expected_error: str): + # Create test directory + test_dir = tmp_path / test_name + + # Copy adapter files + shutil.copytree(zephyr_lora_files, test_dir) + + # Load and modify configuration + config_path = test_dir / "adapter_config.json" + with open(config_path) as f: adapter_config = json.load(f) + # Apply configuration changes + adapter_config.update(config_change) - print(adapter_config) - - # assert False - - # Change rank to invalid value - adapter_config["r"] = 1024 - with open(invalid_rank / "adapter_config.json", "w") as f: + # Save modified configuration + with open(config_path, "w") as f: json.dump(adapter_config, f) - with pytest.raises(openai.BadRequestError, - match="is greater than max_lora_rank"): + # Test loading the adapter + with pytest.raises(openai.BadRequestError, match=expected_error): await client.post("load_lora_adapter", cast_to=str, body={ - "lora_name": "invalid-json", - "lora_path": str(invalid_rank) + "lora_name": test_name, + "lora_path": str(test_dir) }) diff --git a/tests/entrypoints/openai/test_metrics.py b/tests/entrypoints/openai/test_metrics.py index 6523c8b6297c6..469a5fb039fb6 100644 --- a/tests/entrypoints/openai/test_metrics.py +++ b/tests/entrypoints/openai/test_metrics.py @@ -16,6 +16,24 @@ MODEL_NAME = "TinyLlama/TinyLlama-1.1B-Chat-v1.0" +@pytest.fixture(scope="module", params=[True, False]) +def use_v1(request): + # Module-scoped variant of run_with_both_engines + # + # Use this fixture to run a test with both v0 and v1, and + # also to conditionalize the test logic e.g. + # + # def test_metrics_exist(use_v1, server, client): + # ... + # expected = EXPECTED_V1_METRICS if use_v1 else EXPECTED_METRICS + # for metric in expected: + # assert metric in response.text + # + # @skip_v1 wouldn't work here because this is a module-level + # fixture - per-function decorators would have no effect + yield request.param + + @pytest.fixture(scope="module") def default_server_args(): return [ @@ -36,10 +54,12 @@ def default_server_args(): "--enable-chunked-prefill", "--disable-frontend-multiprocessing", ]) -def server(default_server_args, request): +def server(use_v1, default_server_args, request): if request.param: default_server_args.append(request.param) - with RemoteOpenAIServer(MODEL_NAME, default_server_args) as remote_server: + env_dict = dict(VLLM_USE_V1='1' if use_v1 else '0') + with RemoteOpenAIServer(MODEL_NAME, default_server_args, + env_dict=env_dict) as remote_server: yield remote_server @@ -84,7 +104,9 @@ async def client(server): @pytest.mark.asyncio async def test_metrics_counts(server: RemoteOpenAIServer, - client: openai.AsyncClient): + client: openai.AsyncClient, use_v1: bool): + if use_v1: + pytest.skip("Skipping test on vllm V1") for _ in range(_NUM_REQUESTS): # sending a request triggers the metrics to be logged. await client.completions.create( @@ -174,10 +196,15 @@ async def test_metrics_counts(server: RemoteOpenAIServer, "swap_space_bytes", ] +EXPECTED_METRICS_V1 = [ + "vllm:num_requests_running", + "vllm:num_requests_waiting", +] + @pytest.mark.asyncio async def test_metrics_exist(server: RemoteOpenAIServer, - client: openai.AsyncClient): + client: openai.AsyncClient, use_v1: bool): # sending a request triggers the metrics to be logged. await client.completions.create(model=MODEL_NAME, prompt="Hello, my name is", @@ -187,11 +214,13 @@ async def test_metrics_exist(server: RemoteOpenAIServer, response = requests.get(server.url_for("metrics")) assert response.status_code == HTTPStatus.OK - for metric in EXPECTED_METRICS: + for metric in (EXPECTED_METRICS_V1 if use_v1 else EXPECTED_METRICS): assert metric in response.text -def test_metrics_exist_run_batch(): +def test_metrics_exist_run_batch(use_v1: bool): + if use_v1: + pytest.skip("Skipping test on vllm V1") input_batch = """{"custom_id": "request-0", "method": "POST", "url": "/v1/embeddings", "body": {"model": "intfloat/e5-mistral-7b-instruct", "input": "You are a helpful assistant."}}""" # noqa: E501 base_url = "0.0.0.0" diff --git a/tests/entrypoints/openai/test_rerank.py b/tests/entrypoints/openai/test_rerank.py new file mode 100644 index 0000000000000..cfd8f33133960 --- /dev/null +++ b/tests/entrypoints/openai/test_rerank.py @@ -0,0 +1,87 @@ +import pytest +import requests + +from vllm.entrypoints.openai.protocol import RerankResponse + +from ...utils import RemoteOpenAIServer + +MODEL_NAME = "BAAI/bge-reranker-base" + + +@pytest.fixture(scope="module") +def server(): + args = ["--enforce-eager", "--max-model-len", "100"] + + with RemoteOpenAIServer(MODEL_NAME, args) as remote_server: + yield remote_server + + +@pytest.mark.asyncio +@pytest.mark.parametrize("model_name", [MODEL_NAME]) +def test_rerank_texts(server: RemoteOpenAIServer, model_name: str): + query = "What is the capital of France?" + documents = [ + "The capital of Brazil is Brasilia.", "The capital of France is Paris." + ] + + rerank_response = requests.post(server.url_for("rerank"), + json={ + "model": model_name, + "query": query, + "documents": documents, + }) + rerank_response.raise_for_status() + rerank = RerankResponse.model_validate(rerank_response.json()) + + assert rerank.id is not None + assert rerank.results is not None + assert len(rerank.results) == 2 + assert rerank.results[0].relevance_score >= 0.9 + assert rerank.results[1].relevance_score <= 0.01 + + +@pytest.mark.asyncio +@pytest.mark.parametrize("model_name", [MODEL_NAME]) +def test_top_n(server: RemoteOpenAIServer, model_name: str): + query = "What is the capital of France?" + documents = [ + "The capital of Brazil is Brasilia.", + "The capital of France is Paris.", "Cross-encoder models are neat" + ] + + rerank_response = requests.post(server.url_for("rerank"), + json={ + "model": model_name, + "query": query, + "documents": documents, + "top_n": 2 + }) + rerank_response.raise_for_status() + rerank = RerankResponse.model_validate(rerank_response.json()) + + assert rerank.id is not None + assert rerank.results is not None + assert len(rerank.results) == 2 + assert rerank.results[0].relevance_score >= 0.9 + assert rerank.results[1].relevance_score <= 0.01 + + +@pytest.mark.asyncio +@pytest.mark.parametrize("model_name", [MODEL_NAME]) +def test_rerank_max_model_len(server: RemoteOpenAIServer, model_name: str): + + query = "What is the capital of France?" * 100 + documents = [ + "The capital of Brazil is Brasilia.", "The capital of France is Paris." + ] + + rerank_response = requests.post(server.url_for("rerank"), + json={ + "model": model_name, + "query": query, + "documents": documents + }) + assert rerank_response.status_code == 400 + # Assert just a small fragments of the response + assert "Please reduce the length of the input." in \ + rerank_response.text \ No newline at end of file diff --git a/tests/entrypoints/openai/test_run_batch.py b/tests/entrypoints/openai/test_run_batch.py index 097d6b1a32349..1f8a56bb43ac6 100644 --- a/tests/entrypoints/openai/test_run_batch.py +++ b/tests/entrypoints/openai/test_run_batch.py @@ -1,3 +1,4 @@ +import json import subprocess import sys import tempfile @@ -21,6 +22,9 @@ {"custom_id": "request-3", "method": "POST", "url": "/v1/embeddings", "body": {"model": "intfloat/e5-mistral-7b-instruct", "input": "Hello world!"}} {"custom_id": "request-4", "method": "POST", "url": "/v1/embeddings", "body": {"model": "NonExistModel", "input": "Hello world!"}}""" +INPUT_SCORE_BATCH = """{"custom_id": "request-1", "method": "POST", "url": "/v1/score", "body": {"model": "BAAI/bge-reranker-v2-m3", "text_1": "What is the capital of France?", "text_2": ["The capital of Brazil is Brasilia.", "The capital of France is Paris."]}} +{"custom_id": "request-2", "method": "POST", "url": "/v1/score", "body": {"model": "BAAI/bge-reranker-v2-m3", "text_1": "What is the capital of France?", "text_2": ["The capital of Brazil is Brasilia.", "The capital of France is Paris."]}}""" + def test_empty_file(): with tempfile.NamedTemporaryFile( @@ -102,3 +106,36 @@ def test_embeddings(): # Ensure that the output format conforms to the openai api. # Validation should throw if the schema is wrong. BatchRequestOutput.model_validate_json(line) + + +def test_score(): + with tempfile.NamedTemporaryFile( + "w") as input_file, tempfile.NamedTemporaryFile( + "r") as output_file: + input_file.write(INPUT_SCORE_BATCH) + input_file.flush() + proc = subprocess.Popen([ + sys.executable, + "-m", + "vllm.entrypoints.openai.run_batch", + "-i", + input_file.name, + "-o", + output_file.name, + "--model", + "BAAI/bge-reranker-v2-m3", + ], ) + proc.communicate() + proc.wait() + assert proc.returncode == 0, f"{proc=}" + + contents = output_file.read() + for line in contents.strip().split("\n"): + # Ensure that the output format conforms to the openai api. + # Validation should throw if the schema is wrong. + BatchRequestOutput.model_validate_json(line) + + # Ensure that there is no error in the response. + line_dict = json.loads(line) + assert isinstance(line_dict, dict) + assert line_dict["error"] is None diff --git a/tests/entrypoints/openai/test_score.py b/tests/entrypoints/openai/test_score.py index a803ea4a8d6ad..0d19615bc0d99 100644 --- a/tests/entrypoints/openai/test_score.py +++ b/tests/entrypoints/openai/test_score.py @@ -10,9 +10,7 @@ @pytest.fixture(scope="module") def server(): - args = [ - "--enforce-eager", - ] + args = ["--enforce-eager", "--max-model-len", "100"] with RemoteOpenAIServer(MODEL_NAME, args) as remote_server: yield remote_server @@ -20,8 +18,7 @@ def server(): @pytest.mark.asyncio @pytest.mark.parametrize("model_name", [MODEL_NAME]) -async def test_text_1_str_text_2_list(server: RemoteOpenAIServer, - model_name: str): +def test_text_1_str_text_2_list(server: RemoteOpenAIServer, model_name: str): text_1 = "What is the capital of France?" text_2 = [ "The capital of Brazil is Brasilia.", "The capital of France is Paris." @@ -45,8 +42,7 @@ async def test_text_1_str_text_2_list(server: RemoteOpenAIServer, @pytest.mark.asyncio @pytest.mark.parametrize("model_name", [MODEL_NAME]) -async def test_text_1_list_text_2_list(server: RemoteOpenAIServer, - model_name: str): +def test_text_1_list_text_2_list(server: RemoteOpenAIServer, model_name: str): text_1 = [ "What is the capital of the United States?", "What is the capital of France?" @@ -73,8 +69,7 @@ async def test_text_1_list_text_2_list(server: RemoteOpenAIServer, @pytest.mark.asyncio @pytest.mark.parametrize("model_name", [MODEL_NAME]) -async def test_text_1_str_text_2_str(server: RemoteOpenAIServer, - model_name: str): +def test_text_1_str_text_2_str(server: RemoteOpenAIServer, model_name: str): text_1 = "What is the capital of France?" text_2 = "The capital of France is Paris." @@ -91,3 +86,36 @@ async def test_text_1_str_text_2_str(server: RemoteOpenAIServer, assert score.data is not None assert len(score.data) == 1 assert score.data[0].score >= 0.9 + + +@pytest.mark.asyncio +@pytest.mark.parametrize("model_name", [MODEL_NAME]) +def test_score_max_model_len(server: RemoteOpenAIServer, model_name: str): + + text_1 = "What is the capital of France?" * 20 + text_2 = [ + "The capital of Brazil is Brasilia.", "The capital of France is Paris." + ] + + score_response = requests.post(server.url_for("score"), + json={ + "model": model_name, + "text_1": text_1, + "text_2": text_2, + }) + assert score_response.status_code == 400 + # Assert just a small fragments of the response + assert "Please reduce the length of the input." in \ + score_response.text + + # Test truncation + score_response = requests.post(server.url_for("score"), + json={ + "model": model_name, + "text_1": text_1, + "text_2": text_2, + "truncate_prompt_tokens": 101 + }) + assert score_response.status_code == 400 + assert "Please, select a smaller truncation size." in \ + score_response.text diff --git a/tests/entrypoints/openai/test_serving_chat.py b/tests/entrypoints/openai/test_serving_chat.py index 85f485364a411..e88d6c3c67829 100644 --- a/tests/entrypoints/openai/test_serving_chat.py +++ b/tests/entrypoints/openai/test_serving_chat.py @@ -103,6 +103,116 @@ def test_serving_chat_should_set_correct_max_tokens(): assert mock_engine.generate.call_args.args[1].max_tokens == 10 + # Setting server's max_tokens in the generation_config.json + # lower than context_window - prompt_tokens + mock_model_config = MockModelConfig() + mock_model_config.diff_sampling_param = { + "max_tokens": 10 # Setting server-side max_tokens limit + } + + # Reinitialize the engine with new settings + mock_engine = MagicMock(spec=MQLLMEngineClient) + mock_engine.get_tokenizer.return_value = get_tokenizer(MODEL_NAME) + mock_engine.errored = False + + # Initialize the serving chat + models = OpenAIServingModels(engine_client=mock_engine, + base_model_paths=BASE_MODEL_PATHS, + model_config=mock_model_config) + serving_chat = OpenAIServingChat(mock_engine, + mock_model_config, + models, + response_role="assistant", + chat_template=CHAT_TEMPLATE, + chat_template_content_format="auto", + request_logger=None) + + # Test Case 1: No max_tokens specified in request + req = ChatCompletionRequest( + model=MODEL_NAME, + messages=[{ + "role": "user", + "content": "what is 1+1?" + }], + guided_decoding_backend="outlines", + ) + + with suppress(Exception): + asyncio.run(serving_chat.create_chat_completion(req)) + + assert mock_engine.generate.call_args.args[1].max_tokens == 10 + + # Test Case 2: Request's max_tokens set higher than server accepts + req.max_tokens = 15 + + with suppress(Exception): + asyncio.run(serving_chat.create_chat_completion(req)) + + assert mock_engine.generate.call_args.args[1].max_tokens == 10 + + # Test Case 3: Request's max_tokens set lower than server accepts + req.max_tokens = 5 + + with suppress(Exception): + asyncio.run(serving_chat.create_chat_completion(req)) + + assert mock_engine.generate.call_args.args[1].max_tokens == 5 + + # Setting server's max_tokens in the generation_config.json + # higher than context_window - prompt_tokens + mock_model_config = MockModelConfig() + mock_model_config.diff_sampling_param = { + "max_tokens": 200 # Setting server-side max_tokens limit + } + + # Reinitialize the engine with new settings + mock_engine = MagicMock(spec=MQLLMEngineClient) + mock_engine.get_tokenizer.return_value = get_tokenizer(MODEL_NAME) + mock_engine.errored = False + + # Initialize the serving chat + models = OpenAIServingModels(engine_client=mock_engine, + base_model_paths=BASE_MODEL_PATHS, + model_config=mock_model_config) + serving_chat = OpenAIServingChat(mock_engine, + mock_model_config, + models, + response_role="assistant", + chat_template=CHAT_TEMPLATE, + chat_template_content_format="auto", + request_logger=None) + + # Test case 1: No max_tokens specified, defaults to context_window + req = ChatCompletionRequest( + model=MODEL_NAME, + messages=[{ + "role": "user", + "content": "what is 1+1?" + }], + guided_decoding_backend="outlines", + ) + + with suppress(Exception): + asyncio.run(serving_chat.create_chat_completion(req)) + + assert mock_engine.generate.call_args.args[1].max_tokens == 93 + + # Test Case 2: Request's max_tokens set higher than server accepts + req.max_tokens = 100 + + with suppress(Exception): + asyncio.run(serving_chat.create_chat_completion(req)) + + assert mock_engine.generate.call_args.args[1].max_tokens == 93 + + # Test Case 3: Request's max_tokens set lower than server accepts + req.max_tokens = 5 + + with suppress(Exception): + asyncio.run(serving_chat.create_chat_completion(req)) + + assert mock_engine.generate.call_args.args[1].max_tokens == 5 + def test_serving_chat_could_load_correct_generation_config(): diff --git a/tests/entrypoints/test_chat_utils.py b/tests/entrypoints/test_chat_utils.py index 8f242df4a60e3..513b466c10d60 100644 --- a/tests/entrypoints/test_chat_utils.py +++ b/tests/entrypoints/test_chat_utils.py @@ -754,6 +754,7 @@ def test_resolve_content_format_hf_defined(model, expected_format): ("template_chatglm.jinja", "string"), ("template_chatglm2.jinja", "string"), ("template_chatml.jinja", "string"), + ("template_deepseek_vl2.jinja", "string"), ("template_falcon_180b.jinja", "string"), ("template_falcon.jinja", "string"), ("template_inkbot.jinja", "string"), diff --git a/tests/kernels/test_activation.py b/tests/kernels/test_activation.py index a84501f9c303f..dac26efe866b8 100644 --- a/tests/kernels/test_activation.py +++ b/tests/kernels/test_activation.py @@ -6,8 +6,9 @@ from tests.kernels.utils import opcheck from vllm.model_executor.layers.activation import (FastGELU, FatreluAndMul, - GeluAndMul, NewGELU, - QuickGELU, SiluAndMul) + GeluAndMul, MulAndSilu, + NewGELU, QuickGELU, + SiluAndMul) from vllm.platforms import current_platform from .allclose_default import get_default_atol, get_default_rtol @@ -21,8 +22,9 @@ ] -@pytest.mark.parametrize("activation", - ["silu", "gelu", "gelu_tanh", "fatrelu"]) +@pytest.mark.parametrize( + "activation", + ["silu_and_mul", "mul_and_silu", "gelu", "gelu_tanh", "fatrelu"]) @pytest.mark.parametrize("num_tokens", NUM_TOKENS) @pytest.mark.parametrize("d", D) @pytest.mark.parametrize("dtype", DTYPES) @@ -40,9 +42,12 @@ def test_act_and_mul( current_platform.seed_everything(seed) torch.set_default_device(device) x = torch.randn(num_tokens, 2 * d, dtype=dtype) - if activation == "silu": + if activation == "silu_and_mul": layer = SiluAndMul() fn = torch.ops._C.silu_and_mul + if activation == "mul_and_silu": + layer = MulAndSilu() + fn = torch.ops._C.mul_and_silu elif activation == "gelu": layer = GeluAndMul(approximate="none") fn = torch.ops._C.gelu_and_mul @@ -55,8 +60,9 @@ def test_act_and_mul( fn = torch.ops._C.fatrelu_and_mul out = layer(x) ref_out = layer.forward_native(x) - # The SiLU, GELU and FatReLU implementations are equivalent to the native - # PyTorch implementations, so we can do exact comparison. + # The SiluAndMul, MulAndSilu, GELU and FatReLU implementations are + # equivalent to the native PyTorch implementations, so we can do exact + # comparison. torch.testing.assert_close(out, ref_out, atol=0.0, rtol=0.0) d = x.shape[-1] // 2 diff --git a/tests/kernels/test_attention.py b/tests/kernels/test_attention.py index 13c92dadcd4ed..16e52d15b0293 100644 --- a/tests/kernels/test_attention.py +++ b/tests/kernels/test_attention.py @@ -7,7 +7,7 @@ from tests.kernels.utils import opcheck from vllm import _custom_ops as ops from vllm.platforms import current_platform -from vllm.utils import get_max_shared_memory_bytes +from vllm.utils import get_max_shared_memory_bytes, is_navi from .allclose_default import get_default_atol, get_default_rtol @@ -31,9 +31,9 @@ NUM_PREFILL_SEQS = [3] # Arbitrary values for testing NUM_HEADS = [(40, 40), (64, 8)] # Arbitrary values for testing -# FlashAttention forward only supports head dimension at most 128 -# https://github.com/ROCmSoftwarePlatform/flash-attention/blob/3d2b6f5d037782cc2c906909a46fb7e2e1b48b25/csrc/flash_attn_rocm/flash_api.cpp#L62 -HEAD_SIZES = [64, 80, 120, 256] +# This should be sync with get_supported_head_sizes() in +# vllm.attention.ops.paged_attn.PagedAttention +HEAD_SIZES = [64, 80, 96, 112, 120, 128, 192, 256] BLOCK_SIZES = [16, 32] USE_ALIBI = [False, True] @@ -182,7 +182,11 @@ def test_paged_attention( key_cache, value_cache = key_caches[0], value_caches[0] # Using default kv_scale - k_scale = v_scale = 1.0 + k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device) + + # additional argument for v1/v2 pa kernel + num_threads = 1024 if current_platform.is_rocm() \ + and not is_navi() else 128 # Call the paged attention kernel. output = torch.empty_like(query) @@ -204,16 +208,16 @@ def test_paged_attention( v_scale, ) - opcheck(torch.ops._C.paged_attention_v1, - (output, query, key_cache, value_cache, num_kv_heads, scale, - block_tables, seq_lens, block_size, max_seq_len, alibi_slopes, - kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0), - cond=(head_size == HEAD_SIZES[0] - and block_size == BLOCK_SIZES[0])) + opcheck( + torch.ops._C.paged_attention_v1, + (output, query, key_cache, value_cache, num_kv_heads, scale, + block_tables, seq_lens, block_size, max_seq_len, alibi_slopes, + kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0, num_threads), + cond=(head_size == HEAD_SIZES[0] and block_size == BLOCK_SIZES[0])) elif version in ("v2", "rocm"): if current_platform.is_rocm(): - PARTITION_SIZE = 1024 if version == "v2" else 512 + PARTITION_SIZE = 1024 if version == "v2" else 256 num_partitions = ((max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE) assert PARTITION_SIZE % block_size == 0 num_seqs, num_heads, head_size = output.shape @@ -248,13 +252,14 @@ def test_paged_attention( v_scale, ) - opcheck(torch.ops._C.paged_attention_v2, - (output, exp_sums, max_logits, tmp_output, query, - key_cache, value_cache, num_kv_heads, scale, block_tables, - seq_lens, block_size, max_seq_len, alibi_slopes, - kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0), - cond=(head_size == HEAD_SIZES[0] - and block_size == BLOCK_SIZES[0])) + opcheck( + torch.ops._C.paged_attention_v2, + (output, exp_sums, max_logits, tmp_output, query, key_cache, + value_cache, num_kv_heads, scale, block_tables, seq_lens, + block_size, max_seq_len, alibi_slopes, kv_cache_dtype, + k_scale, v_scale, 0, 0, 0, 64, 0, num_threads), + cond=(head_size == HEAD_SIZES[0] + and block_size == BLOCK_SIZES[0])) else: ops.paged_attention_rocm( @@ -275,13 +280,15 @@ def test_paged_attention( kv_cache_dtype, k_scale, v_scale, + None, + PARTITION_SIZE, ) opcheck(torch.ops._rocm_C.paged_attention, (output, exp_sums, max_logits, tmp_output, query, key_cache, value_cache, num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len, alibi_slopes, - kv_cache_dtype, k_scale, v_scale), + kv_cache_dtype, k_scale, v_scale, None, PARTITION_SIZE), cond=(head_size == HEAD_SIZES[0] and block_size == BLOCK_SIZES[0])) @@ -433,4 +440,4 @@ def test_multi_query_kv_attention( ) atol = get_default_atol(output) if current_platform.is_rocm() else 1e-3 rtol = get_default_rtol(output) if current_platform.is_rocm() else 1e-5 - torch.testing.assert_close(output, ref_output, atol=atol, rtol=rtol) + torch.testing.assert_close(output, ref_output, atol=atol, rtol=rtol) \ No newline at end of file diff --git a/tests/kernels/test_attention_selector.py b/tests/kernels/test_attention_selector.py index a08c874407e3f..492acb91e8ed9 100644 --- a/tests/kernels/test_attention_selector.py +++ b/tests/kernels/test_attention_selector.py @@ -94,7 +94,12 @@ def test_flash_attn(monkeypatch): def test_invalid_env(monkeypatch): - """Throw an exception if the backend name is invalid.""" + """Ignore the invalid env variable if it is set.""" override_backend_env_variable(monkeypatch, STR_INVALID_VAL) - with pytest.raises(ValueError): - get_attn_backend(16, torch.float16, None, 16, False) + with patch("vllm.attention.selector.current_platform", CudaPlatform()): + backend = get_attn_backend(32, torch.float16, None, 16, False) + assert backend.get_name() == "FLASH_ATTN" + + # when block size == 16, backend will fall back to XFORMERS + backend = get_attn_backend(16, torch.float16, None, 16, False) + assert backend.get_name() == "XFORMERS" diff --git a/tests/kernels/test_blocksparse_attention.py b/tests/kernels/test_blocksparse_attention.py index 1667e53c0749e..2c08e928f9c21 100644 --- a/tests/kernels/test_blocksparse_attention.py +++ b/tests/kernels/test_blocksparse_attention.py @@ -210,7 +210,7 @@ def test_paged_attention( key_cache, value_cache = key_caches[0], value_caches[0] # Using default kv_scale - k_scale = v_scale = 1.0 + k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device) tp_rank = 0 # Call the paged attention kernel. diff --git a/tests/kernels/test_cache.py b/tests/kernels/test_cache.py index 40550ed51e2c7..c848be4f9d807 100644 --- a/tests/kernels/test_cache.py +++ b/tests/kernels/test_cache.py @@ -160,7 +160,7 @@ def test_reshape_and_cache( cloned_value_cache = value_cache.clone() # Using default kv_scale - k_scale = v_scale = 1.0 + k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device) # Call the reshape_and_cache kernel. opcheck(torch.ops._C_cache_ops.reshape_and_cache, @@ -258,8 +258,8 @@ def test_reshape_and_cache_flash( del key_caches del value_caches - k_scale = key.amax().item() / 256 - v_scale = value.amax().item() / 256 + k_scale = (key.amax() / 256.0).to(torch.float32) + v_scale = (value.amax() / 256.0).to(torch.float32) # Clone the KV caches. if kv_cache_dtype == "fp8": @@ -284,12 +284,12 @@ def test_reshape_and_cache_flash( result_key_cache = torch.empty_like(key_cache, dtype=torch.float16) ops.convert_fp8(result_key_cache, key_cache, - k_scale, + k_scale.item(), kv_dtype=kv_cache_dtype) result_value_cache = torch.empty_like(value_cache, dtype=torch.float16) ops.convert_fp8(result_value_cache, value_cache, - v_scale, + v_scale.item(), kv_dtype=kv_cache_dtype) # Run the reference implementation. diff --git a/tests/kernels/test_cascade_flash_attn.py b/tests/kernels/test_cascade_flash_attn.py index 45ec6df4e711e..8edfde42ede74 100644 --- a/tests/kernels/test_cascade_flash_attn.py +++ b/tests/kernels/test_cascade_flash_attn.py @@ -6,7 +6,9 @@ from vllm.platforms import current_platform from vllm.v1.attention.backends.flash_attn import (cascade_attention, merge_attn_states) -from vllm.vllm_flash_attn import flash_attn_varlen_func +from vllm.vllm_flash_attn import (fa_version_unsupported_reason, + flash_attn_varlen_func, + is_fa_version_supported) NUM_HEADS = [(4, 4), (8, 2), (16, 2)] HEAD_SIZES = [128, 192, 256] @@ -78,6 +80,7 @@ def test_merge_kernel( @pytest.mark.parametrize("block_size", BLOCK_SIZES) @pytest.mark.parametrize("soft_cap", [None, 50]) @pytest.mark.parametrize("num_blocks", [2048]) +@pytest.mark.parametrize("fa_version", [2, 3]) @torch.inference_mode() def test_cascade( seq_lens_and_common_prefix: Tuple[List[Tuple[int, int]], int], @@ -87,8 +90,13 @@ def test_cascade( block_size: int, soft_cap: Optional[float], num_blocks: int, + fa_version: int, ) -> None: torch.set_default_device("cuda") + if not is_fa_version_supported(fa_version): + pytest.skip(f"Flash attention version {fa_version} not supported due " + f"to: \"{fa_version_unsupported_reason(fa_version)}\"") + current_platform.seed_everything(0) window_size = (-1, -1) @@ -118,9 +126,7 @@ def test_cascade( cu_query_lens = torch.tensor([0] + query_lens, dtype=torch.int32).cumsum(dim=0, dtype=torch.int32) - cu_kv_lens = torch.tensor([0] + kv_lens, - dtype=torch.int32).cumsum(dim=0, - dtype=torch.int32) + kv_lens_tensor = torch.tensor(kv_lens, dtype=torch.int32) max_num_blocks_per_seq = (max_kv_len + block_size - 1) // block_size block_tables = torch.randint(0, num_blocks, @@ -140,7 +146,7 @@ def test_cascade( k=key_cache, v=value_cache, cu_seqlens_q=cu_query_lens, - cu_seqlens_k=cu_kv_lens, + seqused_k=kv_lens_tensor, max_seqlen_q=max_query_len, max_seqlen_k=max_kv_len, softmax_scale=scale, @@ -154,10 +160,8 @@ def test_cascade( assert all(common_prefix_len < kv_len for kv_len in kv_lens) cu_prefix_query_lens = torch.tensor([0, total_num_query_tokens], dtype=torch.int32) - cu_prefix_kv_lens = torch.tensor([0, common_prefix_len], dtype=torch.int32) - cu_suffix_kv_lens = ( - cu_kv_lens - - torch.arange(num_seqs + 1, dtype=torch.int32) * common_prefix_len) + prefix_kv_lens = torch.tensor([common_prefix_len], dtype=torch.int32) + suffix_kv_lens = kv_lens_tensor - common_prefix_len output = torch.empty_like(query) cascade_attention( output=output, @@ -167,8 +171,8 @@ def test_cascade( cu_query_lens=cu_query_lens, max_query_len=max_query_len, cu_prefix_query_lens=cu_prefix_query_lens, - cu_prefix_kv_lens=cu_prefix_kv_lens, - cu_suffix_kv_lens=cu_suffix_kv_lens, + prefix_kv_lens=prefix_kv_lens, + suffix_kv_lens=suffix_kv_lens, max_kv_len=max_kv_len, softmax_scale=scale, alibi_slopes=None, @@ -176,6 +180,7 @@ def test_cascade( logits_soft_cap=soft_cap if soft_cap is not None else 0, block_table=block_tables, common_prefix_len=common_prefix_len, + fa_version=fa_version, ) # Compare the results. diff --git a/tests/kernels/test_cutlass.py b/tests/kernels/test_cutlass.py index afe53797322f9..c3eddacec2727 100644 --- a/tests/kernels/test_cutlass.py +++ b/tests/kernels/test_cutlass.py @@ -2,7 +2,7 @@ Run `pytest tests/kernels/test_cutlass.py`. """ -from typing import Optional, Type +from typing import Type import pytest import torch @@ -11,6 +11,8 @@ from vllm import _custom_ops as ops from vllm.platforms import current_platform +from .utils import baseline_scaled_mm, to_fp8, to_int8 + MNK_FACTORS = [ (1, 256, 128), (1, 16384, 1024), @@ -41,34 +43,10 @@ capability = capability[0] * 10 + capability[1] -def to_fp8(tensor: torch.Tensor): - finfo = torch.finfo(torch.float8_e4m3fn) - return torch.round(tensor.clamp( - min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn) - - -def to_int8(tensor: torch.Tensor): - return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8) - - def rand_int8(shape: tuple, device: str = "cuda"): return to_int8(torch.rand(shape, device=device) * 255 - 128) -def baseline_scaled_mm(a: torch.Tensor, - b: torch.Tensor, - scale_a: torch.Tensor, - scale_b: torch.Tensor, - out_dtype: Type[torch.dtype], - bias: Optional[torch.Tensor] = None) -> torch.Tensor: - output = (scale_a * (scale_b * (torch.mm( - a.to(dtype=torch.float32), b.to(dtype=torch.float32))))).to(out_dtype) - if bias is not None: - output = output + bias - - return output - - def cutlass_fp8_gemm_helper(m: int, n: int, k: int, diff --git a/tests/kernels/test_cutlass_2of4_sparse.py b/tests/kernels/test_cutlass_2of4_sparse.py new file mode 100644 index 0000000000000..56495df34aa6c --- /dev/null +++ b/tests/kernels/test_cutlass_2of4_sparse.py @@ -0,0 +1,214 @@ +"""Tests for sparse cutlass kernels + +Run `pytest tests/kernels/test_semi_structured.py`. +""" +from typing import Tuple, Type + +import pytest +import torch +import torch.nn.functional as F + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.utils.w8a8_utils import ( + sparse_cutlass_supported) +from vllm.platforms import current_platform + +from .utils import baseline_scaled_mm, to_fp8, to_int8 + +CUDA_DEVICES = [ + f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2) +] + +capability = current_platform.get_device_capability() +capability = capability[0] * 10 + capability[1] + + +def to_bf16(tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(dtype=torch.bfloat16) + + +def to_fp16(tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(dtype=torch.float16) + + +def prune_to_2_4(tensor): + # Reshape tensor to [N, 4] where N is number of groups of 4 + original_shape = tensor.shape + reshaped = tensor.reshape(-1, 4) + + # Get indices of top 2 absolute values in each group of 4 + _, indices = torch.topk(torch.abs(reshaped), k=2, dim=1) + + # Create binary mask + mask = torch.zeros_like(reshaped) + mask.scatter_(dim=1, + index=indices, + src=torch.ones_like(indices, dtype=mask.dtype)) + + # Apply mask and reshape back + pruned = reshaped * mask + + # Turn all -0.0 to 0.0 + pruned[pruned == -0.0] = 0.0 + + return pruned.reshape(original_shape) + + +def make_rand_sparse_tensors( + dtype: torch.dtype, m: int, n: int, k: int +) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + a = torch.randn((m, k), device='cuda') * 5 + b = torch.randn((n, k), device='cuda').t() * 5 + + b = prune_to_2_4(b.t()).t() + + if dtype == torch.int8: + a, b = to_int8(a), to_int8(b) + elif dtype == torch.float8_e4m3fn: + a, b = to_fp8(a), to_fp8(b) + elif dtype == torch.float16: + a, b = to_fp16(a), to_fp16(b) + elif dtype == torch.bfloat16: + a, b = to_bf16(a), to_bf16(b) + else: + raise ValueError("unsupported dtype") + + b_compressed, e = ops.cutlass_sparse_compress(b.t()) + + # Compressed B, Metadata, Original A, B + return b_compressed, e, a, b + + +@pytest.mark.skipif(not sparse_cutlass_supported(), + reason="Sparse CUTLASS is not supported on this GPU type.") +# Test working with a subset of A and B for sparse matmul +def test_cutlass_sparse_subset(): + + big_m = 1024 + m, n, k = 512, 512, 512 + + # Create tensors + b_comp, e, whole_a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, + big_m, n, k) + a = whole_a[0:m, 0:k] + scale_a = torch.randn((1, 1), device="cuda", dtype=torch.float32) / 10 + scale_b = torch.randn((1, 1), device="cuda", dtype=torch.float32) / 10 + + out = ops.cutlass_scaled_sparse_mm(a, + b_comp, + e, + scale_a, + scale_b, + out_dtype=torch.bfloat16) + baseline = baseline_scaled_mm(a, + b, + scale_a, + scale_b, + out_dtype=torch.bfloat16) + + torch.testing.assert_close(out, baseline, rtol=1e-1, atol=1e0) + + +MNK_FACTORS = [ + (1, 256, 128), + (1, 16384, 1024), + (1, 24576, 512), + (16, 256, 512), + (16, 16384, 128), + (16, 24576, 4096), + (32, 8192, 4096), + (32, 16384, 4096), + (33, 1024, 1024), + (33, 8192, 128), + (64, 2048, 512), + (64, 16384, 1024), + (100, 8192, 512), + (128, 32768, 4096), + (256, 4096, 4096), + (512, 256, 1024), + (512, 8192, 4096), + (512, 16384, 128), + (512, 24576, 128), +] + + +# Test working with a subset of A and B for sparse matmul +@pytest.mark.skip(reason="2of4 sparse w16a16 CUTLASS produces bad output.") +@pytest.mark.skipif(not sparse_cutlass_supported(), + reason="Sparse CUTLASS is not supported on this GPU type.") +@pytest.mark.parametrize("m, k, n", MNK_FACTORS) +@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16]) +def test_cutlass_sparse_gemm(m: int, k: int, n: int, dtype: Type[torch.dtype]): + + # Create tensors + b_comp, e, a, b = make_rand_sparse_tensors(dtype, m, n, k) + scale_a = torch.ones((1, 1), device="cuda", dtype=torch.float32) + scale_b = torch.ones((1, 1), device="cuda", dtype=torch.float32) + + out = ops.cutlass_scaled_sparse_mm(a, + b_comp, + e, + scale_a, + scale_b, + out_dtype=dtype) + baseline = F.linear(a, b.T) + + torch.testing.assert_close(out, baseline, rtol=1e-2, atol=1e-2) + + +@pytest.mark.skipif(not sparse_cutlass_supported(), + reason="Sparse CUTLASS is not supported on this GPU type.") +@pytest.mark.parametrize("m, k, n", MNK_FACTORS) +@pytest.mark.skipif(not current_platform.has_device_capability(89), + reason="FP8 is not supported on this GPU type.") +def test_cutlass_sparse_fp8_gemm(m: int, n: int, k: int): + + # Create tensors + b_comp, e, a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, m, n, k) + scale_a = (torch.randn((1, 1), device="cuda", dtype=torch.float32)) + scale_b = (torch.randn((1, 1), device="cuda", dtype=torch.float32)) + + out = ops.cutlass_scaled_sparse_mm(a, + b_comp, + e, + scale_a, + scale_b, + out_dtype=torch.bfloat16) + + baseline = baseline_scaled_mm(a, + b, + scale_a, + scale_b, + out_dtype=torch.bfloat16) + + torch.testing.assert_close(out, baseline, rtol=1e0, atol=2e0) + + +@pytest.mark.skipif(not sparse_cutlass_supported(), + reason="Sparse CUTLASS is not supported on this GPU type.") +@pytest.mark.parametrize("m,k,n", MNK_FACTORS) +@pytest.mark.parametrize("per_act_token", [True, False]) +@pytest.mark.parametrize("per_out_ch", [True, False]) +@pytest.mark.parametrize("use_bias", [True, False]) +def test_cutlass_sparse_int8_gemm(m: int, n: int, k: int, per_act_token: bool, + per_out_ch: bool, use_bias: bool): + + # Create tensors + b_comp, e, a, b = make_rand_sparse_tensors(torch.int8, m, n, k) + scale_a = (torch.randn((1, 1), device="cuda", dtype=torch.float32)) + scale_b = (torch.randn((1, 1), device="cuda", dtype=torch.float32)) + + out = ops.cutlass_scaled_sparse_mm(a, + b_comp, + e, + scale_a, + scale_b, + out_dtype=torch.bfloat16) + + baseline = baseline_scaled_mm(a, + b, + scale_a, + scale_b, + out_dtype=torch.bfloat16) + + torch.testing.assert_close(out, baseline, rtol=1e0, atol=2e0) diff --git a/tests/kernels/test_flash_attn.py b/tests/kernels/test_flash_attn.py index 1ae78d7b46c5b..0ee0bf6c6a374 100644 --- a/tests/kernels/test_flash_attn.py +++ b/tests/kernels/test_flash_attn.py @@ -4,8 +4,10 @@ import torch from vllm.platforms import current_platform -from vllm.vllm_flash_attn import (flash_attn_varlen_func, - flash_attn_with_kvcache) +from vllm.vllm_flash_attn import (fa_version_unsupported_reason, + flash_attn_varlen_func, + flash_attn_with_kvcache, + is_fa_version_supported) NUM_HEADS = [(4, 4), (8, 2), (16, 2)] HEAD_SIZES = [128, 256] @@ -80,6 +82,7 @@ def ref_paged_attn( @pytest.mark.parametrize("soft_cap", [None, 10.0, 50.0]) @pytest.mark.parametrize("num_blocks", NUM_BLOCKS) @pytest.mark.parametrize("sliding_window", [None, 256]) +@pytest.mark.parametrize("fa_version", [2, 3]) @torch.inference_mode() def test_flash_attn_with_paged_kv( use_out: bool, @@ -91,8 +94,13 @@ def test_flash_attn_with_paged_kv( soft_cap: Optional[float], num_blocks: int, sliding_window: Optional[int], + fa_version: int, ) -> None: torch.set_default_device("cuda") + if not is_fa_version_supported(fa_version): + pytest.skip(f"Flash attention version {fa_version} not supported due " + f"to: \"{fa_version_unsupported_reason(fa_version)}\"") + current_platform.seed_everything(0) num_seqs = len(kv_lens) num_query_heads = num_heads[0] @@ -131,6 +139,7 @@ def test_flash_attn_with_paged_kv( cache_seqlens=kv_lens_tensor, softcap=soft_cap if soft_cap is not None else 0, window_size=window_size, + fa_version=fa_version, ) output = output if not use_out else out output = output.squeeze(1) @@ -159,6 +168,7 @@ def test_flash_attn_with_paged_kv( @pytest.mark.parametrize("dtype", DTYPES) @pytest.mark.parametrize("soft_cap", [None, 10.0, 50.0]) @pytest.mark.parametrize("num_blocks", NUM_BLOCKS) +@pytest.mark.parametrize("fa_version", [2, 3]) @torch.inference_mode() def test_varlen_with_paged_kv( use_out: bool, @@ -170,8 +180,12 @@ def test_varlen_with_paged_kv( block_size: int, soft_cap: Optional[float], num_blocks: int, + fa_version: int, ) -> None: torch.set_default_device("cuda") + if not is_fa_version_supported(fa_version): + pytest.skip(f"Flash attention version {fa_version} not supported due " + f"to: \"{fa_version_unsupported_reason(fa_version)}\"") current_platform.seed_everything(0) num_seqs = len(seq_lens) query_lens = [x[0] for x in seq_lens] @@ -198,9 +212,7 @@ def test_varlen_with_paged_kv( cu_query_lens = torch.tensor([0] + query_lens, dtype=torch.int32).cumsum(dim=0, dtype=torch.int32) - cu_kv_lens = torch.tensor([0] + kv_lens, - dtype=torch.int32).cumsum(dim=0, - dtype=torch.int32) + kv_lens = torch.tensor(kv_lens, dtype=torch.int32) max_num_blocks_per_seq = (max_kv_len + block_size - 1) // block_size block_tables = torch.randint(0, @@ -215,7 +227,7 @@ def test_varlen_with_paged_kv( v=value_cache, out=out, cu_seqlens_q=cu_query_lens, - cu_seqlens_k=cu_kv_lens, + seqused_k=kv_lens, max_seqlen_q=max_query_len, max_seqlen_k=max_kv_len, softmax_scale=scale, @@ -223,6 +235,7 @@ def test_varlen_with_paged_kv( window_size=window_size, block_table=block_tables, softcap=soft_cap if soft_cap is not None else 0, + fa_version=fa_version, ) output = output if not use_out else out diff --git a/tests/kernels/test_flashinfer.py b/tests/kernels/test_flashinfer.py index a2c8f71665737..1645ef911d697 100644 --- a/tests/kernels/test_flashinfer.py +++ b/tests/kernels/test_flashinfer.py @@ -133,17 +133,19 @@ def test_flashinfer_decode_with_paged_kv( use_tensor_cores=( (num_query_heads//num_kv_heads) > 4) ) - wrapper.begin_forward(kv_indptr, - kv_indices, - kv_last_page_lens, - num_query_heads, - num_kv_heads, - head_size, - block_size, - "NONE", - data_type=dtype) - - output = wrapper.forward(query, key_value_cache, logits_soft_cap=soft_cap) + wrapper.plan(kv_indptr, + kv_indices, + kv_last_page_lens, + num_query_heads, + num_kv_heads, + head_size, + block_size, + "NONE", + q_data_type=dtype, + kv_data_type=dtype, + logits_soft_cap=soft_cap) + + output = wrapper.run(query, key_value_cache) ref_output = ref_paged_attn(query=query, key_cache=key_cache, @@ -228,7 +230,7 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]], workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.int8) wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper( workspace_buffer, "NHD") - wrapper.begin_forward( + wrapper.plan( qo_indptr, kv_indptr, kv_indices, @@ -237,12 +239,14 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]], num_kv_heads, head_size, block_size, + q_data_type=dtype, + kv_data_type=dtype, + logits_soft_cap=soft_cap, ) - output = wrapper.forward( + output = wrapper.run( query, key_value_cache, - logits_soft_cap=soft_cap, ) ref_output = ref_paged_attn(query=query, @@ -253,7 +257,7 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]], block_tables=block_tables, scale=scale, soft_cap=soft_cap) - torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2), \ + torch.testing.assert_close(output, ref_output, atol=5e-2, rtol=1e-2), \ f"{torch.max(torch.abs(output - ref_output))}" @@ -332,7 +336,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv( workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.int8) wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper( workspace_buffer, "NHD") - wrapper.begin_forward( + wrapper.plan( qo_indptr, kv_indptr, kv_indices, @@ -341,13 +345,12 @@ def test_flashinfer_prefill_with_paged_fp8_kv( num_kv_heads, head_size, block_size, + q_data_type=dtype, + kv_data_type=kv_cache_dtype, + logits_soft_cap=soft_cap, ) - output = wrapper.forward(query, - kv_cache_fp8, - logits_soft_cap=soft_cap, - k_scale=k_scale, - v_scale=v_scale) + output = wrapper.run(query, kv_cache_fp8, k_scale=k_scale, v_scale=v_scale) ref_output = ref_paged_attn(query=query, key_cache=key_cache.squeeze(1), @@ -360,7 +363,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv( del query del block_tables # verify prefill fp8 - torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2), \ + torch.testing.assert_close(output, ref_output, atol=5e-2, rtol=1e-2), \ f"{torch.max(torch.abs(output - ref_output))}" @@ -439,21 +442,18 @@ def test_flashinfer_decode_with_paged_fp8_kv( wrapper = flashinfer.\ BatchDecodeWithPagedKVCacheWrapper(workspace_buffer, "NHD", use_tensor_cores=use_tensor_cores) - wrapper.begin_forward(kv_indptr, - kv_indices, - kv_last_page_lens, - num_query_heads, - num_kv_heads, - head_size, - block_size, - "NONE", - data_type=dtype, - q_data_type=dtype) - output = wrapper.forward(query, - kv_cache_fp8, - logits_soft_cap=soft_cap, - k_scale=k_scale, - v_scale=v_scale) + wrapper.plan(kv_indptr, + kv_indices, + kv_last_page_lens, + num_query_heads, + num_kv_heads, + head_size, + block_size, + "NONE", + q_data_type=dtype, + kv_data_type=kv_cache_dtype, + logits_soft_cap=soft_cap) + output = wrapper.run(query, kv_cache_fp8, k_scale=k_scale, v_scale=v_scale) key_cache = key_value_cache[:, 0, :, :, :].squeeze(1) value_cache = key_value_cache[:, 1, :, :, :].squeeze(1) diff --git a/tests/kernels/test_mha_attn.py b/tests/kernels/test_mha_attn.py new file mode 100644 index 0000000000000..eab874e9e02bb --- /dev/null +++ b/tests/kernels/test_mha_attn.py @@ -0,0 +1,126 @@ +""" +Test: + +* Tests for MultiHeadAttention layer +""" +from unittest.mock import patch + +import pytest +import torch + +from vllm.attention.layer import MultiHeadAttention +from vllm.attention.selector import _Backend, _cached_get_attn_backend +from vllm.platforms import current_platform +from vllm.platforms.cpu import CpuPlatform +from vllm.platforms.cuda import CudaPlatform +from vllm.platforms.rocm import RocmPlatform + + +@pytest.fixture(autouse=True) +def clear_cache(): + """Clear lru cache to ensure each test case runs without caching. + """ + _cached_get_attn_backend.cache_clear() + + +@pytest.mark.parametrize("device", ["cpu", "hip", "cuda"]) +def test_mha_attn_platform(device: str): + """ + Test the attention selector between different platform and device. + """ + torch.set_default_dtype(torch.float16) + + if device == "cpu": + with patch("vllm.attention.selector.current_platform", CpuPlatform()): + attn = MultiHeadAttention(16, 64, scale=1) + assert attn.attn_backend == _Backend.TORCH_SDPA + elif device == "hip": + with patch("vllm.attention.selector.current_platform", RocmPlatform()): + attn = MultiHeadAttention(16, 64, scale=1) + assert attn.attn_backend == _Backend.TORCH_SDPA + else: + with patch("vllm.attention.selector.current_platform", CudaPlatform()): + attn = MultiHeadAttention(16, 64, scale=1) + assert attn.attn_backend == _Backend.XFORMERS + + with patch("vllm.attention.selector.current_platform", CudaPlatform()): + attn = MultiHeadAttention(16, 72, scale=1) + assert attn.attn_backend == _Backend.XFORMERS + + +def ref_attention( + query: torch.Tensor, + key: torch.Tensor, + value: torch.Tensor, + scale: float, +) -> torch.Tensor: + """ + Native implementation of scaled dot product attention without mask: + - query, key, value: [batch_size, seq_len, num_heads, head_size] + - attn_mask: [batch_size, seq_len, seq_len] + """ + query, key, value = (x.transpose(1, 2) for x in (query, key, value)) + attn_weights = scale * torch.matmul(query, key.transpose(2, 3)) + attn_weights = torch.softmax(attn_weights, dim=-1).to(value.dtype) + out = torch.matmul(attn_weights, value).transpose(1, 2) + return out + + +BATCH_SIZES = [1, 16] +SEQ_LENS = [1] +NUM_HEADS = [1, 16] +NUM_KV_HEADS = [1] +HEAD_SIZES = [64, 80] +# flshattF and tritonflashattF supported: {torch.float16, torch.bfloat16} +DTYPES = [ + torch.half, torch.bfloat16, torch.float +] if not current_platform.is_rocm() else [torch.half, torch.bfloat16] +CUDA_DEVICES = ["cuda"] + + +@pytest.mark.parametrize("batch_size", BATCH_SIZES) +@pytest.mark.parametrize("seq_len", SEQ_LENS) +@pytest.mark.parametrize("num_heads", NUM_HEADS) +@pytest.mark.parametrize("num_kv_heads", NUM_KV_HEADS) +@pytest.mark.parametrize("head_size", HEAD_SIZES) +@pytest.mark.parametrize("dtype", DTYPES) +@pytest.mark.parametrize("device", CUDA_DEVICES) +def test_mha_attn_forward( + batch_size: int, + seq_len: int, + num_heads: int, + num_kv_heads: int, + head_size: int, + dtype: torch.dtype, + device: str, +): + current_platform.seed_everything(0) + torch.set_default_device(device) + torch.set_default_dtype(dtype) + + q = torch.randn(batch_size, seq_len, num_heads * head_size) + k = torch.randn(batch_size, seq_len, num_kv_heads * head_size) + v = torch.randn(batch_size, seq_len, num_kv_heads * head_size) + scale = 1.0 / head_size**0.5 + attn = MultiHeadAttention(num_heads, + head_size, + scale=scale, + num_kv_heads=num_kv_heads) + output = attn(q, k, v) + + assert num_heads % num_kv_heads == 0 + num_queries_per_kv = num_heads // num_kv_heads + q = q.reshape(batch_size, seq_len, num_heads, head_size) + k = k.reshape(batch_size, seq_len, num_kv_heads, head_size) + v = v.reshape(batch_size, seq_len, num_kv_heads, head_size) + if num_queries_per_kv > 1: + k = torch.repeat_interleave(k, num_queries_per_kv, dim=2) + v = torch.repeat_interleave(v, num_queries_per_kv, dim=2) + + ref_output = ref_attention( + q, + k, + v, + scale=scale, + ).reshape(batch_size, seq_len, num_heads * head_size) + torch.testing.assert_close(output, ref_output) diff --git a/tests/kernels/test_prefix_prefill.py b/tests/kernels/test_prefix_prefill.py index 3fdb7996ba4e0..10e73ab950b0e 100644 --- a/tests/kernels/test_prefix_prefill.py +++ b/tests/kernels/test_prefix_prefill.py @@ -138,6 +138,7 @@ def test_contexted_kv_attention( # to V_cache[num_blocks, num_kv_heads, head_size, block_size] v_cache = v_cache.view(-1, block_size, num_kv_heads, head_size).permute(0, 2, 3, 1).contiguous() + k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device) # Warm up the Triton kernel by calling it once before actually measuring # generation time @@ -153,6 +154,8 @@ def test_contexted_kv_attention( b_seq_len, b_ctx_len, max_input_len, + k_scale, + v_scale, sliding_window=sliding_window) torch.cuda.synchronize() start_time = time.time() @@ -168,6 +171,8 @@ def test_contexted_kv_attention( b_seq_len, b_ctx_len, max_input_len, + k_scale, + v_scale, sliding_window=sliding_window) torch.cuda.synchronize() end_time = time.time() @@ -366,6 +371,7 @@ def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor: # to V_cache[num_blocks, num_kv_heads, head_size, block_size] v_cache = v_cache.view(-1, block_size, num_kv_heads, head_size).permute(0, 2, 3, 1).contiguous() + k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device) # Warm up the Triton kernel by calling it once before actually measuring # generation time @@ -381,6 +387,8 @@ def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor: b_seq_len, b_ctx_len, max_input_len, + k_scale, + v_scale, alibi_slopes=alibi_slopes) torch.cuda.synchronize() start_time = time.time() @@ -396,6 +404,8 @@ def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor: b_seq_len, b_ctx_len, max_input_len, + k_scale, + v_scale, alibi_slopes=alibi_slopes) torch.cuda.synchronize() end_time = time.time() diff --git a/tests/kernels/test_triton_scaled_mm.py b/tests/kernels/test_triton_scaled_mm.py index 8e96a2f70d751..a5aab3c2ea4b0 100644 --- a/tests/kernels/test_triton_scaled_mm.py +++ b/tests/kernels/test_triton_scaled_mm.py @@ -39,6 +39,23 @@ def get_8bit_types(): return types +# This test is to check regressions for int8 support on ROCm. +@pytest.mark.parametrize("model_path", [ + "neuralmagic/Llama-3.2-1B-quantized.w8a8", +]) +@pytest.mark.parametrize("max_tokens", [32]) +@pytest.mark.parametrize("num_logprobs", [10]) +@pytest.mark.skipif(not current_platform.is_rocm(), + reason="Should only run on ROCm") +def test_rocm_compressed_tensors_w8a8(vllm_runner, example_prompts, model_path, + max_tokens, num_logprobs): + dtype = "bfloat16" + + with vllm_runner(model_path, dtype=dtype) as vllm_model: + vllm_model.generate_greedy_logprobs(example_prompts, max_tokens, + num_logprobs) + + @pytest.mark.parametrize("M", [1, 33, 64, 512]) @pytest.mark.parametrize("N", [256, 971, 20486]) @pytest.mark.parametrize("K", [128, 496, 1024]) diff --git a/tests/kernels/utils.py b/tests/kernels/utils.py index c9f29014fddcf..09ab818ebd52b 100644 --- a/tests/kernels/utils.py +++ b/tests/kernels/utils.py @@ -5,7 +5,7 @@ import unittest from numbers import Number from typing import (Any, Dict, List, NamedTuple, Optional, Sequence, Tuple, - Union) + Type, Union) import pytest import torch @@ -1107,3 +1107,28 @@ def opcheck(op: Union[torch._ops.OpOverload, torch._ops.OpOverloadPacket, kwargs, test_utils=test_utils, raise_exception=raise_exception) if cond else {} + + +# For testing quantized linear kernels +def to_fp8(tensor: torch.Tensor): + finfo = torch.finfo(torch.float8_e4m3fn) + return torch.round(tensor.clamp( + min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn) + + +def to_int8(tensor: torch.Tensor): + return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8) + + +def baseline_scaled_mm(a: torch.Tensor, + b: torch.Tensor, + scale_a: torch.Tensor, + scale_b: torch.Tensor, + out_dtype: Type[torch.dtype], + bias: Optional[torch.Tensor] = None) -> torch.Tensor: + output = (scale_a * (scale_b * (torch.mm( + a.to(dtype=torch.float32), b.to(dtype=torch.float32))))).to(out_dtype) + if bias is not None: + output = output + bias + + return output diff --git a/tests/lora/test_lora_checkpoints.py b/tests/lora/test_lora_checkpoints.py index 537d95b025a9d..b907af47d08d7 100644 --- a/tests/lora/test_lora_checkpoints.py +++ b/tests/lora/test_lora_checkpoints.py @@ -3,6 +3,7 @@ import pytest from vllm.lora.models import LoRAModel +from vllm.lora.peft_helper import PEFTHelper from vllm.model_executor.models.baichuan import BaiChuanBaseForCausalLM from vllm.model_executor.models.utils import WeightsMapper @@ -30,11 +31,14 @@ def test_load_checkpoints( else: expected_lora_modules.append(module) if lora_name == "baichuan7B": + peft_helper = PEFTHelper.from_local_dir(baichuan_lora_files, + max_position_embeddings=4096) # For the baichuan7B model, load it's LoRA, # and the test should pass. LoRAModel.from_local_checkpoint( baichuan_lora_files, expected_lora_modules, + peft_helper=peft_helper, lora_model_id=1, device="cpu", embedding_modules=embedding_modules, @@ -43,9 +47,12 @@ def test_load_checkpoints( # Test that the target_modules contain prefix # such as "model.layers.0.self_atten.W_pack", and # the test should pass. + peft_helper = PEFTHelper.from_local_dir(baichuan_zero_lora_files, + max_position_embeddings=4096) LoRAModel.from_local_checkpoint( baichuan_zero_lora_files, expected_lora_modules, + peft_helper=peft_helper, lora_model_id=1, device="cpu", embedding_modules=embedding_modules, @@ -53,9 +60,12 @@ def test_load_checkpoints( elif lora_name == "baichuan7B-zero-regex": # Test that the `target_modules` in the form of regular expressions, # such as `model\\..*(W_pack|o_proj)`, and the test should pass. + peft_helper = PEFTHelper.from_local_dir(baichuan_regex_lora_files, + max_position_embeddings=4096) LoRAModel.from_local_checkpoint( baichuan_regex_lora_files, expected_lora_modules, + peft_helper=peft_helper, lora_model_id=1, device="cpu", embedding_modules=embedding_modules, @@ -64,10 +74,13 @@ def test_load_checkpoints( # For the baichuan7B model, load chatglm3-6b's LoRA, # and the test should raise the following error. expected_error = "Please verify that the loaded LoRA module is correct" # noqa: E501 + peft_helper = PEFTHelper.from_local_dir(chatglm3_lora_files, + max_position_embeddings=4096) with pytest.raises(ValueError, match=expected_error): LoRAModel.from_local_checkpoint( chatglm3_lora_files, expected_lora_modules, + peft_helper=peft_helper, lora_model_id=1, device="cpu", embedding_modules=embedding_modules, @@ -94,9 +107,12 @@ def test_lora_weights_mapping(baichuan_lora_files): ".layers.": ".baichuan_layers.", }, ) + peft_helper = PEFTHelper.from_local_dir(baichuan_lora_files, + max_position_embeddings=4096) lora_model = LoRAModel.from_local_checkpoint( baichuan_lora_files, expected_lora_modules, + peft_helper=peft_helper, lora_model_id=1, device="cpu", embedding_modules=embedding_modules, diff --git a/tests/lora/test_lora_huggingface.py b/tests/lora/test_lora_huggingface.py index e2daf9d135113..1c0ee01c038d0 100644 --- a/tests/lora/test_lora_huggingface.py +++ b/tests/lora/test_lora_huggingface.py @@ -3,6 +3,7 @@ import pytest from vllm.lora.models import LoRAModel +from vllm.lora.peft_helper import PEFTHelper from vllm.lora.utils import get_adapter_absolute_path from vllm.model_executor.models.llama import LlamaForCausalLM @@ -27,9 +28,11 @@ def test_load_checkpoints_from_huggingface(lora_fixture_name, request): lora_path = get_adapter_absolute_path(lora_name) # lora loading should work for either absolute path and hugggingface id. + peft_helper = PEFTHelper.from_local_dir(lora_path, 4096) lora_model = LoRAModel.from_local_checkpoint( lora_path, expected_lora_modules, + peft_helper=peft_helper, lora_model_id=1, device="cpu", embedding_modules=embedding_modules, diff --git a/tests/lora/test_peft_helper.py b/tests/lora/test_peft_helper.py new file mode 100644 index 0000000000000..a524d5ce5f34a --- /dev/null +++ b/tests/lora/test_peft_helper.py @@ -0,0 +1,109 @@ +import json +import math +import shutil + +import pytest + +from vllm.config import LoRAConfig +from vllm.lora.peft_helper import PEFTHelper + +ERROR_CASES = [ + ( + "test_rank", + { + "r": 1024 + }, + "is greater than max_lora_rank", + ), + ( + "test_bias", + { + "bias": "all" + }, + "Adapter bias cannot be used without bias_enabled", + ), + ("test_dora", { + "use_dora": True + }, "does not yet support DoRA"), + ( + "test_modules_to_save", + { + "modules_to_save": ["lm_head"] + }, + "only supports modules_to_save being None", + ), +] + + +def test_peft_helper_pass(long_context_lora_files_16k_1, tmp_path): + peft_helper = PEFTHelper.from_local_dir(long_context_lora_files_16k_1, + max_position_embeddings=4096) + lora_config = LoRAConfig(max_lora_rank=16, max_cpu_loras=3, max_loras=2) + peft_helper.validate_legal(lora_config) + assert peft_helper.r == 8 + assert peft_helper.lora_alpha == 16 + assert peft_helper.target_modules == [ + "q_proj", + "v_proj", + "k_proj", + "o_proj", + "gate_proj", + "up_proj", + "down_proj", + "embed_tokens", + "lm_head", + ] + assert peft_helper.context_length == 16384 + assert peft_helper.vllm_max_position_embeddings == 4096 + assert peft_helper.vllm_long_context_scaling_factor == float( + math.ceil(peft_helper.context_length / + peft_helper.vllm_max_position_embeddings)) + # test RSLoRA + rslora_config = dict(use_rslora=True) + test_dir = tmp_path / "test_rslora" + shutil.copytree(long_context_lora_files_16k_1, test_dir) + + # Load and modify configuration + config_path = test_dir / "adapter_config.json" + with open(config_path) as f: + adapter_config = json.load(f) + # Apply configuration changes + adapter_config.update(rslora_config) + + # Save modified configuration + with open(config_path, "w") as f: + json.dump(adapter_config, f) + + peft_helper = PEFTHelper.from_local_dir(test_dir, + max_position_embeddings=4096) + peft_helper.validate_legal(lora_config) + scaling = peft_helper.lora_alpha / math.sqrt(peft_helper.r) + assert abs(peft_helper.vllm_lora_scaling_factor - scaling) < 1e-3 + + +@pytest.mark.parametrize("test_name,config_change,expected_error", ERROR_CASES) +def test_peft_helper_error( + sql_lora_files, + tmp_path, + test_name: str, + config_change: dict, + expected_error: str, +): + test_dir = tmp_path / test_name + shutil.copytree(sql_lora_files, test_dir) + + # Load and modify configuration + config_path = test_dir / "adapter_config.json" + with open(config_path) as f: + adapter_config = json.load(f) + # Apply configuration changes + adapter_config.update(config_change) + + # Save modified configuration + with open(config_path, "w") as f: + json.dump(adapter_config, f) + lora_config = LoRAConfig(max_lora_rank=16, max_cpu_loras=3, max_loras=2) + # Test loading the adapter + with pytest.raises(ValueError, match=expected_error): + PEFTHelper.from_local_dir( + test_dir, max_position_embeddings=4096).validate_legal(lora_config) diff --git a/tests/model_executor/test_model_load_with_params.py b/tests/model_executor/test_model_load_with_params.py index 0609fd96825e3..9c1f784c1c93b 100644 --- a/tests/model_executor/test_model_load_with_params.py +++ b/tests/model_executor/test_model_load_with_params.py @@ -25,13 +25,12 @@ def test_model_loading_with_params(vllm_runner): with vllm_runner(model_name=MODEL_NAME, revision=REVISION, dtype="float16", - max_model_len=MAX_MODEL_LEN) as model: - output = model.encode("Write a short story about a robot that" - " dreams for the first time.\n") + max_model_len=MAX_MODEL_LEN) as vllm_model: + output = vllm_model.encode("Write a short story about a robot that" + " dreams for the first time.\n") - model_config = model.model.llm_engine.model_config - - model_tokenizer = model.model.llm_engine.tokenizer + model_config = vllm_model.model.llm_engine.model_config + model_tokenizer = vllm_model.model.llm_engine.tokenizer # asserts on the bert model config file assert model_config.encoder_config["max_seq_length"] == 512 @@ -46,11 +45,13 @@ def test_model_loading_with_params(vllm_runner): assert model_tokenizer.tokenizer_config["do_lower_case"] assert model_tokenizer.tokenizer.model_max_length == 512 - model = model.model.llm_engine.model_executor\ - .driver_worker.model_runner.model - assert isinstance(model, BertEmbeddingModel) - assert model._pooler.pooling_type == PoolingType.CLS - assert model._pooler.normalize + def check_model(model): + assert isinstance(model, BertEmbeddingModel) + assert model._pooler.pooling_type == PoolingType.CLS + assert model._pooler.normalize + + vllm_model.apply_model(check_model) + # assert output assert output @@ -64,13 +65,12 @@ def test_roberta_model_loading_with_params(vllm_runner): with vllm_runner(model_name=MODEL_NAME_ROBERTA, revision=REVISION_ROBERTA, dtype="float16", - max_model_len=MAX_MODEL_LEN) as model: - output = model.encode("Write a short story about a robot that" - " dreams for the first time.\n") + max_model_len=MAX_MODEL_LEN) as vllm_model: + output = vllm_model.encode("Write a short story about a robot that" + " dreams for the first time.\n") - model_config = model.model.llm_engine.model_config - - model_tokenizer = model.model.llm_engine.tokenizer + model_config = vllm_model.model.llm_engine.model_config + model_tokenizer = vllm_model.model.llm_engine.tokenizer # asserts on the bert model config file assert model_config.encoder_config["max_seq_length"] == 512 @@ -84,11 +84,12 @@ def test_roberta_model_loading_with_params(vllm_runner): assert model_tokenizer.tokenizer_id == "intfloat/multilingual-e5-large" assert not model_tokenizer.tokenizer_config["do_lower_case"] - model = model.model.llm_engine.model_executor\ - .driver_worker.model_runner.model - assert isinstance(model, RobertaEmbeddingModel) - assert model._pooler.pooling_type == PoolingType.MEAN - assert model._pooler.normalize + def check_model(model): + assert isinstance(model, RobertaEmbeddingModel) + assert model._pooler.pooling_type == PoolingType.MEAN + assert model._pooler.normalize + + vllm_model.apply_model(check_model) # assert output assert output @@ -103,17 +104,18 @@ def test_facebook_roberta_model_loading_with_params(vllm_runner): model_name = "FacebookAI/roberta-base" with vllm_runner(model_name=model_name, dtype="float16", - max_model_len=MAX_MODEL_LEN) as model: - output = model.encode("Write a short story about a robot that" - " dreams for the first time.\n") + max_model_len=MAX_MODEL_LEN) as vllm_model: + output = vllm_model.encode("Write a short story about a robot that" + " dreams for the first time.\n") - model_tokenizer = model.model.llm_engine.tokenizer + model_tokenizer = vllm_model.model.llm_engine.tokenizer assert model_tokenizer.tokenizer_id == model_name - model = model.model.llm_engine.model_executor\ - .driver_worker.model_runner.model - assert not hasattr(model, "lm_head") - assert isinstance(model, RobertaEmbeddingModel) - assert isinstance(model._pooler, CLSPool) + def check_model(model): + assert isinstance(model, RobertaEmbeddingModel) + assert not hasattr(model, "lm_head") + assert isinstance(model._pooler, CLSPool) + + vllm_model.apply_model(check_model) assert output diff --git a/tests/models/decoder_only/language/test_gguf.py b/tests/models/decoder_only/language/test_gguf.py index 81b93ebdf0fc0..ad8f8a0c320e9 100644 --- a/tests/models/decoder_only/language/test_gguf.py +++ b/tests/models/decoder_only/language/test_gguf.py @@ -66,12 +66,16 @@ def gguf_model(self): gguf_filename="starcoder2-3b.Q6_K.gguf", ) +DOLPHIN_CONFIG = GGUFTestConfig( + # Test VocabParallelEmbedding sharding issue. + original_model="cognitivecomputations/TinyDolphin-2.8-1.1b", + gguf_repo="tsunemoto/TinyDolphin-2.8-1.1b-GGUF", + gguf_filename="tinydolphin-2.8-1.1b.Q6_K.gguf", +) + MODELS = [ - LLAMA_CONFIG, - QWEN2_CONFIG, - PHI3_CONFIG, - GPT2_CONFIG, - STABLELM_CONFIG, + LLAMA_CONFIG, QWEN2_CONFIG, PHI3_CONFIG, GPT2_CONFIG, STABLELM_CONFIG, + DOLPHIN_CONFIG # STARCODER_CONFIG, # broken ] @@ -106,15 +110,18 @@ def test_models( messages, tokenize=False, add_generation_prompt=True) # Run unquantized model. - with vllm_runner(model_name=model.original_model, - dtype=dtype, - max_model_len=MAX_MODEL_LEN, - tensor_parallel_size=tp_size) as original_model: + with vllm_runner( + model_name=model.original_model, + enforce_eager=True, # faster tests + dtype=dtype, + max_model_len=MAX_MODEL_LEN, + tensor_parallel_size=tp_size) as original_model: original_outputs = original_model.generate_greedy_logprobs( example_prompts[:-1], max_tokens, num_logprobs) # Run gguf model. with vllm_runner(model_name=model.gguf_model, + enforce_eager=True, tokenizer_name=model.original_model, dtype=dtype, max_model_len=MAX_MODEL_LEN, diff --git a/tests/models/decoder_only/language/test_jamba.py b/tests/models/decoder_only/language/test_jamba.py index 057b04349e8b7..2e06b10fbb827 100644 --- a/tests/models/decoder_only/language/test_jamba.py +++ b/tests/models/decoder_only/language/test_jamba.py @@ -33,10 +33,13 @@ def test_models( with vllm_runner(model, dtype=dtype) as vllm_model: vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens) + # This test is for verifying whether the model's extra_repr # can be printed correctly. - print(vllm_model.model.llm_engine.model_executor.driver_worker. - model_runner.model) + def print_model(model): + print(model) + + vllm_model.apply_model(print_model) for i in range(len(example_prompts)): hf_output_ids, hf_output_str = hf_outputs[i] diff --git a/tests/models/decoder_only/language/test_mamba.py b/tests/models/decoder_only/language/test_mamba.py index 06739e8f02253..1ad4f5aae8f5b 100644 --- a/tests/models/decoder_only/language/test_mamba.py +++ b/tests/models/decoder_only/language/test_mamba.py @@ -51,10 +51,13 @@ def test_models( with vllm_runner(model, dtype=dtype) as vllm_model: vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens) + # This test is for verifying whether the model's extra_repr # can be printed correctly. - print(vllm_model.model.llm_engine.model_executor.driver_worker. - model_runner.model) + def print_model(model): + print(model) + + vllm_model.apply_model(print_model) for i in range(len(example_prompts)): hf_output_ids, hf_output_str = hf_outputs[i] diff --git a/tests/models/decoder_only/language/test_models.py b/tests/models/decoder_only/language/test_models.py index 4e110366a09f3..c7efa4edbbc0a 100644 --- a/tests/models/decoder_only/language/test_models.py +++ b/tests/models/decoder_only/language/test_models.py @@ -73,10 +73,13 @@ def test_models( with vllm_runner(model, dtype=dtype) as vllm_model: vllm_outputs = vllm_model.generate_greedy_logprobs( example_prompts, max_tokens, num_logprobs) + # This test is for verifying whether the model's extra_repr # can be printed correctly. - print(vllm_model.model.llm_engine.model_executor.driver_worker. - model_runner.model) + def print_model(model): + print(model) + + vllm_model.apply_model(print_model) check_logprobs_close( outputs_0_lst=hf_outputs, diff --git a/tests/models/decoder_only/vision_language/test_models.py b/tests/models/decoder_only/vision_language/test_models.py index 7620ed1107e8f..14d9a739be318 100644 --- a/tests/models/decoder_only/vision_language/test_models.py +++ b/tests/models/decoder_only/vision_language/test_models.py @@ -9,7 +9,7 @@ import pytest from transformers import AutoModelForVision2Seq -from transformers.utils import is_flash_attn_2_available +from transformers import __version__ as TRANSFORMERS_VERSION from vllm.platforms import current_platform from vllm.utils import identity @@ -139,9 +139,7 @@ #### Extended model tests "aria": VLMTestInfo( models=["rhymes-ai/Aria"], - tokenizer_mode="slow", test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), - dtype="bfloat16", prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ", # noqa: E501 img_idx_to_prompt=lambda idx: "<|img|>\n", max_model_len=4096, @@ -157,8 +155,8 @@ max_tokens=64, marks=[ pytest.mark.skipif( - not is_flash_attn_2_available(), - reason="Model needs flash-attn for numeric convergence.", + TRANSFORMERS_VERSION < "4.48.0", + reason="HF model requires transformers>=4.48.0", ), large_gpu_mark(min_gb=64), ], @@ -189,30 +187,27 @@ dtype="bfloat16", ), "deepseek_vl_v2": VLMTestInfo( - models=["deepseek-ai/deepseek-vl2-small"], + models=["Isotr0py/deepseek-vl2-tiny"], # model repo using dynamic module test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), - dtype="bfloat16", prompt_formatter=lambda img_prompt: f"<|User|>: {img_prompt}\n\n<|Assistant|>: ", # noqa: E501 max_model_len=4096, max_num_seqs=2, single_image_prompts=IMAGE_ASSETS.prompts({ - "stop_sign": "\nWhat's the color of the stop sign and car?", - "cherry_blossom": "\nWhat's the color of the tower?", + "stop_sign": "\nWhat's the content in the center of the image?", # noqa: E501 + "cherry_blossom": "\nPlease infer the season with reason in details.", # noqa: E501 }), - multi_image_prompt="image_1:\nimage_2:\nDescribe the two images shortly.", # noqa: E501 + multi_image_prompt="image_1:\nimage_2:\nWhich image can we see the car and the tower?", # noqa: E501 vllm_runner_kwargs={"hf_overrides": {"architectures": ["DeepseekVLV2ForCausalLM"]}}, # noqa: E501 - image_size_factors=[(0.10, 0.15)], patch_hf_runner=model_utils.deepseekvl2_patch_hf_runner, postprocess_inputs=model_utils.cast_dtype_post_processor("images"), hf_output_post_proc=model_utils.deepseekvl2_trunc_hf_output, stop_str=["<｜end▁of▁sentence｜>", "<｜begin▁of▁sentence｜>"], # noqa: E501 - num_logprobs=5, + image_size_factors=[(), (1.0, ), (1.0, 1.0, 1.0), (0.1, 0.5, 1.0)], marks=[ pytest.mark.skipif( - not is_flash_attn_2_available(), - reason="Model needs flash-attn for numeric convergence.", - ), - large_gpu_mark(min_gb=48), + TRANSFORMERS_VERSION >= "4.48.0", + reason="HF model is not compatible with transformers>=4.48.0", + ) ], ), "fuyu": VLMTestInfo( diff --git a/tests/models/decoder_only/vision_language/test_qwen2_vl.py b/tests/models/decoder_only/vision_language/test_qwen2_vl.py index 16e256e040a74..5a485f3d81747 100644 --- a/tests/models/decoder_only/vision_language/test_qwen2_vl.py +++ b/tests/models/decoder_only/vision_language/test_qwen2_vl.py @@ -5,7 +5,6 @@ import torch from PIL import Image -from vllm.entrypoints.llm import LLM from vllm.multimodal.image import rescale_image_size from vllm.multimodal.video import rescale_video_size, sample_frames_from_video @@ -69,7 +68,7 @@ class Qwen2VLPromptVideoEmbeddingInput(TypedDict): def batch_make_image_embeddings( image_batches: List[Union[Image.Image, List[Image.Image]]], processor, - llm: LLM) -> List[Qwen2VLPromptImageEmbeddingInput]: + llm: VllmRunner) -> List[Qwen2VLPromptImageEmbeddingInput]: """batched image embeddings for Qwen2-VL This will infer all images' embeddings in a single batch, @@ -105,17 +104,19 @@ def batch_make_image_embeddings( pixel_values = preprocess_result["pixel_values"] image_grid_thw = preprocess_result["image_grid_thw"] - # pixel values to embeddinds & grid_thws - with torch.no_grad(): - visual = llm.llm_engine.model_executor.driver_worker. \ - model_runner.model.visual + # pixel values to embeddings & grid_thws + def get_image_embeds(model): + with torch.no_grad(): + visual = model.visual - pixel_values_on_device = pixel_values.to(visual.device, - dtype=visual.dtype) - image_grid_thw_on_device = image_grid_thw.to(visual.device, - dtype=torch.int64) - image_embeds = visual(pixel_values_on_device, - grid_thw=image_grid_thw_on_device) + pixel_values_on_device = pixel_values.to(visual.device, + dtype=visual.dtype) + image_grid_thw_on_device = image_grid_thw.to(visual.device, + dtype=torch.int64) + return visual(pixel_values_on_device, + grid_thw=image_grid_thw_on_device) + + image_embeds = torch.concat(llm.apply_model(get_image_embeds)) # split into original batches result: List[Qwen2VLPromptImageEmbeddingInput] = [] @@ -124,11 +125,10 @@ def batch_make_image_embeddings( for image_batch in image_batches_: cur_batch_image_count = len(image_batch) merge_size = image_processor.merge_size - cur_batch_embed_len = sum([ - grid_thw.prod() // merge_size // merge_size + cur_batch_embed_len = sum( + grid_thw.prod(-1) // merge_size // merge_size for grid_thw in image_grid_thw[image_counter:image_counter + - cur_batch_image_count] - ]) + cur_batch_image_count]) result.append({ "image_embeds": @@ -151,7 +151,7 @@ def batch_make_image_embeddings( def batch_make_video_embeddings( video_batches: PromptVideoInput, processor, - llm: LLM) -> List[Qwen2VLPromptVideoEmbeddingInput]: + llm: VllmRunner) -> List[Qwen2VLPromptVideoEmbeddingInput]: """batched video embeddings for Qwen2-VL A NDArray represents a single video's all frames. @@ -187,17 +187,19 @@ def batch_make_video_embeddings( pixel_values = preprocess_result["pixel_values_videos"] video_grid_thw = preprocess_result["video_grid_thw"] - # pixel values to embeddinds & grid_thws - with torch.no_grad(): - visual = llm.llm_engine.model_executor.driver_worker.\ - model_runner.model.visual + # pixel values to embeddings & grid_thws + def get_image_embeds(model): + with torch.no_grad(): + visual = model.visual + + pixel_values_on_device = pixel_values.to(visual.device, + dtype=visual.dtype) + video_grid_thw_on_device = video_grid_thw.to(visual.device, + dtype=torch.int64) + return visual(pixel_values_on_device, + grid_thw=video_grid_thw_on_device) - pixel_values_on_device = pixel_values.to(visual.device, - dtype=visual.dtype) - video_grid_thw_on_device = video_grid_thw.to(visual.device, - dtype=torch.int64) - video_embeds = visual(pixel_values_on_device, - grid_thw=video_grid_thw_on_device) + video_embeds = torch.concat(llm.apply_model(get_image_embeds)) # split into original batches result: List[Qwen2VLPromptVideoEmbeddingInput] = [] @@ -206,11 +208,10 @@ def batch_make_video_embeddings( for video_batch in video_batches_: cur_batch_video_count = len(video_batch) merge_size = image_processor.merge_size - cur_batch_embed_len = sum([ - grid_thw.prod() // merge_size // merge_size + cur_batch_embed_len = sum( + grid_thw.prod(-1) // merge_size // merge_size for grid_thw in video_grid_thw[video_counter:video_counter + - cur_batch_video_count] - ]) + cur_batch_video_count]) result.append({ "video_embeds": @@ -280,9 +281,9 @@ def run_embedding_input_test( max_tokens, num_logprobs=num_logprobs, images=batch_make_image_embeddings( - images, processor, vllm_model.model) if images else None, + images, processor, vllm_model) if images else None, videos=batch_make_video_embeddings( - videos, processor, vllm_model.model) if videos else None) + videos, processor, vllm_model) if videos else None) for prompts, images, videos in inputs ] diff --git a/tests/models/embedding/language/test_cls_models.py b/tests/models/embedding/language/test_cls_models.py index 6673a9fc22f69..0cbe4afe96c0a 100644 --- a/tests/models/embedding/language/test_cls_models.py +++ b/tests/models/embedding/language/test_cls_models.py @@ -24,10 +24,13 @@ def test_classification_models( ) -> None: with vllm_runner(model, dtype=dtype) as vllm_model: vllm_outputs = vllm_model.classify(example_prompts) + # This test is for verifying whether the model's extra_repr # can be printed correctly. - print(vllm_model.model.llm_engine.model_executor.driver_worker. - model_runner.model) + def print_model(model): + print(model) + + vllm_model.apply_model(print_model) with hf_runner(model, dtype=dtype, diff --git a/tests/models/embedding/language/test_embedding.py b/tests/models/embedding/language/test_embedding.py index 04ab4dd7371a3..e17198e385475 100644 --- a/tests/models/embedding/language/test_embedding.py +++ b/tests/models/embedding/language/test_embedding.py @@ -17,14 +17,15 @@ marks=[pytest.mark.core_model, pytest.mark.cpu_model]), pytest.param("sentence-transformers/all-MiniLM-L12-v2"), pytest.param("intfloat/multilingual-e5-large"), - # [Encoder-decoder] - pytest.param("intfloat/e5-mistral-7b-instruct", - marks=[pytest.mark.core_model, pytest.mark.cpu_model]), + # [Decoder-only] pytest.param("BAAI/bge-multilingual-gemma2", marks=[pytest.mark.core_model]), - pytest.param("ssmits/Qwen2-7B-Instruct-embed-base"), + pytest.param("intfloat/e5-mistral-7b-instruct", + marks=[pytest.mark.core_model, pytest.mark.cpu_model]), pytest.param("Alibaba-NLP/gte-Qwen2-1.5B-instruct"), pytest.param("Alibaba-NLP/gte-Qwen2-7B-instruct"), + pytest.param("ssmits/Qwen2-7B-Instruct-embed-base"), + # [Encoder-decoder] pytest.param("sentence-transformers/stsb-roberta-base-v2"), ], ) @@ -61,10 +62,13 @@ def test_models( max_model_len=None, **vllm_extra_kwargs) as vllm_model: vllm_outputs = vllm_model.encode(example_prompts) + # This test is for verifying whether the model's extra_repr # can be printed correctly. - print(vllm_model.model.llm_engine.model_executor.driver_worker. - model_runner.model) + def print_model(model): + print(model) + + vllm_model.apply_model(print_model) check_embeddings_close( embeddings_0_lst=hf_outputs, diff --git a/tests/models/multimodal/processing/test_common.py b/tests/models/multimodal/processing/test_common.py index 0a38779e0e4f0..fe5b733c750a8 100644 --- a/tests/models/multimodal/processing/test_common.py +++ b/tests/models/multimodal/processing/test_common.py @@ -11,6 +11,7 @@ from vllm.multimodal.utils import cached_get_tokenizer from ....multimodal.utils import random_audio, random_image, random_video +from ...registry import HF_EXAMPLE_MODELS def _test_processing_correctness( @@ -20,10 +21,9 @@ def _test_processing_correctness( num_batches: int, simplify_rate: float, ): - if model_id == "TIGER-Lab/Mantis-8B-siglip-llama3": - hf_overrides = {"architectures": ["MantisForConditionalGeneration"]} - else: - hf_overrides = {} + model_info = HF_EXAMPLE_MODELS.find_hf_info(model_id) + model_info.check_available_online(on_fail="skip") + model_info.check_transformers_version(on_fail="skip") limit_mm_per_prompt = { modality: 3 if supports_multi else 1 @@ -35,11 +35,11 @@ def _test_processing_correctness( task="auto", tokenizer=model_id, tokenizer_mode="auto", - trust_remote_code=True, + trust_remote_code=model_info.trust_remote_code, seed=0, dtype="float16", revision=None, - hf_overrides=hf_overrides, + hf_overrides=model_info.hf_overrides, limit_mm_per_prompt=limit_mm_per_prompt, ) @@ -139,6 +139,7 @@ def _test_processing_correctness( ("rhymes-ai/Aria", {"image": True}), ("Salesforce/blip2-opt-2.7b", {"image": False}), ("facebook/chameleon-7b", {"image": False}), + ("deepseek-ai/deepseek-vl2-tiny", {"image": True}), ("adept/fuyu-8b", {"image": False}), ("llava-hf/llava-1.5-7b-hf", {"image": True}), ("llava-hf/llava-v1.6-mistral-7b-hf", {"image": True}), diff --git a/tests/models/multimodal/processing/test_llava_next.py b/tests/models/multimodal/processing/test_llava_next.py index 1eec35d9c3c72..6de649f87204d 100644 --- a/tests/models/multimodal/processing/test_llava_next.py +++ b/tests/models/multimodal/processing/test_llava_next.py @@ -13,6 +13,67 @@ from ...utils import build_model_context +def _validate_image_max_tokens_one( + processor: BaseMultiModalProcessor, + max_tokens: int, + failed_size_excs: list[tuple[ImageSize, Exception]], + image_size: ImageSize, +) -> None: + info = processor.info + feature_size = info.get_num_image_tokens(image_width=image_size.width, + image_height=image_size.height) + + try: + assert feature_size <= max_tokens, f"{feature_size} <= {max_tokens}" + except Exception as exc: + failed_size_excs.append((image_size, exc)) + + +@pytest.mark.skip("This test takes around 5 minutes to run. " + "Comment this out to run it manually.") +@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"]) +def test_processor_max_tokens(model_id): + ctx = build_model_context( + model_name=model_id, + tokenizer_name=model_id, + mm_processor_kwargs=None, + limit_mm_per_prompt={"image": 1}, + ) + processor = MULTIMODAL_REGISTRY.create_processor( + ctx.model_config, + tokenizer=cached_get_tokenizer(ctx.model_config.tokenizer), + ) + info = processor.info + + seen_aspect_ratios = set[float]() + image_sizes = list[ImageSize]() + + # The aspect ratio of the grid layout is between 1 and 2 + # NOTE: Assumes that feature size calculation is the same if we + # swap the width and height of the image + for w, h in itertools.product(range(32, 4096), repeat=2): + aspect_ratio = w / h + if 1 <= aspect_ratio <= 2 and aspect_ratio not in seen_aspect_ratios: + image_sizes.append(ImageSize(w, h)) + seen_aspect_ratios.add(aspect_ratio) + + failed_size_excs = list[tuple[ImageSize, Exception]]() + + validate_one = partial( + _validate_image_max_tokens_one, + processor, + info.get_max_image_tokens(), # type: ignore + failed_size_excs, + ) + pqdm(image_sizes, validate_one, n_jobs=8, desc="Validating image sizes") + + if failed_size_excs: + msg = "Found failing image sizes:" \ + + "\n========\n".join(f"[{size}]\n{exc}" + for size, exc in failed_size_excs) + raise AssertionError(msg) + + def _validate_image_prompt_replacements_one( processor: BaseMultiModalProcessor, num_imgs: int, diff --git a/tests/models/multimodal/processing/test_llava_onevision.py b/tests/models/multimodal/processing/test_llava_onevision.py index 94ea604c58b43..806437d35ec87 100644 --- a/tests/models/multimodal/processing/test_llava_onevision.py +++ b/tests/models/multimodal/processing/test_llava_onevision.py @@ -13,6 +13,68 @@ from ...utils import build_model_context +def _validate_image_max_tokens_one( + processor: BaseMultiModalProcessor, + max_tokens: int, + failed_size_excs: list[tuple[ImageSize, Exception]], + image_size: ImageSize, +) -> None: + info = processor.info + feature_size = info.get_num_image_tokens(image_width=image_size.width, + image_height=image_size.height) + + try: + assert feature_size <= max_tokens, f"{feature_size} <= {max_tokens}" + except Exception as exc: + failed_size_excs.append((image_size, exc)) + + +@pytest.mark.skip("This test takes around 5 minutes to run. " + "Comment this out to run it manually.") +@pytest.mark.parametrize("model_id", + ["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"]) +def test_processor_max_tokens(model_id): + ctx = build_model_context( + model_name=model_id, + tokenizer_name=model_id, + mm_processor_kwargs=None, + limit_mm_per_prompt={"image": 1}, + ) + processor = MULTIMODAL_REGISTRY.create_processor( + ctx.model_config, + tokenizer=cached_get_tokenizer(ctx.model_config.tokenizer), + ) + info = processor.info + + seen_aspect_ratios = set[float]() + image_sizes = list[ImageSize]() + + # The aspect ratio of the grid layout is between 1 and 6 + # NOTE: Assumes that feature size calculation is the same if we + # swap the width and height of the image + for w, h in itertools.product(range(32, 4096), repeat=2): + aspect_ratio = w / h + if 1 <= aspect_ratio <= 6 and aspect_ratio not in seen_aspect_ratios: + image_sizes.append(ImageSize(w, h)) + seen_aspect_ratios.add(aspect_ratio) + + failed_size_excs = list[tuple[ImageSize, Exception]]() + + validate_one = partial( + _validate_image_max_tokens_one, + processor, + info.get_max_image_tokens(), # type: ignore + failed_size_excs, + ) + pqdm(image_sizes, validate_one, n_jobs=8, desc="Validating image sizes") + + if failed_size_excs: + msg = "Found failing image sizes:" \ + + "\n========\n".join(f"[{size}]\n{exc}" + for size, exc in failed_size_excs) + raise AssertionError(msg) + + def _validate_image_prompt_replacements_one( processor: BaseMultiModalProcessor, num_imgs: int, diff --git a/tests/models/registry.py b/tests/models/registry.py index d079725b2f78d..0bd06dea0ec7f 100644 --- a/tests/models/registry.py +++ b/tests/models/registry.py @@ -1,5 +1,9 @@ from dataclasses import dataclass, field -from typing import AbstractSet, Mapping, Optional +from typing import AbstractSet, Any, Literal, Mapping, Optional + +import pytest +from packaging.version import Version +from transformers import __version__ as TRANSFORMERS_VERSION @dataclass(frozen=True) @@ -38,6 +42,50 @@ class _HfExamplesInfo: trust_remote_code: bool = False """The ``trust_remote_code`` level required to load the model.""" + hf_overrides: dict[str, Any] = field(default_factory=dict) + """The ``hf_overrides`` required to load the model.""" + + def check_transformers_version( + self, + *, + on_fail: Literal["error", "skip"], + ) -> None: + """ + If the installed transformers version does not meet the requirements, + perform the given action. + """ + if self.min_transformers_version is None: + return + + current_version = TRANSFORMERS_VERSION + required_version = self.min_transformers_version + if Version(current_version) < Version(required_version): + msg = ( + f"You have `transformers=={current_version}` installed, but " + f"`transformers>={required_version}` is required to run this " + "model") + + if on_fail == "error": + raise RuntimeError(msg) + else: + pytest.skip(msg) + + def check_available_online( + self, + *, + on_fail: Literal["error", "skip"], + ) -> None: + """ + If the model is not available online, perform the given action. + """ + if not self.is_available_online: + msg = "Model is not available online" + + if on_fail == "error": + raise RuntimeError(msg) + else: + pytest.skip(msg) + # yapf: disable _TEXT_GENERATION_EXAMPLE_MODELS = { @@ -48,8 +96,6 @@ class _HfExamplesInfo: trust_remote_code=True), "ArcticForCausalLM": _HfExamplesInfo("Snowflake/snowflake-arctic-instruct", trust_remote_code=True), - "AriaForConditionalGeneration": _HfExamplesInfo("rhymes-ai/Aria", - trust_remote_code=True), "BaiChuanForCausalLM": _HfExamplesInfo("baichuan-inc/Baichuan-7B", trust_remote_code=True), "BaichuanForCausalLM": _HfExamplesInfo("baichuan-inc/Baichuan2-7B-chat", @@ -69,6 +115,7 @@ class _HfExamplesInfo: "DeepseekV3ForCausalLM": _HfExamplesInfo("deepseek-ai/DeepSeek-V3", # noqa: E501 trust_remote_code=True), "ExaoneForCausalLM": _HfExamplesInfo("LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct"), # noqa: E501 + "Fairseq2LlamaForCausalLM": _HfExamplesInfo("mgleize/fairseq2-dummy-Llama-3.2-1B"), # noqa: E501 "FalconForCausalLM": _HfExamplesInfo("tiiuae/falcon-7b"), "GemmaForCausalLM": _HfExamplesInfo("google/gemma-2b"), "Gemma2ForCausalLM": _HfExamplesInfo("google/gemma-2-9b"), @@ -85,6 +132,8 @@ class _HfExamplesInfo: trust_remote_code=True), "InternLM2VEForCausalLM": _HfExamplesInfo("OpenGVLab/Mono-InternVL-2B", trust_remote_code=True), + "InternLM3ForCausalLM": _HfExamplesInfo("internlm/internlm3-8b-instruct", + trust_remote_code=True), "JAISLMHeadModel": _HfExamplesInfo("inceptionai/jais-13b-chat"), "JambaForCausalLM": _HfExamplesInfo("ai21labs/AI21-Jamba-1.5-Mini"), "LlamaForCausalLM": _HfExamplesInfo("meta-llama/Meta-Llama-3-8B"), @@ -152,6 +201,7 @@ class _HfExamplesInfo: "MistralModel": _HfExamplesInfo("intfloat/e5-mistral-7b-instruct"), "Qwen2Model": _HfExamplesInfo("ssmits/Qwen2-7B-Instruct-embed-base"), "Qwen2ForRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-RM-72B"), + "Qwen2ForProcessRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-PRM-7B"), "Qwen2ForSequenceClassification": _HfExamplesInfo("jason9693/Qwen2.5-1.5B-apeach"), # noqa: E501 "RobertaModel": _HfExamplesInfo("sentence-transformers/stsb-roberta-base-v2"), # noqa: E501 "RobertaForMaskedLM": _HfExamplesInfo("sentence-transformers/all-roberta-large-v1"), # noqa: E501 @@ -172,6 +222,8 @@ class _HfExamplesInfo: _MULTIMODAL_EXAMPLE_MODELS = { # [Decoder-only] + "AriaForConditionalGeneration": _HfExamplesInfo("rhymes-ai/Aria", + min_transformers_version="4.48"), "Blip2ForConditionalGeneration": _HfExamplesInfo("Salesforce/blip2-opt-2.7b"), # noqa: E501 "ChameleonForConditionalGeneration": _HfExamplesInfo("facebook/chameleon-7b"), # noqa: E501 "ChatGLMModel": _HfExamplesInfo("THUDM/glm-4v-9b", @@ -179,8 +231,8 @@ class _HfExamplesInfo: trust_remote_code=True), "ChatGLMForConditionalGeneration": _HfExamplesInfo("chatglm2-6b", is_available_online=False), - # TODO(Isotr0py): Use deepseek-vl2-tiny for test after it's supported - "DeepseekVLV2ForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-vl2-small"), # noqa: E501 + "DeepseekVLV2ForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-vl2-tiny", # noqa: E501 + hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]}), # noqa: E501 "FuyuForCausalLM": _HfExamplesInfo("adept/fuyu-8b"), "H2OVLChatModel": _HfExamplesInfo("h2oai/h2ovl-mississippi-800m"), "InternVLChatModel": _HfExamplesInfo("OpenGVLab/InternVL2-1B", @@ -191,7 +243,8 @@ class _HfExamplesInfo: "LlavaNextForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-v1.6-mistral-7b-hf"), # noqa: E501 "LlavaNextVideoForConditionalGeneration": _HfExamplesInfo("llava-hf/LLaVA-NeXT-Video-7B-hf"), # noqa: E501 "LlavaOnevisionForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-onevision-qwen2-0.5b-ov-hf"), # noqa: E501 - "MantisForConditionalGeneration": _HfExamplesInfo("TIGER-Lab/Mantis-8B-siglip-llama3"), # noqa: E501 + "MantisForConditionalGeneration": _HfExamplesInfo("TIGER-Lab/Mantis-8B-siglip-llama3", # noqa: E501 + hf_overrides={"architectures": ["MantisForConditionalGeneration"]}), # noqa: E501 "MiniCPMV": _HfExamplesInfo("openbmb/MiniCPM-Llama3-V-2_5", trust_remote_code=True), "MolmoForCausalLM": _HfExamplesInfo("allenai/Molmo-7B-D-0924", @@ -208,7 +261,8 @@ class _HfExamplesInfo: trust_remote_code=True), "Qwen2AudioForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2-Audio-7B-Instruct"), # noqa: E501 "Qwen2VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2-VL-2B-Instruct"), # noqa: E501 - "UltravoxModel": _HfExamplesInfo("fixie-ai/ultravox-v0_3"), + "UltravoxModel": _HfExamplesInfo("fixie-ai/ultravox-v0_3", + trust_remote_code=True), # [Encoder-decoder] "MllamaForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-3.2-11B-Vision-Instruct"), # noqa: E501 "WhisperForConditionalGeneration": _HfExamplesInfo("openai/whisper-large-v3"), # noqa: E501 @@ -244,5 +298,17 @@ def get_supported_archs(self) -> AbstractSet[str]: def get_hf_info(self, model_arch: str) -> _HfExamplesInfo: return self.hf_models[model_arch] + def find_hf_info(self, model_id: str) -> _HfExamplesInfo: + for info in self.hf_models.values(): + if info.default == model_id: + return info + + # Fallback to extras + for info in self.hf_models.values(): + if any(extra == model_id for extra in info.extras.values()): + return info + + raise ValueError(f"No example model defined for {model_id}") + HF_EXAMPLE_MODELS = HfExampleModels(_EXAMPLE_MODELS) diff --git a/tests/models/test_initialization.py b/tests/models/test_initialization.py index daece7c93c0ef..24a036795f9dd 100644 --- a/tests/models/test_initialization.py +++ b/tests/models/test_initialization.py @@ -1,9 +1,7 @@ from unittest.mock import patch import pytest -from packaging.version import Version from transformers import PretrainedConfig -from transformers import __version__ as TRANSFORMERS_VERSION from vllm import LLM @@ -13,16 +11,8 @@ @pytest.mark.parametrize("model_arch", HF_EXAMPLE_MODELS.get_supported_archs()) def test_can_initialize(model_arch): model_info = HF_EXAMPLE_MODELS.get_hf_info(model_arch) - if not model_info.is_available_online: - pytest.skip("Model is not available online") - if model_info.min_transformers_version is not None: - current_version = TRANSFORMERS_VERSION - required_version = model_info.min_transformers_version - if Version(current_version) < Version(required_version): - pytest.skip( - f"You have `transformers=={current_version}` installed, but " - f"`transformers>={required_version}` is required to run this " - "model") + model_info.check_available_online(on_fail="skip") + model_info.check_transformers_version(on_fail="skip") # Avoid OOM def hf_overrides(hf_config: PretrainedConfig) -> PretrainedConfig: @@ -60,4 +50,4 @@ def _initialize_kv_caches(self) -> None: trust_remote_code=model_info.trust_remote_code, load_format="dummy", hf_overrides=hf_overrides, - ) + ) \ No newline at end of file diff --git a/tests/models/test_registry.py b/tests/models/test_registry.py index 73b70d65e8e0b..ac0366847e334 100644 --- a/tests/models/test_registry.py +++ b/tests/models/test_registry.py @@ -21,6 +21,9 @@ @pytest.mark.parametrize("model_arch", ModelRegistry.get_supported_archs()) def test_registry_imports(model_arch): + model_info = HF_EXAMPLE_MODELS.get_hf_info(model_arch) + model_info.check_transformers_version(on_fail="skip") + # Ensure all model classes can be imported successfully model_cls, _ = ModelRegistry.resolve_model_cls(model_arch) diff --git a/tests/multi_step/test_correctness_async_llm.py b/tests/multi_step/test_correctness_async_llm.py index 8456a463adeeb..b8524ed83026b 100644 --- a/tests/multi_step/test_correctness_async_llm.py +++ b/tests/multi_step/test_correctness_async_llm.py @@ -16,7 +16,8 @@ NUM_PROMPTS = [10] DEFAULT_SERVER_ARGS: List[str] = [ - "--worker-use-ray", + "--distributed-executor-backend", + "ray", "--gpu-memory-utilization", "0.85", "--swap-space", diff --git a/tests/multi_step/test_correctness_llm.py b/tests/multi_step/test_correctness_llm.py index cc1fd19252019..34030d9d6ac60 100644 --- a/tests/multi_step/test_correctness_llm.py +++ b/tests/multi_step/test_correctness_llm.py @@ -5,6 +5,8 @@ import pytest +from tests.kernels.utils import override_backend_env_variable + from ..models.utils import check_logprobs_close, check_outputs_equal MODELS = [ @@ -19,10 +21,11 @@ @pytest.mark.parametrize("tp_size", [1]) @pytest.mark.parametrize("enable_chunked_prefill", [False, True]) @pytest.mark.parametrize("max_tokens", [5]) -@pytest.mark.parametrize("enforce_eager", [True]) +@pytest.mark.parametrize("enforce_eager", [True, False]) @pytest.mark.parametrize("num_scheduler_steps", NUM_SCHEDULER_STEPS) @pytest.mark.parametrize("num_prompts", NUM_PROMPTS) @pytest.mark.parametrize("num_logprobs", [None, 5]) +@pytest.mark.parametrize("attention_backend", ["FLASH_ATTN", "FLASHINFER"]) def test_multi_step_llm( hf_runner, vllm_runner, @@ -36,6 +39,8 @@ def test_multi_step_llm( num_scheduler_steps: int, num_prompts: int, num_logprobs: Optional[int], + attention_backend: str, + monkeypatch, ) -> None: """Test vLLM engine with multi-step scheduling via sync LLM Engine. @@ -63,6 +68,7 @@ def test_multi_step_llm( num_logprobs: corresponds to the `logprobs` argument to the OpenAI completions endpoint; `None` -> 1 logprob returned. """ + override_backend_env_variable(monkeypatch, attention_backend) prompts = example_prompts if len(prompts) < num_prompts: @@ -114,6 +120,7 @@ def test_multi_step_llm( @pytest.mark.parametrize("num_scheduler_steps", NUM_SCHEDULER_STEPS) @pytest.mark.parametrize("num_prompts", NUM_PROMPTS) @pytest.mark.parametrize("num_logprobs,num_prompt_logprobs", [(5, 5)]) +@pytest.mark.parametrize("attention_backend", ["FLASH_ATTN"]) def test_multi_step_llm_w_prompt_logprobs( vllm_runner, example_prompts, @@ -126,6 +133,8 @@ def test_multi_step_llm_w_prompt_logprobs( num_prompts: int, num_logprobs: Optional[int], num_prompt_logprobs: Optional[int], + attention_backend: str, + monkeypatch, ) -> None: """Test prompt logprobs with multi-step scheduling via sync LLM Engine. @@ -155,6 +164,7 @@ def test_multi_step_llm_w_prompt_logprobs( note that this argument is not supported by the OpenAI completions endpoint. """ + override_backend_env_variable(monkeypatch, attention_backend) prompts = example_prompts if len(prompts) < num_prompts: @@ -205,6 +215,7 @@ def test_multi_step_llm_w_prompt_logprobs( @pytest.mark.parametrize("num_scheduler_steps", NUM_SCHEDULER_STEPS) @pytest.mark.parametrize("num_prompts", NUM_PROMPTS) @pytest.mark.parametrize("num_logprobs", [None, 5]) +@pytest.mark.parametrize("attention_backend", ["FLASH_ATTN"]) def test_multi_step_llm_chunked_prefill_prefix_cache( vllm_runner, example_prompts, @@ -216,6 +227,8 @@ def test_multi_step_llm_chunked_prefill_prefix_cache( num_scheduler_steps: int, num_prompts: int, num_logprobs: Optional[int], + attention_backend: str, + monkeypatch, ) -> None: """Test vLLM engine with multi-step+"single-step chunked prefill"+APC. @@ -278,6 +291,8 @@ def test_multi_step_llm_chunked_prefill_prefix_cache( # # The Incorrect scheduling behavior - if it occurs - will cause an exception # in the model runner resulting from `do_sample=False`. + override_backend_env_variable(monkeypatch, attention_backend) + assert len(example_prompts) >= 2 challenge_prompts = copy.deepcopy(example_prompts) challenge_prompts[0] = ('vLLM is a high-throughput and memory-efficient ' diff --git a/tests/multimodal/test_processing.py b/tests/multimodal/test_processing.py index 54269c3ef7ce0..13f820d013e2a 100644 --- a/tests/multimodal/test_processing.py +++ b/tests/multimodal/test_processing.py @@ -7,12 +7,16 @@ from vllm.config import ModelConfig from vllm.multimodal import MULTIMODAL_REGISTRY -from vllm.multimodal.processing import (PlaceholderInfo, PromptReplacement, +# yapf conflicts with isort for this block +# yapf: disable +from vllm.multimodal.processing import (PlaceholderFeaturesInfo, + PromptReplacement, find_mm_placeholders, find_text_matches, find_token_matches, iter_token_matches, replace_text_matches, replace_token_matches) +# yapf: enable from vllm.multimodal.profiling import MultiModalProfiler from vllm.multimodal.utils import cached_get_tokenizer from vllm.transformers_utils.tokenizer import AnyTokenizer @@ -421,6 +425,8 @@ def test_find_replace_tokens( "pattern_1": [32000, 32000], "pattern_2": [], "pattern_3": [1550, 918, 1550], + # Test different modalities having the same tokens (32000) + "pattern_4": [32000], }, ], ) @@ -431,11 +437,19 @@ def test_find_replace_tokens( [1, 9833, 28747, 32000, 9833, 28747, 32000, 32000, 918], { "pattern_1": [ - PlaceholderInfo( + PlaceholderFeaturesInfo( modality="pattern_1", item_idx=0, start_idx=6, - replacement=[32000, 32000], + tokens=[32000, 32000], + ), + ], + "pattern_4": [ + PlaceholderFeaturesInfo( + modality="pattern_4", + item_idx=0, + start_idx=3, + tokens=[32000], ), ], } @@ -445,52 +459,61 @@ def test_find_replace_tokens( [1, 32000, 32000, 9833, 28747, 32000, 32000, 1550, 918, 1550], { "pattern_1": [ - PlaceholderInfo( + PlaceholderFeaturesInfo( modality="pattern_1", item_idx=0, start_idx=1, - replacement=[32000, 32000], + tokens=[32000, 32000], ), - PlaceholderInfo( + PlaceholderFeaturesInfo( modality="pattern_1", item_idx=1, start_idx=5, - replacement=[32000, 32000], + tokens=[32000, 32000], ), ], "pattern_3": [ - PlaceholderInfo( + PlaceholderFeaturesInfo( modality="pattern_3", item_idx=0, start_idx=7, - replacement=[1550, 918, 1550], + tokens=[1550, 918, 1550], ), ], + # No match for pattern_4 as it has lower priority than pattern_1 } ), ( [1, 32000, 32000, 32000, 32000, 32000, 1550, 918, 1550], { "pattern_1": [ - PlaceholderInfo( + PlaceholderFeaturesInfo( modality="pattern_1", item_idx=0, start_idx=1, - replacement=[32000, 32000], + tokens=[32000, 32000], ), - PlaceholderInfo( + PlaceholderFeaturesInfo( modality="pattern_1", item_idx=1, start_idx=3, - replacement=[32000, 32000], + tokens=[32000, 32000], + ), + ], + "pattern_4": [ + PlaceholderFeaturesInfo( + modality="pattern_4", + item_idx=0, + start_idx=5, + tokens=[32000], ), ], "pattern_3": [ - PlaceholderInfo( + PlaceholderFeaturesInfo( modality="pattern_3", item_idx=0, start_idx=6, - replacement=[1550, 918, 1550], + tokens=[1550, 918, 1550], ), ], } diff --git a/tests/plugins/vllm_add_dummy_platform/vllm_add_dummy_platform/dummy_attention_backend.py b/tests/plugins/vllm_add_dummy_platform/vllm_add_dummy_platform/dummy_attention_backend.py new file mode 100644 index 0000000000000..5634be3c8d882 --- /dev/null +++ b/tests/plugins/vllm_add_dummy_platform/vllm_add_dummy_platform/dummy_attention_backend.py @@ -0,0 +1,8 @@ +from vllm.attention.backends.flash_attn import FlashAttentionBackend + + +class DummyAttentionBackend(FlashAttentionBackend): + + @staticmethod + def get_name() -> str: + return "Dummy_Backend" diff --git a/tests/plugins/vllm_add_dummy_platform/vllm_add_dummy_platform/dummy_platform.py b/tests/plugins/vllm_add_dummy_platform/vllm_add_dummy_platform/dummy_platform.py index fde93142f1103..84721d5971ccf 100644 --- a/tests/plugins/vllm_add_dummy_platform/vllm_add_dummy_platform/dummy_platform.py +++ b/tests/plugins/vllm_add_dummy_platform/vllm_add_dummy_platform/dummy_platform.py @@ -3,3 +3,7 @@ class DummyPlatform(CudaPlatform): device_name = "DummyDevice" + + def get_attn_backend_cls(self, backend_name, head_size, dtype, + kv_cache_dtype, block_size, use_v1): + return "vllm_add_dummy_platform.dummy_attention_backend.DummyAttentionBackend" # noqa E501 diff --git a/tests/plugins_tests/test_platform_plugins.py b/tests/plugins_tests/test_platform_plugins.py index 69698b34c71a3..661aa5f649ab9 100644 --- a/tests/plugins_tests/test_platform_plugins.py +++ b/tests/plugins_tests/test_platform_plugins.py @@ -1,3 +1,10 @@ +import torch + +from tests.kernels.utils import override_backend_env_variable +from vllm.attention.selector import get_attn_backend +from vllm.utils import STR_INVALID_VAL + + def test_platform_plugins(): # simulate workload by running an example import runpy @@ -14,3 +21,10 @@ def test_platform_plugins(): f"Expected DummyDevice, got {current_platform.device_name}, " "possibly because current_platform is imported before the plugin" f" is loaded. The first import:\n{_init_trace}") + + +def test_oot_attention_backend(monkeypatch): + # ignore the backend env variable if it is set + override_backend_env_variable(monkeypatch, STR_INVALID_VAL) + backend = get_attn_backend(16, torch.float16, torch.float16, 16, False) + assert backend.get_name() == "Dummy_Backend" diff --git a/tests/quantization/test_compressed_tensors.py b/tests/quantization/test_compressed_tensors.py index 92436889ecffe..bf0d454ad511c 100644 --- a/tests/quantization/test_compressed_tensors.py +++ b/tests/quantization/test_compressed_tensors.py @@ -30,50 +30,55 @@ def test_compressed_tensors_w8a8_static_setup(vllm_runner, model_args): model_path, strategy, quant_type, shape_0, is_symmetric = model_args with vllm_runner(model_path, enforce_eager=True) as llm: - model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 - layer = model.model.layers[0] - - qkv_proj = layer.self_attn.qkv_proj - o_proj = layer.self_attn.o_proj - gate_up_proj = layer.mlp.gate_up_proj - down_proj = layer.mlp.down_proj - - # assert zp for symmetric and asymmetric cases - def zp_valid(zp: Optional[torch.Tensor]): - if is_symmetric: - return zp is None - - return zp is not None and zp.dtype is torch.int32 - - assert zp_valid(qkv_proj.input_zero_point) - assert zp_valid(o_proj.input_zero_point) - assert zp_valid(gate_up_proj.input_zero_point) - assert zp_valid(down_proj.input_zero_point) - - assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod) - assert isinstance(o_proj.quant_method, CompressedTensorsLinearMethod) - assert isinstance(gate_up_proj.quant_method, - CompressedTensorsLinearMethod) - assert isinstance(down_proj.quant_method, - CompressedTensorsLinearMethod) - assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8Int8) - - assert qkv_proj.scheme.strategy == strategy - assert qkv_proj.scheme.is_static_input_scheme - expected_type = torch.int8 - - assert qkv_proj.weight.dtype is expected_type - assert o_proj.weight.dtype is expected_type - assert gate_up_proj.weight.dtype is expected_type - - if qkv_proj.scheme.strategy == "tensor": - # Make sure it is a channelwise buffer - # After running process_weights_after_loading - assert len(qkv_proj.weight_scale.shape) == 2 - assert qkv_proj.weight_scale.shape[0] == shape_0 - assert qkv_proj.weight_scale.shape[1] == 1 - assert qkv_proj.weight_scale.dtype is torch.float32 - assert qkv_proj.input_scale.dtype is torch.float32 + + def check_model(model): + layer = model.model.layers[0] + + qkv_proj = layer.self_attn.qkv_proj + o_proj = layer.self_attn.o_proj + gate_up_proj = layer.mlp.gate_up_proj + down_proj = layer.mlp.down_proj + + # assert zp for symmetric and asymmetric cases + def zp_valid(zp: Optional[torch.Tensor]): + if is_symmetric: + return zp is None + + return zp is not None and zp.dtype is torch.int32 + + assert zp_valid(qkv_proj.input_zero_point) + assert zp_valid(o_proj.input_zero_point) + assert zp_valid(gate_up_proj.input_zero_point) + assert zp_valid(down_proj.input_zero_point) + + assert isinstance(qkv_proj.quant_method, + CompressedTensorsLinearMethod) + assert isinstance(o_proj.quant_method, + CompressedTensorsLinearMethod) + assert isinstance(gate_up_proj.quant_method, + CompressedTensorsLinearMethod) + assert isinstance(down_proj.quant_method, + CompressedTensorsLinearMethod) + assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8Int8) + + assert qkv_proj.scheme.strategy == strategy + assert qkv_proj.scheme.is_static_input_scheme + expected_type = torch.int8 + + assert qkv_proj.weight.dtype is expected_type + assert o_proj.weight.dtype is expected_type + assert gate_up_proj.weight.dtype is expected_type + + if qkv_proj.scheme.strategy == "tensor": + # Make sure it is a channelwise buffer + # After running process_weights_after_loading + assert len(qkv_proj.weight_scale.shape) == 2 + assert qkv_proj.weight_scale.shape[0] == shape_0 + assert qkv_proj.weight_scale.shape[1] == 1 + assert qkv_proj.weight_scale.dtype is torch.float32 + assert qkv_proj.input_scale.dtype is torch.float32 + + llm.apply_model(check_model) output = llm.generate_greedy(["Hello my name is"], max_tokens=20) assert output @@ -129,16 +134,20 @@ def test_compressed_tensors_no_enforce_eager(vllm_runner): def test_compressed_tensors_w8a8_dynamic_per_token(vllm_runner, model_args): model_path, strategy = model_args with vllm_runner(model_path, dtype=torch.float16) as llm: - model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 - layer = model.model.layers[0] - qkv_proj = layer.self_attn.qkv_proj + def check_model(model): + layer = model.model.layers[0] - assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod) - assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8Int8) - assert not qkv_proj.scheme.is_static_input_scheme - assert qkv_proj.scheme.strategy == strategy - assert qkv_proj.weight.dtype is torch.int8 + qkv_proj = layer.self_attn.qkv_proj + + assert isinstance(qkv_proj.quant_method, + CompressedTensorsLinearMethod) + assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8Int8) + assert not qkv_proj.scheme.is_static_input_scheme + assert qkv_proj.scheme.strategy == strategy + assert qkv_proj.weight.dtype is torch.int8 + + llm.apply_model(check_model) output = llm.generate_greedy(["Hello my name is"], max_tokens=20) assert output @@ -152,19 +161,24 @@ def test_compressed_tensors_w8a8_dynamic_per_token(vllm_runner, model_args): def test_compressed_tensors_wNa16(vllm_runner, wNa16_args): model, strategy, group, pack_factor = wNa16_args with vllm_runner(model) as llm: - model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 - layer = model.model.layers[0] - qkv_proj = layer.self_attn.qkv_proj - assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod) - assert isinstance(qkv_proj.scheme, CompressedTensorsWNA16) + def check_model(model): + layer = model.model.layers[0] + + qkv_proj = layer.self_attn.qkv_proj + assert isinstance(qkv_proj.quant_method, + CompressedTensorsLinearMethod) + assert isinstance(qkv_proj.scheme, CompressedTensorsWNA16) - assert qkv_proj.scheme.strategy == strategy - assert qkv_proj.scheme.group_size == (-1 if group is None else group) + assert qkv_proj.scheme.strategy == strategy + assert qkv_proj.scheme.group_size == (-1 + if group is None else group) - assert qkv_proj.weight_packed.dtype is torch.int32 - assert qkv_proj.weight_scale.dtype is torch.float16 - assert qkv_proj.scheme.pack_factor == pack_factor + assert qkv_proj.weight_packed.dtype is torch.int32 + assert qkv_proj.weight_scale.dtype is torch.float16 + assert qkv_proj.scheme.pack_factor == pack_factor + + llm.apply_model(check_model) output = llm.generate_greedy("Hello my name is", max_tokens=20) assert output @@ -173,14 +187,18 @@ def test_compressed_tensors_wNa16(vllm_runner, wNa16_args): def test_compressed_tensors_w4a16_marlin24(vllm_runner): model_path = "nm-testing/llama7b-one-shot-2_4-w4a16-marlin24-t" with vllm_runner(model_path) as llm: - model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 - layer = model.model.layers[0] - qkv_proj = layer.self_attn.qkv_proj + def check_model(model): + layer = model.model.layers[0] + + qkv_proj = layer.self_attn.qkv_proj + + assert isinstance(qkv_proj.quant_method, + CompressedTensorsLinearMethod) + assert isinstance(qkv_proj.scheme, CompressedTensorsW4A16Sparse24) + assert qkv_proj.weight_packed.dtype is torch.int32 - assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod) - assert isinstance(qkv_proj.scheme, CompressedTensorsW4A16Sparse24) - assert qkv_proj.weight_packed.dtype is torch.int32 + llm.apply_model(check_model) output = llm.generate_greedy("Hello my name is", max_tokens=20) assert output @@ -189,23 +207,27 @@ def test_compressed_tensors_w4a16_marlin24(vllm_runner): def test_compressed_tensors_fp8(vllm_runner): model_path = "nm-testing/Meta-Llama-3-8B-FP8-compressed-tensors-test" with vllm_runner(model_path) as llm: - model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 - layer = model.model.layers[0] - qkv_proj = layer.self_attn.qkv_proj + def check_model(model): + layer = model.model.layers[0] - assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod) - assert isinstance( - qkv_proj.scheme, - (CompressedTensorsW8A8Fp8, CompressedTensorsW8A16Fp8)) + qkv_proj = layer.self_attn.qkv_proj - assert qkv_proj.input_scale.dtype is torch.float32 + assert isinstance(qkv_proj.quant_method, + CompressedTensorsLinearMethod) + assert isinstance( + qkv_proj.scheme, + (CompressedTensorsW8A8Fp8, CompressedTensorsW8A16Fp8)) - if isinstance(qkv_proj.scheme, CompressedTensorsW8A8Fp8): - assert len(qkv_proj.input_scale.shape) == 0 - assert qkv_proj.weight.dtype is torch.float8_e4m3fn - assert qkv_proj.weight_scale.dtype is torch.float32 - assert len(qkv_proj.weight_scale.shape) == 0 + assert qkv_proj.input_scale.dtype is torch.float32 + + if isinstance(qkv_proj.scheme, CompressedTensorsW8A8Fp8): + assert len(qkv_proj.input_scale.shape) == 0 + assert qkv_proj.weight.dtype is torch.float8_e4m3fn + assert qkv_proj.weight_scale.dtype is torch.float32 + assert len(qkv_proj.weight_scale.shape) == 0 + + llm.apply_model(check_model) output = llm.generate_greedy("Hello my name is", max_tokens=20) assert output @@ -248,12 +270,15 @@ def _test_2of4_quant_models(qkv_proj, weight_strategy, input_strategy): def test_compressed_tensors_2of4_quant_fp8(vllm_runner, args_2of4): model, weight_strategy, input_strategy = args_2of4 with vllm_runner(model) as llm: - model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 - layer = model.model.layers[0] - qkv_proj = layer.self_attn.qkv_proj - assert qkv_proj.scheme.weights_dtype == torch.float8_e4m3fn - _test_2of4_quant_models(qkv_proj, weight_strategy, input_strategy) + def check_model(model): + layer = model.model.layers[0] + + qkv_proj = layer.self_attn.qkv_proj + assert qkv_proj.scheme.weights_dtype == torch.float8_e4m3fn + _test_2of4_quant_models(qkv_proj, weight_strategy, input_strategy) + + llm.apply_model(check_model) output = llm.generate_greedy("Hello my name is", max_tokens=20) print(output) @@ -273,40 +298,49 @@ def test_compressed_tensors_2of4_quant_fp8(vllm_runner, args_2of4): def test_compressed_tensors_2of4_quant_int8(vllm_runner, args_2of4): model, weight_strategy, input_strategy = args_2of4 with vllm_runner(model) as llm: - model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 - layer = model.model.layers[0] - qkv_proj = layer.self_attn.qkv_proj - assert qkv_proj.scheme.weights_dtype == torch.int8 - _test_2of4_quant_models(qkv_proj, weight_strategy, input_strategy) + def check_model(model): + layer = model.model.layers[0] + + qkv_proj = layer.self_attn.qkv_proj + assert qkv_proj.scheme.weights_dtype == torch.int8 + _test_2of4_quant_models(qkv_proj, weight_strategy, input_strategy) + + llm.apply_model(check_model) output = llm.generate_greedy("Hello my name is", max_tokens=20) print(output) assert output +@pytest.mark.skip(reason="2of4 sparse w16a16 CUTLASS produces bad output.") @pytest.mark.skipif(not sparse_cutlass_supported(), - reason="Sparse FP8 is not yet supported on this GPU type.") + reason="2of4 Sparse is not yet supported on this GPU type." + ) @pytest.mark.parametrize( "args_2of4", [("nm-testing/TinyLlama-1.1B-Chat-v1.0-2of4-Sparse-Dense-Compressor")]) def test_compressed_tensors_2of4_sparse(vllm_runner, args_2of4): model = args_2of4 with vllm_runner(model) as llm: - model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 - layer = model.model.layers[0] - - qkv_proj = layer.self_attn.qkv_proj - assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod) - assert isinstance(qkv_proj.scheme, CompressedTensors24) - - assert qkv_proj.scheme.weight_quant is None - assert qkv_proj.scheme.input_quant is None - assert not qkv_proj.scheme.quantized - assert qkv_proj.quant_method.quantization_config.sparsity_scheme_map - sparsity_map = qkv_proj.quant_method.quantization_config.sparsity_scheme_map # noqa: E501 - assert sparsity_map.get("Linear").format == "dense" - assert sparsity_map.get("Linear").sparsity_structure == "2:4" + + def check_model(model): + layer = model.model.layers[0] + + qkv_proj = layer.self_attn.qkv_proj + assert isinstance(qkv_proj.quant_method, + CompressedTensorsLinearMethod) + assert isinstance(qkv_proj.scheme, CompressedTensors24) + + assert qkv_proj.scheme.weight_quant is None + assert qkv_proj.scheme.input_quant is None + assert not qkv_proj.scheme.quantized + assert qkv_proj.quant_method.quantization_config.sparsity_scheme_map + sparsity_map = qkv_proj.quant_method.quantization_config.sparsity_scheme_map # noqa: E501 + assert sparsity_map.get("Linear").format == "dense" + assert sparsity_map.get("Linear").sparsity_structure == "2:4" + + llm.apply_model(check_model) output = llm.generate_greedy("Hello my name is", max_tokens=20) print(output) diff --git a/tests/quantization/test_fp8.py b/tests/quantization/test_fp8.py index a0c1d7e24c503..4bff734746297 100644 --- a/tests/quantization/test_fp8.py +++ b/tests/quantization/test_fp8.py @@ -49,13 +49,17 @@ def test_model_load_and_run(vllm_runner, model_id: str, force_marlin: bool, def test_kv_cache_model_load_and_run(vllm_runner, model_id: str): with vllm_runner(model_id, kv_cache_dtype="fp8") as llm: - model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 - attn = model.model.layers[0].self_attn.attn - assert isinstance(attn.quant_method, Fp8KVCacheMethod) - # NOTE: it is valid for scales to be 1.0 (default value), but we know - # these checkpoints have scales < 1.0 - assert 0.0 < attn._k_scale < 1.0 - assert 0.0 < attn._v_scale < 1.0 + def check_model(model): + attn = model.model.layers[0].self_attn.attn + + assert isinstance(attn.quant_method, Fp8KVCacheMethod) + + # NOTE: it is valid for scales to be 1.0 (default value), but + # we know these checkpoints have scales < 1.0 + assert 0.0 < attn._k_scale < 1.0 + assert 0.0 < attn._v_scale < 1.0 + + llm.apply_model(check_model) # note: this does not test accuracy, just that we can run through # see lm-eval tests for accuracy @@ -77,22 +81,24 @@ def test_load_fp16_model(vllm_runner, kv_cache_dtype: str, force_marlin: bool, quantization="fp8", kv_cache_dtype=kv_cache_dtype) as llm: - model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 - fc1 = model.model.decoder.layers[0].fc1 - assert isinstance(fc1.quant_method, Fp8LinearMethod) - if kv_cache_dtype == "fp8": - attn = model.model.decoder.layers[0].self_attn.attn - assert isinstance(attn.quant_method, Fp8KVCacheMethod) - assert attn._k_scale == 1.0 - assert attn._v_scale == 1.0 - - if current_platform.has_device_capability(89) and not force_marlin: - # For GPUs with hardware support, we keep weights in fp8 - assert fc1.weight.dtype == torch.float8_e4m3fn - else: - # For GPUs without hardware support, we pack the fp8 weights - # for weight-only quantization using Marlin kernels - assert fc1.weight.dtype == torch.int32 + def check_model(model): + fc1 = model.model.decoder.layers[0].fc1 + assert isinstance(fc1.quant_method, Fp8LinearMethod) + if kv_cache_dtype == "fp8": + attn = model.model.decoder.layers[0].self_attn.attn + assert isinstance(attn.quant_method, Fp8KVCacheMethod) + assert attn._k_scale == 1.0 + assert attn._v_scale == 1.0 + + if current_platform.has_device_capability(89) and not force_marlin: + # For GPUs with hardware support, we keep weights in fp8 + assert fc1.weight.dtype == torch.float8_e4m3fn + else: + # For GPUs without hardware support, we pack the fp8 weights + # for weight-only quantization using Marlin kernels + assert fc1.weight.dtype == torch.int32 + + llm.apply_model(check_model) @pytest.mark.skipif(not is_quant_method_supported("fp8"), diff --git a/tests/quantization/test_lm_head.py b/tests/quantization/test_lm_head.py index ad526a4065101..fa2d9645ea47f 100644 --- a/tests/quantization/test_lm_head.py +++ b/tests/quantization/test_lm_head.py @@ -28,20 +28,23 @@ def test_lm_head( model_lm_head_quant: Tuple[str, bool], ) -> None: model, lm_head_quantized = model_lm_head_quant - vllm_model = vllm_runner(model, dtype=torch.float16, max_model_len=2048) - - lm_head_layer = (vllm_model.model.llm_engine.model_executor.driver_worker. - model_runner.model.lm_head) - - if lm_head_quantized: - assert isinstance( - lm_head_layer.linear_method, - (GPTQLinearMethod, GPTQMarlinLinearMethod, MarlinLinearMethod)) - else: - assert isinstance(lm_head_layer.linear_method, - UnquantizedEmbeddingMethod) - - print( - vllm_model.generate_greedy(prompts=["Hello my name is"], - max_tokens=10)[0][1]) - del vllm_model + + with vllm_runner(model, dtype=torch.float16, + max_model_len=2048) as vllm_model: + + def check_model(model): + lm_head_layer = model.lm_head + + if lm_head_quantized: + assert isinstance(lm_head_layer.linear_method, + (GPTQLinearMethod, GPTQMarlinLinearMethod, + MarlinLinearMethod)) + else: + assert isinstance(lm_head_layer.linear_method, + UnquantizedEmbeddingMethod) + + vllm_model.apply_model(check_model) + + print( + vllm_model.generate_greedy(prompts=["Hello my name is"], + max_tokens=10)[0][1]) diff --git a/tests/quantization/test_quark.py b/tests/quantization/test_quark.py new file mode 100644 index 0000000000000..11382ad708faa --- /dev/null +++ b/tests/quantization/test_quark.py @@ -0,0 +1,33 @@ +"""Test model set-up and weight loading for quark-quantized models. + +Run `pytest tests/quantization/test_quark.py`. +""" + +import torch + +from vllm.model_executor.layers.quantization.quark.quark import ( # noqa: E501 + QuarkLinearMethod, QuarkW8A8Fp8) + + +def test_quark_fp8(vllm_runner): + model_path = "amd/Llama-3.1-8B-Instruct-FP8-KV-Quark-test" + with vllm_runner(model_path) as llm: + + def check_model(model): + layer = model.model.layers[0] + + qkv_proj = layer.self_attn.qkv_proj + + assert isinstance(qkv_proj.quant_method, QuarkLinearMethod) + assert isinstance(qkv_proj.scheme, QuarkW8A8Fp8) + + if isinstance(qkv_proj.scheme, QuarkW8A8Fp8): + assert len(qkv_proj.input_scale.shape) == 0 + assert qkv_proj.weight.dtype is torch.float8_e4m3fn + #assert qkv_proj.weight.dtype is torch.float8_e4m3fnuz + assert len(qkv_proj.weight_scale.shape) == 0 + + llm.apply_model(check_model) + + output = llm.generate_greedy("Hello my name is", max_tokens=20) + assert output diff --git a/tests/quantization/test_register_quantization_config.py b/tests/quantization/test_register_quantization_config.py new file mode 100644 index 0000000000000..8e7f44a399ddf --- /dev/null +++ b/tests/quantization/test_register_quantization_config.py @@ -0,0 +1,117 @@ +"""Tests register custom quantization config. + +See https://github.com/vllm-project/vllm/issues/11926 for more details. + +Run `pytest tests/quantization/test_register_quantization_config.py`. +""" +from typing import Any, Dict, List, Optional + +import pytest +import torch +import torch.nn.functional as F + +from vllm.model_executor.layers.linear import LinearBase # noqa: E501 +from vllm.model_executor.layers.linear import UnquantizedLinearMethod +from vllm.model_executor.layers.quantization import ( + get_quantization_config, register_quantization_config) +from vllm.model_executor.layers.quantization.base_config import ( # noqa: E501 + QuantizationConfig) + + +class FakeQuantLinearMethod(UnquantizedLinearMethod): + """Fake quantization linear method for per-token dynamic quantization.""" + + def __init__(self, num_bits: int = 8) -> None: + """Initialize the quantization method.""" + super().__init__() + self.num_bits = num_bits + + def apply(self, + layer: "torch.nn.Module", + x: "torch.Tensor", + bias: Optional["torch.Tensor"] = None) -> "torch.Tensor": + """Perform fake quantization before the linear layer.""" + + # Calculate the scales dynamically + max_val = torch.amax(x, dim=(0, -1), keepdims=True) + min_val = torch.amin(x, dim=(0, -1), keepdims=True) + scales = (max_val - min_val) / (2**self.num_bits - 1) + + # Fake quantize the input + quant_x = torch.clamp(torch.round(x / scales), -2**(self.num_bits - 1), + 2**(self.num_bits - 1) - 1) + dequant_x = quant_x * scales + + return F.linear(dequant_x, layer.weight, bias) + + +@register_quantization_config("custom_quant") +class CustomQuantConfig(QuantizationConfig): + """Custom quantization config for per-token dynamic fake quantization.""" + + def __init__(self, num_bits: int = 8) -> None: + """Initialize the quantization config.""" + self.num_bits = num_bits + + def get_name(self) -> str: + """Name of the quantization method.""" + return "custom_quant" + + def get_supported_act_dtypes(self) -> List["torch.dtype"]: + """List of supported activation dtypes.""" + return [torch.float16, torch.bfloat16] + + @classmethod + def get_min_capability(cls) -> int: + """Minimum GPU capability to support the quantization method.""" + return -1 + + @staticmethod + def get_config_filenames() -> List[str]: + """List of filenames to search for in the model directory.""" + return [] + + @classmethod + def from_config(cls, config: Dict[str, Any]) -> "CustomQuantConfig": + """Create a config class from the model's quantization config.""" + return CustomQuantConfig(num_bits=config.get("num_bits", 8)) + + def get_quant_method(self, layer: "torch.nn.Module", + prefix: str) -> Optional["FakeQuantLinearMethod"]: + """Get the quantize method to use for the quantized layer.""" + if isinstance(layer, LinearBase): + return FakeQuantLinearMethod(num_bits=self.num_bits) + return None + + +def test_register_quantization_config(): + """Test register custom quantization config.""" + + # The quantization method `custom_quant` should be registered. + assert get_quantization_config("custom_quant") == CustomQuantConfig + + # The quantization method `custom_quant` is already exists, + # should raise an error. + with pytest.raises(ValueError): + register_quantization_config("custom_quant")(CustomQuantConfig) + + +@pytest.mark.parametrize(argnames="model", + argvalues=[ + "meta-llama/Meta-Llama-3-8B-Instruct", + ]) +def test_custom_quant(vllm_runner, model): + """Test infer with the custom quantization method.""" + with vllm_runner(model_name=model, + quantization="custom_quant", + enforce_eager=True) as llm: + + model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501 + layer = model.model.layers[0] + qkv_proj = layer.self_attn.qkv_proj + + # Check the quantization method is FakeQuantLinearMethod + assert isinstance(qkv_proj.quant_method, FakeQuantLinearMethod) + + output = llm.generate_greedy("Hello my name is", max_tokens=20) + assert output diff --git a/tests/samplers/test_seeded_generate.py b/tests/samplers/test_seeded_generate.py index 88067f19c8f07..bf1ee6c397838 100644 --- a/tests/samplers/test_seeded_generate.py +++ b/tests/samplers/test_seeded_generate.py @@ -31,7 +31,7 @@ def test_random_sample_with_seed( sampling_params = SamplingParams( # Parameters to ensure sufficient randomness - temperature=2.0, + temperature=3.0, top_p=min(random.random() + 0.3, 1), top_k=random.randint(5, 20), n=random.randint(1, 10), @@ -75,3 +75,8 @@ def test_random_sample_with_seed( # verify requests with the same seed match assert outputs[1] == outputs[4] assert outputs[2] == outputs[5] + + # verify generations within the same parallel sampling group differ + for output in outputs: + for sub_output_a, sub_output_b in combinations(output, 2): + assert sub_output_a != sub_output_b diff --git a/tests/tensorizer_loader/test_tensorizer.py b/tests/tensorizer_loader/test_tensorizer.py index bf409d2d97aa1..6e7eec1c6ab34 100644 --- a/tests/tensorizer_loader/test_tensorizer.py +++ b/tests/tensorizer_loader/test_tensorizer.py @@ -3,6 +3,7 @@ import os import pathlib import subprocess +from functools import partial from unittest.mock import MagicMock, patch import openai @@ -24,7 +25,6 @@ # yapf: enable from vllm.utils import PlaceholderModule, import_from_path -from ..conftest import VllmRunner from ..utils import VLLM_PATH, RemoteOpenAIServer from .conftest import retry_until_skip @@ -58,16 +58,6 @@ def is_curl_installed(): return False -def get_torch_model(vllm_runner: VllmRunner): - return vllm_runner \ - .model \ - .llm_engine \ - .model_executor \ - .driver_worker \ - .model_runner \ - .model - - def write_keyfile(keyfile_path: str): encryption_params = EncryptionParams.random() pathlib.Path(keyfile_path).parent.mkdir(parents=True, exist_ok=True) @@ -121,8 +111,10 @@ def test_deserialized_encrypted_vllm_model_has_same_outputs( config_for_serializing = TensorizerConfig(tensorizer_uri=model_path, encryption_keyfile=key_path) - serialize_vllm_model(get_torch_model(vllm_model), - config_for_serializing) + + vllm_model.apply_model( + partial(serialize_vllm_model, + tensorizer_config=config_for_serializing)) config_for_deserializing = TensorizerConfig(tensorizer_uri=model_path, encryption_keyfile=key_path) @@ -175,8 +167,10 @@ def test_vllm_model_can_load_with_lora(vllm_runner, tmp_path): with vllm_runner(model_ref, ) as vllm_model: model_path = tmp_path / (model_ref + ".tensors") - serialize_vllm_model(get_torch_model(vllm_model), - TensorizerConfig(tensorizer_uri=model_path)) + vllm_model.apply_model( + partial( + serialize_vllm_model, + tensorizer_config=TensorizerConfig(tensorizer_uri=model_path))) with vllm_runner( model_ref, @@ -215,8 +209,10 @@ def test_openai_apiserver_with_tensorizer(vllm_runner, tmp_path): with vllm_runner(model_ref, ) as vllm_model: model_path = tmp_path / (model_ref + ".tensors") - serialize_vllm_model(get_torch_model(vllm_model), - TensorizerConfig(tensorizer_uri=model_path)) + vllm_model.apply_model( + partial( + serialize_vllm_model, + tensorizer_config=TensorizerConfig(tensorizer_uri=model_path))) model_loader_extra_config = { "tensorizer_uri": str(model_path), @@ -337,7 +333,9 @@ def test_vllm_tensorized_model_has_same_outputs(vllm_runner, tmp_path): with vllm_runner(model_ref) as vllm_model: outputs = vllm_model.generate(prompts, sampling_params) - serialize_vllm_model(get_torch_model(vllm_model), config) + + vllm_model.apply_model( + partial(serialize_vllm_model, tensorizer_config=config)) assert is_vllm_tensorized(config) diff --git a/tests/test_utils.py b/tests/test_utils.py index 93c9a82f8529f..7ebc2c0b857e5 100644 --- a/tests/test_utils.py +++ b/tests/test_utils.py @@ -2,15 +2,17 @@ import os import socket from typing import AsyncIterator, Tuple +from unittest.mock import patch import pytest import torch +from vllm_test_utils import monitor from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config -from vllm.utils import (FlexibleArgumentParser, PlaceholderModule, - StoreBoolean, bind_kv_cache, deprecate_kwargs, - get_open_port, memory_profiling, merge_async_iterators, - supports_kw) +from vllm.utils import (FlexibleArgumentParser, MemorySnapshot, + PlaceholderModule, StoreBoolean, bind_kv_cache, + deprecate_kwargs, get_open_port, memory_profiling, + merge_async_iterators, supports_kw) from .utils import error_on_warning, fork_new_process_for_each_test @@ -299,17 +301,24 @@ def test_memory_profiling(): # 512 MiB allocation outside of this instance handle1 = lib.cudaMalloc(512 * 1024 * 1024) - baseline_memory_in_bytes = \ - torch.cuda.mem_get_info()[1] - torch.cuda.mem_get_info()[0] + baseline_snapshot = MemorySnapshot() # load weights weights = torch.randn(128, 1024, 1024, device='cuda', dtype=torch.float32) - weights_memory_in_bytes = 128 * 1024 * 1024 * 4 # 512 MiB + weights_memory = 128 * 1024 * 1024 * 4 # 512 MiB - with memory_profiling(baseline_memory_in_bytes=baseline_memory_in_bytes, - weights_memory_in_bytes=weights_memory_in_bytes) as result: + def measure_current_non_torch(): + free, total = torch.cuda.mem_get_info() + current_used = total - free + current_torch = torch.cuda.memory_reserved() + current_non_torch = current_used - current_torch + return current_non_torch + + with memory_profiling(baseline_snapshot=baseline_snapshot, + weights_memory=weights_memory) as result, \ + monitor(measure_current_non_torch) as monitored_values: # make a memory spike, 1 GiB spike = torch.randn(256, 1024, 1024, device='cuda', dtype=torch.float32) del spike @@ -317,11 +326,18 @@ def test_memory_profiling(): # Add some extra non-torch memory 256 MiB (simulate NCCL) handle2 = lib.cudaMalloc(256 * 1024 * 1024) + # this is an analytic value, it is exact, + # we only have 256 MiB non-torch memory increase + measured_diff = monitored_values.values[-1] - monitored_values.values[0] + assert measured_diff == 256 * 1024 * 1024 + # Check that the memory usage is within 5% of the expected values - non_torch_ratio = result.non_torch_increase_in_bytes / (256 * 1024 * 1024) # noqa - torch_peak_ratio = result.torch_peak_increase_in_bytes / (1024 * 1024 * 1024) # noqa + # 5% tolerance is caused by cuda runtime. + # we cannot control cuda runtime in the granularity of bytes, + # which causes a small error (<10 MiB in practice) + non_torch_ratio = result.non_torch_increase / (256 * 1024 * 1024) # noqa assert abs(non_torch_ratio - 1) <= 0.05 - assert abs(torch_peak_ratio - 1) <= 0.05 + assert result.torch_peak_increase == 1024 * 1024 * 1024 del weights lib.cudaFree(handle1) lib.cudaFree(handle2) @@ -390,7 +406,10 @@ def test_bind_kv_cache_encoder_decoder(): def test_bind_kv_cache_pp(): - cfg = VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=2)) + with patch("vllm.utils.cuda_device_count_stateless", lambda: 2): + # this test runs with 1 GPU, but we simulate 2 GPUs + cfg = VllmConfig( + parallel_config=ParallelConfig(pipeline_parallel_size=2)) with set_current_vllm_config(cfg): from vllm.attention import Attention diff --git a/tests/tracing/test_tracing.py b/tests/tracing/test_tracing.py index fe5fc979c66a3..49a16d16eb840 100644 --- a/tests/tracing/test_tracing.py +++ b/tests/tracing/test_tracing.py @@ -100,32 +100,32 @@ def test_traces(trace_service): attributes = decode_attributes( request.resource_spans[0].scope_spans[0].spans[0].attributes) - assert attributes.get(SpanAttributes.LLM_RESPONSE_MODEL) == model + assert attributes.get(SpanAttributes.GEN_AI_RESPONSE_MODEL) == model assert attributes.get( - SpanAttributes.LLM_REQUEST_ID) == outputs[0].request_id + SpanAttributes.GEN_AI_REQUEST_ID) == outputs[0].request_id + assert attributes.get(SpanAttributes.GEN_AI_REQUEST_TEMPERATURE + ) == sampling_params.temperature assert attributes.get( - SpanAttributes.LLM_REQUEST_TEMPERATURE) == sampling_params.temperature + SpanAttributes.GEN_AI_REQUEST_TOP_P) == sampling_params.top_p assert attributes.get( - SpanAttributes.LLM_REQUEST_TOP_P) == sampling_params.top_p - assert attributes.get( - SpanAttributes.LLM_REQUEST_MAX_TOKENS) == sampling_params.max_tokens - assert attributes.get(SpanAttributes.LLM_REQUEST_N) == sampling_params.n - assert attributes.get(SpanAttributes.LLM_USAGE_PROMPT_TOKENS) == len( + SpanAttributes.GEN_AI_REQUEST_MAX_TOKENS) == sampling_params.max_tokens + assert attributes.get(SpanAttributes.GEN_AI_REQUEST_N) == sampling_params.n + assert attributes.get(SpanAttributes.GEN_AI_USAGE_PROMPT_TOKENS) == len( outputs[0].prompt_token_ids) completion_tokens = sum(len(o.token_ids) for o in outputs[0].outputs) assert attributes.get( - SpanAttributes.LLM_USAGE_COMPLETION_TOKENS) == completion_tokens + SpanAttributes.GEN_AI_USAGE_COMPLETION_TOKENS) == completion_tokens metrics = outputs[0].metrics assert attributes.get( - SpanAttributes.LLM_LATENCY_TIME_IN_QUEUE) == metrics.time_in_queue + SpanAttributes.GEN_AI_LATENCY_TIME_IN_QUEUE) == metrics.time_in_queue ttft = metrics.first_token_time - metrics.arrival_time assert attributes.get( - SpanAttributes.LLM_LATENCY_TIME_TO_FIRST_TOKEN) == ttft + SpanAttributes.GEN_AI_LATENCY_TIME_TO_FIRST_TOKEN) == ttft e2e_time = metrics.finished_time - metrics.arrival_time - assert attributes.get(SpanAttributes.LLM_LATENCY_E2E) == e2e_time + assert attributes.get(SpanAttributes.GEN_AI_LATENCY_E2E) == e2e_time assert metrics.scheduler_time > 0 - assert attributes.get( - SpanAttributes.LLM_LATENCY_TIME_IN_SCHEDULER) == metrics.scheduler_time + assert attributes.get(SpanAttributes.GEN_AI_LATENCY_TIME_IN_SCHEDULER + ) == metrics.scheduler_time # Model forward and model execute should be none, since detailed traces is # not enabled. assert metrics.model_forward_time is None @@ -166,37 +166,37 @@ def test_traces_with_detailed_steps(trace_service): attributes = decode_attributes( request.resource_spans[0].scope_spans[0].spans[0].attributes) - assert attributes.get(SpanAttributes.LLM_RESPONSE_MODEL) == model + assert attributes.get(SpanAttributes.GEN_AI_RESPONSE_MODEL) == model assert attributes.get( - SpanAttributes.LLM_REQUEST_ID) == outputs[0].request_id + SpanAttributes.GEN_AI_REQUEST_ID) == outputs[0].request_id + assert attributes.get(SpanAttributes.GEN_AI_REQUEST_TEMPERATURE + ) == sampling_params.temperature assert attributes.get( - SpanAttributes.LLM_REQUEST_TEMPERATURE) == sampling_params.temperature + SpanAttributes.GEN_AI_REQUEST_TOP_P) == sampling_params.top_p assert attributes.get( - SpanAttributes.LLM_REQUEST_TOP_P) == sampling_params.top_p - assert attributes.get( - SpanAttributes.LLM_REQUEST_MAX_TOKENS) == sampling_params.max_tokens - assert attributes.get(SpanAttributes.LLM_REQUEST_N) == sampling_params.n - assert attributes.get(SpanAttributes.LLM_USAGE_PROMPT_TOKENS) == len( + SpanAttributes.GEN_AI_REQUEST_MAX_TOKENS) == sampling_params.max_tokens + assert attributes.get(SpanAttributes.GEN_AI_REQUEST_N) == sampling_params.n + assert attributes.get(SpanAttributes.GEN_AI_USAGE_PROMPT_TOKENS) == len( outputs[0].prompt_token_ids) completion_tokens = sum(len(o.token_ids) for o in outputs[0].outputs) assert attributes.get( - SpanAttributes.LLM_USAGE_COMPLETION_TOKENS) == completion_tokens + SpanAttributes.GEN_AI_USAGE_COMPLETION_TOKENS) == completion_tokens metrics = outputs[0].metrics assert attributes.get( - SpanAttributes.LLM_LATENCY_TIME_IN_QUEUE) == metrics.time_in_queue + SpanAttributes.GEN_AI_LATENCY_TIME_IN_QUEUE) == metrics.time_in_queue ttft = metrics.first_token_time - metrics.arrival_time assert attributes.get( - SpanAttributes.LLM_LATENCY_TIME_TO_FIRST_TOKEN) == ttft + SpanAttributes.GEN_AI_LATENCY_TIME_TO_FIRST_TOKEN) == ttft e2e_time = metrics.finished_time - metrics.arrival_time - assert attributes.get(SpanAttributes.LLM_LATENCY_E2E) == e2e_time + assert attributes.get(SpanAttributes.GEN_AI_LATENCY_E2E) == e2e_time assert metrics.scheduler_time > 0 - assert attributes.get( - SpanAttributes.LLM_LATENCY_TIME_IN_SCHEDULER) == metrics.scheduler_time + assert attributes.get(SpanAttributes.GEN_AI_LATENCY_TIME_IN_SCHEDULER + ) == metrics.scheduler_time assert metrics.model_forward_time > 0 assert attributes.get( - SpanAttributes.LLM_LATENCY_TIME_IN_MODEL_FORWARD) == pytest.approx( + SpanAttributes.GEN_AI_LATENCY_TIME_IN_MODEL_FORWARD) == pytest.approx( metrics.model_forward_time / 1000) assert metrics.model_execute_time > 0 - assert attributes.get(SpanAttributes.LLM_LATENCY_TIME_IN_MODEL_EXECUTE + assert attributes.get(SpanAttributes.GEN_AI_LATENCY_TIME_IN_MODEL_EXECUTE ) == metrics.model_execute_time assert metrics.model_forward_time < 1000 * metrics.model_execute_time diff --git a/tests/v1/core/test_prefix_caching.py b/tests/v1/core/test_prefix_caching.py index b97f55b8c6535..f434fa8c61a80 100644 --- a/tests/v1/core/test_prefix_caching.py +++ b/tests/v1/core/test_prefix_caching.py @@ -49,9 +49,10 @@ def test_prefill(): unique_token_ids = [3] * 7 all_token_ids = common_token_ids + unique_token_ids req0 = make_request("0", all_token_ids) - computed_blocks = manager.get_computed_blocks(req0) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req0) assert len(req0.kv_block_hashes) == 3 assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req0, 55, computed_blocks) assert [b.block_id for b in blocks] == [0, 1, 2, 3, 4] @@ -73,9 +74,10 @@ def test_prefill(): # Incomplete 1 block (5 tokens) unique_token_ids = [3] * 5 req1 = make_request("1", common_token_ids + unique_token_ids) - computed_blocks = manager.get_computed_blocks(req1) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req1) assert len(req1.kv_block_hashes) == 3 assert [b.block_id for b in computed_blocks] == [0, 1, 2] + assert num_computed_tokens == 3 * 16 num_new_tokens = 53 - 3 * 16 blocks = manager.allocate_slots(req1, num_new_tokens, computed_blocks) assert [b.block_id for b in blocks] == [5, 6] @@ -91,7 +93,7 @@ def test_prefill(): # All blocks should be available. assert manager.free_block_queue.num_free_blocks == 10 # The order should be - # [unallocated (7, 8)] + # [unallocated (7, 8, 9)] # [unique_req0 (4, 3)] # [unique_req1 (6, 5)] # [common (2, 1, 0)] @@ -103,9 +105,10 @@ def test_prefill(): # Incomplete 1 block (6 tokens) unique_token_ids = [3] * 6 req2 = make_request("2", common_token_ids + unique_token_ids) - computed_blocks = manager.get_computed_blocks(req2) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req2) assert len(req2.kv_block_hashes) == 3 assert [b.block_id for b in computed_blocks] == [0, 1, 2] + assert num_computed_tokens == 3 * 16 num_new_tokens = 53 - 3 * 16 blocks = manager.allocate_slots(req2, num_new_tokens, computed_blocks) assert [b.block_id for b in blocks] == [7, 8] @@ -123,8 +126,9 @@ def test_prefill(): # Cache miss and eviction. req3 = make_request("3", [99] * (16 * 9)) - computed_blocks = manager.get_computed_blocks(req3) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req3) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req3, 16 * 9, computed_blocks) # This block ID order also checks the eviction order. assert [b.block_id for b in blocks] == [9, 4, 3, 6, 5, 8, 7, 2, 1, 0] @@ -150,8 +154,9 @@ def test_decode(): # Incomplete 1 block (7 tokens) unique_token_ids = [3] * 7 req0 = make_request("0", common_token_ids + unique_token_ids) - computed_blocks = manager.get_computed_blocks(req0) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req0) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req0, 55, computed_blocks) assert [b.block_id for b in blocks] == [0, 1, 2, 3, 4] @@ -197,16 +202,18 @@ def test_evict(): last_token_id = 5 * 16 + 7 req0 = make_request("0", list(range(last_token_id))) - computed_blocks = manager.get_computed_blocks(req0) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req0) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req0, 5 * 16 + 7, computed_blocks) assert len(blocks) == 7 # 5 full + 1 partial + 1 preallocated # 3 blocks. req1 = make_request("1", list(range(last_token_id, last_token_id + 3 * 16))) - computed_blocks = manager.get_computed_blocks(req1) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req1) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req1, 3 * 16, computed_blocks) assert len(blocks) == 3 # 3 full blocks last_token_id += 3 * 16 @@ -222,8 +229,9 @@ def test_evict(): # Touch the first 2 blocks. req2 = make_request("2", list(range(2 * 16 + 3))) - computed_blocks = manager.get_computed_blocks(req2) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req2) assert [b.block_id for b in computed_blocks] == [0, 1] + assert num_computed_tokens == 2 * 16 blocks = manager.allocate_slots(req2, 3, computed_blocks) assert [b.block_id for b in blocks] == [6, 5] assert manager.free_block_queue.num_free_blocks == 6 @@ -247,8 +255,9 @@ def test_hash_block_correct_reuse(): # Allocate 1 block and cache it. num_tokens = block_size * 1 req = make_request("0", list(range(num_tokens))) - computed_blocks = manager.get_computed_blocks(req) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req, num_tokens, computed_blocks) assert len(blocks) == 1 @@ -258,8 +267,9 @@ def test_hash_block_correct_reuse(): # Allocate a new block that's not full, make sure hash info on the # block is cleared. req = make_request("1", list(range(num_tokens - 1))) - computed_blocks = manager.get_computed_blocks(req) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req, num_tokens - 1, computed_blocks) assert len(blocks) == 1 @@ -284,16 +294,18 @@ def test_computed_blocks_not_evicted(): # Allocate a block and cache it. num_tokens = block_size * 1 req0 = make_request("0", list(range(num_tokens))) - computed_blocks = manager.get_computed_blocks(req0) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req0) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req0, num_tokens, computed_blocks) assert len(blocks) == 1 assert blocks[0].block_id == 0 # Allocate another block. req1 = make_request("1", list(range(num_tokens, num_tokens * 2))) - computed_blocks = manager.get_computed_blocks(req1) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req1) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req1, num_tokens, computed_blocks) assert len(blocks) == 1 assert blocks[0].block_id == 1 @@ -305,9 +317,10 @@ def test_computed_blocks_not_evicted(): # Now if we have a cache hit on the first block, we should evict the second # cached block rather than the first one. req2 = make_request("2", list(range(num_tokens * 2))) - computed_blocks = manager.get_computed_blocks(req2) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req2) assert len(computed_blocks) == 1 assert computed_blocks[0].block_id == 0 + assert num_computed_tokens == block_size blocks = manager.allocate_slots(req2, num_tokens * 2 - num_tokens, computed_blocks) @@ -331,8 +344,9 @@ def test_basic_prefix_caching_disabled(): req1 = make_request("1", list(range(10))) # 2 blocks and some more - computed_blocks = manager.get_computed_blocks(req1) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req1) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req1, 10, computed_blocks) assert len(blocks) == 3 @@ -341,15 +355,17 @@ def test_basic_prefix_caching_disabled(): # No caching. req2 = make_request("2", list(range(16))) # shared prefix - computed_blocks = manager.get_computed_blocks(req2) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req2) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req2, 16, computed_blocks) assert len(blocks) == 4 # New requests should not have any blocks. req3 = make_request("3", list(range(4))) - computed_blocks = manager.get_computed_blocks(req3) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req3) assert not computed_blocks + assert num_computed_tokens == 0 blocks = manager.allocate_slots(req3, 4, computed_blocks) assert not blocks @@ -371,8 +387,9 @@ def test_preallocate_blocks(num_preallocate_tokens: int, block_size: int): num_preallocated_blocks = cdiv(num_preallocate_tokens, block_size) req = make_request("0", list(range(block_size * 30))) - computed_blocks = manager.get_computed_blocks(req) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req) assert not computed_blocks + assert num_computed_tokens == 0 # Just ask for 1 block. blocks = manager.allocate_slots(req, block_size, computed_blocks) req.num_computed_tokens = block_size @@ -469,10 +486,11 @@ def test_mm_prefix_caching(): all_token_ids, mm_positions=mm_positions, mm_hashes=mm_hashes) - computed_blocks = manager.get_computed_blocks(req0) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req0) # Completed block should have hashes with extra keys. assert not computed_blocks + assert num_computed_tokens == 0 assert len(req0.kv_block_hashes) == 3 assert req0.kv_block_hashes[0].extra_keys == ("aaa", ) assert req0.kv_block_hashes[1].extra_keys == ("aaa", "bbb") @@ -503,8 +521,9 @@ def test_mm_prefix_caching(): all_token_ids, mm_positions=mm_positions, mm_hashes=mm_hashes) - computed_blocks = manager.get_computed_blocks(req1) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req1) assert len(computed_blocks) == 3 + assert num_computed_tokens == 3 * 16 def test_prefill_not_enough_free_blocks_with_computed_blocks(): @@ -527,15 +546,17 @@ def test_prefill_not_enough_free_blocks_with_computed_blocks(): # | Common-0 | Common-1 | Common-2 | ... | common_token_ids = [i for i in range(3) for _ in range(16)] req0 = make_request("0", common_token_ids) - computed_blocks = manager.get_computed_blocks(req0) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req0) assert not computed_blocks + assert num_computed_tokens == 0 manager.allocate_slots(req0, 48, computed_blocks) block_part0 = manager.req_to_blocks[req0.request_id] # | Common-0 | Common-1 | Common-2 | Req1-3 | Req1-4 | Req1-5 | ... | req1 = make_request("1", common_token_ids * 2) - computed_blocks = manager.get_computed_blocks(req1) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req1) assert computed_blocks == block_part0 + assert num_computed_tokens == 3 * 16 manager.allocate_slots(req1, 48, computed_blocks) block_part1 = manager.req_to_blocks[req1.request_id] # | Common-0 | Common-1 | Common-2 | Req1-3 (F) | Req1-4 (F) | @@ -547,8 +568,9 @@ def test_prefill_not_enough_free_blocks_with_computed_blocks(): # | Common-0 | Common-1 | Common-2 | Req1-3 (F) | Req1-4 (F) | # | Req1-5(F)| Req2-0 | Req2-1 | ... | req2 = make_request("2", [7] * block_size * 2) - computed_blocks = manager.get_computed_blocks(req2) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req2) assert not computed_blocks + assert num_computed_tokens == 0 manager.allocate_slots(req2, block_size * 2, computed_blocks) # Req3 is Req2 + 3 new blocks, so the first 6 blocks are computed, @@ -556,11 +578,81 @@ def test_prefill_not_enough_free_blocks_with_computed_blocks(): # In this case, the ref_cnt of the computed blocks should not be changed. assert manager.free_block_queue.num_free_blocks == 5 req3 = make_request("3", common_token_ids * 3) - computed_blocks = manager.get_computed_blocks(req3) + computed_blocks, num_computed_tokens = manager.get_computed_blocks(req3) assert computed_blocks == block_part1 + assert num_computed_tokens == 6 * 16 # Req3 cannot be allocated. assert manager.allocate_slots(req3, 48, computed_blocks) is None # Block 0-2 are used by Req 1. assert {block.ref_cnt for block in block_part1[:3]} == {1} # Block 3-5 are free. assert {block.ref_cnt for block in block_part1[3:]} == {0} + + +def test_reset_prefix_cache(): + manager = KVCacheManager( + block_size=16, + num_gpu_blocks=10, + max_model_len=8192, + sliding_window=None, + enable_caching=True, + num_preallocate_tokens=0, + ) + + full_block_token_ids = [i for i in range(3) for _ in range(16)] + unique_token_ids = [3] * 7 + all_token_ids = full_block_token_ids + unique_token_ids + req0 = make_request("0", all_token_ids) + blocks = manager.allocate_slots(req0, 55, []) + assert [b.block_id for b in blocks] == [0, 1, 2, 3] + + unique_token_ids = [4] * 7 + all_token_ids = full_block_token_ids + unique_token_ids + req1 = make_request("1", all_token_ids) + computed_blocks, _ = manager.get_computed_blocks(req1) + assert len(req1.kv_block_hashes) == 3 + assert len(computed_blocks) == 3 + blocks = manager.allocate_slots(req1, 7, computed_blocks) + assert [b.block_id for b in blocks] == [4] + + # Failed to reset prefix cache because some blocks are not freed yet. + assert not manager.reset_prefix_cache() + assert manager.cached_block_hash_to_block + + # Free the blocks. + manager.free(req0) + manager.free(req1) + + assert manager.reset_prefix_cache() + assert not manager.cached_block_hash_to_block + assert all([blk.block_hash is None for blk in manager.block_pool]) + + +def test_uncache_blocks(): + manager = KVCacheManager( + block_size=16, + num_gpu_blocks=10, + max_model_len=8192, + sliding_window=None, + enable_caching=True, + num_preallocate_tokens=0, + ) + + req0 = make_request("0", list(range(30))) + blocks = manager.allocate_slots(req0, 30, []) + assert [b.block_id for b in blocks] == [0, 1] + assert len(manager.cached_block_hash_to_block) == 1 + + req0.num_computed_tokens = 30 + + # Simulate speculative tokens. + for _ in range(5): + req0.append_output_token_ids(8) + manager.append_slots(req0, 5) + assert len(manager.cached_block_hash_to_block) == 2 + + # After sampling, assuming only 1 token is accepted. + req0.num_computed_tokens = 31 + num_uncached_blocks = manager.uncache_blocks(req0) + assert num_uncached_blocks == 1 + assert len(manager.cached_block_hash_to_block) == 1 diff --git a/tests/v1/engine/test_async_llm.py b/tests/v1/engine/test_async_llm.py index 2c805e18eebae..10f783b21a9ec 100644 --- a/tests/v1/engine/test_async_llm.py +++ b/tests/v1/engine/test_async_llm.py @@ -1,4 +1,5 @@ import asyncio +from contextlib import ExitStack from typing import List, Tuple import pytest @@ -6,6 +7,7 @@ from vllm import SamplingParams from vllm.engine.arg_utils import AsyncEngineArgs from vllm.platforms import current_platform +from vllm.sampling_params import RequestOutputKind from vllm.v1.engine.async_llm import AsyncLLM if not current_platform.is_cuda(): @@ -18,28 +20,39 @@ async def generate(engine: AsyncLLM, request_id: str, + output_kind: RequestOutputKind, max_tokens: int) -> Tuple[int, str]: count = 0 - async for _ in engine.generate(request_id=request_id, - prompt="Hello my name is Robert and", - sampling_params=SamplingParams( - max_tokens=max_tokens, temperature=0)): + sampling_params = SamplingParams(max_tokens=max_tokens, + output_kind=output_kind, + temperature=0) + async for out in engine.generate(request_id=request_id, + prompt="Hello my name is Robert and", + sampling_params=sampling_params): + + num_tokens = len(out.outputs[0].token_ids) + if output_kind == RequestOutputKind.DELTA: + count += num_tokens + else: + count = num_tokens - count += 1 await asyncio.sleep(0.) return count, request_id +@pytest.mark.parametrize( + "output_kind", [RequestOutputKind.DELTA, RequestOutputKind.FINAL_ONLY]) @pytest.mark.asyncio -async def test_load(monkeypatch): +async def test_load(monkeypatch, output_kind: RequestOutputKind): # TODO(rickyx): Remove monkeypatch once we have a better way to test V1 # so that in the future when we switch, we don't have to change all the # tests. - with monkeypatch.context() as m: + with monkeypatch.context() as m, ExitStack() as after: m.setenv("VLLM_USE_V1", "1") engine = AsyncLLM.from_engine_args(ENGINE_ARGS) + after.callback(engine.shutdown) NUM_REQUESTS = 10000 NUM_EXPECTED_TOKENS = 10 @@ -51,26 +64,33 @@ async def test_load(monkeypatch): for request_id in request_ids: tasks.append( asyncio.create_task( - generate(engine, request_id, NUM_EXPECTED_TOKENS))) + generate(engine, request_id, output_kind, + NUM_EXPECTED_TOKENS))) # Confirm that we got all the EXPECTED tokens from the requests. - for task in tasks: + done, pending = await asyncio.wait(tasks, + return_when=asyncio.FIRST_EXCEPTION) + for task in pending: + task.cancel() + for task in done: num_generated_tokens, request_id = await task assert num_generated_tokens == NUM_EXPECTED_TOKENS, ( f"{request_id} generated {num_generated_tokens} but " f"expected {NUM_EXPECTED_TOKENS}") assert not engine.output_processor.has_unfinished_requests() - engine.shutdown() +@pytest.mark.parametrize( + "output_kind", [RequestOutputKind.DELTA, RequestOutputKind.FINAL_ONLY]) @pytest.mark.asyncio -async def test_abort(monkeypatch): +async def test_abort(monkeypatch, output_kind: RequestOutputKind): - with monkeypatch.context() as m: + with monkeypatch.context() as m, ExitStack() as after: m.setenv("VLLM_USE_V1", "1") engine = AsyncLLM.from_engine_args(ENGINE_ARGS) + after.callback(engine.shutdown) NUM_REQUESTS = 100 NUM_EXPECTED_TOKENS = 100 @@ -83,7 +103,8 @@ async def test_abort(monkeypatch): for request_id in request_ids: tasks.append( asyncio.create_task( - generate(engine, request_id, NUM_EXPECTED_TOKENS))) + generate(engine, request_id, output_kind, + NUM_EXPECTED_TOKENS))) # API server cancels requests when they disconnect. for idx in REQUEST_IDS_TO_ABORT: @@ -108,9 +129,7 @@ async def test_abort(monkeypatch): # Confirm we can do another generation. request_id = f"request-{REQUEST_IDS_TO_ABORT[0]}" task = asyncio.create_task( - generate(engine, request_id, NUM_EXPECTED_TOKENS)) + generate(engine, request_id, output_kind, NUM_EXPECTED_TOKENS)) num_generated_tokens, request_id = await task assert num_generated_tokens == NUM_EXPECTED_TOKENS assert not engine.output_processor.has_unfinished_requests() - - engine.shutdown() diff --git a/tests/v1/engine/test_engine_core.py b/tests/v1/engine/test_engine_core.py index cccfd305ac604..033bbcfce564e 100644 --- a/tests/v1/engine/test_engine_core.py +++ b/tests/v1/engine/test_engine_core.py @@ -144,7 +144,7 @@ def test_engine_core(monkeypatch): def test_engine_core_advanced_sampling(monkeypatch): """ A basic end-to-end test to verify that the engine functions correctly - when additional sampling parameters, such as min_tokens and + when additional sampling parameters, such as top_p, min_tokens, and presence_penalty, are set. """ with monkeypatch.context() as m: @@ -167,11 +167,23 @@ def test_engine_core_advanced_sampling(monkeypatch): stop_token_ids=[1001, 1002], ) engine_core.add_request(request) - assert len(engine_core.scheduler.waiting) == 1 - assert len(engine_core.scheduler.running) == 0 - # Loop through until they are all done. - while len(engine_core.step().outputs) > 0: - pass - assert len(engine_core.scheduler.waiting) == 0 - assert len(engine_core.scheduler.running) == 0 + def _check_engine_state(): + assert len(engine_core.scheduler.waiting) == 1 + assert len(engine_core.scheduler.running) == 0 + # Loop through until they are all done. + while len(engine_core.step().outputs) > 0: + pass + assert len(engine_core.scheduler.waiting) == 0 + assert len(engine_core.scheduler.running) == 0 + + _check_engine_state() + + # Second request. + request2 = make_request() + request2.sampling_params = SamplingParams( + top_p=0.99, + top_k=50, + ) + engine_core.add_request(request2) + _check_engine_state() diff --git a/tests/v1/test_stats.py b/tests/v1/test_stats.py new file mode 100644 index 0000000000000..580392ac5f446 --- /dev/null +++ b/tests/v1/test_stats.py @@ -0,0 +1,300 @@ +import pytest + +from vllm.sampling_params import SamplingParams +from vllm.v1.stats.common import RequestStats, RequestStatsUpdate + + +def make_update( + request_id: str, + update_type: RequestStatsUpdate.Type, + monotonic_ts_s: float, + **kwargs, +): + if update_type == RequestStatsUpdate.Type.INPUT_PROCESSED: + kwargs.setdefault("sampling_params", SamplingParams(n=1)) + kwargs.setdefault("num_prompt_tokens", 10) + elif update_type == RequestStatsUpdate.Type.PREFILLING: + kwargs.setdefault("num_computed_tokens", 10) + kwargs.setdefault("num_cached_tokens", 10) + elif update_type == RequestStatsUpdate.Type.DETOKENIZED: + kwargs.setdefault("num_new_tokens", 10) + elif update_type == RequestStatsUpdate.Type.FINISHED: + kwargs.setdefault("finish_reason", "test_reason") + + return RequestStatsUpdate( + request_id=request_id, + type=update_type, + monotonic_ts_s=monotonic_ts_s, + **kwargs, + ) + + +def test_invalid_request_update(): + request_id = "test_request" + update_specific_required_fields = { + RequestStatsUpdate.Type.INPUT_PROCESSED: [ + "sampling_params", + "num_prompt_tokens", + ], + RequestStatsUpdate.Type.PREFILLING: [ + "num_computed_tokens", + "num_cached_tokens", + ], + RequestStatsUpdate.Type.DETOKENIZED: ["num_new_tokens"], + RequestStatsUpdate.Type.FINISHED: ["finish_reason"], + } + + # Missing a required field should raise an assertion error. + for update_type in RequestStatsUpdate.Type: + required_fields = update_specific_required_fields.get(update_type, []) + + # Try to miss one of the required fields. + kwargs = {field: object() for field in required_fields} + for field in required_fields: + copy_kwargs = kwargs.copy() + copy_kwargs.pop(field) + with pytest.raises(ValueError): + RequestStatsUpdate( + request_id=request_id, + type=update_type, + **copy_kwargs, + ) + + +def test_invalid_request_update_transition(): + # Test invalid transition type. + for src in RequestStatsUpdate.Type: + for dst in RequestStatsUpdate.Type: + if dst not in RequestStatsUpdate._VALID_TRANSITIONS[src]: + with pytest.raises(AssertionError): + RequestStatsUpdate.check_valid_update( + make_update( + update_type=dst, + request_id="test_request", + monotonic_ts_s=1, + ), + last_update_type=src, + last_updated_ts_s=0, + ) + else: + RequestStatsUpdate.check_valid_update( + make_update( + request_id="test_request", + update_type=dst, + monotonic_ts_s=1, + ), + last_update_type=src, + last_updated_ts_s=0, + ) + + # Test invalid timestamp. + with pytest.raises(AssertionError): + RequestStatsUpdate.check_valid_update( + make_update( + request_id="test_request", + update_type=RequestStatsUpdate.Type.ARRIVED, + monotonic_ts_s=1, + ), + last_update_type=None, + last_updated_ts_s=2, + ) + + +def test_lifecycle_updates(): + request_id = "test_request" + stats = RequestStats(request_id=request_id) + + # Test the below scenario: + arrived_ts = 0 + input_processed_ts = 1 + queued_ts = 2 + prefilling_ts = 3 + decoded_ts = 5 + detokenized_ts = 6 + decoded_2_ts = 7 + detokenized_2_ts = 8 + preempted_ts = 9 + resumed_ts = 10 + decoded_3_ts = 11 + detokenized_3_ts = 12 + finished_ts = 13 + + # Test ARRIVED + arrived_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.ARRIVED, + monotonic_ts_s=arrived_ts, + ) + stats.update_from(arrived_update) + assert stats.arrival_ts_s == arrived_ts + assert stats.last_updated_ts_s == arrived_ts + + # Test INPUT_PROCESSED + sampling_params = SamplingParams(n=1) + input_processed_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.INPUT_PROCESSED, + monotonic_ts_s=input_processed_ts, + sampling_params=sampling_params, + num_prompt_tokens=6, + ) + stats.update_from(input_processed_update) + assert stats.input_processor_end_ts_s == input_processed_ts + assert stats.last_updated_ts_s == input_processed_ts + assert stats.num_prompt_tokens == 6 + assert stats.sampling_params == sampling_params + + assert stats.first_token_ts_s is None + assert stats.prefill_ts_s is None + + # Test QUEUED + queued_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.QUEUED, + monotonic_ts_s=queued_ts, + ) + stats.update_from(queued_update) + assert stats.queued_ts_s == queued_ts + assert stats.last_updated_ts_s == queued_ts + + # Test PREFILLING + prefilling_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.PREFILLING, + monotonic_ts_s=prefilling_ts, + num_computed_tokens=3, + num_cached_tokens=1, + ) + stats.update_from(prefilling_update) + assert stats.prefill_ts_s == prefilling_ts + assert stats.num_computed_tokens == 3 + assert stats.num_cached_tokens == 1 + assert stats.queue_duration_s == prefilling_ts - queued_ts + + # Test DECODING + decoded_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.DECODING, + monotonic_ts_s=decoded_ts, + ) + stats.update_from(decoded_update) + assert stats.last_updated_ts_s == decoded_ts + + # Test DETOKENIZED + detokenized_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.DETOKENIZED, + monotonic_ts_s=detokenized_ts, + num_new_tokens=1, + ) + stats.update_from(detokenized_update) + assert stats.last_updated_ts_s == detokenized_ts + assert stats.num_output_tokens == 1 + # Since arrival + assert stats.first_token_latency_s == detokenized_ts - arrived_ts + # Since first scheduled + assert stats.prefill_latency_s == detokenized_ts - prefilling_ts + + # Test another DECODING and DETOKENIZED should + # yield correct inter token latency + decoded_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.DECODING, + monotonic_ts_s=decoded_2_ts, + ) + stats.update_from(decoded_update) + + detokenized_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.DETOKENIZED, + monotonic_ts_s=detokenized_2_ts, + num_new_tokens=1, + ) + stats.update_from(detokenized_update) + assert stats.output_token_latency_s_lst == [ + detokenized_2_ts - detokenized_ts, + ] + assert stats.num_output_tokens == 2 + + # Test PREEMPTED + preempted_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.PREEMPTED, + monotonic_ts_s=preempted_ts, + ) + stats.update_from(preempted_update) + assert stats.last_updated_ts_s == preempted_ts + assert stats.preempted_ts_s_lst == [preempted_ts] + # States should be reset + assert stats.num_computed_tokens == 0 + assert stats.num_cached_tokens == 0 + # These states should not be reset + assert stats.num_output_tokens == 2 + assert stats.output_token_latency_s_lst == [ + detokenized_2_ts - detokenized_ts, + ] + assert stats.prefill_latency_s == prefilling_ts - arrived_ts + assert stats.num_prompt_tokens == 6 + assert stats.prefill_start_ts_s_lst == [prefilling_ts] + + # Test resumed + resumed_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.PREFILLING, + monotonic_ts_s=resumed_ts, + num_computed_tokens=6, + num_cached_tokens=2, + ) + stats.update_from(resumed_update) + # prefill timestamp should not be updated since it's a resumed prefill + assert stats.prefill_ts_s == prefilling_ts + assert stats.num_computed_tokens == 6 + assert stats.num_cached_tokens == 2 + assert stats.prefill_start_ts_s_lst == [ + prefilling_ts, + resumed_ts, + ] + assert stats.last_updated_ts_s == resumed_ts + + # Test another DECODED/DETOKENIZED should yield correct first token latency. + decoded_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.DECODING, + monotonic_ts_s=decoded_3_ts, + ) + detokenized_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.DETOKENIZED, + monotonic_ts_s=detokenized_3_ts, + num_new_tokens=1, + ) + stats.update_from(decoded_update) + stats.update_from(detokenized_update) + assert stats.first_token_ts_s == detokenized_ts - arrived_ts + assert stats.num_output_tokens == 3 + assert stats.output_token_latency_s_lst == [ + detokenized_2_ts - detokenized_ts, + detokenized_3_ts - detokenized_2_ts, + ] + + # Test FINISHED + finished_update = RequestStatsUpdate( + request_id=request_id, + type=RequestStatsUpdate.Type.FINISHED, + monotonic_ts_s=finished_ts, + finish_reason="test_reason", + ) + stats.update_from(finished_update) + assert stats.last_updated_ts_s == finished_ts + assert stats.e2e_latency_s == finished_ts - arrived_ts + assert stats.inference_latency_s == finished_ts - prefilling_ts + assert stats.prefill_latency_s == detokenized_ts - prefilling_ts + assert stats.decode_latency_s == finished_ts - detokenized_ts + assert stats.first_token_latency_s == detokenized_ts - arrived_ts + assert stats.queue_duration_s == prefilling_ts - queued_ts + assert stats.is_finished + assert stats.finish_reason == "test_reason" + + # TODO(rickyx): Add model forward/execute time. + assert stats.model_forward_duration_s == 0.0 + assert stats.model_execute_duration_s == 0.0 diff --git a/tests/v1/test_utils.py b/tests/v1/test_utils.py new file mode 100644 index 0000000000000..ac773b611f406 --- /dev/null +++ b/tests/v1/test_utils.py @@ -0,0 +1,62 @@ +from typing import List + +import torch + +from vllm.v1.utils import bind_kv_cache + + +def test_bind_kv_cache(): + from vllm.attention import Attention + + ctx = { + 'layers.0.self_attn': Attention(32, 128, 0.1), + 'layers.1.self_attn': Attention(32, 128, 0.1), + 'layers.2.self_attn': Attention(32, 128, 0.1), + 'layers.3.self_attn': Attention(32, 128, 0.1), + } + kv_cache = { + 'layers.0.self_attn': torch.zeros((1, )), + 'layers.1.self_attn': torch.zeros((1, )), + 'layers.2.self_attn': torch.zeros((1, )), + 'layers.3.self_attn': torch.zeros((1, )), + } + runner_kv_caches: List[torch.Tensor] = [] + bind_kv_cache(kv_cache, ctx, runner_kv_caches) + assert ctx['layers.0.self_attn'].kv_cache[0] is kv_cache[ + 'layers.0.self_attn'] + assert ctx['layers.1.self_attn'].kv_cache[0] is kv_cache[ + 'layers.1.self_attn'] + assert ctx['layers.2.self_attn'].kv_cache[0] is kv_cache[ + 'layers.2.self_attn'] + assert ctx['layers.3.self_attn'].kv_cache[0] is kv_cache[ + 'layers.3.self_attn'] + + assert runner_kv_caches[0] is kv_cache['layers.0.self_attn'] + assert runner_kv_caches[1] is kv_cache['layers.1.self_attn'] + assert runner_kv_caches[2] is kv_cache['layers.2.self_attn'] + assert runner_kv_caches[3] is kv_cache['layers.3.self_attn'] + + +def test_bind_kv_cache_non_attention(): + from vllm.attention import Attention + + # example from Jamba PP=2 + ctx = { + 'model.layers.20.attn': Attention(32, 128, 0.1), + 'model.layers.28.attn': Attention(32, 128, 0.1), + } + kv_cache = { + 'model.layers.20.attn': torch.zeros((1, )), + 'model.layers.28.attn': torch.zeros((1, )), + } + + runner_kv_caches: List[torch.Tensor] = [] + bind_kv_cache(kv_cache, ctx, runner_kv_caches) + + assert ctx['model.layers.20.attn'].kv_cache[0] is kv_cache[ + 'model.layers.20.attn'] + assert ctx['model.layers.28.attn'].kv_cache[0] is kv_cache[ + 'model.layers.28.attn'] + + assert runner_kv_caches[0] is kv_cache['model.layers.20.attn'] + assert runner_kv_caches[1] is kv_cache['model.layers.28.attn'] diff --git a/tests/weight_loading/models.txt b/tests/weight_loading/models.txt index a06956ce18a93..272206d4502e9 100644 --- a/tests/weight_loading/models.txt +++ b/tests/weight_loading/models.txt @@ -30,4 +30,5 @@ marlin, nm-testing/zephyr-beta-7b-marlin-g128, main marlin, robertgshaw2/zephyr-7b-beta-channelwise-marlin, main qqq, HandH1998/QQQ-Llama-3-8b-g128, main qqq, HandH1998/QQQ-Llama-3-8b, main -hqq, nm-testing/Llama-3.2-1B-Instruct-HQQ, main \ No newline at end of file +hqq, nm-testing/Llama-3.2-1B-Instruct-HQQ, main +None, mgleize/fairseq2-dummy-Llama-3.2-1B, main \ No newline at end of file diff --git a/tests/weight_loading/test_weight_loading.py b/tests/weight_loading/test_weight_loading.py index 199731bdc21fe..7a3786456d0d6 100644 --- a/tests/weight_loading/test_weight_loading.py +++ b/tests/weight_loading/test_weight_loading.py @@ -20,12 +20,13 @@ def test_weight_loading(vllm_runner): """ Test parameter weight loading with tp>1. """ - with vllm_runner(model_name=MODEL_NAME, - revision=REVISION, - dtype=torch.half if QUANTIZATION == "gptq" else "auto", - quantization=QUANTIZATION, - max_model_len=MAX_MODEL_LEN, - tensor_parallel_size=2) as model: + with vllm_runner( + model_name=MODEL_NAME, + revision=REVISION, + dtype=torch.half if QUANTIZATION == "gptq" else "auto", + quantization=None if QUANTIZATION == "None" else QUANTIZATION, + max_model_len=MAX_MODEL_LEN, + tensor_parallel_size=2) as model: output = model.generate_greedy("Hello world!", max_tokens=20) print(output) diff --git a/tools/actionlint.sh b/tools/actionlint.sh deleted file mode 100755 index f6a8b5e83a2de..0000000000000 --- a/tools/actionlint.sh +++ /dev/null @@ -1,13 +0,0 @@ -#!/bin/bash - -if command -v actionlint &> /dev/null; then - actionlint "$@" - exit 0 -elif [ -x ./actionlint ]; then - ./actionlint "$@" - exit 0 -fi - -# download a binary to the current directory - v1.7.3 -bash <(curl https://raw.githubusercontent.com/rhysd/actionlint/aa0a7be8e566b096e64a5df8ff290ec24fa58fbc/scripts/download-actionlint.bash) -./actionlint "$@" diff --git a/tools/mypy.sh b/tools/mypy.sh index bf95e4c526fd1..77d342da1ec82 100755 --- a/tools/mypy.sh +++ b/tools/mypy.sh @@ -1,12 +1,16 @@ #!/bin/bash CI=${1:-0} -PYTHON_VERSION=${2:-3.9} +PYTHON_VERSION=${2:-local} if [ "$CI" -eq 1 ]; then set -e fi +if [ $PYTHON_VERSION == "local" ]; then + PYTHON_VERSION=$(python -c 'import sys; print(f"{sys.version_info.major}.{sys.version_info.minor}")') +fi + run_mypy() { echo "Running mypy on $1" if [ "$CI" -eq 1 ] && [ -z "$1" ]; then diff --git a/tools/shellcheck.sh b/tools/shellcheck.sh index d99fa77b96351..7efb3cabc64fe 100755 --- a/tools/shellcheck.sh +++ b/tools/shellcheck.sh @@ -19,4 +19,4 @@ if ! [ -x "$(command -v shellcheck)" ]; then fi # TODO - fix warnings in .buildkite/run-amd-test.sh -find . -name "*.sh" -not -path "./.buildkite/run-amd-test.sh" -print0 | xargs -0 -I {} sh -c 'git check-ignore -q "{}" || shellcheck "{}"' +find . -name "*.sh" ".git" -prune -not -path "./.buildkite/run-amd-test.sh" -print0 | xargs -0 -I {} sh -c 'git check-ignore -q "{}" || shellcheck -s bash "{}"' diff --git a/vllm/__init__.py b/vllm/__init__.py index 521393fde98ec..84ada481b4f23 100644 --- a/vllm/__init__.py +++ b/vllm/__init__.py @@ -1,10 +1,13 @@ """vLLM: a high-throughput and memory-efficient inference engine for LLMs""" +import os + +import torch from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs from vllm.engine.async_llm_engine import AsyncLLMEngine from vllm.engine.llm_engine import LLMEngine -from vllm.entrypoints.llm import LLM from vllm.entrypoints.fast_sync_llm import FastSyncLLM +from vllm.entrypoints.llm import LLM from vllm.executor.ray_utils import initialize_ray_cluster from vllm.inputs import PromptType, TextPrompt, TokensPrompt from vllm.model_executor.models import ModelRegistry @@ -18,6 +21,19 @@ from .version import __version__, __version_tuple__ +# set some common config/environment variables that should be set +# for all processes created by vllm and all processes +# that interact with vllm workers. +# they are executed whenever `import vllm` is called. + +# see https://github.com/NVIDIA/nccl/issues/1234 +os.environ['NCCL_CUMEM_ENABLE'] = '0' + +# see https://github.com/vllm-project/vllm/issues/10480 +os.environ['TORCHINDUCTOR_COMPILE_THREADS'] = '1' +# see https://github.com/vllm-project/vllm/issues/10619 +torch._inductor.config.compile_threads = 1 + __all__ = [ "__version__", "__version_tuple__", diff --git a/vllm/assets/image.py b/vllm/assets/image.py index cb831cb0b5bb4..0a55506f88255 100644 --- a/vllm/assets/image.py +++ b/vllm/assets/image.py @@ -26,4 +26,4 @@ def image_embeds(self) -> torch.Tensor: """ image_path = get_vllm_public_assets(filename=f"{self.name}.pt", s3_prefix=VLM_IMAGES_DIR) - return torch.load(image_path, map_location="cpu") + return torch.load(image_path, map_location="cpu", weights_only=True) diff --git a/vllm/attention/backends/abstract.py b/vllm/attention/backends/abstract.py index 901bdb2ff2013..fe097b346f3c6 100644 --- a/vllm/attention/backends/abstract.py +++ b/vllm/attention/backends/abstract.py @@ -1,8 +1,8 @@ from abc import ABC, abstractmethod from contextlib import contextmanager from dataclasses import dataclass, fields -from typing import (TYPE_CHECKING, Any, Dict, Generic, List, Optional, Set, - Tuple, Type, TypeVar) +from typing import (TYPE_CHECKING, Any, Dict, Generic, List, Optional, + Protocol, Set, Tuple, Type, TypeVar) import torch @@ -31,6 +31,10 @@ class AttentionType: class AttentionBackend(ABC): """Abstract class for attention backends.""" + # For some attention backends, we allocate an output tensor before + # calling the custom op. When piecewise cudagraph is enabled, this + # makes sure the output tensor is allocated inside the cudagraph. + accept_output_buffer: bool = False @staticmethod @abstractmethod @@ -61,11 +65,6 @@ def make_metadata(cls, *args, **kwargs) -> "AttentionMetadata": def get_builder_cls() -> Type["AttentionMetadataBuilder"]: raise NotImplementedError - @classmethod - def make_metadata_builder(cls, *args, - **kwargs) -> "AttentionMetadataBuilder": - return cls.get_builder_cls()(*args, **kwargs) - @staticmethod @abstractmethod def get_kv_cache_shape( @@ -214,6 +213,12 @@ class AttentionMetadataBuilder(ABC, Generic[T]): @abstractmethod def __init__(self, input_builder: "ModelRunnerInputBuilderBase") -> None: + """Create the builder, remember some configuration and parameters.""" + raise NotImplementedError + + @abstractmethod + def prepare(self) -> None: + """Prepare for one batch.""" raise NotImplementedError @abstractmethod @@ -223,6 +228,26 @@ def build(self, seq_lens: List[int], query_lens: List[int], raise NotImplementedError +class AttentionLayer(Protocol): + + _k_scale: torch.Tensor + _v_scale: torch.Tensor + _k_scale_float: torch.Tensor + _v_scale_float: torch.Tensor + _q_scale: torch.Tensor + _prob_scale: torch.Tensor + + def forward( + self, + query: torch.Tensor, + key: torch.Tensor, + value: torch.Tensor, + kv_cache: torch.Tensor, + attn_metadata: AttentionMetadata, + ) -> torch.Tensor: + ... + + class AttentionImpl(ABC, Generic[T]): @abstractmethod @@ -244,15 +269,12 @@ def __init__( @abstractmethod def forward( self, + layer: AttentionLayer, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, kv_cache: torch.Tensor, attn_metadata: T, - k_scale: torch.Tensor, - v_scale: torch.Tensor, - q_scale: Optional[torch.Tensor] = None, - prob_scale: Optional[torch.Tensor] = None, fp8_out_scale: Optional[torch.Tensor] = None, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: diff --git a/vllm/attention/backends/blocksparse_attn.py b/vllm/attention/backends/blocksparse_attn.py index 7840c9f076778..2d7f16c42a448 100644 --- a/vllm/attention/backends/blocksparse_attn.py +++ b/vllm/attention/backends/blocksparse_attn.py @@ -4,6 +4,7 @@ import torch from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, + AttentionLayer, AttentionMetadata, AttentionType) from vllm.attention.backends.utils import (CommonAttentionState, CommonMetadataBuilder) @@ -89,8 +90,7 @@ class BlocksparseFlashAttentionBackend(AttentionBackend): @staticmethod def get_name() -> str: - # For attention layer compatibility - return "FLASH_ATTN" + return "BLOCK_SPARSE_FLASH_ATTN" @staticmethod def get_impl_cls() -> Type["BlocksparseFlashAttentionImpl"]: @@ -361,15 +361,12 @@ def __init__( def forward( self, + layer: AttentionLayer, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, kv_cache: torch.Tensor, attn_metadata: BlocksparseFlashAttentionMetadata, - k_scale: torch.Tensor, - v_scale: torch.Tensor, - q_scale: Optional[torch.Tensor] = None, - prob_scale: Optional[torch.Tensor] = None, fp8_out_scale: Optional[torch.Tensor] = None, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: @@ -407,8 +404,8 @@ def forward( value_cache, attn_metadata.slot_mapping, self.kv_cache_dtype, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) if prefill_meta := attn_metadata.prefill_metadata: @@ -445,8 +442,8 @@ def forward( self.num_kv_heads, self.scale, self.alibi_slopes, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, tp_rank=self.tp_rank, blocksparse_local_blocks=self.local_blocks, blocksparse_vert_stride=self.vert_stride, diff --git a/vllm/attention/backends/flash_attn.py b/vllm/attention/backends/flash_attn.py old mode 100644 new mode 100755 index 35fb78db67daf..8bfb5e461d6f6 --- a/vllm/attention/backends/flash_attn.py +++ b/vllm/attention/backends/flash_attn.py @@ -8,6 +8,7 @@ from vllm import _custom_ops as ops from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, + AttentionLayer, AttentionMetadata, AttentionMetadataBuilder, AttentionType) @@ -16,19 +17,27 @@ compute_slot_mapping_start_idx, get_num_prefill_decode_query_kv_tokens, get_seq_len_block_table_args, is_all_cross_attn_metadata_set, is_all_encoder_attn_metadata_set, is_block_tables_empty) +from vllm.envs import VLLM_FLASH_ATTN_VERSION +from vllm.logger import init_logger from vllm.multimodal import MultiModalPlaceholderMap +from vllm.platforms import current_platform from vllm.utils import async_tensor_h2d, make_tensor_with_pad +from vllm.vllm_flash_attn import (fa_version_unsupported_reason, + flash_attn_varlen_func, + flash_attn_with_kvcache, + is_fa_version_supported) if TYPE_CHECKING: from vllm.worker.model_runner import (ModelInputForGPUBuilder, ModelInputForGPUWithSamplingMetadata) -from vllm.vllm_flash_attn import (flash_attn_varlen_func, - flash_attn_with_kvcache) +logger = init_logger(__name__) class FlashAttentionBackend(AttentionBackend): + accept_output_buffer: bool = True + @staticmethod def get_supported_head_sizes() -> List[int]: return [32, 64, 96, 128, 160, 192, 224, 256] @@ -374,6 +383,12 @@ class FlashAttentionMetadataBuilder( AttentionMetadataBuilder[FlashAttentionMetadata]): def __init__(self, input_builder: "ModelInputForGPUBuilder"): + self.input_builder = input_builder + self.runner = input_builder.runner + self.sliding_window = input_builder.sliding_window + self.block_size = input_builder.block_size + + def prepare(self): self.slot_mapping: List[int] = [] self.prefill_seq_lens: List[int] = [] self.context_lens: List[int] = [] @@ -387,11 +402,6 @@ def __init__(self, input_builder: "ModelInputForGPUBuilder"): self.num_decode_tokens = 0 self.has_prefix_cache_hit = False - self.input_builder = input_builder - self.runner = input_builder.runner - self.sliding_window = input_builder.sliding_window - self.block_size = input_builder.block_size - def _add_seq_group( self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup", chunked_prefill_enabled: bool, prefix_cache_hit: bool): @@ -633,17 +643,33 @@ def __init__( f"Supported head sizes are: {support_head_sizes}.") self.attn_type = attn_type + # if hopper default to FA3, otherwise stick to FA2 for now + # TODO(lucas): profile FA3 on ampere to see if it makes sense to + # use FA3 as default for both + if current_platform.get_device_capability()[0] >= 9: + self.fa_version = 3 if is_fa_version_supported(3) else 2 + else: + self.fa_version = 2 + + if VLLM_FLASH_ATTN_VERSION is not None: + assert VLLM_FLASH_ATTN_VERSION in [2, 3] + self.fa_version = VLLM_FLASH_ATTN_VERSION + + if not is_fa_version_supported(self.fa_version): + logger.error("Cannot use FA version %d is not supported due to %s", + self.fa_version, + fa_version_unsupported_reason(self.fa_version)) + + assert is_fa_version_supported(self.fa_version) + def forward( self, + layer: AttentionLayer, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, kv_cache: torch.Tensor, attn_metadata: FlashAttentionMetadata, - k_scale: float = 1.0, - v_scale: float = 1.0, - q_scale: Optional[torch.Tensor] = None, - prob_scale: Optional[torch.Tensor] = None, fp8_out_scale: Optional[torch.Tensor] = None, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: @@ -661,7 +687,7 @@ def forward( NOTE: It in-place updates the output tensor. """ # NOTE(woosuk): FlashAttention does not support FP8 KV cache. - assert k_scale == 1.0 and v_scale == 1.0, ( + assert layer._k_scale_float == 1.0 and layer._v_scale_float == 1.0, ( "key/v_scale is not supported in FlashAttention.") assert output is not None, "Output tensor must be provided." @@ -713,8 +739,8 @@ def forward( kv_cache[1], updated_slot_mapping.flatten(), # type: ignore[union-attr] kv_cache_dtype, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) (num_prefill_query_tokens, num_prefill_kv_tokens, @@ -755,6 +781,7 @@ def forward( alibi_slopes=alibi_slopes, softcap=logits_soft_cap, out=prefill_output, + fa_version=self.fa_version, ) else: # prefix-enabled attention @@ -768,7 +795,7 @@ def forward( v=value_cache, cu_seqlens_q=prefill_meta.query_start_loc, max_seqlen_q=prefill_meta.max_query_len, - cu_seqlens_k=prefill_meta.seq_start_loc, + seqused_k=prefill_meta.seq_lens_tensor, max_seqlen_k=max_seq_len, softmax_scale=softmax_scale, causal=True, @@ -777,6 +804,7 @@ def forward( block_table=prefill_meta.block_tables, softcap=logits_soft_cap, out=prefill_output, + fa_version=self.fa_version, ) if decode_meta := attn_metadata.decode_metadata: @@ -796,7 +824,7 @@ def forward( v=value_cache, cu_seqlens_q=decode_meta.query_start_loc, max_seqlen_q=decode_meta.max_decode_query_len, - cu_seqlens_k=decode_meta.seq_start_loc, + seqused_k=decode_meta.seq_lens_tensor, max_seqlen_k=decode_meta.max_decode_seq_len, softmax_scale=softmax_scale, causal=True, @@ -805,6 +833,7 @@ def forward( softcap=logits_soft_cap, block_table=decode_meta.block_tables, out=decode_output, + fa_version=self.fa_version, ) else: # Use flash_attn_with_kvcache for normal decoding. @@ -825,6 +854,7 @@ def forward( alibi_slopes=alibi_slopes, softcap=logits_soft_cap, out=decode_output.unsqueeze(1), + fa_version=self.fa_version, ) return output diff --git a/vllm/attention/backends/flashinfer.py b/vllm/attention/backends/flashinfer.py index 8342e43c1f7bf..6c10e23dc434c 100644 --- a/vllm/attention/backends/flashinfer.py +++ b/vllm/attention/backends/flashinfer.py @@ -1,3 +1,4 @@ +import dataclasses from collections import defaultdict from contextlib import contextmanager from dataclasses import dataclass @@ -13,9 +14,11 @@ from vllm.vllm_flash_attn import flash_attn_varlen_func FLASHINFER_WORKSPACE_BUFFER_SIZE = 256 * 1024 * 1024 except ImportError: - BatchDecodeWithPagedKVCacheWrapper = None - CUDAGraphBatchDecodeWithPagedKVCacheWrapper = None - BatchPrefillWithPagedKVCacheWrapper = None + # Avoid turning these types into variables during type checking + if not TYPE_CHECKING: + BatchDecodeWithPagedKVCacheWrapper = None + CUDAGraphBatchDecodeWithPagedKVCacheWrapper = None + BatchPrefillWithPagedKVCacheWrapper = None FLASHINFER_WORKSPACE_BUFFER_SIZE = 0 import torch @@ -23,13 +26,16 @@ import vllm.envs as envs from vllm import _custom_ops as ops from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, + AttentionLayer, AttentionMetadata, AttentionMetadataBuilder, AttentionState, AttentionType) from vllm.attention.backends.utils import (PAD_SLOT_ID, compute_slot_mapping, compute_slot_mapping_start_idx, is_block_tables_empty) +from vllm.attention.layer import Attention from vllm.attention.ops.paged_attn import PagedAttention +from vllm.config import VllmConfig, get_current_vllm_config from vllm.utils import (async_tensor_h2d, get_kv_cache_torch_dtype, make_tensor_with_pad) @@ -98,6 +104,72 @@ def get_fp8_dtype_for_flashinfer(kv_cache_dtype: str) -> torch.dtype: raise ValueError(f"Unrecognized FP8 dtype: {kv_cache_dtype}") +@dataclass +class PerLayerParameters: + """ + Currently, FlashInfer backend only support models in which all layers share + the same values for the following hyperparameters. + """ + + window_left: int + logits_soft_cap: Optional[float] + sm_scale: float + + +def get_per_layer_parameters( + vllm_config: VllmConfig) -> Dict[str, PerLayerParameters]: + """ + Scan all attention layers and determine some hyperparameters + to use during `plan`. + """ + + layers = vllm_config.compilation_config.static_forward_context + per_layer_params: Dict[str, PerLayerParameters] = {} + + for key, layer in layers.items(): + assert isinstance(layer, Attention) + + impl = layer.impl + assert isinstance(impl, FlashInferImpl) + + # Infer hyperparameters from the attention layer + window_size = impl.sliding_window + window_left = window_size[0] if window_size is not None else -1 + logits_soft_cap = impl.logits_soft_cap + sm_scale = impl.scale + + per_layer_params[key] = PerLayerParameters(window_left, + logits_soft_cap, sm_scale) + + return per_layer_params + + +def infer_global_hyperparameters( + per_layer_params: Dict[str, PerLayerParameters]) -> PerLayerParameters: + """ + Currently, FlashInfer backend only support models in which all layers share + the same values for the following hyperparameters: + - `window_left` + - `logits_soft_cap` + - `sm_scale` + + So this function asserts that all layers share the same values for these + hyperparameters and returns the global values. + """ + + assert len(per_layer_params) > 0, "No attention layers found in the model." + + param_sets = list(per_layer_params.values()) + global_params = param_sets[0] + for params in param_sets: + assert params == global_params, ( + "FlashInfer backend currently only supports models in which all " + "layers share the same values for the following hyperparameters: " + "`window_left`, `logits_soft_cap`, `sm_scale`.") + + return global_params + + class FlashInferState(AttentionState): def __init__(self, runner): @@ -107,6 +179,11 @@ def __init__(self, runner): self._decode_wrapper = None self._prefill_wrapper = None + # Global hyperparameters shared by all attention layers + self.global_hyperparameters: Optional[PerLayerParameters] = None + + self.vllm_config = get_current_vllm_config() + def _get_workspace_buffer(self): if self._workspace_buffer is None: self._workspace_buffer = torch.empty( @@ -214,6 +291,9 @@ def graph_capture_get_metadata_for_batch( batch_size + 1, dtype=torch.int32) + global_params = infer_global_hyperparameters( + get_per_layer_parameters(self.vllm_config)) + attn_metadata = self.runner.attn_backend.make_metadata( num_prefills=0, slot_mapping=self._graph_slot_mapping[:batch_size], @@ -237,7 +317,9 @@ def graph_capture_get_metadata_for_batch( q_data_type=self.runner.model_config.dtype, use_cuda_graph=True, decode_wrapper=self._graph_decode_wrapper, - prefill_wrapper=None) + prefill_wrapper=None, + **dataclasses.asdict(global_params), + ) attn_metadata.begin_forward() return attn_metadata @@ -257,7 +339,12 @@ def prepare_graph_input_buffers(self, def begin_forward(self, model_input): assert not self._is_graph_capturing state = self - if model_input.attn_metadata.use_cuda_graph: + use_cuda_graph = model_input.attn_metadata.use_cuda_graph + is_decode = model_input.attn_metadata.num_prefills == 0 + # In case of multistep chunked-prefill, there might be prefill requests + # scheduled while CUDA graph mode is enabled. We don't run graph in that + # case. + if use_cuda_graph and is_decode: batch_size = model_input.input_tokens.shape[0] state = (self.runner.graph_runners[model_input.virtual_engine] [batch_size].attn_state) @@ -319,9 +406,28 @@ class FlashInferMetadata(AttentionMetadata): data_type: torch.dtype = None # The data type of the query q_data_type: torch.dtype = None - device: torch.device = torch.device("cuda") + # FlashInfer 0.2 encourages passing host tensors + device: torch.device = torch.device("cpu") is_profile_run: bool = False + # The FlashInfer backend currently supports only models in which all layers + # share the same following hyperparameters: + + # The left (inclusive) window size for the attention window, when + # set to `-1`, the window size will be set to the full length of + # the sequence. Defaults to `-1`. + window_left: int = -1 + # The attention logits soft capping value (used in Gemini, Grok and + # Gemma-2, etc.), if not provided, will be set to `0`. If greater + # than 0, the logits will be capped according to formula: + # $$\texttt{logits\_soft\_cap} \times + # \mathrm{tanh}(x / \texttt{logits\_soft\_cap})$$, + # where $x$ is the input logits. + logits_soft_cap: Optional[float] = None + # The scale used in softmax, if not provided, will be set to + # `1.0 / sqrt(head_dim)`. + sm_scale: Optional[float] = None + def __post_init__(self): # Refer to # https://github.com/flashinfer-ai/flashinfer/blob/3d55c71a62052c590c130897d3a3db49b14fcc34/include/flashinfer/utils.cuh#L157 @@ -357,14 +463,21 @@ def begin_forward(self): self.block_table_bound = self.block_table_bound.to(self.device) self.seq_lens_tensor = self.seq_lens_tensor.to(self.device) self.paged_kv_indices = self.paged_kv_indices.to(self.device) - self.prefill_wrapper.end_forward() - self.prefill_wrapper.begin_forward( + self.prefill_wrapper.plan( self.query_start_loc, self.paged_kv_indptr[:self.num_prefills + 1], self.paged_kv_indices, self.paged_kv_last_page_len[:self.num_prefills], - self.num_qo_heads, self.num_kv_heads, self.head_dim, - self.page_size) + self.num_qo_heads, + self.num_kv_heads, + self.head_dim, + self.page_size, + causal=True, + sm_scale=self.sm_scale, + window_left=self.window_left, + logits_soft_cap=self.logits_soft_cap, + q_data_type=self.q_data_type, + kv_data_type=self.data_type) if self.num_decode_tokens > 0: assert self.paged_kv_indices is not None assert self.paged_kv_indptr is not None @@ -380,8 +493,7 @@ def begin_forward(self): self.seq_lens_tensor = self.seq_lens_tensor.to(self.device) assert self.decode_wrapper is not None - self.decode_wrapper.end_forward() - self.decode_wrapper.begin_forward( + self.decode_wrapper.plan( self.paged_kv_indptr[self.num_prefills:], self.paged_kv_indices, self.paged_kv_last_page_len[self.num_prefills:], @@ -391,8 +503,11 @@ def begin_forward(self): self.page_size, # Disable flashinfer's pos encoding and use vllm's rope. pos_encoding_mode="NONE", + window_left=self.window_left, + logits_soft_cap=self.logits_soft_cap, + sm_scale=self.sm_scale, # kv-cache data type. - data_type=self.data_type, + kv_data_type=self.data_type, # query data type. q_data_type=self.q_data_type) @@ -430,10 +545,24 @@ def advance_step(self, Update metadata in-place to advance one decode step. """ - assert not turn_prefills_into_decodes, \ - ("Chunked prefill is not supported with flashinfer yet." - "turn_prefills_into_decodes is a Multi-Step + Chunked-Prefill " - "specific parameter.") + if turn_prefills_into_decodes: + # When Multi-Step is enabled with Chunked-Prefill, prefills and + # decodes are scheduled together. In the first step, all the + # prefills turn into decodes. This update reflects that + # conversion. + assert self.num_decode_tokens + self.num_prefills == num_seqs + # Flashinfer doesn't support speculative decoding + chunked-prefill + # + multi-step scheduling yet. + assert self.decode_query_len == 1 + self.num_decode_tokens += self.num_prefills + self.num_prefills = 0 + self.num_prefill_tokens = 0 + self.max_prefill_seq_len = 0 + self.max_query_len = 1 + + self.slot_mapping = self.slot_mapping[:num_seqs] + else: + assert self.seq_lens_tensor is not None assert num_seqs > 0 assert num_queries > 0 @@ -469,6 +598,19 @@ def advance_step(self, class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]): def __init__(self, input_builder: "ModelInputForGPUBuilder"): + + self.input_builder = input_builder + self.runner = input_builder.runner + + self.sliding_window = input_builder.sliding_window + self.block_size = input_builder.block_size + + # Global hyperparameters shared by all attention layers + self.global_hyperparameters: Optional[PerLayerParameters] = None + + self.vllm_config = get_current_vllm_config() + + def prepare(self): self.slot_mapping: List[int] = [] self.prefill_seq_lens: List[int] = [] self.context_lens: List[int] = [] @@ -481,12 +623,6 @@ def __init__(self, input_builder: "ModelInputForGPUBuilder"): self.num_prefill_tokens = 0 self.num_decode_tokens = 0 - self.input_builder = input_builder - self.runner = input_builder.runner - - self.sliding_window = input_builder.sliding_window - self.block_size = input_builder.block_size - # Please follow https://docs.flashinfer.ai/tutorials/kv_layout.html#page-layout # for the precise definition of the following fields. # An example: @@ -506,6 +642,20 @@ def __init__(self, input_builder: "ModelInputForGPUBuilder"): self.total_blocks = 0 self.is_profile_run: bool = False + if self.global_hyperparameters is None: + # Infer global hyperparameters, since currently we only support + # models in which all layers share the same values for the + # following hyperparameters: + # - `window_left` + # - `logits_soft_cap` + # - `sm_scale` + inferred_params = infer_global_hyperparameters( + get_per_layer_parameters(self.vllm_config)) + self.global_hyperparameters = inferred_params + self.window_left = inferred_params.window_left + self.logits_soft_cap = inferred_params.logits_soft_cap + self.sm_scale = inferred_params.sm_scale + def _add_seq_group( self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup", chunked_prefill_enabled: bool): @@ -734,7 +884,11 @@ def build(self, seq_lens: List[int], query_lens: List[int], data_type=kv_cache_dtype, q_data_type=self.runner.model_config.dtype, use_cuda_graph=use_captured_graph, - is_profile_run=self.is_profile_run) + is_profile_run=self.is_profile_run, + window_left=self.window_left, + logits_soft_cap=self.logits_soft_cap, + sm_scale=self.sm_scale, + ) class FlashInferImpl(AttentionImpl): @@ -775,15 +929,12 @@ def __init__( def forward( self, + layer: AttentionLayer, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, kv_cache: torch.Tensor, attn_metadata: FlashInferMetadata, - k_scale: float = 1.0, - v_scale: float = 1.0, - q_scale: Optional[torch.Tensor] = None, - prob_scale: Optional[torch.Tensor] = None, fp8_out_scale: Optional[torch.Tensor] = None, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: @@ -812,8 +963,8 @@ def forward( kv_cache[:, 1], attn_metadata.slot_mapping.flatten(), kv_cache_dtype, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) # The FlashInfer api requires data to be in fp8_e4m3 or fp8_e5m2 # to process the cache when the kv_cache_dtype is fp8 @@ -867,25 +1018,34 @@ def forward( else: assert prefill_meta is not None assert prefill_meta.prefill_wrapper is not None - prefill_output = prefill_meta.prefill_wrapper.forward( + + assert prefill_meta.prefill_wrapper._causal + assert prefill_meta.prefill_wrapper._window_left == window_left + assert prefill_meta.prefill_wrapper._logits_soft_cap == ( + logits_soft_cap or 0.0) + assert prefill_meta.prefill_wrapper._sm_scale == softmax_scale + + prefill_output = prefill_meta.prefill_wrapper.run( query, kv_cache, - logits_soft_cap=logits_soft_cap, - causal=True, - k_scale=k_scale, - v_scale=v_scale, - window_left=window_left) + k_scale=layer._k_scale_float, + v_scale=layer._v_scale_float, + ) if decode_meta := attn_metadata.decode_metadata: assert decode_meta is not None assert decode_meta.decode_wrapper is not None - decode_output = decode_meta.decode_wrapper.forward( + + assert decode_meta.decode_wrapper._window_left == window_left + assert decode_meta.decode_wrapper._logits_soft_cap == ( + logits_soft_cap or 0.0) + assert decode_meta.decode_wrapper._sm_scale == softmax_scale + + decode_output = decode_meta.decode_wrapper.run( decode_query, kv_cache, - sm_scale=softmax_scale, - logits_soft_cap=logits_soft_cap, - k_scale=k_scale, - v_scale=v_scale, - window_left=window_left) + k_scale=layer._k_scale_float, + v_scale=layer._v_scale_float, + ) if prefill_output is None and decode_output is not None: # Decode only batch. diff --git a/vllm/attention/backends/hpu_attn.py b/vllm/attention/backends/hpu_attn.py index f19ee9540683f..0937bde68e84f 100644 --- a/vllm/attention/backends/hpu_attn.py +++ b/vllm/attention/backends/hpu_attn.py @@ -11,6 +11,7 @@ from vllm_hpu_extension.utils import Matmul, Softmax, VLLMKVCache from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, + AttentionLayer, AttentionMetadata, AttentionType) from vllm.attention.backends.utils import CommonAttentionState from vllm.attention.ops.hpu_paged_attn import (HPUPagedAttention, @@ -152,15 +153,12 @@ def __init__( def forward( self, + layer: AttentionLayer, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, kv_cache: torch.Tensor, attn_metadata: HPUAttentionMetadata, - k_scale: float = 1.0, - v_scale: float = 1.0, - q_scale: Optional[torch.Tensor] = None, - prob_scale: Optional[torch.Tensor] = None, fp8_out_scale: Optional[torch.Tensor] = None, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: diff --git a/vllm/attention/backends/ipex_attn.py b/vllm/attention/backends/ipex_attn.py index e9e8428e4d261..c060d2f6772bb 100644 --- a/vllm/attention/backends/ipex_attn.py +++ b/vllm/attention/backends/ipex_attn.py @@ -7,6 +7,7 @@ from vllm._ipex_ops import ipex_ops from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, + AttentionLayer, AttentionMetadata, AttentionType) from vllm.attention.backends.utils import CommonAttentionState from vllm.attention.ops.paged_attn import (PagedAttention, @@ -171,15 +172,12 @@ def split_kv_cache( def forward( self, + layer: AttentionLayer, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, kv_cache: torch.Tensor, attn_metadata: IpexAttnMetadata, # type: ignore - k_scale: float = 1.0, - v_scale: float = 1.0, - q_scale: Optional[torch.Tensor] = None, - prob_scale: Optional[torch.Tensor] = None, fp8_out_scale: Optional[torch.Tensor] = None, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: @@ -196,7 +194,7 @@ def forward( Returns: shape = [num_tokens, num_heads * head_size] """ - assert k_scale == 1.0 and v_scale == 1.0 + assert layer._k_scale_float == 1.0 and layer._v_scale_float == 1.0 num_tokens, hidden_size = query.shape # Reshape the query, key, and value tensors. query = query.view(-1, self.num_heads, self.head_size) @@ -213,8 +211,8 @@ def forward( value_cache, attn_metadata.slot_mapping.flatten(), self.kv_cache_dtype, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) if attn_metadata.is_prompt: @@ -299,8 +297,8 @@ def forward( max_seq_len, self.alibi_slopes, self.kv_cache_dtype, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) else: # Run PagedAttention V2. @@ -332,8 +330,8 @@ def forward( max_seq_len, self.alibi_slopes, self.kv_cache_dtype, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) # Reshape the output tensor. diff --git a/vllm/attention/backends/pallas.py b/vllm/attention/backends/pallas.py index bac8c0b07d8fe..1c902d394464d 100644 --- a/vllm/attention/backends/pallas.py +++ b/vllm/attention/backends/pallas.py @@ -5,6 +5,7 @@ import torch_xla.experimental.custom_kernel # Required to register custom ops. from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, + AttentionLayer, AttentionMetadata, AttentionType) from vllm.attention.backends.utils import CommonAttentionState @@ -150,15 +151,12 @@ def __init__( def forward( self, + layer: AttentionLayer, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, kv_cache: Tuple[torch.Tensor, torch.Tensor], attn_metadata: PallasMetadata, - k_scale: float = 1.0, - v_scale: float = 1.0, - q_scale: Optional[torch.Tensor] = None, - prob_scale: Optional[torch.Tensor] = None, fp8_out_scale: Optional[torch.Tensor] = None, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: @@ -176,7 +174,7 @@ def forward( Returns: shape = [batch_size, seq_len, num_heads * head_size] """ - assert k_scale == 1.0 and v_scale == 1.0 + assert layer._k_scale_float == 1.0 and layer._v_scale_float == 1.0 batch_size, seq_len, hidden_size = query.shape query = query.view(batch_size, seq_len, self.num_heads, self.head_size) key = key.view(batch_size, seq_len, self.num_kv_heads, self.head_size) diff --git a/vllm/attention/backends/placeholder_attn.py b/vllm/attention/backends/placeholder_attn.py index d2dc0d6cf0a5f..826311896d1d2 100644 --- a/vllm/attention/backends/placeholder_attn.py +++ b/vllm/attention/backends/placeholder_attn.py @@ -255,6 +255,11 @@ class PlaceholderAttentionMetadataBuilder( AttentionMetadataBuilder[PlaceholderAttentionMetadata]): def __init__(self, input_builder: "ModelInputForGPUBuilder"): + + self.input_builder = input_builder + self.runner = input_builder.runner + + def prepare(self): self.prefill_seq_lens: List[int] = [] self.context_lens: List[int] = [] self.curr_seq_lens: List[int] = [] @@ -265,9 +270,6 @@ def __init__(self, input_builder: "ModelInputForGPUBuilder"): self.num_prefill_tokens = 0 self.num_decode_tokens = 0 - self.input_builder = input_builder - self.runner = input_builder.runner - def _add_seq_group( self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup", chunked_prefill_enabled: bool): diff --git a/vllm/attention/backends/rocm_flash_attn.py b/vllm/attention/backends/rocm_flash_attn.py index 174539d3146b7..5f6e7c7cca94d 100644 --- a/vllm/attention/backends/rocm_flash_attn.py +++ b/vllm/attention/backends/rocm_flash_attn.py @@ -7,6 +7,7 @@ import vllm.envs as envs from vllm import _custom_ops as ops from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, + AttentionLayer, AttentionMetadata, AttentionType) from vllm.attention.backends.utils import (CommonAttentionState, CommonMetadataBuilder) @@ -543,15 +544,12 @@ def repeat_kv(self, x: torch.Tensor, n_rep: int) -> torch.Tensor: def forward( self, + layer: AttentionLayer, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, kv_cache: torch.Tensor, attn_metadata: ROCmFlashAttentionMetadata, - k_scale: torch.Tensor, - v_scale: torch.Tensor, - q_scale: Optional[torch.Tensor] = None, - prob_scale: Optional[torch.Tensor] = None, fp8_out_scale: Optional[torch.Tensor] = None, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: @@ -629,8 +627,8 @@ def forward( if self.attn_type != AttentionType.ENCODER_DECODER else attn_metadata.cross_slot_mapping, self.kv_cache_dtype, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) if self.attn_type != AttentionType.ENCODER: @@ -683,10 +681,11 @@ def forward( seq_lens, make_attn_mask=False) # type: ignore full_scales = ( - 1.0 / q_scale.item(), 1.0 / k_scale.item(), - 1.0 / v_scale.item(), 1.0 / prob_scale.item(), + layer._q_scale.item(), layer._k_scale.item(), + layer._v_scale.item(), layer._prob_scale.item(), fp8_out_scale.item()) if ( - fp8_out_scale and q_scale and prob_scale + fp8_out_scale and layer._q_scale + and layer._prob_scale and envs.VLLM_USE_ROCM_FP8_FLASH_ATTN) else None out, _ = self.attn_func( query, @@ -768,8 +767,8 @@ def forward( prefill_meta.max_query_len, self.alibi_slopes, self.sliding_window[0], - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) if decode_meta := attn_metadata.decode_metadata: @@ -831,8 +830,8 @@ def forward( max_seq_len, self.alibi_slopes, self.kv_cache_dtype, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, fp8_out_scale if cpa_fp8_out else None, _PARTITION_SIZE_ROCM, ) @@ -856,8 +855,8 @@ def forward( self.num_kv_heads, self.scale, self.alibi_slopes, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) # Reshape the output tensor. diff --git a/vllm/attention/backends/torch_sdpa.py b/vllm/attention/backends/torch_sdpa.py index 1ae73d37ed891..4ae89617d4a55 100644 --- a/vllm/attention/backends/torch_sdpa.py +++ b/vllm/attention/backends/torch_sdpa.py @@ -7,6 +7,7 @@ from torch.nn.functional import scaled_dot_product_attention from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, + AttentionLayer, AttentionMetadata, AttentionMetadataBuilder, AttentionType) @@ -281,7 +282,10 @@ class TorchSDPAMetadataBuilder(AttentionMetadataBuilder[TorchSDPAMetadata]): def __init__(self, input_builder: ModelInputForCPUBuilder) -> None: self.chunked_prefill = input_builder.chunked_prefill - self.input_data = input_builder.input_data + self.input_builder = input_builder + + def prepare(self): + self.input_data = self.input_builder.input_data def build(self, seq_lens: List[int], query_lens: List[int], cuda_graph_pad_size: int, batch_size: int) -> TorchSDPAMetadata: @@ -430,15 +434,12 @@ def __init__( def forward( self, + layer: AttentionLayer, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, kv_cache: torch.Tensor, attn_metadata: TorchSDPAMetadata, # type: ignore - k_scale: float = 1.0, - v_scale: float = 1.0, - q_scale: Optional[torch.Tensor] = None, - prob_scale: Optional[torch.Tensor] = None, fp8_out_scale: Optional[torch.Tensor] = None, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: @@ -455,7 +456,6 @@ def forward( Returns: shape = [num_tokens, num_heads * head_size] """ - assert k_scale == 1.0 and v_scale == 1.0 attn_type = self.attn_type if (attn_type == AttentionType.ENCODER and (not attn_metadata.is_all_encoder_attn_metadata_set)): @@ -497,11 +497,9 @@ def forward( # Update self-attention KV cache (prefill/decode) updated_slot_mapping = attn_metadata.slot_mapping - PagedAttention.write_to_paged_cache(key, value, key_cache, - value_cache, - updated_slot_mapping, - self.kv_cache_dtype, - k_scale, v_scale) + PagedAttention.write_to_paged_cache( + key, value, key_cache, value_cache, updated_slot_mapping, + self.kv_cache_dtype, layer._k_scale, layer._v_scale) if attn_type != AttentionType.ENCODER: # Decoder self-attention supports chunked prefill. @@ -575,8 +573,8 @@ def forward( self.num_kv_heads, self.scale, self.alibi_slopes, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) # Reshape the output tensor. diff --git a/vllm/attention/backends/utils.py b/vllm/attention/backends/utils.py index d7d4a5166975c..bee50f38df4ce 100644 --- a/vllm/attention/backends/utils.py +++ b/vllm/attention/backends/utils.py @@ -123,6 +123,13 @@ class CommonMetadataBuilder(AttentionMetadataBuilder[TAttentionMetadata]): _metadata_cls: Type[TAttentionMetadata] def __init__(self, input_builder: "ModelInputForGPUBuilder"): + self.input_builder = input_builder + self.runner = input_builder.runner + + self.sliding_window = input_builder.sliding_window + self.block_size = input_builder.block_size + + def prepare(self): self.slot_mapping: List[int] = [] self.prefill_seq_lens: List[int] = [] self.context_lens: List[int] = [] @@ -135,12 +142,6 @@ def __init__(self, input_builder: "ModelInputForGPUBuilder"): self.num_prefill_tokens = 0 self.num_decode_tokens = 0 - self.input_builder = input_builder - self.runner = input_builder.runner - - self.sliding_window = input_builder.sliding_window - self.block_size = input_builder.block_size - def _add_seq_group( self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup", chunked_prefill_enabled: bool): diff --git a/vllm/attention/backends/xformers.py b/vllm/attention/backends/xformers.py index d28738a6d6f1a..4cd43947f2111 100644 --- a/vllm/attention/backends/xformers.py +++ b/vllm/attention/backends/xformers.py @@ -10,6 +10,7 @@ LowerTriangularMaskWithTensorBias) from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, + AttentionLayer, AttentionMetadata, AttentionType) from vllm.attention.backends.utils import ( CommonAttentionState, CommonMetadataBuilder, @@ -198,6 +199,8 @@ def prefill_metadata(self) -> Optional["XFormersMetadata"]: # Compute some attn_metadata fields which default to None query_start_loc = (None if self.query_start_loc is None else self.query_start_loc[:self.num_prefills + 1]) + seq_start_loc = (None if self.seq_start_loc is None else + self.seq_start_loc[:self.num_prefills + 1]) slot_mapping = (None if self.slot_mapping is None else self.slot_mapping[:self.num_prefill_tokens]) seq_lens = (None if self.seq_lens is None else @@ -224,6 +227,7 @@ def prefill_metadata(self) -> Optional["XFormersMetadata"]: max_prefill_seq_len=self.max_prefill_seq_len, max_decode_seq_len=0, query_start_loc=query_start_loc, + seq_start_loc=seq_start_loc, context_lens_tensor=context_lens_tensor, block_tables=block_tables, use_cuda_graph=False, @@ -414,15 +418,12 @@ def __init__( def forward( self, + layer: AttentionLayer, query: torch.Tensor, key: Optional[torch.Tensor], value: Optional[torch.Tensor], kv_cache: torch.Tensor, attn_metadata: "XFormersMetadata", - k_scale: float = 1.0, - v_scale: float = 1.0, - q_scale: Optional[torch.Tensor] = None, - prob_scale: Optional[torch.Tensor] = None, fp8_out_scale: Optional[torch.Tensor] = None, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: @@ -529,11 +530,9 @@ def forward( # If kv_cache is not provided, the new key and value tensors are # not cached. This happens during the initial memory # profiling run. - PagedAttention.write_to_paged_cache(key, value, key_cache, - value_cache, - updated_slot_mapping, - self.kv_cache_dtype, - k_scale, v_scale) + PagedAttention.write_to_paged_cache( + key, value, key_cache, value_cache, updated_slot_mapping, + self.kv_cache_dtype, layer._k_scale, layer._v_scale) (num_prefill_query_tokens, num_prefill_kv_tokens, num_decode_query_tokens) = \ get_num_prefill_decode_query_kv_tokens(attn_metadata, attn_type) @@ -585,8 +584,8 @@ def forward( prefill_meta.max_query_len, self.alibi_slopes, self.sliding_window, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) assert output[:num_prefill_query_tokens].shape == out.shape output[:num_prefill_query_tokens] = out @@ -612,8 +611,8 @@ def forward( self.num_kv_heads, self.scale, self.alibi_slopes, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) # Reshape the output tensor. diff --git a/vllm/attention/layer.py b/vllm/attention/layer.py index 7b612df21791c..dd6d039f9e6f8 100644 --- a/vllm/attention/layer.py +++ b/vllm/attention/layer.py @@ -78,6 +78,12 @@ def __init__( self._v_scale = torch.tensor(1.0, dtype=torch.float32) self._q_scale = torch.tensor(1.0, dtype=torch.float32) self._prob_scale = torch.tensor(1.0, dtype=torch.float32) + + # We also keep the float32 versions of k/v_scale for attention + # backends that don't support tensors (Flashinfer) + self._k_scale_float = torch.tensor(1.0, dtype=torch.float32) + self._v_scale_float = torch.tensor(1.0, dtype=torch.float32) + quant_method = quant_config.get_quant_method( self, prefix=prefix) if quant_config else None if quant_method is not None: @@ -107,7 +113,9 @@ def __init__( self.num_heads = num_heads self.head_size = head_size self.num_kv_heads = num_kv_heads + self.sliding_window = sliding_window self.backend = backend_name_to_enum(attn_backend.get_name()) + self.dtype = dtype # For cuda and cpu platforms, we control how # torch.compile works by registering the attention as one giant @@ -116,11 +124,7 @@ def __init__( self.use_direct_call = not current_platform.is_cuda( ) and not current_platform.is_cpu() - # For some attention backends, we allocate an output tensor before - # calling the custom op. When piecewise cudagraph is enabled, this - # makes sure the output tensor is allocated inside the cudagraph. - self.use_output = self.backend == _Backend.FLASH_ATTN or \ - self.backend == _Backend.FLASH_ATTN_VLLM_V1 + self.use_output = attn_backend.accept_output_buffer compilation_config = get_current_vllm_config().compilation_config if prefix in compilation_config.static_forward_context: raise ValueError(f"Duplicate layer name: {prefix}") @@ -144,12 +148,12 @@ def forward( query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, - _kv_cache: torch.Tensor, - _attn_metadata: AttentionMetadata, + kv_cache: torch.Tensor, + attn_metadata: AttentionMetadata, fp8_out_scale: Optional[torch.Tensor] = None, ) -> torch.Tensor: if self.calculate_kv_scales and \ - _attn_metadata.enable_kv_scales_calculation: + attn_metadata.enable_kv_scales_calculation: self.calc_kv_scales(query, key, value) if self.use_output: output = torch.empty_like(query) @@ -182,6 +186,8 @@ def calc_kv_scales(self, query, key, value): self._q_scale.copy_(torch.abs(query).max() / self.q_range) self._k_scale.copy_(torch.abs(key).max() / self.k_range) self._v_scale.copy_(torch.abs(value).max() / self.v_range) + self._k_scale_float = self._k_scale.item() + self._v_scale_float = self._v_scale.item() # We only calculate the scales once self.calculate_kv_scales = False @@ -210,18 +216,22 @@ def __init__( self.scale = scale self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads + assert self.num_heads % self.num_kv_heads == 0 + self.num_queries_per_kv = self.num_heads // self.num_kv_heads + dtype = torch.get_default_dtype() attn_backend = get_attn_backend(head_size, dtype, kv_cache_dtype=None, block_size=16, is_attention_free=False) - attn_backend = backend_name_to_enum(attn_backend.get_name()) - if attn_backend in {_Backend.FLASH_ATTN, _Backend.FLASH_ATTN_VLLM_V1}: - attn_backend = _Backend.XFORMERS + backend = backend_name_to_enum(attn_backend.get_name()) + if backend in {_Backend.FLASH_ATTN, _Backend.FLASH_ATTN_VLLM_V1}: + backend = _Backend.XFORMERS - self.attn_backend = attn_backend if attn_backend in { - _Backend.TORCH_SDPA, _Backend.XFORMERS + self.attn_backend = backend if backend in { + _Backend.TORCH_SDPA, + _Backend.XFORMERS, } else _Backend.TORCH_SDPA def forward( @@ -231,7 +241,7 @@ def forward( value: torch.Tensor, ) -> torch.Tensor: """Input shape: batch_size x seq_len x hidden_size""" - # TODO(Isotr0py): Use existing backend implementations and support FA2 + # TODO(Isotr0py): Use existing backend implementations and support FA3 bsz, q_len, _ = query.size() kv_len = key.size(1) @@ -239,6 +249,11 @@ def forward( key = key.view(bsz, kv_len, self.num_kv_heads, self.head_size) value = value.view(bsz, kv_len, self.num_kv_heads, self.head_size) + if (num_repeat := self.num_queries_per_kv) > 1: + # Handle MQA and GQA + key = torch.repeat_interleave(key, num_repeat, dim=2) + value = torch.repeat_interleave(value, num_repeat, dim=2) + if self.attn_backend == _Backend.XFORMERS: from xformers import ops as xops @@ -268,9 +283,8 @@ def unified_attention( attn_metadata = forward_context.attn_metadata self = forward_context.attn_layers[layer_name] kv_cache = self.kv_cache[forward_context.virtual_engine] - return self.impl.forward(query, key, value, kv_cache, attn_metadata, - self._k_scale, self._v_scale, self._q_scale, - self._prob_scale, fp8_out_scale) + return self.impl.forward(self, query, key, value, kv_cache, attn_metadata, + fp8_out_scale) def unified_attention_fake( @@ -304,15 +318,12 @@ def unified_attention_with_output( attn_metadata = forward_context.attn_metadata self = forward_context.attn_layers[layer_name] kv_cache = self.kv_cache[forward_context.virtual_engine] - self.impl.forward(query, + self.impl.forward(self, + query, key, value, kv_cache, attn_metadata, - self._k_scale, - self._v_scale, - self._q_scale, - self._prob_scale, fp8_out_scale, output=output) diff --git a/vllm/attention/ops/ipex_attn.py b/vllm/attention/ops/ipex_attn.py index cbc6c74acf09a..3a07184ed31f0 100644 --- a/vllm/attention/ops/ipex_attn.py +++ b/vllm/attention/ops/ipex_attn.py @@ -52,8 +52,8 @@ def write_to_paged_cache( value_cache: torch.Tensor, slot_mapping: torch.Tensor, kv_cache_dtype: str, - k_scale: float, - v_scale: float, + k_scale: torch.Tensor, + v_scale: torch.Tensor, *args, ) -> None: ops.reshape_and_cache( @@ -80,8 +80,8 @@ def forward_decode( num_kv_heads: int, scale: float, alibi_slopes: Optional[torch.Tensor], - k_scale: float, - v_scale: float, + k_scale: torch.Tensor, + v_scale: torch.Tensor, *args, ) -> None: tp_rank: int = 0 @@ -149,8 +149,8 @@ def write_to_paged_cache( value_cache: torch.Tensor, slot_mapping: torch.Tensor, kv_cache_dtype: str, - k_scale: float, - v_scale: float, + k_scale: torch.Tensor, + v_scale: torch.Tensor, *args, ) -> None: ipex_modules.PagedAttention.reshape_and_cache( @@ -170,8 +170,8 @@ def forward_decode( num_kv_heads: int, scale: float, alibi_slopes: Optional[torch.Tensor], - k_scale: float, - v_scale: float, + k_scale: torch.Tensor, + v_scale: torch.Tensor, *args, ) -> None: block_size = value_cache.shape[2] diff --git a/vllm/attention/ops/paged_attn.py b/vllm/attention/ops/paged_attn.py index 65e6aaedfb739..58ed82f3ba929 100644 --- a/vllm/attention/ops/paged_attn.py +++ b/vllm/attention/ops/paged_attn.py @@ -206,8 +206,8 @@ def forward_prefix( max_query_len: int, alibi_slopes: Optional[torch.Tensor], sliding_window: Optional[int], - k_scale: float, - v_scale: float, + k_scale: torch.Tensor, + v_scale: torch.Tensor, ) -> torch.Tensor: output = torch.empty_like(query) context_attention_fwd( diff --git a/vllm/attention/ops/prefix_prefill.py b/vllm/attention/ops/prefix_prefill.py index e8d6854c043af..77f92999232c2 100644 --- a/vllm/attention/ops/prefix_prefill.py +++ b/vllm/attention/ops/prefix_prefill.py @@ -564,7 +564,7 @@ def _fwd_kernel_alibi( other=0.0) # [D,N] if k_load.dtype.is_fp8(): - k = (k_load.to(tl.float32) * k_scale).to(q.dtype) + k = (k_load.to(tl.float32) * tl.load(k_scale)).to(q.dtype) else: k = k_load @@ -604,7 +604,7 @@ def _fwd_kernel_alibi( ((start_n + offs_n[:, None]) < cur_batch_ctx_len), other=0.0) if v_load.dtype.is_fp8(): - v = (v_load.to(tl.float32) * v_scale).to(q.dtype) + v = (v_load.to(tl.float32) * tl.load(v_scale)).to(q.dtype) else: v = v_load p = p.to(v.dtype) @@ -713,8 +713,8 @@ def context_attention_fwd(q, b_seq_len, b_ctx_len, max_input_len, - k_scale: float = 1.0, - v_scale: float = 1.0, + k_scale: torch.Tensor, + v_scale: torch.Tensor, alibi_slopes=None, sliding_window=None): diff --git a/vllm/attention/ops/triton_flash_attention.py b/vllm/attention/ops/triton_flash_attention.py index 828bdc2905957..bc550a85f5a92 100644 --- a/vllm/attention/ops/triton_flash_attention.py +++ b/vllm/attention/ops/triton_flash_attention.py @@ -390,6 +390,8 @@ def get_autotune_configs(): autotune_configs, autotune_keys = get_autotune_configs() +float8_info = torch.finfo(torch.float8_e4m3fnuz) + @triton.autotune( configs=autotune_configs, @@ -451,6 +453,8 @@ def attn_fwd( BIAS_TYPE: tl.constexpr, ENABLE_DROPOUT: tl.constexpr, RETURN_ENCODED_SOFTMAX: tl.constexpr, + FP8_MIN: tl.constexpr = float8_info.min, + FP8_MAX: tl.constexpr = float8_info.max, ): start_m = tl.program_id(0) off_h_q = tl.program_id(1) @@ -733,6 +737,7 @@ def attn_fwd( causal_start_idx = seqlen_q - seqlen_k if USE_FP8: acc *= o_descale + acc = tl.clamp(acc, FP8_MIN, FP8_MAX) acc = acc.to(Out.type.element_ty) if IS_CAUSAL: # noqa: SIM102 if causal_start_idx > start_m_idx and causal_start_idx < end_m_idx: @@ -832,9 +837,9 @@ def forward( def check_and_convert(t, scale): if t.dtype != float8: - finfo = torch.finfo(float8) descale = 1.0 / scale - ts = (t * descale).clamp(min=finfo.min, max=finfo.max) + ts = (t * descale).clamp(min=float8_info.min, + max=float8_info.max) return ts.to(float8) else: return t diff --git a/vllm/attention/selector.py b/vllm/attention/selector.py index 0ff007c87b1c9..81ea6eefb5410 100644 --- a/vllm/attention/selector.py +++ b/vllm/attention/selector.py @@ -14,16 +14,18 @@ logger = init_logger(__name__) -def backend_name_to_enum(backend_name: str) -> _Backend: - assert backend_name is not None - - backend_members = _Backend.__members__ - if backend_name not in backend_members: - raise ValueError(f"Invalid attention backend '{backend_name}'. " - f"Available backends: {', '.join(backend_members)} " - "(case-sensitive).") +def backend_name_to_enum(backend_name: str) -> Optional[_Backend]: + """ + Convert a string backend name to a _Backend enum value. - return _Backend[backend_name] + Returns: + * _Backend: enum value if backend_name is a valid in-tree type + * None: otherwise it's an invalid in-tree type or an out-of-tree platform is + loaded. + """ + assert backend_name is not None + return _Backend[backend_name] if backend_name in _Backend.__members__ else \ + None def get_env_variable_attn_backend() -> Optional[_Backend]: diff --git a/vllm/compilation/backends.py b/vllm/compilation/backends.py index 87655530cead4..7f4f97466d503 100644 --- a/vllm/compilation/backends.py +++ b/vllm/compilation/backends.py @@ -25,23 +25,30 @@ logger = init_logger(__name__) +@dataclasses.dataclass +class InductorArtifact: + hash_str: str = "" + file_path: str = "" + + class InductorHashCache: """ Disk format: a Python list of tuples, each tuple is - (runtime_shape, graph_index, hash_str) + (runtime_shape, graph_index, hash_str, file_path) We use list of tuple for readability. In-memory format: a defaultdict of dict, where the key is runtime_shape, and the value is a dict of graph_index to hash_str. - The data is essentially `Dict[Optional[int], Dict[int, str]]`, + The data is essentially `Dict[Optional[int], Dict[int, InductorArtifact]]`, we don't use json here because json doesn't support int as key. TODO: better off-the-shelf solution to serialize the data? """ def __init__(self, cache_dir: str, disabled: bool = False): - self.cache: defaultdict = defaultdict(dict) + self.cache: Dict[Optional[int], + Dict[int, InductorArtifact]] = defaultdict(dict) self.disabled = disabled self.cache_dir = cache_dir self.cache_file_path = os.path.join(cache_dir, @@ -66,14 +73,25 @@ def deserialize(self, data: str): # because it is a safe way to parse Python literals. # do not use eval(), it is unsafe. list_data = ast.literal_eval(data) - for runtime_shape, graph_index, hash_str in list_data: - self.cache[runtime_shape][graph_index] = hash_str + for item in list_data: + runtime_shape = item[0] + graph_index = item[1] + hash_str = item[2] + # for compatibility of old version, + # where we don't have file_path. + # NOTE: after running the new code, the file_path + # will be updated. + file_path = "" if len(item) == 3 else item[3] + self.cache[runtime_shape][graph_index] = InductorArtifact( + hash_str=hash_str, file_path=file_path) def serialize(self) -> str: data = [] - for runtime_shape, graph_index_to_hash_str in self.cache.items(): - for graph_index, hash_str in graph_index_to_hash_str.items(): - data.append((runtime_shape, graph_index, hash_str)) + for runtime_shape, value in self.cache.items(): + for graph_index, inductor_artifact in value.items(): + data.append( + (runtime_shape, graph_index, inductor_artifact.hash_str, + inductor_artifact.file_path)) printer = pprint.PrettyPrinter(indent=4) return printer.pformat(data) @@ -90,13 +108,14 @@ def __contains__(self, key: Tuple[Optional[int], int]) -> bool: return runtime_shape in self.cache and graph_index in self.cache[ runtime_shape] - def __getitem__(self, key: Tuple[Optional[int], int]) -> str: + def __getitem__(self, key: Tuple[Optional[int], int]) -> InductorArtifact: if self.disabled: raise KeyError("cannot read from disabled cache") runtime_shape, graph_index = key return self.cache[runtime_shape][graph_index] - def __setitem__(self, key: Tuple[Optional[int], int], value: str): + def __setitem__(self, key: Tuple[Optional[int], int], + value: InductorArtifact): # setitem for disabled cache is fine, because we # don't actually write to the disk runtime_shape, graph_index = key @@ -181,7 +200,8 @@ def wrap_inductor(graph: fx.GraphModule, if (runtime_shape, graph_index) in cache_data: # we compiled this graph before # so we can directly lookup the compiled graph via hash - hash_str = cache_data[(runtime_shape, graph_index)] + inductor_artifact = cache_data[(runtime_shape, graph_index)] + hash_str = inductor_artifact.hash_str if graph_index == 0: # adds some info logging for the first graph logger.info( @@ -199,6 +219,7 @@ def wrap_inductor(graph: fx.GraphModule, "Inductor cache lookup failed. Please remove" f"the cache file {cache_data.cache_file_path} and try again." # noqa ) + inductor_artifact.file_path = inductor_compiled_graph.current_callable.__code__.co_filename # noqa # Inductor calling convention (function signature): # f(list) -> tuple @@ -224,19 +245,20 @@ def compiled_graph(*args): # the assumption is that we don't have nested Inductor compilation. # compiled_fx_graph_hash will only be called once, and we can hook # it to get the hash of the compiled graph directly. - from torch._inductor.codecache import compiled_fx_graph_hash + + inductor_artifact = InductorArtifact() + from torch._inductor.codecache import (FxGraphCache, + compiled_fx_graph_hash) + original_load = FxGraphCache.load + + def hijack_load(*args, **kwargs): + inductor_compiled_graph = original_load(*args, **kwargs) + inductor_artifact.file_path = inductor_compiled_graph.current_callable.__code__.co_filename # noqa + return inductor_compiled_graph def hijack_compiled_fx_graph_hash(*args, **kwargs): out = compiled_fx_graph_hash(*args, **kwargs) - # store the hash in the cache - nonlocal cache_data - cache_data[(runtime_shape, graph_index)] = out[0] - if graph_index == 0: - # adds some info logging for the first graph - logger.info("Cache the graph of shape %s for later use", - str(runtime_shape)) - logger.debug("store the %s-th graph for shape %s via hash %s", - graph_index, str(runtime_shape), out[0]) + inductor_artifact.hash_str = out[0] return out def _check_can_cache(*args, **kwargs): @@ -251,19 +273,45 @@ def _check_can_cache(*args, **kwargs): def _get_shape_env() -> AlwaysHitShapeEnv: return AlwaysHitShapeEnv() - with patch(# for hijacking the hash of the compiled graph - "torch._inductor.codecache.compiled_fx_graph_hash", - hijack_compiled_fx_graph_hash), \ - patch(# for providing a dummy shape environment - "torch._inductor.codecache.FxGraphCache._get_shape_env", - _get_shape_env), \ - patch(# for forcing the graph to be cached - "torch._inductor.codecache.FxGraphCache._check_can_cache", - _check_can_cache): + with ExitStack() as stack: + if not cache_data.disabled: + # compilation cache is enabled, patch several functions + + # hijack to get the compiled graph itself + stack.enter_context( + patch("torch._inductor.codecache.FxGraphCache.load", + hijack_load)) + + # for hijacking the hash of the compiled graph + stack.enter_context( + patch("torch._inductor.codecache.compiled_fx_graph_hash", + hijack_compiled_fx_graph_hash)) + + # for providing a dummy shape environment + stack.enter_context( + patch( + "torch._inductor.codecache.FxGraphCache._get_shape_env", + _get_shape_env)) + + # for forcing the graph to be cached + stack.enter_context( + patch( + "torch._inductor.codecache.FxGraphCache._check_can_cache", + _check_can_cache)) + compiled_graph = compile_fx(graph, example_inputs, config_patches=current_config) - + # store the inductor_artifact in the cache + cache_data[(runtime_shape, graph_index)] = inductor_artifact + if graph_index == 0: + # adds some info logging for the first graph + logger.info("Cache the graph of shape %s for later use", + str(runtime_shape)) + logger.debug( + "store the %s-th graph for shape %s via hash %s from file %s", + graph_index, str(runtime_shape), inductor_artifact.hash_str, + inductor_artifact.file_path) # after compiling the last graph, record the end time if graph_index == num_graphs - 1: now = time.time() @@ -476,6 +524,7 @@ def configure_post_pass(self): def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable: + vllm_config = self.vllm_config if not self.compilation_config.cache_dir: # no provided cache dir, generate one based on the known factors # that affects the compilation. if none of the factors change, @@ -484,7 +533,6 @@ def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable: # 1. factors come from the vllm_config (it mainly summarizes how the # model is created) - vllm_config = self.vllm_config config_hash = vllm_config.compute_hash() # 2. factors come from the code files that are traced by Dynamo ( @@ -508,20 +556,26 @@ def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable: hash_key = hashlib.md5( f"{config_hash}_{code_hash}".encode()).hexdigest()[:10] cache_dir = os.path.join( - envs.VLLM_CACHE_ROOT, "torch_compile_cache", hash_key, - f"rank_{vllm_config.parallel_config.rank}") - else: - cache_dir = self.compilation_config.cache_dir + envs.VLLM_CACHE_ROOT, + "torch_compile_cache", + hash_key, + ) + self.compilation_config.cache_dir = cache_dir + + cache_dir = self.compilation_config.cache_dir os.makedirs(cache_dir, exist_ok=True) + local_cache_dir = os.path.join( + cache_dir, f"rank_{vllm_config.parallel_config.rank}") + self.compilation_config.local_cache_dir = local_cache_dir disabled = envs.VLLM_DISABLE_COMPILE_CACHE self.inductor_hash_cache: InductorHashCache = InductorHashCache( - cache_dir, disabled=disabled) + local_cache_dir, disabled=disabled) if disabled: logger.info("vLLM's torch.compile cache is disabled.") else: logger.info("Using cache directory: %s for vLLM's torch.compile", - cache_dir) + local_cache_dir) # when dynamo calls the backend, it means the bytecode # transform and analysis are done @@ -561,6 +615,18 @@ def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable: self.vllm_config, self.graph_pool, self).run(*example_inputs) + graph_path = os.path.join(local_cache_dir, "computation_graph.py") + if not os.path.exists(graph_path): + # code adapted from https://github.com/thuml/depyf/blob/dab831108a752d1facc00acdd6d4243891845c37/depyf/explain/patched_lazy_format_graph_code.py#L30 # noqa + # use `print_readable` because it can include submodules + src = "from __future__ import annotations\nimport torch\n" + \ + self.split_gm.print_readable(print_output=False) + src = src.replace("", "GraphModule") + with open(graph_path, "w") as f: + f.write(src) + + logger.debug("Computation graph saved to %s", graph_path) + self._called = True if not self.compilation_config.use_cudagraph or \ @@ -576,9 +642,13 @@ def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable: ] # index of tensors that have symbolic shapes (batch size) + # for weights and static buffers, they will have concrete shapes. + # symbolic shape only happens for input tensors. + from torch.fx.experimental.symbolic_shapes import is_symbolic self.sym_tensor_indices = [ i for i, x in enumerate(fake_args) - if isinstance(x, torch._subclasses.fake_tensor.FakeTensor) + if isinstance(x, torch._subclasses.fake_tensor.FakeTensor) and \ + any(is_symbolic(d) for d in x.size()) ] # compiler managed cudagraph input buffers @@ -610,7 +680,7 @@ def copy_and_call(*args): class ConcreteSizeEntry: runtime_shape: int need_to_compile: bool # the size is in compile_sizes - use_cudagraph: bool # the size is in capture_sizes + use_cudagraph: bool # the size is in cudagraph_capture_sizes compiled: bool = False runnable: Callable = None # type: ignore @@ -657,8 +727,8 @@ def __init__(self, graph: fx.GraphModule, vllm_config: VllmConfig, self.compile_sizes: Set[int] = set( self.compilation_config.compile_sizes) - self.capture_sizes: Set[int] = set( - self.compilation_config.capture_sizes + self.cudagraph_capture_sizes: Set[int] = set( + self.compilation_config.cudagraph_capture_sizes ) if self.compilation_config.use_cudagraph else set() self.first_run_finished = False @@ -676,11 +746,11 @@ def __init__(self, graph: fx.GraphModule, vllm_config: VllmConfig, # to_be_compiled_sizes tracks the remaining sizes to compile, # and updates during the compilation process, so we need to copy it self.to_be_compiled_sizes: Set[int] = self.compile_sizes.copy() - for shape in self.compile_sizes.union(self.capture_sizes): + for shape in self.compile_sizes.union(self.cudagraph_capture_sizes): self.concrete_size_entries[shape] = ConcreteSizeEntry( runtime_shape=shape, need_to_compile=shape in self.compile_sizes, - use_cudagraph=shape in self.capture_sizes, + use_cudagraph=shape in self.cudagraph_capture_sizes, ) def check_for_ending_compilation(self): diff --git a/vllm/compilation/decorators.py b/vllm/compilation/decorators.py index 10513111ea7f1..17eb0592ced6d 100644 --- a/vllm/compilation/decorators.py +++ b/vllm/compilation/decorators.py @@ -76,8 +76,8 @@ def forward(self, x: torch.Tensor, y: Optional[torch.Tensor]): During runtime, when we actually mark dimensions of tensors, it depends on the value of arguments: - - if it is a single integer, the corresponding dimension of the argument - will be marked as dynamic. + - if it is a single integer (can be negative), the corresponding dimension + of the argument will be marked as dynamic. - if it is `None`, ignored. - if it is `IntermediateTensors`, all the tensors in the intermediate tensors will be marked as dynamic. @@ -177,10 +177,20 @@ def __call__(self, *args, **kwargs): for k, dims in dynamic_arg_dims.items(): arg = bound_args.arguments.get(k) if arg is not None: + dims = [dims] if isinstance(dims, int) else dims if isinstance(arg, torch.Tensor): + # In case dims is specified with negative indexing + dims = [ + arg.ndim + dim if dim < 0 else dim for dim in dims + ] torch._dynamo.mark_dynamic(arg, dims) elif isinstance(arg, IntermediateTensors): for tensor in arg.tensors.values(): + # In case dims is specified with negative indexing + dims = [ + tensor.ndim + dim if dim < 0 else dim + for dim in dims + ] torch._dynamo.mark_dynamic(tensor, dims) else: raise ValueError( @@ -188,6 +198,8 @@ def __call__(self, *args, **kwargs): f" {dims} for argument {k} with type {type(arg)}.") # here, it is the starting point of the `torch.compile` process start_monitoring_torch_compile(self.vllm_config) + logger.debug("Start compiling function %s", + self.original_code_object) # if we don't use custom dispatcher, we can directly call the # compiled function and let torch.compile handle the dispatching, diff --git a/vllm/compilation/wrapper.py b/vllm/compilation/wrapper.py index e3260a10c02ae..f17c84204c09c 100644 --- a/vllm/compilation/wrapper.py +++ b/vllm/compilation/wrapper.py @@ -9,6 +9,9 @@ import vllm.envs as envs from vllm.config import CompilationLevel, get_current_vllm_config +from vllm.logger import init_logger + +logger = init_logger(__name__) class TorchCompileWrapperWithCustomDispatcher: @@ -82,6 +85,32 @@ def bytecode_hook(self, old_code: CodeType, new_code: CodeType): return self.compiled_codes.append(new_code) + local_cache_dir = self.vllm_config.compilation_config.local_cache_dir + if isinstance(local_cache_dir, str): + decompiled_file = os.path.join(local_cache_dir, + "transformed_code.py") + if not os.path.exists(decompiled_file): + try: + # usually the decompilation will succeed for most models, + # as we guarantee a full-graph compilation in Dynamo. + # but there's no 100% guarantee, since decompliation is + # not a reversible process. + import depyf + src = depyf.decompile(new_code) + with open(decompiled_file, "w") as f: + f.write(src) + + logger.debug("Dynamo transformed code saved to %s", + decompiled_file) + except Exception: + pass + + if self.vllm_config.compilation_config.use_cudagraph and \ + "update" in new_code.co_names: + import depyf + src = depyf.decompile(new_code) + msg = "Assigning / modifying buffers of nn.Module during forward pass is not allowed when using cudagraph inside the compiler because it will cause silent errors. Please use eager mode or fix the code. The following code contains clues about which buffer is being modified (please search for the usage of the function `update`):\n" + src # noqa + raise RuntimeError(msg) if self.vllm_config.compilation_config.use_cudagraph and \ "update" in new_code.co_names: diff --git a/vllm/config.py b/vllm/config.py index 46b8b4c62d694..45820f5b641d5 100644 --- a/vllm/config.py +++ b/vllm/config.py @@ -190,39 +190,42 @@ def compute_hash(self) -> str: factors.append(self.rope_theta) return hashlib.sha256(str(factors).encode()).hexdigest() - def __init__(self, - model: str, - task: Union[TaskOption, Literal["draft"]], - tokenizer: str, - tokenizer_mode: str, - trust_remote_code: bool, - dtype: Union[str, torch.dtype], - seed: int, - allowed_local_media_path: str = "", - revision: Optional[str] = None, - code_revision: Optional[str] = None, - rope_scaling: Optional[Dict[str, Any]] = None, - rope_theta: Optional[float] = None, - tokenizer_revision: Optional[str] = None, - max_model_len: Optional[int] = None, - spec_target_max_model_len: Optional[int] = None, - quantization: Optional[str] = None, - enforce_eager: Optional[bool] = None, - max_seq_len_to_capture: Optional[int] = None, - max_logprobs: int = 20, - disable_sliding_window: bool = False, - skip_tokenizer_init: bool = False, - served_model_name: Optional[Union[str, List[str]]] = None, - limit_mm_per_prompt: Optional[Mapping[str, int]] = None, - use_async_output_proc: bool = True, - config_format: ConfigFormat = ConfigFormat.AUTO, - hf_overrides: Optional[HfOverrides] = None, - mm_processor_kwargs: Optional[Dict[str, Any]] = None, - disable_mm_preprocessor_cache: bool = False, - override_neuron_config: Optional[Dict[str, Any]] = None, - override_pooler_config: Optional["PoolerConfig"] = None, - logits_processor_pattern: Optional[str] = None, - generation_config: Optional[str] = None) -> None: + def __init__( + self, + model: str, + task: Union[TaskOption, Literal["draft"]], + tokenizer: str, + tokenizer_mode: str, + trust_remote_code: bool, + dtype: Union[str, torch.dtype], + seed: int, + allowed_local_media_path: str = "", + revision: Optional[str] = None, + code_revision: Optional[str] = None, + rope_scaling: Optional[Dict[str, Any]] = None, + rope_theta: Optional[float] = None, + tokenizer_revision: Optional[str] = None, + max_model_len: Optional[int] = None, + spec_target_max_model_len: Optional[int] = None, + quantization: Optional[str] = None, + enforce_eager: Optional[bool] = None, + max_seq_len_to_capture: Optional[int] = None, + max_logprobs: int = 20, + disable_sliding_window: bool = False, + skip_tokenizer_init: bool = False, + served_model_name: Optional[Union[str, List[str]]] = None, + limit_mm_per_prompt: Optional[Mapping[str, int]] = None, + use_async_output_proc: bool = True, + config_format: ConfigFormat = ConfigFormat.AUTO, + hf_overrides: Optional[HfOverrides] = None, + mm_processor_kwargs: Optional[Dict[str, Any]] = None, + disable_mm_preprocessor_cache: bool = False, + override_neuron_config: Optional[Dict[str, Any]] = None, + override_pooler_config: Optional["PoolerConfig"] = None, + logits_processor_pattern: Optional[str] = None, + generation_config: Optional[str] = None, + enable_sleep_mode: bool = False, + ) -> None: self.model = model self.tokenizer = tokenizer self.tokenizer_mode = tokenizer_mode @@ -270,6 +273,12 @@ def __init__(self, self.max_logprobs = max_logprobs self.disable_sliding_window = disable_sliding_window self.skip_tokenizer_init = skip_tokenizer_init + self.enable_sleep_mode = enable_sleep_mode + + from vllm.platforms import current_platform + + if self.enable_sleep_mode and not current_platform.is_cuda(): + raise ValueError("Sleep mode is only supported on CUDA devices.") hf_config = get_config(self.model, trust_remote_code, revision, code_revision, config_format) @@ -302,14 +311,15 @@ def __init__(self, (self.hf_text_config.model_type in ["gemma2", "cohere2"])) if (not self.disable_sliding_window and has_interleaved_attention): - if envs.VLLM_ATTENTION_BACKEND == "XFORMERS": + if (backend := + envs.VLLM_ATTENTION_BACKEND) in ("XFORMERS", "FLASHINFER"): sliding_window_len_min = get_min_sliding_window( self.hf_text_config.sliding_window) logger.warning_once( f"{self.hf_text_config.model_type} has interleaved " "attention, which is currently not supported by the " - "XFORMERS backend. Disabling sliding window and capping " + f"{backend} backend. Disabling sliding window and capping " "the max length to the sliding window size " f"({sliding_window_len_min}).") self.disable_sliding_window = True @@ -350,6 +360,10 @@ def __init__(self, supported_tasks, task = self._resolve_task(task, self.hf_config) self.supported_tasks = supported_tasks self.task: Final = task + if self.task in ("draft", "generate"): + self.truncation_side = "left" + else: + self.truncation_side = "right" self.pooler_config = self._init_pooler_config(override_pooler_config) self.logits_processor_pattern = logits_processor_pattern @@ -546,7 +560,7 @@ def _verify_quantization(self) -> None: optimized_quantization_methods = [ "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin", "awq_marlin", "fbgemm_fp8", "compressed_tensors", - "compressed-tensors", "experts_int8" + "compressed-tensors", "experts_int8", "quark" ] if self.quantization is not None: self.quantization = self.quantization.lower() @@ -598,10 +612,12 @@ def _verify_cuda_graph(self) -> None: self.max_seq_len_to_capture = min(self.max_seq_len_to_capture, self.max_model_len) - if (self.hf_config.model_type == 'deepseek_v3' + MODEL_NOT_SUPPORT_CUDA_GRAPH = ['mllama'] + if (self.hf_config.model_type in MODEL_NOT_SUPPORT_CUDA_GRAPH and not self.enforce_eager): - logger.warning("CUDA graph is not supported for Deepseek V3 yet, " - "fallback to the eager mode.") + logger.warning( + "CUDA graph is not supported for %s yet, fallback to the eager " + "mode.", self.hf_config.model_type) self.enforce_eager = True def _verify_bnb_config(self) -> None: @@ -724,9 +740,12 @@ def get_head_size(self) -> int: if hasattr(self.hf_text_config, "model_type") and (self.hf_text_config.model_type in ('deepseek_v2', 'deepseek_v3')): - # FlashAttention supports only head_size 32, 64, 128, 256, - # we need to pad head_size 192 to 256 - return 256 + qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim", + 0) + qk_nope_head_dim = getattr(self.hf_text_config, "qk_nope_head_dim", + 0) + if qk_rope_head_dim and qk_nope_head_dim: + return qk_rope_head_dim + qk_nope_head_dim if self.is_attention_free: return 0 @@ -896,12 +915,18 @@ def get_diff_sampling_param(self) -> Dict[str, Any]: "top_k", "top_p", "min_p", + "max_new_tokens", ] if any(p in config for p in available_params): diff_sampling_param = { p: config.get(p) for p in available_params if config.get(p) is not None } + # Huggingface definition of max_new_tokens is equivalent + # to vLLM's max_tokens + if "max_new_tokens" in diff_sampling_param: + diff_sampling_param["max_tokens"] = diff_sampling_param.pop( + "max_new_tokens") else: diff_sampling_param = {} return diff_sampling_param @@ -1213,9 +1238,6 @@ class ParallelConfig: pipeline_parallel_size: int = 1 # Number of pipeline parallel groups. tensor_parallel_size: int = 1 # Number of tensor parallel groups. - # Deprecated, use distributed_executor_backend instead. - worker_use_ray: Optional[bool] = None - # Maximum number of multiple batches # when load model sequentially. To avoid RAM OOM when using tensor # parallel and large models. @@ -1269,14 +1291,7 @@ def __post_init__(self) -> None: self.world_size = self.pipeline_parallel_size * \ self.tensor_parallel_size - if self.worker_use_ray: - if self.distributed_executor_backend is None: - self.distributed_executor_backend = "ray" - elif not self.use_ray: - raise ValueError(f"worker-use-ray can't be used with " - f"distributed executor backend " - f"'{self.distributed_executor_backend}'.") - ray_only_devices = ["tpu", "hpu"] + ray_only_devices = ["tpu"] from vllm.platforms import current_platform if (current_platform.device_type in ray_only_devices and self.world_size > 1): @@ -1294,8 +1309,11 @@ def __post_init__(self) -> None: from vllm.executor import ray_utils backend = "mp" ray_found = ray_utils.ray_is_available() - if (current_platform.is_cuda() - and cuda_device_count_stateless() < self.world_size): + if current_platform.is_neuron(): + # neuron uses single process to control multiple devices + backend = "uni" + elif (current_platform.is_cuda() + and cuda_device_count_stateless() < self.world_size): if not ray_found: raise ValueError("Unable to load Ray which is " "required for multi-node inference, " @@ -1351,13 +1369,15 @@ def _verify_args(self) -> None: # Lazy import to avoid circular import from vllm.executor.executor_base import ExecutorBase if self.distributed_executor_backend not in ( - "ray", "mp", None) and not (isinstance( + "ray", "mp", "uni", + "external_launcher", None) and not (isinstance( self.distributed_executor_backend, type) and issubclass( self.distributed_executor_backend, ExecutorBase)): raise ValueError( "Unrecognized distributed executor backend " f"{self.distributed_executor_backend}. Supported " - "values are 'ray', 'mp' or custom ExecutorBase subclass.") + "values are 'ray', 'mp' 'uni', 'external_launcher' or" + " custom ExecutorBase subclass.") if self.use_ray: from vllm.executor import ray_utils ray_utils.assert_ray_available() @@ -1403,13 +1423,15 @@ class SchedulerConfig: is_multimodal_model: bool = False - # FIXME(woosuk & ywang96): Below are placeholder values. We need to - # calculate the actual values from the configurations. - # Multimodal encoder run compute budget, only used in V1 - max_num_encoder_input_tokens = 16384 + # NOTE: The following multimodal encoder budget will be initialized to + # max_num_batched_tokens and overridden in case max multimodal embedding + # size is larger. + # TODO (ywang96): Make these configurable. + # Multimodal encoder compute budget, only used in V1 + max_num_encoder_input_tokens: int = field(default=None) # type: ignore # Multimodal encoder cache size, only used in V1 - encoder_cache_size = 16384 + encoder_cache_size: int = field(default=None) # type: ignore # Whether to perform preemption by swapping or # recomputation. If not specified, we determine the mode as follows: @@ -1483,6 +1505,9 @@ def __post_init__(self) -> None: _MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS, ) + self.max_num_encoder_input_tokens = self.max_num_batched_tokens + self.encoder_cache_size = self.max_num_batched_tokens + if self.enable_chunked_prefill: logger.info( "Chunked prefill is enabled with max_num_batched_tokens=%d.", @@ -2708,10 +2733,11 @@ class CompilationConfig(BaseModel): - use_inductor: whether to use inductor compilation. - False: inductor compilation is not used. graph runs in eager. - True: inductor compilation is used. one graph for symbolic shape - is compiled. In addition, compile for cudagraph sizes that are - in candidate_compile_sizes, using configurations - in inductor_compile_config. - - candidate_compile_sizes: sizes to compile for inductor. + is compiled. In addition, compile for compile_sizes, + using configurations in inductor_compile_config. + - compile_sizes: sizes to compile for inductor. In addition + to integers, it also supports "cudagraph_capture_sizes" to + specify the sizes for cudagraph capture. - inductor_compile_config: additional configurations for inductor. - None: use default configurations. - inductor_passes: additional passes for inductor. It is a dictionary @@ -2739,7 +2765,7 @@ class CompilationConfig(BaseModel): splitting_ops: List[str] = Field(default=None) # type: ignore use_inductor: bool = True - candidate_compile_sizes: Optional[List[int]] = Field(default=None) + compile_sizes: Optional[List[Union[int, str]]] = Field(default=None) inductor_compile_config: Dict = Field(default_factory=dict) inductor_passes: Dict[str, str] = Field(default_factory=dict) @@ -2787,9 +2813,8 @@ def model_post_init(self, __context: Any) -> None: pass_config: PassConfig = Field(default_factory=PassConfig) # not configurable, computed after init - compile_sizes: List[int] = PrivateAttr - capture_sizes: List[int] = PrivateAttr max_capture_size: int = PrivateAttr + local_cache_dir: str = PrivateAttr # local cache dir for each rank # optimization: # Intuitively, bs_to_padded_graph_size should be Dict[int, int]. # since we know all keys are in a range [0, max_capture_size], @@ -2867,17 +2892,8 @@ def model_post_init(self, __context: Any) -> None: "vllm.unified_attention_with_output", ] else: - # v0 can use full graph compilation without splitting, - # splitting is optional. - # right now we still need it. kv cache shape - # will be included in the graph if we don't split - # the graph. - # TODO: hide kv cache in static forward context - # so that inductor does not see it. - self.splitting_ops = [ - "vllm.unified_attention", - "vllm.unified_attention_with_output", - ] + # v0 uses full graph compilation + self.splitting_ops = [] for k, v in self.inductor_passes.items(): if not isinstance(v, str): @@ -2923,43 +2939,47 @@ def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]: from vllm.compilation.backends import VllmBackend return VllmBackend(vllm_config) - def init_with_cudagraph_sizes(self, sizes_to_specialize: List[int]): + def init_with_cudagraph_sizes(self, + cudagraph_capture_sizes: List[int]) -> None: """To complete the initialization of config, we need to know the cudagraph sizes.""" if self.cudagraph_capture_sizes is None: - self.capture_sizes = sizes_to_specialize + self.cudagraph_capture_sizes = cudagraph_capture_sizes else: - self.capture_sizes = self.cudagraph_capture_sizes + # de-duplicate the sizes provided by the config + self.cudagraph_capture_sizes = list( + set(self.cudagraph_capture_sizes)) logger.info(("cudagraph sizes specified by model runner" " %s is overridden by config %s"), - sizes_to_specialize, self.cudagraph_capture_sizes) - - if self.candidate_compile_sizes is None: - self.candidate_compile_sizes = [] - self.compile_sizes = [ - x for x in self.candidate_compile_sizes if x in self.capture_sizes - ] - ignored_sizes = [ - x for x in self.candidate_compile_sizes - if x not in self.capture_sizes - ] - if ignored_sizes: - logger.warning(("candidate_compile_sizes %s are ignored " - "because they are not cudagraph capture sizes."), - ignored_sizes) + cudagraph_capture_sizes, self.cudagraph_capture_sizes) + + computed_compile_sizes = [] + if self.compile_sizes is not None: + # de-duplicate the sizes provided by the config + self.compile_sizes = list(set(self.compile_sizes)) + for x in self.compile_sizes: + if isinstance(x, str): + assert x == "cudagraph_capture_sizes", \ + "Unrecognized size type in compile_sizes, " \ + f"expect 'cudagraph_capture_sizes', got {x}" + computed_compile_sizes.extend(self.cudagraph_capture_sizes) + else: + assert isinstance(x, int) + computed_compile_sizes.append(x) + self.compile_sizes = computed_compile_sizes # type: ignore # sort to make sure cudagraph capture sizes are in descending order - self.capture_sizes.sort(reverse=True) - self.max_capture_size = self.capture_sizes[ - 0] if self.capture_sizes else 0 + self.cudagraph_capture_sizes.sort(reverse=True) + self.max_capture_size = self.cudagraph_capture_sizes[ + 0] if self.cudagraph_capture_sizes else 0 # pre-compute the mapping from batch size to padded graph size self.bs_to_padded_graph_size = [ 0 for i in range(self.max_capture_size + 1) ] - for end, start in zip(self.capture_sizes, - self.capture_sizes[1:] + [0]): + for end, start in zip(self.cudagraph_capture_sizes, + self.cudagraph_capture_sizes[1:] + [0]): for bs in range(start, end): if bs == start: self.bs_to_padded_graph_size[bs] = start @@ -3179,7 +3199,8 @@ def __post_init__(self): if self.compilation_config is None: self.compilation_config = CompilationConfig() - if envs.VLLM_USE_V1 and not self.model_config.enforce_eager: + if envs.VLLM_USE_V1 and self.model_config is not None and \ + not self.model_config.enforce_eager: # NOTE(woosuk): Currently, we use inductor because the piecewise # CUDA graphs do not work properly with the custom CUDA kernels. # FIXME(woosuk): Disable inductor to reduce the compilation time @@ -3229,14 +3250,14 @@ def _set_cudagraph_sizes(self): However, if users specify the cudagraph capture sizes through compilation config, we will use the specified sizes instead. - In the end, `vllm_config.compilation_config.capture_sizes` will be the - final sizes to capture cudagraph (in descending order). + In the end, `vllm_config.compilation_config.cudagraph_capture_sizes` + will be the final sizes to capture cudagraph (in descending order). During runtime, if batchsize is larger than - `vllm_config.compilation_config.capture_sizes`, + `vllm_config.compilation_config.cudagraph_capture_sizes`, no cudagraph will be used. If the batch size is no larger than - `vllm_config.compilation_config.capture_sizes`, + `vllm_config.compilation_config.cudagraph_capture_sizes`, we can quickly find the padded graph size for a given batch size by looking up `vllm_config.compilation_config.bs_to_padded_graph_size`. """ @@ -3318,7 +3339,7 @@ def __str__(self): @contextmanager -def set_current_vllm_config(vllm_config: VllmConfig): +def set_current_vllm_config(vllm_config: VllmConfig, check_compile=False): """ Temporarily set the current VLLM config. Used during model initialization. @@ -3338,7 +3359,8 @@ def set_current_vllm_config(vllm_config: VllmConfig): vllm_config.compilation_config.enabled_custom_ops) logger.debug("disabled custom ops: %s", vllm_config.compilation_config.disabled_custom_ops) - if vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \ + if check_compile and \ + vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \ and compilation_counter.num_models_seen == num_models_seen: # If the model supports compilation, # compilation_counter.num_models_seen should be increased diff --git a/vllm/connections.py b/vllm/connections.py index e785a0b3ebd74..4c9f4f40cf640 100644 --- a/vllm/connections.py +++ b/vllm/connections.py @@ -29,7 +29,7 @@ def get_sync_client(self) -> requests.Session: # required, so that the client is only accessible inside async event loop async def get_async_client(self) -> aiohttp.ClientSession: if self._async_client is None or not self.reuse_client: - self._async_client = aiohttp.ClientSession() + self._async_client = aiohttp.ClientSession(trust_env=True) return self._async_client diff --git a/vllm/core/block/cpu_gpu_block_allocator.py b/vllm/core/block/cpu_gpu_block_allocator.py index 3a57487a6cd8a..c3e1665b4464e 100644 --- a/vllm/core/block/cpu_gpu_block_allocator.py +++ b/vllm/core/block/cpu_gpu_block_allocator.py @@ -339,6 +339,13 @@ def get_prefix_cache_hit_rate(self, device: Device) -> float: assert device in self._allocators return self._allocators[device].get_prefix_cache_hit_rate() + def reset_prefix_cache(self) -> bool: + """Reset prefix cache for all devices.""" + success = True + for allocator in self._allocators.values(): + success = success and allocator.reset_prefix_cache() + return success + def get_and_reset_swaps(self) -> List[Tuple[int, int]]: """Returns and clears the mapping of source to destination block IDs. Will be called after every swapping operations for now, and after every diff --git a/vllm/core/block/interfaces.py b/vllm/core/block/interfaces.py index 985a1098b6cd1..cb432db919c73 100644 --- a/vllm/core/block/interfaces.py +++ b/vllm/core/block/interfaces.py @@ -192,6 +192,11 @@ def get_prefix_cache_hit_rate(self) -> float: """Prefix cache hit rate. -1 means not supported or disabled.""" pass + @abstractmethod + def reset_prefix_cache(self) -> bool: + """Reset prefix cache.""" + pass + class NoFreeBlocksError(ValueError): pass @@ -297,6 +302,11 @@ def get_prefix_cache_hit_rate(self, device: Device) -> float: """Prefix cache hit rate. -1 means not supported or disabled.""" pass + @abstractmethod + def reset_prefix_cache(self) -> bool: + """Reset prefix cache.""" + pass + @abstractmethod def find_cached_blocks_prefix( self, diff --git a/vllm/core/block/naive_block.py b/vllm/core/block/naive_block.py index 9b94918ab38ef..c38ae2dd6761b 100644 --- a/vllm/core/block/naive_block.py +++ b/vllm/core/block/naive_block.py @@ -1,5 +1,5 @@ from collections import deque -from typing import Deque, FrozenSet, Iterable, List, Optional, Tuple +from typing import Deque, FrozenSet, Iterable, List, Optional, Tuple, Union from vllm.core.block.common import (BlockPool, CopyOnWriteTracker, RefCounter, get_all_blocks_recursively) @@ -136,16 +136,18 @@ def _allocate_block_id(self) -> BlockId: self._refcounter.incr(block_id) return block_id - def _free_block_id(self, block: Block) -> None: - block_id = block.block_id + def _free_block_id(self, block: Union[Block, BlockId]) -> None: + if isinstance(block, Block): + block_id = block.block_id + block.block_id = None + else: + block_id = block assert block_id is not None refcount = self._refcounter.decr(block_id) if refcount == 0: self._free_block_indices.appendleft(block_id) - block.block_id = None - def free(self, block: Block, keep_block_object: bool = False) -> None: # Release the physical block id self._free_block_id(block) @@ -154,6 +156,9 @@ def free(self, block: Block, keep_block_object: bool = False) -> None: if not keep_block_object: self._block_pool.free_block(block) + def free_block_id(self, block_id: BlockId) -> None: + self._free_block_id(block_id) + def fork(self, last_block: Block) -> List[Block]: """Creates a new sequence of blocks that shares the same underlying memory as the original sequence. @@ -325,6 +330,10 @@ def swap_in(self, blocks: List[Block]) -> None: def get_prefix_cache_hit_rate(self) -> float: return -1 + def reset_prefix_cache(self) -> bool: + """No prefix cache for naive block allocator.""" + return True + def find_cached_blocks_prefix(self, block_hashes: List[int]) -> List[int]: # Not applicable for naive block allocator. return [] diff --git a/vllm/core/block/prefix_caching_block.py b/vllm/core/block/prefix_caching_block.py index 1238303234deb..ccdc5daa9595c 100644 --- a/vllm/core/block/prefix_caching_block.py +++ b/vllm/core/block/prefix_caching_block.py @@ -12,6 +12,7 @@ from vllm.core.block.naive_block import (BlockPool, NaiveBlock, NaiveBlockAllocator) from vllm.core.evictor import EvictionPolicy, Evictor, make_evictor +from vllm.logger import init_logger from vllm.sequence import Sequence PrefixHash = int @@ -21,6 +22,8 @@ # then we know this block hasn't been accessed yet. _DEFAULT_LAST_ACCESSED_TIME = -1 +logger = init_logger(__name__) + class BlockTracker: """Used to track the status of a block inside the prefix caching allocator @@ -105,7 +108,8 @@ def __init__( # Evitor used to maintain how we want to handle those computed blocks # if we find memory pressure is high. - self.evictor: Evictor = make_evictor(eviction_policy) + self.eviction_policy = eviction_policy + self.evictor: Evictor = make_evictor(self.eviction_policy) # We share the refcounter between allocators. This allows us to promote # blocks originally allocated in the hashless allocator to immutable @@ -428,6 +432,44 @@ def all_block_ids(self) -> FrozenSet[int]: def get_prefix_cache_hit_rate(self) -> float: return self.metric_data.get_hit_rate() + def reset_prefix_cache(self) -> bool: + """Reset prefix cache. This function may be used in RLHF + flows to invalid prefix caching after the weights are updated, + or used for resetting prefix caching status for benchmarking. + + Returns: + bool: True if the prefix cache is successfully reset, + False otherwise. + """ + num_used_blocks = (self.get_num_total_blocks() - + self.get_num_free_blocks()) + if num_used_blocks > 0: + logger.warning( + "Failed to reset prefix cache because some " + "blocks (%d) are not freed yet", num_used_blocks) + return False + + # Free all blocks in the evictor. + while (block_id := + self._maybe_allocate_evicted_block_id()) is not None: + self._hashless_allocator.free_block_id(block_id) + + # Should not have any cached blocks because all blocks are evicted. + assert not self._cached_blocks + + # Reset the evictor. + self.evictor = make_evictor(self.eviction_policy) + + # Reset the block tracker. + for block_id in self._block_tracker: + self._block_tracker[block_id] = BlockTracker() + + # Reset the metrics. + self.metric_data = CacheMetricData() + + logger.info("Successfully reset prefix cache") + return True + def is_block_cached(self, block: Block) -> bool: assert block.content_hash is not None return block.content_hash in self._cached_blocks diff --git a/vllm/core/block_manager.py b/vllm/core/block_manager.py index b41e848221882..62a5f0bda061a 100644 --- a/vllm/core/block_manager.py +++ b/vllm/core/block_manager.py @@ -455,6 +455,9 @@ def get_num_free_cpu_blocks(self) -> int: def get_prefix_cache_hit_rate(self, device: Device) -> float: return self.block_allocator.get_prefix_cache_hit_rate(device) + def reset_prefix_cache(self) -> bool: + return self.block_allocator.reset_prefix_cache() + def _can_swap(self, seq_group: SequenceGroup, device: Device, diff --git a/vllm/core/interfaces.py b/vllm/core/interfaces.py index b10b8d3f4a5bf..9c7e246e3c4ed 100644 --- a/vllm/core/interfaces.py +++ b/vllm/core/interfaces.py @@ -122,6 +122,11 @@ def get_prefix_cache_hit_rate(self, device: Device) -> float: """Prefix cache hit rate. -1 means not supported or disabled.""" pass + @abstractmethod + def reset_prefix_cache(self) -> bool: + """Reset prefix cache for all devices.""" + pass + @abstractmethod def get_num_cached_tokens(self, seq: Sequence) -> int: pass diff --git a/vllm/core/placeholder_block_space_manager.py b/vllm/core/placeholder_block_space_manager.py index a47e594518534..f9924be4a3835 100644 --- a/vllm/core/placeholder_block_space_manager.py +++ b/vllm/core/placeholder_block_space_manager.py @@ -90,5 +90,8 @@ def mark_blocks_as_computed(self, seq_group: SequenceGroup, def get_prefix_cache_hit_rate(self, device: Device) -> float: return -1 + def reset_prefix_cache(self) -> bool: + return True + def get_num_cached_tokens(self, seq: Sequence) -> int: return 0 diff --git a/vllm/core/scheduler.py b/vllm/core/scheduler.py index 5601b724efe8c..4c8d47fc02aeb 100644 --- a/vllm/core/scheduler.py +++ b/vllm/core/scheduler.py @@ -517,6 +517,9 @@ def has_unfinished_seqs(self) -> bool: def get_prefix_cache_hit_rate(self, device: Device) -> float: return self.block_manager.get_prefix_cache_hit_rate(device) + def reset_prefix_cache(self) -> bool: + return self.block_manager.reset_prefix_cache() + def get_num_unfinished_seq_groups(self) -> int: return len(self.waiting) + len(self.running) + len(self.swapped) diff --git a/vllm/device_allocator/__init__.py b/vllm/device_allocator/__init__.py new file mode 100644 index 0000000000000..e69de29bb2d1d diff --git a/vllm/device_allocator/cumem.py b/vllm/device_allocator/cumem.py new file mode 100644 index 0000000000000..a43418dbb3b46 --- /dev/null +++ b/vllm/device_allocator/cumem.py @@ -0,0 +1,254 @@ +# cumem-based pytorch pluggable allocator to implement sleep mode. +# other approaches tried but failed: +# - cuda-python package binding +# - custom libcuda driver ctypes wrapper +# both of them failed because of cuda context mismatch. +# not sure why, they are created from a different context. +# the only successful approach is to call cuda driver API in C. +import dataclasses +from contextlib import contextmanager +from typing import Callable, Dict, Optional, Tuple, Union + +import torch + +from vllm.utils import is_pin_memory_available + + +def find_loaded_library(lib_name) -> Optional[str]: + """ + According to according to https://man7.org/linux/man-pages/man5/proc_pid_maps.5.html, + the file `/proc/self/maps` contains the memory maps of the process, which includes the + shared libraries loaded by the process. We can use this file to find the path of the + a loaded library. + """ # noqa + found_line = None + with open("/proc/self/maps") as f: + for line in f: + if lib_name in line: + found_line = line + break + if found_line is None: + # the library is not loaded in the current process + return None + # if lib_name is libcudart, we need to match a line with: + # address /path/to/libcudart-hash.so.11.0 + start = found_line.index("/") + path = found_line[start:].strip() + filename = path.split("/")[-1] + assert filename.rpartition(".so")[0].startswith(lib_name), \ + f"Unexpected filename: {filename} for library {lib_name}" + return path + + +cumem_available = False +try: + from vllm.cumem_allocator import (init_module, python_create_and_map, + python_unmap_and_release) + from vllm.distributed.device_communicators.cuda_wrapper import ( + CudaRTLibrary) + lib_name = find_loaded_library("cumem_allocator") + libcudart = CudaRTLibrary() + cumem_available = True +except ModuleNotFoundError: + # rocm platform does not support cumem allocator + init_module = None + python_create_and_map = None + python_unmap_and_release = None + CudaRTLibrary = None + lib_name = None + libcudart = None + +# py_device, py_alignedSize, py_d_mem, py_p_memHandle +HandleType = Tuple[int, int, int, int] + + +@dataclasses.dataclass +class AllocationData: + handle: HandleType + tag: str + cpu_backup_tensor: Optional[torch.Tensor] = None + + +def create_and_map(allocation_handle: HandleType) -> None: + python_create_and_map(*allocation_handle) + + +def unmap_and_release(allocation_handle: HandleType) -> None: + python_unmap_and_release(*allocation_handle) + + +def get_pluggable_allocator( + python_malloc_fn: Callable[[int], + int], python_free_func: Callable[[int, int], + None] +) -> torch.cuda.memory.CUDAPluggableAllocator: + init_module(python_malloc_fn, python_free_func) + new_alloc = torch.cuda.memory.CUDAPluggableAllocator( + lib_name, 'my_malloc', 'my_free') + return new_alloc + + +@contextmanager +def use_memory_pool_with_allocator( + python_malloc_fn: Callable[[int], int], + python_free_func: Callable[[int, int], None]) -> None: + new_alloc = get_pluggable_allocator(python_malloc_fn, python_free_func) + mem_pool = torch.cuda.memory.MemPool(new_alloc._allocator) + with torch.cuda.memory.use_mem_pool(mem_pool): + yield mem_pool + + +class CuMemAllocator: + """ + A singleton class that manages a memory pool for CUDA tensors. + The memory in this pool can be offloaded or discarded when the + allocator sleeps. + + Inside the `use_memory_pool(tag)` context, all tensors created will + be allocated in the memory pool, and has the same tag as the + tag passed to the context. + + When we call `sleep`, all tensors with the specified tag will be + offloaded to CPU memory, and the rest of the tensors will be discarded. + When we call `wake_up`, all tensors that are previously offloaded + will be loaded back to GPU memory, and the rest of the tensors will + have empty memory. + + Why it needs to be a singleton? + When allocated tensors are garbage collected, PyTorch will call + the free callback, which will call the `python_free_callback` method. + The C-extension uses a global variable to store the function of an + instance of this class. If we create multiple instances of this class, + the global variable will be overwritten and the free callback will + not work as expected. + """ + instance: "CuMemAllocator" = None + default_tag: str = "default" + + @staticmethod + def get_instance() -> "CuMemAllocator": + """ + CuMemAllocator is a singleton class. + We cannot call the constructor directly. + Call this method to get the instance. + """ + assert cumem_available, "cumem allocator is not available" + if CuMemAllocator.instance is None: + CuMemAllocator.instance = CuMemAllocator() + return CuMemAllocator.instance + + def __init__(self): + self.pointer_to_data: Dict[int, AllocationData] = {} + self.current_tag: str = CuMemAllocator.default_tag + + def python_malloc_callback(self, allocation_handle: HandleType) -> None: + """ + Internal method to store the allocation data + when memory is allocated in the memory pool.""" + py_d_mem = allocation_handle[2] + self.pointer_to_data[py_d_mem] = AllocationData( + allocation_handle, self.current_tag) + return + + def python_free_callback(self, ptr: int) -> HandleType: + """ + Internal method to look up the allocation data + when memory is freed in the memory pool.""" + data = self.pointer_to_data.pop(ptr) + if data.cpu_backup_tensor is not None: + data.cpu_backup_tensor = None + return data.handle + + def sleep( + self, + offload_tags: Optional[Union[Tuple[str, ...], + str]] = None) -> None: + """ + Put the allocator in sleep mode. + All data in the memory allocation with the specified tag will be + offloaded to CPU memory, and others will be discarded. + + :param offload_tags: The tags of the memory allocation that will be + offloaded. The rest of the memory allocation will be discarded. + """ + if offload_tags is None: + # by default, allocated tensors are offloaded + # when the allocator sleeps + offload_tags = (CuMemAllocator.default_tag, ) + elif isinstance(offload_tags, str): + offload_tags = (offload_tags, ) + + assert isinstance(offload_tags, tuple) + + for ptr, data in self.pointer_to_data.items(): + handle = data.handle + if data.tag in offload_tags: + size_in_bytes = handle[1] + cpu_backup_tensor = torch.empty( + size_in_bytes, + dtype=torch.uint8, + device='cpu', + pin_memory=is_pin_memory_available()) + cpu_ptr = cpu_backup_tensor.data_ptr() + libcudart.cudaMemcpy(cpu_ptr, ptr, size_in_bytes) + data.cpu_backup_tensor = cpu_backup_tensor + unmap_and_release(handle) + + def wake_up(self): + """ + Wake up the allocator from sleep mode. + All data that is previously offloaded will be loaded back to GPU + memory, and the rest of the data will have empty memory.""" + for ptr, data in self.pointer_to_data.items(): + handle = data.handle + create_and_map(handle) + if data.cpu_backup_tensor is not None: + cpu_backup_tensor = data.cpu_backup_tensor + if cpu_backup_tensor is not None: + size_in_bytes = cpu_backup_tensor.numel( + ) * cpu_backup_tensor.element_size() + cpu_ptr = cpu_backup_tensor.data_ptr() + libcudart.cudaMemcpy(ptr, cpu_ptr, size_in_bytes) + data.cpu_backup_tensor = None + + @contextmanager + def use_memory_pool(self, tag: Optional[str] = None): + """ + A context manager to use the memory pool. + All memory allocation created inside the context will be allocated + in the memory pool, and has the specified tag. + + :param tag: The tag of the memory allocation. If None, the default tag + will be used. + """ + if tag is None: + tag = CuMemAllocator.default_tag + + assert isinstance(tag, str) + + old_tag = self.current_tag + self.current_tag = tag + with use_memory_pool_with_allocator(self.python_malloc_callback, + self.python_free_callback): + yield + # PyTorch's bug, calling torch.cuda.empty_cache() will error + # when using pluggable allocator, see + # https://github.com/pytorch/pytorch/issues/145168 . + # if we have some memory allocated and then freed, + # the memory will not be released. + # right now it is fine, because we only use this allocator + # during weight loading and kv cache creation, where we only + # allocate memory. + # TODO: we need to find a way to release the memory, + # i.e. calling torch.cuda.empty_cache() + self.current_tag = old_tag + + def get_current_usage(self) -> int: + """ + Get the total number of bytes allocated in the memory pool. + """ + sum_bytes: int = 0 + for ptr, data in self.pointer_to_data.items(): + handle = data.handle + sum_bytes += handle[1] + return sum_bytes diff --git a/vllm/distributed/kv_transfer/kv_connector/simple_connector.py b/vllm/distributed/kv_transfer/kv_connector/simple_connector.py index 4ace03ff1184e..7780e2dfa317d 100644 --- a/vllm/distributed/kv_transfer/kv_connector/simple_connector.py +++ b/vllm/distributed/kv_transfer/kv_connector/simple_connector.py @@ -35,6 +35,7 @@ def __init__( ): self.config = config.kv_transfer_config + self.tp_size = config.parallel_config.tensor_parallel_size if self.config.kv_connector == "PyNcclConnector": from vllm.distributed.kv_transfer.kv_pipe.pynccl_pipe import ( @@ -161,7 +162,7 @@ def send_kv_caches_and_hidden_states( end_layer = model_executable.model.end_layer model_config = model_executable.model.config - num_heads = model_config.num_key_value_heads + num_heads = int(model_config.num_key_value_heads / self.tp_size) hidden_size = model_config.hidden_size num_attention_heads = model_config.num_attention_heads head_size = int(hidden_size / num_attention_heads) diff --git a/vllm/distributed/parallel_state.py b/vllm/distributed/parallel_state.py index bd13b8d64b605..0552cbf2faad9 100644 --- a/vllm/distributed/parallel_state.py +++ b/vllm/distributed/parallel_state.py @@ -39,7 +39,8 @@ import vllm.envs as envs from vllm.distributed.utils import StatelessProcessGroup from vllm.logger import init_logger -from vllm.utils import direct_register_custom_op, supports_custom_op +from vllm.utils import (current_stream, direct_register_custom_op, + supports_custom_op) if TYPE_CHECKING: from vllm.config import VllmConfig @@ -365,7 +366,7 @@ def _all_reduce_out_place(self, input_: torch.Tensor) -> torch.Tensor: return out pynccl_comm = self.pynccl_comm assert pynccl_comm is not None - out = pynccl_comm.all_reduce(input_) + out = pynccl_comm.all_reduce(input_, stream=current_stream()) if out is None: # fall back to the default all-reduce using PyTorch. # this usually happens during testing. @@ -870,12 +871,14 @@ def init_model_parallel_group( ) -> GroupCoordinator: if use_custom_allreduce is None: use_custom_allreduce = _ENABLE_CUSTOM_ALL_REDUCE + from vllm.platforms import current_platform return GroupCoordinator( group_ranks=group_ranks, local_rank=local_rank, torch_distributed_backend=backend, - use_pynccl=True, - use_custom_allreduce=use_custom_allreduce, + use_pynccl=current_platform.is_cuda_alike(), + use_custom_allreduce=current_platform.is_cuda_alike() + and use_custom_allreduce, use_tpu_communicator=True, use_hpu_communicator=True, use_xpu_communicator=True, @@ -1189,6 +1192,11 @@ def cleanup_dist_env_and_memory(shutdown_ray: bool = False): from vllm.platforms import current_platform if not current_platform.is_cpu(): torch.cuda.empty_cache() + try: + torch._C._host_emptyCache() + except AttributeError: + logger.warning( + "torch._C._host_emptyCache() only available in Pytorch >=2.5") def in_the_same_node_as(pg: Union[ProcessGroup, StatelessProcessGroup], diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py index 75ee1e40c9321..f5ff388451567 100644 --- a/vllm/engine/arg_utils.py +++ b/vllm/engine/arg_utils.py @@ -100,7 +100,6 @@ class EngineArgs: kv_cache_dtype: str = 'auto' seed: int = 0 max_model_len: Optional[int] = None - worker_use_ray: bool = False # Note: Specifying a custom executor backend by passing a class # is intended for expert use only. The API may change without # notice. @@ -196,6 +195,8 @@ class EngineArgs: kv_transfer_config: Optional[KVTransferConfig] = None generation_config: Optional[str] = None + enable_sleep_mode: bool = False + calculate_kv_scales: Optional[bool] = None def __post_init__(self): @@ -238,7 +239,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: choices=get_args(TaskOption), help='The task to use the model for. Each vLLM instance only ' 'supports one task, even if the same model can be used for ' - 'multiple tasks. When the model only supports one task, "auto" ' + 'multiple tasks. When the model only supports one task, ``"auto"`` ' 'can be used to select it; otherwise, you must specify explicitly ' 'which task to use.') parser.add_argument( @@ -250,7 +251,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: parser.add_argument( '--skip-tokenizer-init', action='store_true', - help='Skip initialization of tokenizer and detokenizer') + help='Skip initialization of tokenizer and detokenizer.') parser.add_argument( '--revision', type=nullable_str, @@ -377,7 +378,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: # Parallel arguments parser.add_argument( '--distributed-executor-backend', - choices=['ray', 'mp'], + choices=['ray', 'mp', 'uni', 'external_launcher'], default=EngineArgs.distributed_executor_backend, help='Backend to use for distributed model ' 'workers, either "ray" or "mp" (multiprocessing). If the product ' @@ -385,12 +386,8 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: 'or equal to the number of GPUs available, "mp" will be used to ' 'keep processing on a single host. Otherwise, this will default ' 'to "ray" if Ray is installed and fail otherwise. Note that tpu ' - 'and hpu only support Ray for distributed inference.') + 'only supports Ray for distributed inference.') - parser.add_argument( - '--worker-use-ray', - action='store_true', - help='Deprecated, use --distributed-executor-backend=ray.') parser.add_argument('--pipeline-parallel-size', '-pp', type=int, @@ -419,7 +416,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: choices=[8, 16, 32, 64, 128], help='Token block size for contiguous chunks of ' 'tokens. This is ignored on neuron devices and ' - 'set to max-model-len. On CUDA devices, ' + 'set to ``--max-model-len``. On CUDA devices, ' 'only block sizes up to 32 are supported. ' 'On HPU devices, block size defaults to 128.') @@ -428,12 +425,12 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: action=argparse.BooleanOptionalAction, default=EngineArgs.enable_prefix_caching, help="Enables automatic prefix caching. " - "Use --no-enable-prefix-caching to disable explicitly.", + "Use ``--no-enable-prefix-caching`` to disable explicitly.", ) parser.add_argument('--disable-sliding-window', action='store_true', help='Disables sliding window, ' - 'capping to sliding window size') + 'capping to sliding window size.') parser.add_argument('--use-v2-block-manager', action='store_true', default=True, @@ -527,7 +524,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: default=None, type=json.loads, help='RoPE scaling configuration in JSON format. ' - 'For example, {"rope_type":"dynamic","factor":2.0}') + 'For example, ``{"rope_type":"dynamic","factor":2.0}``') parser.add_argument('--rope-theta', default=None, type=float, @@ -596,7 +593,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: default=None, type=json.loads, help=('Overrides for the multimodal input mapping/processing, ' - 'e.g., image processor. For example: {"num_crops": 4}.')) + 'e.g., image processor. For example: ``{"num_crops": 4}``.')) parser.add_argument( '--disable-mm-preprocessor-cache', action='store_true', @@ -850,7 +847,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: "of the provided names. The model name in the model " "field of a response will be the first name in this " "list. If not specified, the model name will be the " - "same as the `--model` argument. Noted that this name(s) " + "same as the ``--model`` argument. Noted that this name(s) " "will also be used in `model_name` tag content of " "prometheus metrics, if multiple names provided, metrics " "tag will take the first one.") @@ -870,7 +867,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: default=None, help="Valid choices are " + ",".join(ALLOWED_DETAILED_TRACE_MODULES) + - ". It makes sense to set this only if --otlp-traces-endpoint is" + ". It makes sense to set this only if ``--otlp-traces-endpoint`` is" " set. If set, it will collect detailed traces for the specified " "modules. This involves use of possibly costly and or blocking " "operations and hence might have a performance impact.") @@ -897,13 +894,13 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: type=json.loads, default=None, help="Override or set neuron device configuration. " - "e.g. {\"cast_logits_dtype\": \"bloat16\"}.'") + "e.g. ``{\"cast_logits_dtype\": \"bloat16\"}``.") parser.add_argument( '--override-pooler-config', type=PoolerConfig.from_json, default=None, help="Override or set the pooling method for pooling models. " - "e.g. {\"pooling_type\": \"mean\", \"normalize\": false}.'") + "e.g. ``{\"pooling_type\": \"mean\", \"normalize\": false}``.") parser.add_argument('--compilation-config', '-O', @@ -942,7 +939,15 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: "Defaults to None, will use the default generation config in vLLM. " "If set to 'auto', the generation config will be automatically " "loaded from model. If set to a folder path, the generation config " - "will be loaded from the specified folder path.") + "will be loaded from the specified folder path. If " + "`max_new_tokens` is specified, then it sets a server-wide limit " + "on the number of output tokens for all requests.") + + parser.add_argument("--enable-sleep-mode", + action="store_true", + default=False, + help="Enable sleep mode for the engine. " + "(only cuda platform is supported)") parser.add_argument( '--calculate-kv-scales', @@ -996,7 +1001,9 @@ def create_model_config(self) -> ModelConfig: override_neuron_config=self.override_neuron_config, override_pooler_config=self.override_pooler_config, logits_processor_pattern=self.logits_processor_pattern, - generation_config=self.generation_config) + generation_config=self.generation_config, + enable_sleep_mode=self.enable_sleep_mode, + ) def create_load_config(self) -> LoadConfig: return LoadConfig( @@ -1061,7 +1068,6 @@ def create_engine_config(self, parallel_config = ParallelConfig( pipeline_parallel_size=self.pipeline_parallel_size, tensor_parallel_size=self.tensor_parallel_size, - worker_use_ray=self.worker_use_ray, max_parallel_loading_workers=self.max_parallel_loading_workers, disable_custom_all_reduce=self.disable_custom_all_reduce, tokenizer_pool_config=TokenizerPoolConfig.create_config( @@ -1269,11 +1275,22 @@ def _override_v1_engine_args(self, usage_context: UsageContext) -> None: self.enable_chunked_prefill = True # When no user override, set the default values based on the usage # context. - # TODO(woosuk): Tune the default values for different hardware. - default_max_num_batched_tokens = { - UsageContext.LLM_CLASS: 8192, - UsageContext.OPENAI_API_SERVER: 2048, - } + # Use different default values for different hardware. + from vllm.platforms import current_platform + device_name = current_platform.get_device_name().lower() + if "h100" in device_name or "h200" in device_name: + # For H100 and H200, we use larger default values. + default_max_num_batched_tokens = { + UsageContext.LLM_CLASS: 16384, + UsageContext.OPENAI_API_SERVER: 8192, + } + else: + # TODO(woosuk): Tune the default values for other hardware. + default_max_num_batched_tokens = { + UsageContext.LLM_CLASS: 8192, + UsageContext.OPENAI_API_SERVER: 2048, + } + if (self.max_num_batched_tokens is None and usage_context in default_max_num_batched_tokens): self.max_num_batched_tokens = default_max_num_batched_tokens[ diff --git a/vllm/engine/async_llm_engine.py b/vllm/engine/async_llm_engine.py index da23ed19ef7be..739ea06ae3818 100644 --- a/vllm/engine/async_llm_engine.py +++ b/vllm/engine/async_llm_engine.py @@ -18,9 +18,7 @@ from vllm.engine.llm_engine import LLMEngine, SchedulerOutputState from vllm.engine.metrics_types import StatLoggerBase from vllm.engine.protocol import EngineClient -from vllm.executor.executor_base import ExecutorAsyncBase -from vllm.executor.gpu_executor import GPUExecutorAsync -from vllm.executor.ray_utils import initialize_ray_cluster +from vllm.executor.executor_base import ExecutorBase from vllm.inputs import PromptType from vllm.inputs.preprocess import InputPreprocessor from vllm.logger import init_logger @@ -620,69 +618,9 @@ def __del__(self): rt.new_requests_event.set() @classmethod - def _get_executor_cls( - cls, engine_config: VllmConfig) -> Type[ExecutorAsyncBase]: - distributed_executor_backend = ( - engine_config.parallel_config.distributed_executor_backend) - if isinstance(distributed_executor_backend, type): - if not issubclass(distributed_executor_backend, ExecutorAsyncBase): - raise TypeError( - "distributed_executor_backend must be a subclass of " - f"ExecutorAsyncBase. Got {distributed_executor_backend}.") - executor_class = distributed_executor_backend - elif engine_config.device_config.device_type == "neuron": - from vllm.executor.neuron_executor import NeuronExecutorAsync - executor_class = NeuronExecutorAsync - elif engine_config.device_config.device_type == "tpu": - if distributed_executor_backend == "ray": - from vllm.executor.ray_tpu_executor import RayTPUExecutorAsync - executor_class = RayTPUExecutorAsync - else: - assert distributed_executor_backend is None - from vllm.executor.tpu_executor import TPUExecutorAsync - executor_class = TPUExecutorAsync - elif engine_config.device_config.device_type == "cpu": - from vllm.executor.cpu_executor import CPUExecutorAsync - executor_class = CPUExecutorAsync - elif engine_config.device_config.device_type == "hpu": - if distributed_executor_backend == "ray": - initialize_ray_cluster(engine_config.parallel_config) - from vllm.executor.ray_hpu_executor import RayHPUExecutorAsync - executor_class = RayHPUExecutorAsync - else: - from vllm.executor.hpu_executor import HPUExecutorAsync - executor_class = HPUExecutorAsync - elif engine_config.device_config.device_type == "openvino": - assert distributed_executor_backend is None, ( - "Distributed execution is not supported with " - "the OpenVINO backend.") - from vllm.executor.openvino_executor import OpenVINOExecutorAsync - executor_class = OpenVINOExecutorAsync - elif engine_config.device_config.device_type == "xpu": - if distributed_executor_backend is None: - from vllm.executor.xpu_executor import XPUExecutorAsync - executor_class = XPUExecutorAsync - elif distributed_executor_backend == "ray": - from vllm.executor.ray_xpu_executor import RayXPUExecutorAsync - executor_class = RayXPUExecutorAsync - elif distributed_executor_backend == "mp": - from vllm.executor.multiproc_xpu_executor import ( - MultiprocessingXPUExecutorAsync) - executor_class = MultiprocessingXPUExecutorAsync - else: - raise RuntimeError( - "Not supported distributed execution model on XPU device.") - elif distributed_executor_backend == "ray": - from vllm.executor.ray_gpu_executor import RayGPUExecutorAsync - executor_class = RayGPUExecutorAsync - elif distributed_executor_backend == "mp": - from vllm.executor.multiproc_gpu_executor import ( - MultiprocessingGPUExecutorAsync) - executor_class = MultiprocessingGPUExecutorAsync - else: - from vllm.executor.gpu_executor import GPUExecutorAsync - executor_class = GPUExecutorAsync - return executor_class + def _get_executor_cls(cls, + engine_config: VllmConfig) -> Type[ExecutorBase]: + return LLMEngine._get_executor_cls(engine_config) @classmethod def from_engine_args( @@ -700,9 +638,6 @@ def from_engine_args( executor_class = cls._get_executor_cls(engine_config) - if executor_class.uses_ray: - initialize_ray_cluster(engine_config.parallel_config) - # Create the async LLM engine. engine = cls( vllm_config=engine_config, @@ -1242,23 +1177,15 @@ def remove_logger(self, logger_name: str) -> None: self.engine.remove_logger(logger_name=logger_name) async def start_profile(self) -> None: - # using type instead of isinstance to check to avoid capturing - # inherited classes - if type(self.engine.model_executor) == GPUExecutorAsync: # noqa: E721 - self.engine.model_executor.start_profile() - else: - self.engine.model_executor._run_workers("start_profile") + self.engine.start_profile() async def stop_profile(self) -> None: - # using type instead of isinstance to check to avoid capturing - # inherited classes - if type(self.engine.model_executor) == GPUExecutorAsync: # noqa: E721 - self.engine.model_executor.stop_profile() - else: - self.engine.model_executor._run_workers("stop_profile") + self.engine.stop_profile() + + async def reset_prefix_cache(self) -> None: + self.engine.reset_prefix_cache() async def add_lora(self, lora_request: LoRARequest) -> None: - """Load a new LoRA adapter into the engine for future requests.""" self.engine.add_lora(lora_request) diff --git a/vllm/engine/llm_engine.py b/vllm/engine/llm_engine.py index 16deef91b93b2..8f6dc87289399 100644 --- a/vllm/engine/llm_engine.py +++ b/vllm/engine/llm_engine.py @@ -28,8 +28,6 @@ from vllm.entrypoints.openai.logits_processors import ( get_logits_processors as get_openai_logits_processors) from vllm.executor.executor_base import ExecutorBase -from vllm.executor.gpu_executor import GPUExecutor -from vllm.executor.ray_utils import initialize_ray_cluster from vllm.inputs import (INPUT_REGISTRY, InputRegistry, ProcessorInputs, PromptType, SingletonInputsAdapter) from vllm.inputs.parse import is_encoder_decoder_inputs, is_token_prompt @@ -442,64 +440,31 @@ def _get_executor_cls(cls, raise TypeError( "distributed_executor_backend must be a subclass of " f"ExecutorBase. Got {distributed_executor_backend}.") - if distributed_executor_backend.uses_ray: # type: ignore - initialize_ray_cluster(engine_config.parallel_config) executor_class = distributed_executor_backend - elif engine_config.device_config.device_type == "neuron": - from vllm.executor.neuron_executor import NeuronExecutor - executor_class = NeuronExecutor - elif engine_config.device_config.device_type == "tpu": + elif engine_config.parallel_config.world_size > 1: if distributed_executor_backend == "ray": - initialize_ray_cluster(engine_config.parallel_config) - from vllm.executor.ray_tpu_executor import RayTPUExecutor - executor_class = RayTPUExecutor - else: - assert distributed_executor_backend is None - from vllm.executor.tpu_executor import TPUExecutor - executor_class = TPUExecutor - elif engine_config.device_config.device_type == "cpu": - from vllm.executor.cpu_executor import CPUExecutor - executor_class = CPUExecutor - elif engine_config.device_config.device_type == "hpu": - if distributed_executor_backend == "ray": - initialize_ray_cluster(engine_config.parallel_config) - from vllm.executor.ray_hpu_executor import RayHPUExecutor - executor_class = RayHPUExecutor - else: - from vllm.executor.hpu_executor import HPUExecutor - executor_class = HPUExecutor - elif engine_config.device_config.device_type == "openvino": - from vllm.executor.openvino_executor import OpenVINOExecutor - executor_class = OpenVINOExecutor - elif engine_config.device_config.device_type == "xpu": - if distributed_executor_backend == "ray": - initialize_ray_cluster(engine_config.parallel_config) - from vllm.executor.ray_xpu_executor import RayXPUExecutor - executor_class = RayXPUExecutor + from vllm.executor.ray_distributed_executor import ( + RayDistributedExecutor) + executor_class = RayDistributedExecutor elif distributed_executor_backend == "mp": - # FIXME(kunshang): - # spawn needs calling `if __name__ == '__main__':`` - # fork is not supported for xpu start new process. - logger.error( - "Both start methods (spawn and fork) have issue " - "on XPU if you use mp backend, Please try ray instead.") - else: - from vllm.executor.xpu_executor import XPUExecutor - executor_class = XPUExecutor - elif distributed_executor_backend == "ray": - initialize_ray_cluster(engine_config.parallel_config) - from vllm.executor.ray_gpu_executor import RayGPUExecutor - executor_class = RayGPUExecutor - elif distributed_executor_backend == "mp": - from vllm.executor.multiproc_gpu_executor import ( - MultiprocessingGPUExecutor) - assert not envs.VLLM_USE_RAY_SPMD_WORKER, ( - "multiprocessing distributed executor backend does not " - "support VLLM_USE_RAY_SPMD_WORKER=1") - executor_class = MultiprocessingGPUExecutor + from vllm.executor.mp_distributed_executor import ( + MultiprocessingDistributedExecutor) + assert not envs.VLLM_USE_RAY_SPMD_WORKER, ( + "multiprocessing distributed executor backend does not " + "support VLLM_USE_RAY_SPMD_WORKER=1") + executor_class = MultiprocessingDistributedExecutor + elif distributed_executor_backend == "uni": + # JAX-style, single-process, multi-device executor. + from vllm.executor.uniproc_executor import UniProcExecutor + executor_class = UniProcExecutor + elif distributed_executor_backend == "external_launcher": + # executor with external launcher + from vllm.executor.uniproc_executor import ( # noqa + ExecutorWithExternalLauncher) + executor_class = ExecutorWithExternalLauncher else: - from vllm.executor.gpu_executor import GPUExecutor - executor_class = GPUExecutor + from vllm.executor.uniproc_executor import UniProcExecutor + executor_class = UniProcExecutor return executor_class @classmethod @@ -724,7 +689,9 @@ def add_request( :class:`~vllm.PoolingParams` for pooling. arrival_time: The arrival time of the request. If None, we use the current monotonic time. + lora_request: The LoRA request to add. trace_headers: OpenTelemetry trace headers. + prompt_adapter_request: The prompt adapter request to add. priority: The priority of the request. Only applicable with priority scheduling. @@ -947,6 +914,14 @@ def has_unfinished_requests_for_virtual_engine( """ return self.scheduler[virtual_engine].has_unfinished_seqs() + def reset_prefix_cache(self) -> bool: + """Reset prefix cache for all devices.""" + + success = True + for scheduler in self.scheduler: + success = success and scheduler.reset_prefix_cache() + return success + @staticmethod def _process_sequence_group_outputs( seq_group: SequenceGroup, @@ -1845,27 +1820,27 @@ def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool: def list_prompt_adapters(self) -> List[int]: return self.model_executor.list_prompt_adapters() + def start_profile(self) -> None: + self.model_executor.start_profile() + + def stop_profile(self) -> None: + self.model_executor.stop_profile() + + def sleep(self, level: int = 1) -> None: + assert self.vllm_config.model_config.enable_sleep_mode, ( + "Sleep mode is not enabled in the model config") + self.model_executor.sleep(level=level) + + def wake_up(self) -> None: + assert self.vllm_config.model_config.enable_sleep_mode, ( + "Sleep mode is not enabled in the model config") + self.model_executor.wake_up() + def check_health(self) -> None: if self.tokenizer: self.tokenizer.check_health() self.model_executor.check_health() - def start_profile(self) -> None: - # using type instead of isinstance to check to avoid capturing - # inherited classes (MultiprocessingGPUExecutor) - if type(self.model_executor) == GPUExecutor: # noqa: E721 - self.model_executor.start_profile() - else: - self.model_executor._run_workers("start_profile") - - def stop_profile(self) -> None: - # using type instead of isinstance to check to avoid capturing - # inherited classes (MultiprocessingGPUExecutor) - if type(self.model_executor) == GPUExecutor: # noqa: E721 - self.model_executor.stop_profile() - else: - self.model_executor._run_workers("stop_profile") - def is_tracing_enabled(self) -> bool: return self.tracer is not None @@ -1900,46 +1875,44 @@ def create_trace_span(self, seq_group: SequenceGroup) -> None: metrics = seq_group.metrics ttft = metrics.first_token_time - metrics.arrival_time e2e_time = metrics.finished_time - metrics.arrival_time - # attribute names are based on - # https://github.com/open-telemetry/semantic-conventions/blob/main/docs/gen-ai/llm-spans.md - seq_span.set_attribute(SpanAttributes.LLM_RESPONSE_MODEL, + seq_span.set_attribute(SpanAttributes.GEN_AI_RESPONSE_MODEL, self.model_config.model) - seq_span.set_attribute(SpanAttributes.LLM_REQUEST_ID, + seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_ID, seq_group.request_id) - seq_span.set_attribute(SpanAttributes.LLM_REQUEST_TEMPERATURE, + seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_TEMPERATURE, seq_group.sampling_params.temperature) - seq_span.set_attribute(SpanAttributes.LLM_REQUEST_TOP_P, + seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_TOP_P, seq_group.sampling_params.top_p) - seq_span.set_attribute(SpanAttributes.LLM_REQUEST_MAX_TOKENS, + seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_MAX_TOKENS, seq_group.sampling_params.max_tokens) - seq_span.set_attribute(SpanAttributes.LLM_REQUEST_N, + seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_N, seq_group.sampling_params.n) - seq_span.set_attribute(SpanAttributes.LLM_USAGE_NUM_SEQUENCES, + seq_span.set_attribute(SpanAttributes.GEN_AI_USAGE_NUM_SEQUENCES, seq_group.num_seqs()) - seq_span.set_attribute(SpanAttributes.LLM_USAGE_PROMPT_TOKENS, + seq_span.set_attribute(SpanAttributes.GEN_AI_USAGE_PROMPT_TOKENS, len(seq_group.prompt_token_ids)) seq_span.set_attribute( - SpanAttributes.LLM_USAGE_COMPLETION_TOKENS, + SpanAttributes.GEN_AI_USAGE_COMPLETION_TOKENS, sum([ seq.get_output_len() for seq in seq_group.get_finished_seqs() ])) - seq_span.set_attribute(SpanAttributes.LLM_LATENCY_TIME_IN_QUEUE, + seq_span.set_attribute(SpanAttributes.GEN_AI_LATENCY_TIME_IN_QUEUE, metrics.time_in_queue) seq_span.set_attribute( - SpanAttributes.LLM_LATENCY_TIME_TO_FIRST_TOKEN, ttft) - seq_span.set_attribute(SpanAttributes.LLM_LATENCY_E2E, e2e_time) + SpanAttributes.GEN_AI_LATENCY_TIME_TO_FIRST_TOKEN, ttft) + seq_span.set_attribute(SpanAttributes.GEN_AI_LATENCY_E2E, e2e_time) if metrics.scheduler_time is not None: seq_span.set_attribute( - SpanAttributes.LLM_LATENCY_TIME_IN_SCHEDULER, + SpanAttributes.GEN_AI_LATENCY_TIME_IN_SCHEDULER, metrics.scheduler_time) if metrics.model_forward_time is not None: seq_span.set_attribute( - SpanAttributes.LLM_LATENCY_TIME_IN_MODEL_FORWARD, + SpanAttributes.GEN_AI_LATENCY_TIME_IN_MODEL_FORWARD, metrics.model_forward_time / 1000.0) if metrics.model_execute_time is not None: seq_span.set_attribute( - SpanAttributes.LLM_LATENCY_TIME_IN_MODEL_EXECUTE, + SpanAttributes.GEN_AI_LATENCY_TIME_IN_MODEL_EXECUTE, metrics.model_execute_time) def _validate_model_inputs(self, inputs: ProcessorInputs, diff --git a/vllm/engine/metrics.py b/vllm/engine/metrics.py index c8aec8dd3afa3..b771c190dd82a 100644 --- a/vllm/engine/metrics.py +++ b/vllm/engine/metrics.py @@ -120,7 +120,8 @@ def __init__(self, labelnames: List[str], vllm_config: VllmConfig): labelnames=labelnames) buckets = [1, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8096] if not vllm_config.model_config.enforce_eager: - buckets = vllm_config.compilation_config.capture_sizes.copy() + buckets = vllm_config.compilation_config.\ + cudagraph_capture_sizes.copy() buckets.sort() self.histogram_iteration_tokens = self._histogram_cls( name="vllm:iteration_tokens_total", @@ -258,21 +259,6 @@ def __init__(self, labelnames: List[str], vllm_config: VllmConfig): documentation="Number of emitted tokens.", labelnames=labelnames)) - # Deprecated in favor of vllm:prompt_tokens_total - self.gauge_avg_prompt_throughput = self._gauge_cls( - name="vllm:avg_prompt_throughput_toks_per_s", - documentation="Average prefill throughput in tokens/s.", - labelnames=labelnames, - multiprocess_mode="sum", - ) - # Deprecated in favor of vllm:generation_tokens_total - self.gauge_avg_generation_throughput = self._gauge_cls( - name="vllm:avg_generation_throughput_toks_per_s", - documentation="Average generation throughput in tokens/s.", - labelnames=labelnames, - multiprocess_mode="sum", - ) - # end-metrics-definitions @@ -634,20 +620,6 @@ def _log_prometheus(self, stats: Stats) -> None: self._log_histogram(self.metrics.histogram_max_tokens_request, stats.max_tokens_requests) - def _log_prometheus_interval(self, prompt_throughput: float, - generation_throughput: float) -> None: - # Logs metrics to prometheus that are computed every logging_interval. - # Support legacy gauge metrics that make throughput calculations on - # the vLLM side. Moving forward, we should use counters like - # counter_prompt_tokens, counter_generation_tokens - # Which log raw data and calculate summaries using rate() on the - # grafana/prometheus side. See - # https://github.com/vllm-project/vllm/pull/2316#discussion_r1464204666 - self.metrics.gauge_avg_prompt_throughput.labels( - **self.labels).set(prompt_throughput) - self.metrics.gauge_avg_generation_throughput.labels( - **self.labels).set(generation_throughput) - def log(self, stats: Stats): """Logs to prometheus and tracked stats every iteration.""" # Log to prometheus. @@ -663,20 +635,6 @@ def log(self, stats: Stats): # Log locally every local_interval seconds. if local_interval_elapsed(stats.now, self.last_local_log, self.local_interval): - # Compute summary metrics for tracked stats (and log them - # to promethus if applicable). - prompt_throughput = get_throughput(self.num_prompt_tokens, - now=stats.now, - last_log=self.last_local_log) - generation_throughput = get_throughput( - self.num_generation_tokens, - now=stats.now, - last_log=self.last_local_log) - - self._log_prometheus_interval( - prompt_throughput=prompt_throughput, - generation_throughput=generation_throughput) - if self.spec_decode_metrics is not None: self._log_gauge( self.metrics.gauge_spec_decode_draft_acceptance_rate, diff --git a/vllm/engine/multiprocessing/__init__.py b/vllm/engine/multiprocessing/__init__.py index 7132f9840001a..d9703b820a779 100644 --- a/vllm/engine/multiprocessing/__init__.py +++ b/vllm/engine/multiprocessing/__init__.py @@ -121,6 +121,10 @@ class RPCUProfileRequest(Enum): STOP_PROFILE = 2 +class RPCResetPrefixCacheRequest(Enum): + RESET_PREFIX_CACHE = 1 + + @dataclass class RPCLoadAdapterRequest: lora_request: LoRARequest @@ -134,7 +138,8 @@ class RPCAdapterLoadedResponse: RPC_REQUEST_T = Union[RPCProcessRequest, RPCAbortRequest, RPCStartupRequest, - RPCUProfileRequest, RPCLoadAdapterRequest] + RPCUProfileRequest, RPCLoadAdapterRequest, + RPCResetPrefixCacheRequest] REQUEST_OUTPUTS_T = Union[List[RequestOutput], RPCAdapterLoadedResponse, RPCError] diff --git a/vllm/engine/multiprocessing/client.py b/vllm/engine/multiprocessing/client.py index a9ab899535180..5237f63c34c01 100644 --- a/vllm/engine/multiprocessing/client.py +++ b/vllm/engine/multiprocessing/client.py @@ -27,8 +27,9 @@ VLLM_RPC_SUCCESS_STR, RPCAbortRequest, RPCAdapterLoadedResponse, RPCError, RPCLoadAdapterRequest, - RPCProcessRequest, RPCStartupRequest, - RPCStartupResponse, + RPCProcessRequest, + RPCResetPrefixCacheRequest, + RPCStartupRequest, RPCStartupResponse, RPCUProfileRequest) from vllm.engine.protocol import EngineClient # yapf: enable @@ -262,7 +263,14 @@ async def setup(self): """Setup the client before it starts sending server requests.""" # Start output_loop - self.output_loop = asyncio.create_task(self.run_output_handler_loop()) + if self.output_loop is None: + # only generate once to avoid multiple concurrent output_loops + # this will lead to race conditions and wrong orders of tokens + # returned by the engine + # setup will be called multiple times during the startup of + # the engine + self.output_loop = asyncio.create_task( + self.run_output_handler_loop()) with self.get_data_socket() as socket: # Wait until server is ready. @@ -271,8 +279,9 @@ async def setup(self): self.tracing_flag = response.tracing_enabled # Start health_loop. - self.health_loop = asyncio.create_task( - self.run_heartbeat_loop(timeout=VLLM_RPC_TIMEOUT)) + if self.health_loop is None: + self.health_loop = asyncio.create_task( + self.run_heartbeat_loop(timeout=VLLM_RPC_TIMEOUT)) def close(self): """Destroy the ZeroMQ Context.""" @@ -667,6 +676,13 @@ async def stop_profile(self) -> None: await self._send_one_way_rpc_request( request=RPCUProfileRequest.STOP_PROFILE, socket=self.input_socket) + async def reset_prefix_cache(self) -> None: + """Reset the prefix cache""" + + await self._send_one_way_rpc_request( + request=RPCResetPrefixCacheRequest.RESET_PREFIX_CACHE, + socket=self.input_socket) + async def add_lora(self, lora_request: LoRARequest) -> None: """Load a new LoRA adapter into the engine for future requests.""" # Uses the same I/O as generate requests diff --git a/vllm/engine/multiprocessing/engine.py b/vllm/engine/multiprocessing/engine.py index 36f4df4b02731..166f89743b3cd 100644 --- a/vllm/engine/multiprocessing/engine.py +++ b/vllm/engine/multiprocessing/engine.py @@ -16,11 +16,11 @@ VLLM_RPC_SUCCESS_STR, RPCAbortRequest, RPCAdapterLoadedResponse, RPCError, RPCLoadAdapterRequest, - RPCProcessRequest, RPCStartupRequest, - RPCStartupResponse, + RPCProcessRequest, + RPCResetPrefixCacheRequest, + RPCStartupRequest, RPCStartupResponse, RPCUProfileRequest) # yapf: enable -from vllm.executor.gpu_executor import GPUExecutor from vllm.logger import init_logger from vllm.outputs import RequestOutput from vllm.usage.usage_lib import UsageContext @@ -238,6 +238,8 @@ def handle_new_input(self): self.stop_profile() elif isinstance(request, RPCLoadAdapterRequest): self._handle_load_adapter_request(request) + elif isinstance(request, RPCResetPrefixCacheRequest): + self.reset_prefix_cache() else: raise ValueError("Unknown RPCRequest Type: " f"{type(request)}") @@ -297,6 +299,7 @@ def _handle_load_adapter_request(self, request: RPCLoadAdapterRequest): is_engine_errored=False, exception=e) self._send_outputs(rpc_err) + return # Otherwise, send back the successful load message self._send_outputs( RPCAdapterLoadedResponse(request_id=request.request_id)) @@ -356,16 +359,13 @@ def _set_errored(self, e: BaseException): self._errored_with = e def start_profile(self) -> None: - if type(self.engine.model_executor) is GPUExecutor: - self.engine.model_executor.start_profile() - else: - self.engine.model_executor._run_workers("start_profile") + self.engine.start_profile() def stop_profile(self) -> None: - if type(self.engine.model_executor) is GPUExecutor: - self.engine.model_executor.stop_profile() - else: - self.engine.model_executor._run_workers("stop_profile") + self.engine.stop_profile() + + def reset_prefix_cache(self) -> bool: + return self.engine.reset_prefix_cache() def signal_handler(*_) -> None: diff --git a/vllm/engine/output_processor/multi_step.py b/vllm/engine/output_processor/multi_step.py index c8b282b1a7676..99c2baf3f4df4 100644 --- a/vllm/engine/output_processor/multi_step.py +++ b/vllm/engine/output_processor/multi_step.py @@ -144,7 +144,7 @@ def process_outputs(self, def _process_decode_and_stop(self, seq: Sequence, sampling_params: SamplingParams) -> None: new_char_count = 0 - if sampling_params.detokenize: + if sampling_params.detokenize and self.detokenizer: new_char_count = self.detokenizer.decode_sequence_inplace( seq, sampling_params) diff --git a/vllm/engine/output_processor/single_step.py b/vllm/engine/output_processor/single_step.py index da3185f33dbe9..55c56abea0da3 100644 --- a/vllm/engine/output_processor/single_step.py +++ b/vllm/engine/output_processor/single_step.py @@ -102,9 +102,9 @@ def process_prompt_logprob(self, seq_group: SequenceGroup, Args: seq_group: the output is associated with this :class:`SequenceGroup` - output: the :class:`SequenceGroupOutput` for a single scheduler step + outputs: the :class:`SequenceGroupOutput` for a single scheduler step """ - assert len(outputs) == 1, ("Single step should only has 1 output.") + assert len(outputs) == 1, "Single step should only have 1 output." output = outputs[0] assert isinstance(output, CompletionSequenceGroupOutput) single_step_process_prompt_logprob(self, seq_group, output) diff --git a/vllm/engine/protocol.py b/vllm/engine/protocol.py index f05ff62c4766b..de7b2c1b91f50 100644 --- a/vllm/engine/protocol.py +++ b/vllm/engine/protocol.py @@ -271,6 +271,11 @@ async def stop_profile(self) -> None: """Start profiling the engine""" ... + @abstractmethod + async def reset_prefix_cache(self) -> None: + """Reset the prefix cache""" + ... + @abstractmethod async def add_lora(self, lora_request: LoRARequest) -> None: """Load a new LoRA adapter into the engine for future requests.""" diff --git a/vllm/entrypoints/fast_sync_llm.py b/vllm/entrypoints/fast_sync_llm.py index a3c1a455eac39..3f09b148dc92b 100644 --- a/vllm/entrypoints/fast_sync_llm.py +++ b/vllm/entrypoints/fast_sync_llm.py @@ -6,8 +6,9 @@ from vllm.distributed.communication_op import broadcast_tensor_dict from vllm.engine.arg_utils import EngineArgs from vllm.engine.llm_engine import LLMEngine -from vllm.executor.multiproc_gpu_executor import MultiprocessingGPUExecutor -from vllm.executor.ray_gpu_executor import RayGPUExecutor +from vllm.executor.mp_distributed_executor import ( + MultiprocessingDistributedExecutor) +from vllm.executor.ray_distributed_executor import RayDistributedExecutor from vllm.inputs import PromptType, TokensPrompt from vllm.logger import init_logger from vllm.pooling_params import PoolingParams @@ -55,7 +56,7 @@ def _poll_requests(self): (request_id, prompt, sampling_params) = self.input_queue.get() if self.need_restart and isinstance( self.llm_engine.model_executor, - MultiprocessingGPUExecutor): + MultiprocessingDistributedExecutor): logger.info("Restarting worker loops") for worker in self.llm_engine.model_executor.workers: worker.execute_method("start_worker_execution_loop") @@ -74,7 +75,7 @@ def run_engine(self): self.engine_args, usage_context=UsageContext.LLM_CLASS) assert not isinstance( self.llm_engine.model_executor, - RayGPUExecutor), "Ray is not supported in sync openai mode" + RayDistributedExecutor), "Ray is not supported in sync openai mode" self.result_queue.put(("Ready", None, None)) prompt_lens = {} diff --git a/vllm/entrypoints/llm.py b/vllm/entrypoints/llm.py index f65c1106cc5bc..3845fde83265d 100644 --- a/vllm/entrypoints/llm.py +++ b/vllm/entrypoints/llm.py @@ -1,11 +1,13 @@ import itertools import warnings from contextlib import contextmanager -from typing import (Any, ClassVar, Dict, List, Optional, Sequence, Tuple, Type, - Union, cast, overload) +from typing import (Any, Callable, ClassVar, Dict, List, Optional, Sequence, + Tuple, Type, Union, cast, overload) +import cloudpickle +import torch.nn as nn from tqdm import tqdm -from typing_extensions import deprecated +from typing_extensions import TypeVar, deprecated from vllm import envs from vllm.beam_search import (BeamSearchInstance, BeamSearchOutput, @@ -41,6 +43,8 @@ logger = init_logger(__name__) +_R = TypeVar("_R", default=Any) + class LLM: """An LLM for generating texts from given prompts and sampling parameters. @@ -186,6 +190,13 @@ def __init__( if "disable_log_stats" not in kwargs: kwargs["disable_log_stats"] = True + if "worker_cls" in kwargs: + worker_cls = kwargs["worker_cls"] + # if the worker_cls is not qualified string name, + # we serialize it using cloudpickle to avoid pickling issues + if isinstance(worker_cls, type): + kwargs["worker_cls"] = cloudpickle.dumps(worker_cls) + if compilation_config is not None: if isinstance(compilation_config, (int, dict)): compilation_config_instance = CompilationConfig.from_cli( @@ -455,6 +466,44 @@ def generate( outputs = self._run_engine(use_tqdm=use_tqdm) return self.engine_class.validate_outputs(outputs, RequestOutput) + def collective_rpc(self, + method: Union[str, Callable[..., _R]], + timeout: Optional[float] = None, + args: Tuple = (), + kwargs: Optional[Dict[str, Any]] = None) -> List[_R]: + """ + Execute an RPC call on all workers. + + Args: + method: Name of the worker method to execute, or a callable that + is serialized and sent to all workers to execute. + + If the method is a callable, it should accept an additional + `self` argument, in addition to the arguments passed in `args` + and `kwargs`. The `self` argument will be the worker object. + timeout: Maximum time in seconds to wait for execution. Raises a + :exc:`TimeoutError` on timeout. `None` means wait indefinitely. + args: Positional arguments to pass to the worker method. + kwargs: Keyword arguments to pass to the worker method. + + Returns: + A list containing the results from each worker. + + Note: + It is recommended to use this API to only pass control messages, + and set up data-plane communication to pass data. + """ + executor = self.llm_engine.model_executor + return executor.collective_rpc(method, timeout, args, kwargs) + + def apply_model(self, func: Callable[[nn.Module], _R]) -> list[_R]: + """ + Run a function directly on the model inside each worker, + returning the result for each of them. + """ + executor = self.llm_engine.model_executor + return executor.apply_model(func) + def beam_search( self, prompts: List[Union[TokensPrompt, TextPrompt]], @@ -1083,6 +1132,36 @@ def start_profile(self) -> None: def stop_profile(self) -> None: self.llm_engine.stop_profile() + def reset_prefix_cache(self) -> bool: + return self.llm_engine.reset_prefix_cache() + + def sleep(self, level: int = 1): + """ + Put the engine to sleep. The engine should not process any requests. + The caller should guarantee that no requests are being processed + during the sleep period, before `wake_up` is called. + + :param level: The sleep level. Level 1 sleep will offload the model + weights and discard the kv cache. The content of kv cache is + forgotten. Level 1 sleep is good for sleeping and waking up the + engine to run the same model again. The model weights are backed + up in CPU memory. Please make sure there's enough CPU memory to + store the model weights. Level 2 sleep will discard both the model + weights and the kv cache. The content of both the model weights + and kv cache is forgotten. Level 2 sleep is good for sleeping and + waking up the engine to run a different model or update the model, + where previous model weights are not needed. It reduces CPU memory + pressure. + """ + self.reset_prefix_cache() + self.llm_engine.sleep(level=level) + + def wake_up(self): + """ + Wake up the engine from sleep mode. See the :meth:`sleep` method + for more details.""" + self.llm_engine.wake_up() + # LEGACY def _convert_v1_inputs( self, diff --git a/vllm/entrypoints/openai/api_server.py b/vllm/entrypoints/openai/api_server.py index 1aeefe86cd05e..45cf06566faaa 100644 --- a/vllm/entrypoints/openai/api_server.py +++ b/vllm/entrypoints/openai/api_server.py @@ -1,5 +1,6 @@ import asyncio import atexit +import gc import importlib import inspect import multiprocessing @@ -55,6 +56,7 @@ PoolingChatRequest, PoolingCompletionRequest, PoolingRequest, PoolingResponse, + RerankRequest, RerankResponse, ScoreRequest, ScoreResponse, TokenizeRequest, TokenizeResponse, @@ -67,6 +69,7 @@ from vllm.entrypoints.openai.serving_models import (BaseModelPath, OpenAIServingModels) from vllm.entrypoints.openai.serving_pooling import OpenAIServingPooling +from vllm.entrypoints.openai.serving_rerank import JinaAIServingRerank from vllm.entrypoints.openai.serving_score import OpenAIServingScores from vllm.entrypoints.openai.serving_tokenization import ( OpenAIServingTokenization) @@ -104,6 +107,11 @@ async def _force_log(): task.add_done_callback(_running_tasks.remove) else: task = None + + # Mark the startup heap as static so that it's ignored by GC. + # Reduces pause times of oldest generation collections. + gc.collect() + gc.freeze() try: yield finally: @@ -300,6 +308,10 @@ def score(request: Request) -> Optional[OpenAIServingScores]: return request.app.state.openai_serving_scores +def rerank(request: Request) -> Optional[JinaAIServingRerank]: + return request.app.state.jinaai_serving_reranking + + def tokenization(request: Request) -> OpenAIServingTokenization: return request.app.state.openai_serving_tokenization @@ -496,6 +508,40 @@ async def create_score_v1(request: ScoreRequest, raw_request: Request): return await create_score(request, raw_request) +@router.post("/rerank") +@with_cancellation +async def do_rerank(request: RerankRequest, raw_request: Request): + handler = rerank(raw_request) + if handler is None: + return base(raw_request).create_error_response( + message="The model does not support Rerank (Score) API") + generator = await handler.do_rerank(request, raw_request) + if isinstance(generator, ErrorResponse): + return JSONResponse(content=generator.model_dump(), + status_code=generator.code) + elif isinstance(generator, RerankResponse): + return JSONResponse(content=generator.model_dump()) + + assert_never(generator) + + +@router.post("/v1/rerank") +@with_cancellation +async def do_rerank_v1(request: RerankRequest, raw_request: Request): + logger.warning( + "To indicate that the rerank API is not part of the standard OpenAI" + " API, we have located it at `/rerank`. Please update your client" + "accordingly. (Note: Conforms to JinaAI rerank API)") + + return await do_rerank(request, raw_request) + + +@router.post("/v2/rerank") +@with_cancellation +async def do_rerank_v2(request: RerankRequest, raw_request: Request): + return await do_rerank(request, raw_request) + + TASK_HANDLERS: Dict[str, Dict[str, tuple]] = { "generate": { "messages": (ChatCompletionRequest, create_chat_completion), @@ -506,7 +552,10 @@ async def create_score_v1(request: ScoreRequest, raw_request: Request): "default": (EmbeddingCompletionRequest, create_embedding), }, "score": { - "default": (ScoreRequest, create_score), + "default": (RerankRequest, do_rerank) + }, + "rerank": { + "default": (RerankRequest, do_rerank) }, "reward": { "messages": (PoolingChatRequest, create_pooling), @@ -518,6 +567,18 @@ async def create_score_v1(request: ScoreRequest, raw_request: Request): }, } +if envs.VLLM_SERVER_DEV_MODE: + + @router.post("/reset_prefix_cache") + async def reset_prefix_cache(raw_request: Request): + """ + Reset the prefix cache. Note that we currently do not check if the + prefix cache is successfully reset in the API server. + """ + logger.info("Resetting prefix cache...") + await engine_client(raw_request).reset_prefix_cache() + return Response(status_code=200) + @router.post("/invocations") async def invocations(raw_request: Request): @@ -741,6 +802,12 @@ async def init_app_state( state.openai_serving_models, request_logger=request_logger ) if model_config.task == "score" else None + state.jinaai_serving_reranking = JinaAIServingRerank( + engine_client, + model_config, + state.openai_serving_models, + request_logger=request_logger + ) if model_config.task == "score" else None state.openai_serving_tokenization = OpenAIServingTokenization( engine_client, model_config, diff --git a/vllm/entrypoints/openai/cli_args.py b/vllm/entrypoints/openai/cli_args.py index 22206ef8dbfe6..4df75a665bab9 100644 --- a/vllm/entrypoints/openai/cli_args.py +++ b/vllm/entrypoints/openai/cli_args.py @@ -79,29 +79,29 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: parser.add_argument("--host", type=nullable_str, default=None, - help="host name") - parser.add_argument("--port", type=int, default=8000, help="port number") + help="Host name.") + parser.add_argument("--port", type=int, default=8000, help="Port number.") parser.add_argument( "--uvicorn-log-level", type=str, default="info", choices=['debug', 'info', 'warning', 'error', 'critical', 'trace'], - help="log level for uvicorn") + help="Log level for uvicorn.") parser.add_argument("--allow-credentials", action="store_true", - help="allow credentials") + help="Allow credentials.") parser.add_argument("--allowed-origins", type=json.loads, default=["*"], - help="allowed origins") + help="Allowed origins.") parser.add_argument("--allowed-methods", type=json.loads, default=["*"], - help="allowed methods") + help="Allowed methods.") parser.add_argument("--allowed-headers", type=json.loads, default=["*"], - help="allowed headers") + help="Allowed headers.") parser.add_argument("--api-key", type=nullable_str, default=None, @@ -115,10 +115,10 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: action=LoRAParserAction, help="LoRA module configurations in either 'name=path' format" "or JSON format. " - "Example (old format): 'name=path' " + "Example (old format): ``'name=path'`` " "Example (new format): " - "'{\"name\": \"name\", \"local_path\": \"path\", " - "\"base_model_name\": \"id\"}'") + "``{\"name\": \"name\", \"path\": \"lora_path\", " + "\"base_model_name\": \"id\"}``") parser.add_argument( "--prompt-adapters", type=nullable_str, @@ -132,7 +132,7 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: default=None, help="The file path to the chat template, " "or the template in single-line form " - "for the specified model") + "for the specified model.") parser.add_argument( '--chat-template-content-format', type=str, @@ -141,38 +141,39 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: help='The format to render message content within a chat template.' '\n\n' '* "string" will render the content as a string. ' - 'Example: "Hello World"\n' + 'Example: ``"Hello World"``\n' '* "openai" will render the content as a list of dictionaries, ' 'similar to OpenAI schema. ' - 'Example: [{"type": "text", "text": "Hello world!"}]') + 'Example: ``[{"type": "text", "text": "Hello world!"}]``') parser.add_argument("--response-role", type=nullable_str, default="assistant", help="The role name to return if " - "`request.add_generation_prompt=true`.") + "``request.add_generation_prompt=true``.") parser.add_argument("--ssl-keyfile", type=nullable_str, default=None, - help="The file path to the SSL key file") + help="The file path to the SSL key file.") parser.add_argument("--ssl-certfile", type=nullable_str, default=None, - help="The file path to the SSL cert file") + help="The file path to the SSL cert file.") parser.add_argument("--ssl-ca-certs", type=nullable_str, default=None, - help="The CA certificates file") + help="The CA certificates file.") parser.add_argument( "--ssl-cert-reqs", type=int, default=int(ssl.CERT_NONE), - help="Whether client certificate is required (see stdlib ssl module's)" + help="Whether client certificate is required (see stdlib ssl module's)." ) parser.add_argument( "--root-path", type=nullable_str, default=None, - help="FastAPI root_path when app is behind a path based routing proxy") + help="FastAPI root_path when app is behind a path based routing proxy." + ) parser.add_argument( "--middleware", type=nullable_str, @@ -182,15 +183,15 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: "We accept multiple --middleware arguments. " "The value should be an import path. " "If a function is provided, vLLM will add it to the server " - "using @app.middleware('http'). " + "using ``@app.middleware('http')``. " "If a class is provided, vLLM will add it to the server " - "using app.add_middleware(). ") + "using ``app.add_middleware()``. ") parser.add_argument( "--return-tokens-as-token-ids", action="store_true", - help="When --max-logprobs is specified, represents single tokens as " - "strings of the form 'token_id:{token_id}' so that tokens that " - "are not JSON-encodable can be identified.") + help="When ``--max-logprobs`` is specified, represents single tokens " + " as strings of the form 'token_id:{token_id}' so that tokens " + "that are not JSON-encodable can be identified.") parser.add_argument( "--disable-frontend-multiprocessing", action="store_true", @@ -205,9 +206,8 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: "--enable-auto-tool-choice", action="store_true", default=False, - help= - "Enable auto tool choice for supported models. Use --tool-call-parser" - " to specify which parser to use") + help="Enable auto tool choice for supported models. Use " + "``--tool-call-parser`` to specify which parser to use.") valid_tool_parsers = ToolParserManager.tool_parsers.keys() parser.add_argument( @@ -219,7 +219,7 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: help= "Select the tool call parser depending on the model that you're using." " This is used to parse the model-generated tool call into OpenAI API " - "format. Required for --enable-auto-tool-choice.") + "format. Required for ``--enable-auto-tool-choice``.") parser.add_argument( "--tool-parser-plugin", @@ -228,7 +228,7 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: help= "Special the tool parser plugin write to parse the model-generated tool" " into OpenAI API format, the name register in this plugin can be used " - "in --tool-call-parser.") + "in ``--tool-call-parser``.") parser = AsyncEngineArgs.add_cli_args(parser) @@ -243,7 +243,7 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: "--disable-fastapi-docs", action='store_true', default=False, - help="Disable FastAPI's OpenAPI schema, Swagger UI, and ReDoc endpoint" + help="Disable FastAPI's OpenAPI schema, Swagger UI, and ReDoc endpoint." ) parser.add_argument( "--enable-prompt-tokens-details", diff --git a/vllm/entrypoints/openai/protocol.py b/vllm/entrypoints/openai/protocol.py index 14e41346df775..387a4e17900e0 100644 --- a/vllm/entrypoints/openai/protocol.py +++ b/vllm/entrypoints/openai/protocol.py @@ -3,7 +3,7 @@ import re import time from argparse import Namespace -from typing import Any, Dict, List, Literal, Optional, Union +from typing import Any, ClassVar, Dict, List, Literal, Optional, Set, Union import torch from pydantic import BaseModel, ConfigDict, Field, model_validator @@ -42,23 +42,31 @@ class OpenAIBaseModel(BaseModel): # OpenAI API does allow extra fields model_config = ConfigDict(extra="allow") + # Cache class field names + field_names: ClassVar[Optional[Set[str]]] = None + @model_validator(mode="before") @classmethod def __log_extra_fields__(cls, data): - if isinstance(data, dict): + + field_names = cls.field_names + if field_names is None: + if not isinstance(data, dict): + return data # Get all class field names and their potential aliases field_names = set() for field_name, field in cls.model_fields.items(): field_names.add(field_name) - if hasattr(field, 'alias') and field.alias: - field_names.add(field.alias) - - # Compare against both field names and aliases - extra_fields = data.keys() - field_names - if extra_fields: - logger.warning( - "The following fields were present in the request " - "but ignored: %s", extra_fields) + if alias := getattr(field, 'alias', None): + field_names.add(alias) + cls.field_names = field_names + + # Compare against both field names and aliases + if any(k not in field_names for k in data): + logger.warning( + "The following fields were present in the request " + "but ignored: %s", + data.keys() - field_names) return data @@ -372,13 +380,17 @@ def to_beam_search_params( ) -> BeamSearchParams: # TODO(#9845): remove max_tokens when field is removed from OpenAI API max_tokens = self.max_completion_tokens or self.max_tokens - if max_tokens is None: - max_tokens = default_max_tokens if default_sampling_params is None: default_sampling_params = {} n = self.n if self.n is not None else 1 + # Use minimum of context window, user request & server limit. + max_tokens = min( + val for val in (default_max_tokens, max_tokens, + default_sampling_params.get("max_tokens", None)) + if val is not None) + if (temperature := self.temperature) is None: temperature = default_sampling_params.get( "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]) @@ -398,8 +410,34 @@ def to_sampling_params( default_sampling_params: Optional[dict] = None) -> SamplingParams: # TODO(#9845): remove max_tokens when field is removed from OpenAI API max_tokens = self.max_completion_tokens or self.max_tokens - if max_tokens is None: - max_tokens = default_max_tokens + + if default_sampling_params is None: + default_sampling_params = {} + + # Use minimum of context window, user request & server limit. + max_tokens = min( + val for val in (default_max_tokens, max_tokens, + default_sampling_params.get("max_tokens", None)) + if val is not None) + + # Default parameters + if (repetition_penalty := self.repetition_penalty) is None: + repetition_penalty = default_sampling_params.get( + "repetition_penalty", + self._DEFAULT_SAMPLING_PARAMS["repetition_penalty"], + ) + if (temperature := self.temperature) is None: + temperature = default_sampling_params.get( + "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]) + if (top_p := self.top_p) is None: + top_p = default_sampling_params.get( + "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"]) + if (top_k := self.top_k) is None: + top_k = default_sampling_params.get( + "top_k", self._DEFAULT_SAMPLING_PARAMS["top_k"]) + if (min_p := self.min_p) is None: + min_p = default_sampling_params.get( + "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"]) if default_sampling_params is None: default_sampling_params = {} @@ -732,13 +770,17 @@ def to_beam_search_params( default_sampling_params: Optional[dict] = None ) -> BeamSearchParams: max_tokens = self.max_tokens - if max_tokens is None: - max_tokens = default_max_tokens if default_sampling_params is None: default_sampling_params = {} n = self.n if self.n is not None else 1 + # Use minimum of context window, user request & server limit. + max_tokens = min( + val for val in (default_max_tokens, max_tokens, + default_sampling_params.get("max_tokens", None)) + if val is not None) + if (temperature := self.temperature) is None: temperature = default_sampling_params.get("temperature", 1.0) @@ -756,8 +798,34 @@ def to_sampling_params( logits_processor_pattern: Optional[str], default_sampling_params: Optional[dict] = None) -> SamplingParams: max_tokens = self.max_tokens - if max_tokens is None: - max_tokens = default_max_tokens + + if default_sampling_params is None: + default_sampling_params = {} + + # Use minimum of context window, user request & server limit. + max_tokens = min( + val for val in (default_max_tokens, max_tokens, + default_sampling_params.get("max_tokens", None)) + if val is not None) + + # Default parameters + if (repetition_penalty := self.repetition_penalty) is None: + repetition_penalty = default_sampling_params.get( + "repetition_penalty", + self._DEFAULT_SAMPLING_PARAMS["repetition_penalty"], + ) + if (temperature := self.temperature) is None: + temperature = default_sampling_params.get( + "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]) + if (top_p := self.top_p) is None: + top_p = default_sampling_params.get( + "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"]) + if (top_k := self.top_k) is None: + top_k = default_sampling_params.get( + "top_k", self._DEFAULT_SAMPLING_PARAMS["top_k"]) + if (min_p := self.min_p) is None: + min_p = default_sampling_params.get( + "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"]) if default_sampling_params is None: default_sampling_params = {} @@ -992,6 +1060,52 @@ def to_pooling_params(self): return PoolingParams(additional_data=self.additional_data) +class RerankRequest(OpenAIBaseModel): + model: str + query: str + documents: List[str] + top_n: int = Field(default_factory=lambda: 0) + truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None + + # doc: begin-rerank-pooling-params + additional_data: Optional[Any] = None + # doc: end-rerank-pooling-params + + # doc: begin-rerank-extra-params + priority: int = Field( + default=0, + description=( + "The priority of the request (lower means earlier handling; " + "default: 0). Any priority other than 0 will raise an error " + "if the served model does not use priority scheduling.")) + + # doc: end-rerank-extra-params + + def to_pooling_params(self): + return PoolingParams(additional_data=self.additional_data) + + +class RerankDocument(BaseModel): + text: str + + +class RerankResult(BaseModel): + index: int + document: RerankDocument + relevance_score: float + + +class RerankUsage(BaseModel): + total_tokens: int + + +class RerankResponse(OpenAIBaseModel): + id: str + model: str + usage: RerankUsage + results: List[RerankResult] + + class CompletionLogProbs(OpenAIBaseModel): text_offset: List[int] = Field(default_factory=list) token_logprobs: List[Optional[float]] = Field(default_factory=list) @@ -1211,7 +1325,7 @@ class BatchRequestInput(OpenAIBaseModel): url: str # The parameters of the request. - body: Union[ChatCompletionRequest, EmbeddingRequest] + body: Union[ChatCompletionRequest, EmbeddingRequest, ScoreRequest] class BatchResponseData(OpenAIBaseModel): @@ -1222,7 +1336,8 @@ class BatchResponseData(OpenAIBaseModel): request_id: str # The body of the response. - body: Optional[Union[ChatCompletionResponse, EmbeddingResponse]] = None + body: Optional[Union[ChatCompletionResponse, EmbeddingResponse, + ScoreResponse]] = None class BatchRequestOutput(OpenAIBaseModel): diff --git a/vllm/entrypoints/openai/run_batch.py b/vllm/entrypoints/openai/run_batch.py index f8f136f9d5024..37ae23506acea 100644 --- a/vllm/entrypoints/openai/run_batch.py +++ b/vllm/entrypoints/openai/run_batch.py @@ -16,12 +16,14 @@ BatchRequestOutput, BatchResponseData, ChatCompletionResponse, - EmbeddingResponse, ErrorResponse) + EmbeddingResponse, ErrorResponse, + ScoreResponse) # yapf: enable from vllm.entrypoints.openai.serving_chat import OpenAIServingChat from vllm.entrypoints.openai.serving_embedding import OpenAIServingEmbedding from vllm.entrypoints.openai.serving_models import (BaseModelPath, OpenAIServingModels) +from vllm.entrypoints.openai.serving_score import OpenAIServingScores from vllm.usage.usage_lib import UsageContext from vllm.utils import FlexibleArgumentParser, random_uuid from vllm.version import __version__ as VLLM_VERSION @@ -167,7 +169,8 @@ async def run_request(serving_engine_func: Callable, tracker: BatchProgressTracker) -> BatchRequestOutput: response = await serving_engine_func(request.body) - if isinstance(response, (ChatCompletionResponse, EmbeddingResponse)): + if isinstance(response, + (ChatCompletionResponse, EmbeddingResponse, ScoreResponse)): batch_output = BatchRequestOutput( id=f"vllm-{random_uuid()}", custom_id=request.custom_id, @@ -239,6 +242,12 @@ async def main(args): chat_template=None, chat_template_content_format="auto", ) if model_config.task == "embed" else None + openai_serving_scores = (OpenAIServingScores( + engine, + model_config, + openai_serving_models, + request_logger=request_logger, + ) if model_config.task == "score" else None) tracker = BatchProgressTracker() logger.info("Reading batch from %s...", args.input_file) @@ -279,14 +288,28 @@ async def main(args): )) continue + response_futures.append(run_request(handler_fn, request, tracker)) + tracker.submitted() + elif request.url == "/v1/score": + handler_fn = (None if openai_serving_scores is None else + openai_serving_scores.create_score) + if handler_fn is None: + response_futures.append( + make_async_error_request_output( + request, + error_msg="The model does not support Scores API", + )) + continue + response_futures.append(run_request(handler_fn, request, tracker)) tracker.submitted() else: response_futures.append( make_async_error_request_output( request, - error_msg="Only /v1/chat/completions and " - "/v1/embeddings are supported in the batch endpoint.", + error_msg= + "Only /v1/chat/completions, /v1/embeddings, and /v1/score " + "are supported in the batch endpoint.", )) with tracker.pbar(): diff --git a/vllm/entrypoints/openai/serving_engine.py b/vllm/entrypoints/openai/serving_engine.py index 88859255f202a..8d54164e500eb 100644 --- a/vllm/entrypoints/openai/serving_engine.py +++ b/vllm/entrypoints/openai/serving_engine.py @@ -26,7 +26,8 @@ DetokenizeRequest, EmbeddingChatRequest, EmbeddingCompletionRequest, - ErrorResponse, ScoreRequest, + ErrorResponse, RerankRequest, + ScoreRequest, TokenizeChatRequest, TokenizeCompletionRequest) from vllm.entrypoints.openai.serving_models import OpenAIServingModels @@ -203,15 +204,19 @@ def _validate_input( ) -> TextTokensPrompt: token_num = len(input_ids) - # Note: EmbeddingRequest doesn't have max_tokens + # Note: EmbeddingRequest and ScoreRequest doesn't have max_tokens if isinstance(request, - (EmbeddingChatRequest, EmbeddingCompletionRequest)): + (EmbeddingChatRequest, EmbeddingCompletionRequest, + ScoreRequest, RerankRequest)): + + operation = "score" if isinstance(request, ScoreRequest) \ + else "embedding generation" if token_num > self.max_model_len: raise ValueError( f"This model's maximum context length is " f"{self.max_model_len} tokens. However, you requested " - f"{token_num} tokens in the input for embedding " - f"generation. Please reduce the length of the input.") + f"{token_num} tokens in the input for {operation}. " + f"Please reduce the length of the input.") return TextTokensPrompt(prompt=input_text, prompt_token_ids=input_ids) diff --git a/vllm/entrypoints/openai/serving_models.py b/vllm/entrypoints/openai/serving_models.py index a222eafadcb68..fc422f0917bd5 100644 --- a/vllm/entrypoints/openai/serving_models.py +++ b/vllm/entrypoints/openai/serving_models.py @@ -157,24 +157,16 @@ async def load_lora_adapter( # This will also pre-load it for incoming requests try: await self.engine_client.add_lora(lora_request) - except ValueError as e: - # Adapter not found or lora configuration errors - if "No adapter found" in str(e): - return create_error_response(message=str(e), - err_type="NotFoundError", - status_code=HTTPStatus.NOT_FOUND) - else: - return create_error_response( - message=str(e), - err_type="BadRequestError", - status_code=HTTPStatus.BAD_REQUEST) except BaseException as e: - # Some other unexpected problem loading the adapter, e.g. malformed - # input files. - # More detailed error messages for the user would be nicer here + error_type = "BadRequestError" + status_code = HTTPStatus.BAD_REQUEST + if isinstance(e, ValueError) and "No adapter found" in str(e): + error_type = "NotFoundError" + status_code = HTTPStatus.NOT_FOUND + return create_error_response(message=str(e), - err_type="BadRequestError", - status_code=HTTPStatus.BAD_REQUEST) + err_type=error_type, + status_code=status_code) self.lora_requests.append(lora_request) logger.info("Loaded new LoRA adapter: name '%s', path '%s'", lora_name, diff --git a/vllm/entrypoints/openai/serving_rerank.py b/vllm/entrypoints/openai/serving_rerank.py new file mode 100644 index 0000000000000..be4420261afe3 --- /dev/null +++ b/vllm/entrypoints/openai/serving_rerank.py @@ -0,0 +1,206 @@ +import asyncio +from typing import Any, AsyncGenerator, Dict, List, Optional, Union, cast + +from fastapi import Request + +from vllm.config import ModelConfig +from vllm.engine.protocol import EngineClient +from vllm.entrypoints.logger import RequestLogger +from vllm.entrypoints.openai.protocol import (ErrorResponse, RerankDocument, + RerankRequest, RerankResponse, + RerankResult, RerankUsage) +from vllm.entrypoints.openai.serving_engine import OpenAIServing +from vllm.entrypoints.openai.serving_models import OpenAIServingModels +from vllm.inputs.data import TokensPrompt +from vllm.logger import init_logger +from vllm.outputs import PoolingRequestOutput, ScoringRequestOutput +from vllm.transformers_utils.tokenizers.mistral import MistralTokenizer +from vllm.utils import make_async, merge_async_iterators + +logger = init_logger(__name__) + + +class JinaAIServingRerank(OpenAIServing): + + def __init__( + self, + engine_client: EngineClient, + model_config: ModelConfig, + models: OpenAIServingModels, + *, + request_logger: Optional[RequestLogger], + ) -> None: + super().__init__(engine_client=engine_client, + model_config=model_config, + models=models, + request_logger=request_logger) + + async def do_rerank( + self, + request: RerankRequest, + raw_request: Optional[Request] = None + ) -> Union[RerankResponse, ErrorResponse]: + """ + Rerank API based on JinaAI's rerank API; implements the same + API interface. Designed for compatibility with off-the-shelf + tooling, since this is a common standard for reranking APIs + + See example client implementations at + https://github.com/infiniflow/ragflow/blob/main/rag/llm/rerank_model.py + numerous clients use this standard. + """ + error_check_ret = await self._check_model(request) + if error_check_ret is not None: + return error_check_ret + + model_name = request.model + request_id = f"rerank-{self._base_request_id(raw_request)}" + truncate_prompt_tokens = request.truncate_prompt_tokens + query = request.query + documents = request.documents + request_prompts = [] + engine_prompts = [] + top_n = request.top_n if request.top_n > 0 else len(documents) + + try: + ( + lora_request, + prompt_adapter_request, + ) = self._maybe_get_adapters(request) + + tokenizer = await self.engine_client.get_tokenizer(lora_request) + + if prompt_adapter_request is not None: + raise NotImplementedError("Prompt adapter is not supported " + "for scoring models") + + if isinstance(tokenizer, MistralTokenizer): + raise ValueError( + "MistralTokenizer not supported for cross-encoding") + + if not self.model_config.is_cross_encoder: + raise ValueError("Model is not cross encoder.") + + if truncate_prompt_tokens is not None and \ + truncate_prompt_tokens > self.max_model_len: + raise ValueError( + f"truncate_prompt_tokens value ({truncate_prompt_tokens}) " + f"is greater than max_model_len ({self.max_model_len})." + f" Please, select a smaller truncation size.") + for doc in documents: + request_prompt = f"{query}{tokenizer.sep_token}{doc}" + tokenization_kwargs: Dict[str, Any] = {} + if truncate_prompt_tokens is not None: + tokenization_kwargs["truncation"] = True + tokenization_kwargs["max_length"] = truncate_prompt_tokens + + tokenize_async = make_async(tokenizer.__call__, + executor=self._tokenizer_executor) + prompt_inputs = await tokenize_async(text=query, + text_pair=doc, + **tokenization_kwargs) + + input_ids = prompt_inputs["input_ids"] + text_token_prompt = \ + self._validate_input(request, input_ids, request_prompt) + engine_prompt = TokensPrompt( + prompt_token_ids=text_token_prompt["prompt_token_ids"], + token_type_ids=prompt_inputs.get("token_type_ids")) + + request_prompts.append(request_prompt) + engine_prompts.append(engine_prompt) + + except ValueError as e: + logger.exception("Error in preprocessing prompt inputs") + return self.create_error_response(str(e)) + + # Schedule the request and get the result generator. + generators: List[AsyncGenerator[PoolingRequestOutput, None]] = [] + + try: + pooling_params = request.to_pooling_params() + + for i, engine_prompt in enumerate(engine_prompts): + request_id_item = f"{request_id}-{i}" + + self._log_inputs(request_id_item, + request_prompts[i], + params=pooling_params, + lora_request=lora_request, + prompt_adapter_request=prompt_adapter_request) + + trace_headers = (None if raw_request is None else await + self._get_trace_headers(raw_request.headers)) + + generator = self.engine_client.encode( + engine_prompt, + pooling_params, + request_id_item, + lora_request=lora_request, + trace_headers=trace_headers, + priority=request.priority, + ) + + generators.append(generator) + except ValueError as e: + # TODO: Use a vllm-specific Validation Error + return self.create_error_response(str(e)) + result_generator = merge_async_iterators(*generators) + + num_prompts = len(engine_prompts) + + # Non-streaming response + final_res_batch: List[Optional[PoolingRequestOutput]] + final_res_batch = [None] * num_prompts + + try: + async for i, res in result_generator: + final_res_batch[i] = res + + assert all(final_res is not None for final_res in final_res_batch) + + final_res_batch_checked = cast(List[PoolingRequestOutput], + final_res_batch) + + response = self.request_output_to_rerank_response( + final_res_batch_checked, request_id, model_name, documents, + top_n) + except asyncio.CancelledError: + return self.create_error_response("Client disconnected") + except ValueError as e: + # TODO: Use a vllm-specific Validation Error + return self.create_error_response(str(e)) + + return response + + def request_output_to_rerank_response( + self, final_res_batch: List[PoolingRequestOutput], request_id: str, + model_name: str, documents: List[str], + top_n: int) -> RerankResponse: + """ + Convert the output of do_rank to a RerankResponse + """ + results: List[RerankResult] = [] + num_prompt_tokens = 0 + for idx, final_res in enumerate(final_res_batch): + classify_res = ScoringRequestOutput.from_base(final_res) + + result = RerankResult( + index=idx, + document=RerankDocument(text=documents[idx]), + relevance_score=classify_res.outputs.score, + ) + results.append(result) + prompt_token_ids = final_res.prompt_token_ids + num_prompt_tokens += len(prompt_token_ids) + + # sort by relevance, then return the top n if set + results.sort(key=lambda x: x.relevance_score, reverse=True) + if top_n < len(documents): + results = results[:top_n] + + return RerankResponse( + id=request_id, + model=model_name, + results=results, + usage=RerankUsage(total_tokens=num_prompt_tokens)) diff --git a/vllm/entrypoints/openai/serving_score.py b/vllm/entrypoints/openai/serving_score.py index 5d3e7139d7a17..381edf8fac49e 100644 --- a/vllm/entrypoints/openai/serving_score.py +++ b/vllm/entrypoints/openai/serving_score.py @@ -101,6 +101,38 @@ async def create_score( if not self.model_config.is_cross_encoder: raise ValueError("Model is not cross encoder.") + if truncate_prompt_tokens is not None and \ + truncate_prompt_tokens > self.max_model_len: + raise ValueError( + f"truncate_prompt_tokens value ({truncate_prompt_tokens}) " + f"is greater than max_model_len ({self.max_model_len})." + f" Please, select a smaller truncation size.") + + input_pairs = make_pairs(request.text_1, request.text_2) + for q, t in input_pairs: + request_prompt = f"{q}{tokenizer.sep_token}{t}" + + tokenization_kwargs: Dict[str, Any] = {} + if truncate_prompt_tokens is not None: + tokenization_kwargs["truncation"] = True + tokenization_kwargs["max_length"] = truncate_prompt_tokens + + tokenize_async = make_async(tokenizer.__call__, + executor=self._tokenizer_executor) + prompt_inputs = await tokenize_async(text=q, + text_pair=t, + **tokenization_kwargs) + + input_ids = prompt_inputs["input_ids"] + text_token_prompt = \ + self._validate_input(request, input_ids, request_prompt) + engine_prompt = TokensPrompt( + prompt_token_ids=text_token_prompt["prompt_token_ids"], + token_type_ids=prompt_inputs.get("token_type_ids")) + + request_prompts.append(request_prompt) + engine_prompts.append(engine_prompt) + except ValueError as e: logger.exception("Error in preprocessing prompt inputs") return self.create_error_response(str(e)) @@ -108,28 +140,6 @@ async def create_score( # Schedule the request and get the result generator. generators: List[AsyncGenerator[PoolingRequestOutput, None]] = [] - input_pairs = make_pairs(request.text_1, request.text_2) - - for q, t in input_pairs: - request_prompt = f"{q}{tokenizer.sep_token}{t}" - - tokenization_kwargs: Dict[str, Any] = {} - if truncate_prompt_tokens is not None: - tokenization_kwargs["truncation"] = True - tokenization_kwargs["max_length"] = truncate_prompt_tokens - - tokenize_async = make_async(tokenizer.__call__, - executor=self._tokenizer_executor) - prompt_inputs = await tokenize_async(text=q, - text_pair=t, - **tokenization_kwargs) - engine_prompt = TokensPrompt( - prompt_token_ids=prompt_inputs["input_ids"], - token_type_ids=prompt_inputs.get("token_type_ids")) - - request_prompts.append(request_prompt) - engine_prompts.append(engine_prompt) - try: pooling_params = request.to_pooling_params() diff --git a/vllm/envs.py b/vllm/envs.py index e53e7108f953c..0445447dd9df0 100644 --- a/vllm/envs.py +++ b/vllm/envs.py @@ -19,6 +19,7 @@ VLLM_USE_ROCM_CUSTOM_PAGED_ATTN_FP8_OUT: bool = True VLLM_USE_ROCM_FP8_FLASH_ATTN: bool = False RANK: int = 0 + VLLM_FLASH_ATTN_VERSION: Optional[int] = None LOCAL_RANK: int = 0 CUDA_VISIBLE_DEVICES: Optional[str] = None VLLM_ENGINE_ITERATION_TIMEOUT_S: int = 60 @@ -89,6 +90,8 @@ Q_SCALE_CONSTANT: int = 20 K_SCALE_CONSTANT: int = 20 V_SCALE_CONSTANT: int = 10 + VLLM_SERVER_DEV_MODE: bool = False + VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128 def get_default_cache_root(): @@ -105,6 +108,12 @@ def get_default_config_root(): ) +def maybe_convert_int(value: Optional[str]) -> Optional[int]: + if value is None: + return None + return int(value) + + # The begin-* and end* here are used by the documentation generator # to extract the used env vars. @@ -233,6 +242,11 @@ def get_default_config_root(): lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "True").lower() in ("true", "1")), + # Force vllm to use a specific flash-attention version (2 or 3), only valid + # when using the flash-attention backend. + "VLLM_FLASH_ATTN_VERSION": + lambda: maybe_convert_int(os.environ.get("VLLM_FLASH_ATTN_VERSION", None)), + # Internal flag to enable Dynamo fullgraph capture "VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE": lambda: bool( @@ -527,8 +541,6 @@ def get_default_config_root(): # Try to accumulate this many requests before proceeding "VLLM_SYNC_SERVER_ACCUM_REQUESTS": lambda: int(os.getenv("VLLM_SYNC_SERVER_ACCUM_REQUESTS", "1")), - - # Poll for new requests every this many steps "VLLM_SYNC_SERVER_ENGINE_STEPS_BETWEEN_POLLS": lambda: int(os.getenv("VLLM_SYNC_SERVER_ENGINE_STEPS_BETWEEN_POLLS", "1")), @@ -561,7 +573,6 @@ def get_default_config_root(): # for FP8 KV Cache and attention "V_SCALE_CONSTANT": lambda: int(os.getenv("V_SCALE_CONSTANT", "10")), - # If set, enable multiprocessing in LLM for the V1 code path. "VLLM_ENABLE_V1_MULTIPROCESSING": lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))), diff --git a/vllm/executor/cpu_executor.py b/vllm/executor/cpu_executor.py deleted file mode 100644 index b9a6bee5720fd..0000000000000 --- a/vllm/executor/cpu_executor.py +++ /dev/null @@ -1,299 +0,0 @@ -import os -from functools import partial -from typing import Any, Awaitable, List, Optional, Set, Tuple, Union - -from vllm.executor.executor_base import ExecutorAsyncBase, ExecutorBase -from vllm.executor.multiproc_worker_utils import (ProcessWorkerWrapper, - ResultHandler, WorkerMonitor) -from vllm.logger import init_logger -from vllm.lora.request import LoRARequest -from vllm.model_executor.layers.sampler import SamplerOutput -from vllm.prompt_adapter.request import PromptAdapterRequest -from vllm.sequence import ExecuteModelRequest -from vllm.utils import get_distributed_init_method, get_open_port, make_async -from vllm.worker.worker_base import WorkerWrapperBase - -logger = init_logger(__name__) - - -class CPUExecutor(ExecutorBase): - - uses_ray: bool = False - - def _init_executor(self) -> None: - assert self.device_config.device_type == "cpu" - - # - # Environment variables for CPU executor - # - - # Disable torch async compiling which won't work with daemonic processes - os.environ["TORCHINDUCTOR_COMPILE_THREADS"] = "1" - - # Intel OpenMP setting - ld_prealod_str = os.getenv("LD_PRELOAD", "") - if "libiomp5.so" in ld_prealod_str: - # The time(milliseconds) that a thread should wait after - # completing the execution of a parallel region, before sleeping. - os.environ['KMP_BLOCKTIME'] = "1" - # Prevents the CPU to run into low performance state - os.environ['KMP_TPAUSE'] = "0" - # Provides fine granularity parallelism - os.environ['KMP_FORKJOIN_BARRIER_PATTERN'] = "dist,dist" - os.environ['KMP_PLAIN_BARRIER_PATTERN'] = "dist,dist" - os.environ['KMP_REDUCTION_BARRIER_PATTERN'] = "dist,dist" - - # To hint IPEX uses shared memory based AllReduce - os.environ["LOCAL_WORLD_SIZE"] = str( - self.parallel_config.tensor_parallel_size) - - # Multiprocessing-based executor does not support multi-node setting. - # Since it only works for single node, we can use the loopback address - # 127.0.0.1 for communication. - ip = "127.0.0.1" - port = get_open_port() - self.distributed_init_method = get_distributed_init_method(ip, port) - - is_async = isinstance(self, CPUExecutorAsync) - - world_size = self.parallel_config.tensor_parallel_size - result_handler = ResultHandler() - self.parallel_worker_tasks: Optional[Union[Any, Awaitable[Any]]] = None - self.workers = [] - - if is_async: - self.workers = [ - ProcessWorkerWrapper( - result_handler, - partial( - self._create_worker, - rank=rank, - local_rank=rank, - )) for rank in range(0, world_size) - ] - self.driver_worker = self.workers[0] - self.workers = self.workers[1:] - self.driver_method_invoker = _async_driver_method_invoker - else: - self.driver_worker = self._create_worker() - self.driver_method_invoker = _driver_method_invoker - - if world_size != 1: - self.workers = [ - ProcessWorkerWrapper( - result_handler, - partial( - self._create_worker, - rank=rank, - local_rank=rank, - )) for rank in range(1, world_size) - ] - - self.worker_monitor = None - if world_size != 1 or is_async: - if is_async: - async_worker_list = self.workers + [self.driver_worker] - else: - async_worker_list = self.workers - self.worker_monitor = WorkerMonitor(async_worker_list, - result_handler) - result_handler.start() - self.worker_monitor.start() - - self._run_workers("init_device") - self._run_workers("load_model") - - def _create_worker( - self, - local_rank: int = 0, - rank: int = 0, - ): - - wrapper = WorkerWrapperBase(vllm_config=self.vllm_config) - - assert self.distributed_init_method is not None - - kwargs = dict( - vllm_config=self.vllm_config, - local_rank=local_rank, - rank=rank, - distributed_init_method=self.distributed_init_method, - kv_cache_dtype=self.cache_config.cache_dtype, - is_driver_worker=rank == 0, - ) - wrapper.init_worker(**kwargs) - - return wrapper.worker - - def _run_workers( - self, - method: str, - *args, - async_run_remote_workers_only: bool = False, - max_concurrent_workers: Optional[int] = None, - **kwargs, - ) -> Any: - """Runs the given method on all workers. - - Args: - async_run_remote_workers_only: If True the method will be run only - in the remote workers, not the driver worker. It will also be - run asynchronously and return a list of futures rather than - blocking on the results. - """ - - if max_concurrent_workers: - raise NotImplementedError( - "max_concurrent_workers is not supported yet.") - - # Start the workers first. - worker_outputs = [ - worker.execute_method(method, *args, **kwargs) - for worker in self.workers - ] - - if async_run_remote_workers_only: - # Just return futures - return worker_outputs - - driver_worker_output = self.driver_method_invoker( - self.driver_worker, method, *args, **kwargs) - - # Get the results of the workers. - return [driver_worker_output - ] + [output.get() for output in worker_outputs] - - def determine_num_available_blocks(self) -> Tuple[int, int]: - """Determine the number of available KV blocks by invoking the - underlying worker. - """ - return self.driver_method_invoker(self.driver_worker, - "determine_num_available_blocks") - - def initialize_cache(self, num_gpu_blocks: int, - num_cpu_blocks: int) -> None: - """Initialize the KV cache by invoking the underlying worker. - """ - # NOTE: We log here to avoid multiple logs when number of workers is - # greater than one. We could log in the engine, but not all executors - # have GPUs. - # NOTE: `cpu block` for CPU backend is located on CPU memory but is - # referred as `gpu block`. Because we want to reuse the existing block - # management procedure. - logger.info("# CPU blocks: %d", num_gpu_blocks) - - self._run_workers("initialize_cache", - num_gpu_blocks=num_gpu_blocks, - num_cpu_blocks=num_cpu_blocks) - - def execute_model( - self, - execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: - if (self.parallel_config.tensor_parallel_size > 1 - and self.parallel_worker_tasks is None): - self.parallel_worker_tasks = self._run_workers( - "start_worker_execution_loop", - async_run_remote_workers_only=True, - ) - output = self.driver_method_invoker(self.driver_worker, - "execute_model", execute_model_req) - return output - - def stop_remote_worker_execution_loop(self) -> None: - if self.parallel_worker_tasks is None: - return - """ - Passing None will cause the driver to stop the model execution - loop running in each of the remote workers. - """ - self.driver_method_invoker(self.driver_worker, "execute_model", None) - parallel_worker_tasks = self.parallel_worker_tasks - self.parallel_worker_tasks = None - # Ensure that workers exit model loop cleanly - # (this will raise otherwise) - self._wait_for_tasks_completion(parallel_worker_tasks) - - def add_lora(self, lora_request: LoRARequest) -> bool: - return all(self._run_workers("add_lora", lora_request)) - - def remove_lora(self, lora_id: int) -> bool: - return all(self._run_workers("remove_lora", lora_id)) - - def pin_lora(self, lora_id: int) -> bool: - assert lora_id > 0, "lora_id must be greater than 0." - return all(self._run_workers( - "pin_lora", - lora_id=lora_id, - )) - - def list_loras(self) -> Set[int]: - return self.driver_method_invoker(self.driver_worker, "list_loras") - - def add_prompt_adapter( - self, prompt_adapter_request: PromptAdapterRequest) -> bool: - return all( - self._run_workers( - "add_prompt_adapter", - prompt_adapter_request, - )) - - def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool: - return all( - self._run_workers( - "remove_prompt_adapter", - prompt_adapter_id, - )) - - def list_prompt_adapters(self) -> Set[int]: - return self.driver_method_invoker(self.driver_worker, - "list_prompt_adapters") - - def pin_prompt_adapter(self, prompt_adapter_id: int) -> bool: - return all(self._run_workers( - "pin_prompt_adapter", - prompt_adapter_id, - )) - - def check_health(self) -> None: - """Raises an error if engine is unhealthy.""" - if self.worker_monitor is not None and not self.worker_monitor.is_alive( - ): - raise RuntimeError("Worker processes are not running") - - def shutdown(self): - if (worker_monitor := getattr(self, "worker_monitor", - None)) is not None: - worker_monitor.close() - - def _wait_for_tasks_completion(self, parallel_worker_tasks: Any) -> None: - """Wait for futures returned from _run_workers() with - async_run_remote_workers_only to complete.""" - for result in parallel_worker_tasks: - result.get() - - def start_profile(self) -> None: - self.driver_method_invoker(self.driver_worker, "start_profile") - - def stop_profile(self) -> None: - self.driver_method_invoker(self.driver_worker, "stop_profile") - - -class CPUExecutorAsync(CPUExecutor, ExecutorAsyncBase): - - async def execute_model_async( - self, - execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: - output = await make_async(self.execute_model - )(execute_model_req=execute_model_req, ) - return output - - async def check_health_async(self) -> None: - self.check_health() - - -def _driver_method_invoker(driver, method: str, *args, **kwargs): - return getattr(driver, method)(*args, **kwargs) - - -def _async_driver_method_invoker(driver, method: str, *args, **kwargs): - return driver.execute_method(method, *args, **kwargs).get() diff --git a/vllm/executor/distributed_gpu_executor.py b/vllm/executor/distributed_gpu_executor.py deleted file mode 100644 index deb7cb1c97ef5..0000000000000 --- a/vllm/executor/distributed_gpu_executor.py +++ /dev/null @@ -1,212 +0,0 @@ -import asyncio -from abc import abstractmethod -from typing import Any, Awaitable, Dict, List, Optional, Set, Tuple, Union - -from vllm.executor.executor_base import ExecutorAsyncBase -from vllm.executor.gpu_executor import GPUExecutor -from vllm.logger import init_logger -from vllm.lora.request import LoRARequest -from vllm.model_executor.layers.sampler import SamplerOutput -from vllm.sequence import ExecuteModelRequest - -logger = init_logger(__name__) - - -class DistributedGPUExecutor(GPUExecutor): - """Abstract superclass of multi-GPU executor implementations.""" - - def __init__(self, *args, **kwargs): - # This is non-None when the execute model loop is running - # in the parallel workers. It's a coroutine in the AsyncLLMEngine case. - self.parallel_worker_tasks: Optional[Union[Any, Awaitable[Any]]] = None - # Updated by implementations that require additional args to be passed - # to the _run_workers execute_model call - self.extra_execute_model_run_workers_kwargs: Dict[str, Any] = {} - - super().__init__(*args, **kwargs) - - def determine_num_available_blocks(self) -> Tuple[int, int]: - """Determine the number of available KV blocks. - - This invokes `determine_num_available_blocks` on each worker and takes - the min of the results, guaranteeing that the selected cache sizes are - compatible with all workers. - - Returns: - - tuple[num_gpu_blocks, num_cpu_blocks] - """ - # Get the maximum number of blocks that can be allocated on GPU and CPU. - num_blocks = self._run_workers("determine_num_available_blocks", ) - - # Since we use a shared centralized controller, we take the minimum - # number of blocks across all workers to make sure all the memory - # operators can be applied to all workers. - num_gpu_blocks = min(b[0] for b in num_blocks) - num_cpu_blocks = min(b[1] for b in num_blocks) - - return num_gpu_blocks, num_cpu_blocks - - def initialize_cache(self, num_gpu_blocks: int, - num_cpu_blocks: int) -> None: - """Initialize the KV cache in all workers. - """ - - # NOTE: We log here to avoid multiple logs when number of workers is - # greater than one. We could log in the engine, but not all executors - # have GPUs. - logger.info("# GPU blocks: %d, # CPU blocks: %d", num_gpu_blocks, - num_cpu_blocks) - max_concurrency = (num_gpu_blocks * self.cache_config.block_size / - self.model_config.max_model_len) - logger.info("Maximum concurrency for %s tokens per request: %.2fx", - self.model_config.max_model_len, max_concurrency) - - self.cache_config.num_gpu_blocks = num_gpu_blocks - self.cache_config.num_cpu_blocks = num_cpu_blocks - - self._run_workers("initialize_cache", - num_gpu_blocks=num_gpu_blocks, - num_cpu_blocks=num_cpu_blocks) - - def execute_model( - self, - execute_model_req: ExecuteModelRequest, - ) -> List[SamplerOutput]: - if self.parallel_worker_tasks is None: - self.parallel_worker_tasks = self._run_workers( - "start_worker_execution_loop", - async_run_tensor_parallel_workers_only=True, - **self.extra_execute_model_run_workers_kwargs) - - # Only the driver worker returns the sampling results. - driver_outputs = self._driver_execute_model(execute_model_req) - assert driver_outputs is not None - return driver_outputs - - def stop_remote_worker_execution_loop(self) -> None: - if self.parallel_worker_tasks is None: - return - - self._driver_execute_model(execute_model_req=None) - parallel_worker_tasks = self.parallel_worker_tasks - self.parallel_worker_tasks = None - # Ensure that workers exit model loop cleanly - # (this will raise otherwise) - self._wait_for_tasks_completion(parallel_worker_tasks) - - def add_lora(self, lora_request: LoRARequest) -> bool: - assert lora_request.lora_int_id > 0, "lora_id must be greater than 0." - return self._run_workers( - "add_lora", - lora_request=lora_request, - ) - - def remove_lora(self, lora_id: int) -> bool: - assert lora_id > 0, "lora_id must be greater than 0." - return self._run_workers( - "remove_lora", - lora_id=lora_id, - ) - - def pin_lora(self, lora_id: int) -> bool: - assert lora_id > 0, "lora_id must be greater than 0." - return self._run_workers( - "pin_lora", - lora_id=lora_id, - ) - - def list_loras(self) -> Set[int]: - return self._run_workers("list_loras") - - def save_sharded_state( - self, - path: str, - pattern: Optional[str] = None, - max_size: Optional[int] = None, - ) -> None: - self._run_workers("save_sharded_state", - path=path, - pattern=pattern, - max_size=max_size) - - @abstractmethod - def _driver_execute_model( - self, execute_model_req: Optional[ExecuteModelRequest] - ) -> Optional[List[SamplerOutput]]: - """Run execute_model in the driver worker. - - Passing None will cause the driver to stop the model execution loop - running in each of the remote workers. In this case, this method - returns None. Otherwise, this method returns the model output. - """ - raise NotImplementedError - - @abstractmethod - def _run_workers( - self, - method: str, - *args, - async_run_tensor_parallel_workers_only: bool = False, - max_concurrent_workers: Optional[int] = None, - **kwargs, - ) -> Any: - """Runs the given method on all workers. - - Args: - async_run_tensor_parallel_workers_only: If True the method will be - run only in the remote TP workers, not the driver worker. - It will also be run asynchronously and return a list of futures - rather than blocking on the results. - """ - raise NotImplementedError - - @abstractmethod - def _wait_for_tasks_completion(self, parallel_worker_tasks: Any) -> None: - """Wait for futures returned from _run_workers() with - async_run_remote_workers_only to complete.""" - raise NotImplementedError - - -class DistributedGPUExecutorAsync(DistributedGPUExecutor, ExecutorAsyncBase): - - async def execute_model_async( - self, - execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: - if self.parallel_worker_tasks is None: - # Start model execution loop running in the parallel workers - self.parallel_worker_tasks = asyncio.create_task( - self._start_worker_execution_loop()) - - # Only the driver worker returns the sampling results. - return await self._driver_execute_model_async(execute_model_req) - - async def stop_remote_worker_execution_loop_async(self) -> None: - if self.parallel_worker_tasks is None: - return - - await self._driver_execute_model_async() - parallel_worker_tasks = self.parallel_worker_tasks - self.parallel_worker_tasks = None - # Ensure that workers exit model loop cleanly - # (this will raise otherwise) - await parallel_worker_tasks - - @abstractmethod - async def _driver_execute_model_async( - self, - execute_model_req: Optional[ExecuteModelRequest] = None, - ) -> List[SamplerOutput]: - """Execute the model asynchronously in the driver worker. - - Passing None will cause the driver to stop the model execution - loop running in each of the remote workers. - """ - raise NotImplementedError - - @abstractmethod - async def _start_worker_execution_loop(self): - """Run execution loop on all workers. It guarantees all workers run - the loop or None of them is running the loop. Loop can be stopped by - `stop_remote_worker_execution_loop`. - The API is idempotent (guarantee only 1 loop run at any moment).""" - raise NotImplementedError diff --git a/vllm/executor/executor_base.py b/vllm/executor/executor_base.py index 9cba189dd57f9..471d1bfac3119 100644 --- a/vllm/executor/executor_base.py +++ b/vllm/executor/executor_base.py @@ -1,18 +1,31 @@ +import asyncio from abc import ABC, abstractmethod -from typing import List, Optional, Set, Tuple +from typing import (Any, Awaitable, Callable, Dict, List, Optional, Set, Tuple, + Union) + +import torch.nn as nn +from typing_extensions import TypeVar from vllm.config import VllmConfig +from vllm.logger import init_logger from vllm.lora.request import LoRARequest from vllm.model_executor.layers.sampler import SamplerOutput +from vllm.platforms import current_platform from vllm.prompt_adapter.request import PromptAdapterRequest -from vllm.sequence import ExecuteModelRequest +from vllm.sequence import ExecuteModelRequest, PoolerOutput +from vllm.utils import make_async +from vllm.worker.worker_base import WorkerBase + +logger = init_logger(__name__) + +_R = TypeVar("_R", default=Any) class ExecutorBase(ABC): """Base class for all executors. - An executor is responsible for executing the model on a specific device - type (e.g., CPU, GPU, Neuron, etc.). Or it can be a distributed executor + An executor is responsible for executing the model on one device, + or it can be a distributed executor that can execute the model on multiple devices. """ @@ -34,12 +47,42 @@ def __init__( self.prompt_adapter_config = vllm_config.prompt_adapter_config self.observability_config = vllm_config.observability_config self._init_executor() + self.is_sleeping = False @abstractmethod def _init_executor(self) -> None: - pass + raise NotImplementedError @abstractmethod + def collective_rpc(self, + method: Union[str, Callable[..., _R]], + timeout: Optional[float] = None, + args: Tuple = (), + kwargs: Optional[Dict[str, Any]] = None) -> List[_R]: + """ + Execute an RPC call on all workers. + + Args: + method: Name of the worker method to execute, or a callable that + is serialized and sent to all workers to execute. + + If the method is a callable, it should accept an additional + `self` argument, in addition to the arguments passed in `args` + and `kwargs`. The `self` argument will be the worker object. + timeout: Maximum time in seconds to wait for execution. Raises a + :exc:`TimeoutError` on timeout. `None` means wait indefinitely. + args: Positional arguments to pass to the worker method. + kwargs: Keyword arguments to pass to the worker method. + + Returns: + A list containing the results from each worker. + + Note: + It is recommended to use this API to only pass control messages, + and set up data-plane communication to pass data. + """ + raise NotImplementedError + def determine_num_available_blocks(self) -> Tuple[int, int]: """Determine the number of available blocks for the GPU KV cache and swappable CPU KV cache. @@ -53,58 +96,128 @@ def determine_num_available_blocks(self) -> Tuple[int, int]: num_cpu_blocks refers to "swapped" blocks in CPU memory and cannot be appended to. """ - raise NotImplementedError + results = self.collective_rpc("determine_num_available_blocks") + a = min([r[0] for r in results]) + b = min([r[1] for r in results]) + return a, b - @abstractmethod - def initialize_cache(self, num_gpu_blocks: int, - num_cpu_blocks: int) -> None: - """Initialize the KV cache with the given size in blocks. + def initialize_cache(self, num_gpu_blocks: int, num_cpu_blocks) -> None: + """Initialize the KV cache by invoking the underlying worker. """ - raise NotImplementedError + # NOTE: This is logged in the executor because there can be >1 workers. + logger.info("# %s blocks: %d, # CPU blocks: %d", + current_platform.dispatch_key, num_gpu_blocks, + num_cpu_blocks) + max_concurrency = (num_gpu_blocks * self.cache_config.block_size / + self.model_config.max_model_len) + logger.info("Maximum concurrency for %s tokens per request: %.2fx", + self.model_config.max_model_len, max_concurrency) + + self.cache_config.num_gpu_blocks = num_gpu_blocks + self.cache_config.num_cpu_blocks = num_cpu_blocks + + self.collective_rpc("initialize_cache", + args=(num_gpu_blocks, num_cpu_blocks)) + + def apply_model(self, func: Callable[[nn.Module], _R]) -> list[_R]: + """ + Run a function directly on the model inside each worker, + returning the result for each of them. + """ + + def rpc_func(worker: WorkerBase) -> _R: + return func(worker.get_model()) + + return self.collective_rpc(rpc_func) - @abstractmethod def execute_model( self, execute_model_req: ExecuteModelRequest - ) -> Optional[List[SamplerOutput]]: - """Executes at least one model step on the given sequences.""" - raise NotImplementedError + ) -> Optional[List[Union[SamplerOutput, PoolerOutput]]]: + output = self.collective_rpc("execute_model", + args=(execute_model_req, )) + return output[0] def stop_remote_worker_execution_loop(self) -> None: """Releases parallel workers from model loop.""" return - @abstractmethod def add_lora(self, lora_request: LoRARequest) -> bool: - raise NotImplementedError + assert lora_request.lora_int_id > 0, "lora_id must be greater than 0." + return all(self.collective_rpc("add_lora", args=(lora_request, ))) - @abstractmethod def remove_lora(self, lora_id: int) -> bool: - raise NotImplementedError + assert lora_id > 0, "lora_id must be greater than 0." + return all(self.collective_rpc("remove_lora", args=(lora_id, ))) - @abstractmethod def pin_lora(self, lora_id: int) -> bool: - raise NotImplementedError # type: ignore + assert lora_id > 0, "lora_id must be greater than 0." + return all(self.collective_rpc("pin_lora", args=(lora_id, ))) - @abstractmethod def list_loras(self) -> Set[int]: - raise NotImplementedError + sets = self.collective_rpc("list_loras") + for s in sets: + assert s == sets[0], "All workers should have the same LORAs." + return sets[0] - @abstractmethod def add_prompt_adapter( self, prompt_adapter_request: PromptAdapterRequest) -> bool: - raise NotImplementedError + assert prompt_adapter_request.prompt_adapter_id > 0, \ + "prompt_adapter_id must be greater than 0." + return all( + self.collective_rpc("add_prompt_adapter", + args=(prompt_adapter_request, ))) - @abstractmethod def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool: - raise NotImplementedError + assert prompt_adapter_id > 0, \ + "prompt_adapter_id must be greater than 0." + return all( + self.collective_rpc("remove_prompt_adapter", + args=(prompt_adapter_id, ))) - @abstractmethod def pin_prompt_adapter(self, prompt_adapter_id: int) -> bool: - raise NotImplementedError # type: ignore + assert prompt_adapter_id > 0, \ + "prompt_adapter_id must be greater than 0." + return all( + self.collective_rpc("pin_prompt_adapter", + args=(prompt_adapter_id, ))) - @abstractmethod def list_prompt_adapters(self) -> Set[int]: - raise NotImplementedError + sets = self.collective_rpc("list_prompt_adapters") + for s in sets: + assert (s == sets[0] + ), "All workers should have the same prompt adapters." + return sets[0] + + def start_profile(self) -> None: + self.collective_rpc("start_profile") + + def stop_profile(self) -> None: + self.collective_rpc("stop_profile") + + def sleep(self, level: int = 1): + if self.is_sleeping: + logger.warning("Executor is already sleeping.") + return + self.collective_rpc("sleep", kwargs=dict(level=level)) + self.is_sleeping = True + + def wake_up(self): + if not self.is_sleeping: + logger.warning("Executor is not sleeping.") + return + self.collective_rpc("wake_up") + self.is_sleeping = False + + def save_sharded_state( + self, + path: str, + pattern: Optional[str] = None, + max_size: Optional[int] = None, + ) -> None: + self.collective_rpc("save_sharded_state", + kwargs=dict(path=path, + pattern=pattern, + max_size=max_size)) @abstractmethod def check_health(self) -> None: @@ -119,15 +232,12 @@ def shutdown(self) -> None: def __del__(self): self.shutdown() - -class ExecutorAsyncBase(ExecutorBase): - - @abstractmethod async def execute_model_async( self, execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: """Executes one model step on the given sequences.""" - raise NotImplementedError + output = await make_async(self.execute_model)(execute_model_req) + return output async def stop_remote_worker_execution_loop_async(self) -> None: """Releases parallel workers from model loop.""" @@ -137,3 +247,128 @@ async def check_health_async(self) -> None: """Checks if the executor is healthy. If not, it should raise an exception.""" self.check_health() + + +class DistributedExecutorBase(ExecutorBase): + """Abstract superclass of distributed executor implementations.""" + + def __init__(self, *args, **kwargs): + # This is non-None when the execute model loop is running + # in the parallel workers. It's a coroutine in the AsyncLLMEngine case. + self.parallel_worker_tasks: Optional[Union[Any, Awaitable[Any]]] = None + + super().__init__(*args, **kwargs) + + def execute_model( + self, + execute_model_req: ExecuteModelRequest, + ) -> List[SamplerOutput]: + # TODO: unify into collective_rpc + if self.parallel_worker_tasks is None: + self.parallel_worker_tasks = self._run_workers( + "start_worker_execution_loop", + async_run_tensor_parallel_workers_only=True) + + # Only the driver worker returns the sampling results. + driver_outputs = self._driver_execute_model(execute_model_req) + assert driver_outputs is not None + return driver_outputs + + def stop_remote_worker_execution_loop(self) -> None: + if self.parallel_worker_tasks is None: + return + + self._driver_execute_model(execute_model_req=None) + parallel_worker_tasks = self.parallel_worker_tasks + self.parallel_worker_tasks = None + # Ensure that workers exit model loop cleanly + # (this will raise otherwise) + self._wait_for_tasks_completion(parallel_worker_tasks) + + @abstractmethod + def _driver_execute_model( + self, execute_model_req: Optional[ExecuteModelRequest] + ) -> Optional[List[SamplerOutput]]: + """Run execute_model in the driver worker. + + Passing None will cause the driver to stop the model execution loop + running in each of the remote workers. In this case, this method + returns None. Otherwise, this method returns the model output. + """ + raise NotImplementedError + + def collective_rpc(self, + method: Union[str, Callable], + timeout: Optional[float] = None, + args: Tuple = (), + kwargs: Optional[Dict] = None) -> List[Any]: + return self._run_workers(method, *args, **(kwargs or {})) + + @abstractmethod + def _run_workers( + self, + method: Union[str, Callable], + *args, + async_run_tensor_parallel_workers_only: bool = False, + max_concurrent_workers: Optional[int] = None, + **kwargs, + ) -> Any: + """Runs the given method on all workers. + + Args: + async_run_tensor_parallel_workers_only: If True the method will be + run only in the remote TP workers, not the driver worker. + It will also be run asynchronously and return a list of futures + rather than blocking on the results. + + # TODO: simplify and merge with collective_rpc + """ + raise NotImplementedError + + @abstractmethod + def _wait_for_tasks_completion(self, parallel_worker_tasks: Any) -> None: + """Wait for futures returned from _run_workers() with + async_run_remote_workers_only to complete.""" + raise NotImplementedError + + async def execute_model_async( + self, + execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: + if self.parallel_worker_tasks is None: + # Start model execution loop running in the parallel workers + self.parallel_worker_tasks = asyncio.create_task( + self._start_worker_execution_loop()) + + # Only the driver worker returns the sampling results. + return await self._driver_execute_model_async(execute_model_req) + + async def stop_remote_worker_execution_loop_async(self) -> None: + if self.parallel_worker_tasks is None: + return + + await self._driver_execute_model_async() + parallel_worker_tasks = self.parallel_worker_tasks + self.parallel_worker_tasks = None + # Ensure that workers exit model loop cleanly + # (this will raise otherwise) + await parallel_worker_tasks + + @abstractmethod + async def _driver_execute_model_async( + self, + execute_model_req: Optional[ExecuteModelRequest] = None, + ) -> List[SamplerOutput]: + """Execute the model asynchronously in the driver worker. + + Passing None will cause the driver to stop the model execution + loop running in each of the remote workers. + """ + raise NotImplementedError + + @abstractmethod + async def _start_worker_execution_loop(self): + """Run execution loop on all workers. It guarantees all workers run + the loop or None of them is running the loop. Loop can be stopped by + `stop_remote_worker_execution_loop`. + The API is idempotent (guarantee only 1 loop run at any moment).""" + raise NotImplementedError diff --git a/vllm/executor/gpu_executor.py b/vllm/executor/gpu_executor.py deleted file mode 100644 index 7fa34456028dd..0000000000000 --- a/vllm/executor/gpu_executor.py +++ /dev/null @@ -1,145 +0,0 @@ -from typing import Any, Dict, List, Optional, Set, Tuple, Union - -from vllm.executor.executor_base import ExecutorAsyncBase, ExecutorBase -from vllm.logger import init_logger -from vllm.lora.request import LoRARequest -from vllm.model_executor.layers.sampler import SamplerOutput -from vllm.prompt_adapter.request import PromptAdapterRequest -from vllm.sequence import ExecuteModelRequest, PoolerOutput -from vllm.utils import (get_distributed_init_method, get_ip, get_open_port, - make_async) -from vllm.worker.worker_base import WorkerWrapperBase - -logger = init_logger(__name__) - - -def create_worker(**kwargs): - vllm_config = kwargs.get("vllm_config") - wrapper = WorkerWrapperBase(vllm_config=vllm_config) - wrapper.init_worker(**kwargs) - return wrapper.worker - - -class GPUExecutor(ExecutorBase): - - uses_ray: bool = False - - def _init_executor(self) -> None: - """Initialize the worker and load the model. - """ - assert self.parallel_config.world_size == 1, ( - "GPUExecutor only supports single GPU.") - - self.driver_worker = self._create_worker() - self.driver_worker.init_device() - self.driver_worker.load_model() - - def _get_worker_kwargs( - self, - local_rank: int = 0, - rank: int = 0, - distributed_init_method: Optional[str] = None) -> Dict[str, Any]: - """Return worker init args for a given rank.""" - if distributed_init_method is None: - distributed_init_method = get_distributed_init_method( - get_ip(), get_open_port()) - return dict( - vllm_config=self.vllm_config, - local_rank=local_rank, - rank=rank, - distributed_init_method=distributed_init_method, - is_driver_worker=(not self.parallel_config) - or (rank % self.parallel_config.tensor_parallel_size == 0), - ) - - def _create_worker(self, - local_rank: int = 0, - rank: int = 0, - distributed_init_method: Optional[str] = None): - return create_worker(**self._get_worker_kwargs( - local_rank=local_rank, - rank=rank, - distributed_init_method=distributed_init_method)) - - def determine_num_available_blocks(self) -> Tuple[int, int]: - """Determine the number of available KV blocks by invoking the - underlying worker. - """ - return self.driver_worker.determine_num_available_blocks() - - def initialize_cache(self, num_gpu_blocks: int, num_cpu_blocks) -> None: - """Initialize the KV cache by invoking the underlying worker. - """ - # NOTE: This is logged in the executor because there can be >1 worker - # with other executors. We could log in the engine level, but work - # remains to abstract away the device for non-GPU configurations. - logger.info("# GPU blocks: %d, # CPU blocks: %d", num_gpu_blocks, - num_cpu_blocks) - max_concurrency = (num_gpu_blocks * self.cache_config.block_size / - self.model_config.max_model_len) - logger.info("Maximum concurrency for %s tokens per request: %.2fx", - self.model_config.max_model_len, max_concurrency) - - self.driver_worker.initialize_cache(num_gpu_blocks, num_cpu_blocks) - - def execute_model( - self, execute_model_req: ExecuteModelRequest - ) -> Optional[List[Union[SamplerOutput, PoolerOutput]]]: - output = self.driver_worker.execute_model(execute_model_req) - return output - - def add_lora(self, lora_request: LoRARequest) -> bool: - assert lora_request.lora_int_id > 0, "lora_id must be greater than 0." - return self.driver_worker.add_lora(lora_request) - - def remove_lora(self, lora_id: int) -> bool: - assert lora_id > 0, "lora_id must be greater than 0." - return self.driver_worker.remove_lora(lora_id) - - def pin_lora(self, lora_id: int) -> bool: - assert lora_id > 0, "lora_id must be greater than 0." - return self.driver_worker.pin_lora(lora_id) - - def list_loras(self) -> Set[int]: - return self.driver_worker.list_loras() - - def add_prompt_adapter( - self, prompt_adapter_request: PromptAdapterRequest) -> bool: - assert prompt_adapter_request.prompt_adapter_id > 0, \ - "prompt_adapter_id must be greater than 0." - return self.driver_worker.add_prompt_adapter(prompt_adapter_request) - - def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool: - assert prompt_adapter_id > 0, \ - "prompt_adapter_id must be greater than 0." - return self.driver_worker.remove_prompt_adapter(prompt_adapter_id) - - def pin_prompt_adapter(self, prompt_adapter_id: int) -> bool: - assert prompt_adapter_id > 0, \ - "prompt_adapter_id must be greater than 0." - return self.driver_worker.pin_prompt_adapter(prompt_adapter_id) - - def list_prompt_adapters(self) -> Set[int]: - return self.driver_worker.list_prompt_adapters() - - def check_health(self) -> None: - # GPUExecutor will always be healthy as long as - # it's running. - return - - def start_profile(self) -> None: - self.driver_worker.start_profile() - - def stop_profile(self) -> None: - self.driver_worker.stop_profile() - - -class GPUExecutorAsync(GPUExecutor, ExecutorAsyncBase): - - async def execute_model_async( - self, - execute_model_req: ExecuteModelRequest, - ) -> List[Union[SamplerOutput, PoolerOutput]]: - output = await make_async(self.driver_worker.execute_model - )(execute_model_req=execute_model_req) - return output diff --git a/vllm/executor/hpu_executor.py b/vllm/executor/hpu_executor.py deleted file mode 100644 index c9b7bfa71edfa..0000000000000 --- a/vllm/executor/hpu_executor.py +++ /dev/null @@ -1,202 +0,0 @@ -############################################################################### -# Copyright (C) 2024 Habana Labs, Ltd. an Intel Company -############################################################################### - -import contextlib -import os -from typing import Any, Dict, List, Optional, Set, Tuple - -from vllm.executor.executor_base import ExecutorAsyncBase, ExecutorBase -from vllm.logger import init_logger -from vllm.lora.request import LoRARequest -from vllm.model_executor.layers.sampler import SamplerOutput -from vllm.prompt_adapter.request import PromptAdapterRequest -from vllm.sequence import ExecuteModelRequest -from vllm.utils import (get_distributed_init_method, get_ip, get_open_port, - make_async) -from vllm.worker.worker_base import WorkerWrapperBase - -logger = init_logger(__name__) - - -class HPUExecutor(ExecutorBase): - - uses_ray: bool = False - - def _init_executor(self) -> None: - """Initialize the worker and load the model.""" - self._init_worker() - - def _get_worker_kwargs( - self, - local_rank: int = 0, - rank: int = 0, - distributed_init_method: Optional[str] = None) -> Dict[str, Any]: - """Return worker init args for a given rank.""" - if distributed_init_method is None: - distributed_init_method = get_distributed_init_method( - get_ip(), get_open_port()) - return dict( - vllm_config=self.vllm_config, - local_rank=local_rank, - rank=rank, - distributed_init_method=distributed_init_method, - is_driver_worker=rank == 0, - ) - - def _create_worker(self, - local_rank: int = 0, - rank: int = 0, - distributed_init_method: Optional[str] = None): - wrapper = WorkerWrapperBase(vllm_config=self.vllm_config) - wrapper.init_worker(**self._get_worker_kwargs(local_rank, rank, - distributed_init_method)) - return wrapper.worker - - def _init_worker(self): - assert self.parallel_config.world_size == 1, ( - "GPUExecutor only supports single GPU.") - - self.driver_worker = self._create_worker() - self.driver_worker.init_device() - self.driver_worker.load_model() - - def determine_num_available_blocks(self) -> Tuple[int, int]: - """Determine the number of available KV blocks by invoking the - underlying worker. - """ - return self.driver_worker.determine_num_available_blocks() - - def initialize_cache(self, num_gpu_blocks: int, num_cpu_blocks) -> None: - """Initialize the KV cache by invoking the underlying worker. - """ - # NOTE: This is logged in the executor because there can be >1 worker - # with other executors. We could log in the engine level, but work - # remains to abstract away the device for non-GPU configurations. - logger.info("# HPU blocks: %d, # CPU blocks: %d", num_gpu_blocks, - num_cpu_blocks) - from vllm_hpu_extension.profiler import HabanaMemoryProfiler - with HabanaMemoryProfiler() as cache_init_m: - self.driver_worker.initialize_cache(num_gpu_blocks, num_cpu_blocks) - msg = f"init_cache_engine took {cache_init_m.get_summary_string()}" - logger.info(msg) - - def finish_measurements(self): - self.driver_worker.finish_measurements() - - def execute_model( - self, - execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: - # VLLM_HPU_LOG_STEP_GRAPH_COMPILATION - will log graph compilations per engine step, only when there was any - highly recommended to use alongside PT_HPU_METRICS_GC_DETAILS! # noqa:E501 - # VLLM_HPU_LOG_STEP_GRAPH_COMPILATION_ALL - will log graph compilations per engine step, always, even if there were none # noqa:E501 - # VLLM_HPU_LOG_STEP_CPU_FALLBACKS - will log cpu fallbacks per engine step, only when there was any # noqa:E501 - # VLLM_HPU_LOG_STEP_CPU_FALLBACKS_ALL - will log cpu fallbacks per engine step, always, even if there were none # noqa:E501 - log_graph_compilation_all = os.environ.get( - 'VLLM_HPU_LOG_STEP_GRAPH_COMPILATION_ALL', '0') != '0' - log_graph_compilation = os.environ.get( - 'VLLM_HPU_LOG_STEP_GRAPH_COMPILATION', - '0') != '0' or log_graph_compilation_all - log_cpu_fallbacks_all = os.environ.get( - 'VLLM_HPU_LOG_STEP_CPU_FALLBACKS_ALL', '0') != '0' - log_cpu_fallbacks = os.environ.get('VLLM_HPU_LOG_STEP_CPU_FALLBACKS', - '0') != '0' or log_cpu_fallbacks_all - if log_graph_compilation or log_cpu_fallbacks: - from habana_frameworks.torch.hpu.metrics import metric_localcontext - seq_group_metadata_list = execute_model_req.seq_group_metadata_list - is_prompt = any([ - seq_group_metadata.is_prompt - for seq_group_metadata in seq_group_metadata_list - ]) - max_context_len = max([ - max([ - len(v.prompt_token_ids) + len(v.output_token_ids) - for v in seq_group_metadata.seq_data.values() - ]) for seq_group_metadata in seq_group_metadata_list - ]) # whoa, that's some spicy stuff right here - max_num_blocks = ( - (max_context_len - 1) // self.cache_config.block_size) + 1 - input_stats = (f'is_prompt: {is_prompt}, ' - f'num_seqs: {len(seq_group_metadata_list)}, ' - f'max_context_len: {max_context_len}, ' - f'max_num_blocks {max_num_blocks}') - gc_ctx = metric_localcontext( - "graph_compilation" - ) if log_graph_compilation else contextlib.nullcontext() - cpu_fallback_ctx = metric_localcontext( - "cpu_fallback" - ) if log_cpu_fallbacks else contextlib.nullcontext() - with gc_ctx as gc_local_metric, \ - cpu_fallback_ctx as cpu_fallback_local_metric: - output = self.driver_worker.execute_model(execute_model_req) - if (log_graph_compilation and gc_local_metric.stats()[0][1] > 0 - ) or log_graph_compilation_all: - msg = ("VLLM_HPU_STEP_GRAPH_COMPILATION: " - f"{gc_local_metric.stats()}, {input_stats}") - logger.warning(msg) - if (log_cpu_fallbacks and cpu_fallback_local_metric.stats()[0][1] > - 0) or log_cpu_fallbacks_all: - msg = ("VLLM_HPU_STEP_CPU_FALLBACK: " - f"{cpu_fallback_local_metric.stats()}, {input_stats}") - logger.warning(msg) - - return output - - output = self.driver_worker.execute_model(execute_model_req) - return output - - def add_lora(self, lora_request: LoRARequest) -> bool: - assert lora_request.lora_int_id > 0, "lora_id must be greater than 0." - return self.driver_worker.add_lora(lora_request) - - def remove_lora(self, lora_id: int) -> bool: - assert lora_id > 0, "lora_id must be greater than 0." - return self.driver_worker.remove_lora(lora_id) - - def pin_lora(self, lora_id: int) -> bool: - assert lora_id > 0, "lora_id must be greater than 0." - return self.driver_worker.pin_lora(lora_id) - - def list_loras(self) -> Set[int]: - return self.driver_worker.list_loras() - - def add_prompt_adapter( - self, prompt_adapter_request: PromptAdapterRequest) -> bool: - raise NotImplementedError( - "Prompt Adapter is not implemented for HPU backend.") - - def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool: - raise NotImplementedError( - "Prompt Adapter is not implemented for HPU backend.") - - def pin_prompt_adapter(self, prompt_adapter_id: int) -> bool: - raise NotImplementedError( - "Prompt Adapter is not implemented for HPU backend.") - - def list_prompt_adapters(self) -> Set[int]: - raise NotImplementedError( - "Prompt Adapter is not implemented for HPU backend.") - - def check_health(self) -> None: - # GPUExecutor will always be healthy as long as - # it's running. - return - - def start_profile(self) -> None: - self.driver_worker.start_profile() - - def stop_profile(self) -> None: - self.driver_worker.stop_profile() - - def shutdown(self) -> None: - self.driver_worker.shutdown_inc() - - -class HPUExecutorAsync(HPUExecutor, ExecutorAsyncBase): - - async def execute_model_async( - self, - execute_model_req: ExecuteModelRequest, - ) -> List[SamplerOutput]: - output = await make_async(self.driver_worker.execute_model - )(execute_model_req=execute_model_req, ) - return output diff --git a/vllm/executor/multiproc_gpu_executor.py b/vllm/executor/mp_distributed_executor.py similarity index 71% rename from vllm/executor/multiproc_gpu_executor.py rename to vllm/executor/mp_distributed_executor.py index fc58163cade64..78c86321d861d 100644 --- a/vllm/executor/multiproc_gpu_executor.py +++ b/vllm/executor/mp_distributed_executor.py @@ -1,11 +1,10 @@ import asyncio import os -from functools import partial -from typing import Any, List, Optional +from typing import Any, Callable, List, Optional, Union -from vllm.executor.distributed_gpu_executor import ( # yapf: disable - DistributedGPUExecutor, DistributedGPUExecutorAsync) -from vllm.executor.gpu_executor import create_worker +import cloudpickle + +from vllm.executor.executor_base import DistributedExecutorBase from vllm.executor.multiproc_worker_utils import ( ProcessWorkerWrapper, ResultHandler, WorkerMonitor, set_multiprocessing_worker_envs) @@ -13,19 +12,50 @@ from vllm.model_executor.layers.sampler import SamplerOutput from vllm.sequence import ExecuteModelRequest from vllm.utils import (_run_task_with_lock, cuda_device_count_stateless, - get_distributed_init_method, get_open_port, make_async, - update_environment_variables) + get_distributed_init_method, get_ip, get_open_port, + make_async, run_method, update_environment_variables) +from vllm.worker.worker_base import WorkerWrapperBase logger = init_logger(__name__) -class MultiprocessingGPUExecutor(DistributedGPUExecutor): - """Python multiprocessing-based multi-GPU executor""" +class MultiprocessingDistributedExecutor(DistributedExecutorBase): + """Python multiprocessing-based distributed executor""" uses_ray: bool = False + def _check_cuda(self) -> None: + """Check that the number of GPUs is sufficient for the parallel + configuration. Separate from _init_executor to reduce the number of + indented blocks. + """ + parallel_config = self.parallel_config + world_size = parallel_config.world_size + tensor_parallel_size = parallel_config.tensor_parallel_size + + cuda_device_count = cuda_device_count_stateless() + # Use confusing message for more common TP-only case. + if tensor_parallel_size > cuda_device_count: + raise RuntimeError( + f"please set tensor_parallel_size ({tensor_parallel_size}) " + f"to less than max local gpu count ({cuda_device_count})") + + if world_size > cuda_device_count: + raise RuntimeError( + f"please ensure that world_size ({world_size}) " + f"is less than than max local gpu count ({cuda_device_count})") + + # Set CUDA_VISIBLE_DEVICES for the driver, inherited by workers + if "CUDA_VISIBLE_DEVICES" not in os.environ: + update_environment_variables({ + "CUDA_VISIBLE_DEVICES": (",".join(map(str, range(world_size)))) + }) + def _init_executor(self) -> None: - self._check_executor_parameters() + + from vllm.platforms import current_platform + if current_platform.is_cuda_alike(): + self._check_cuda() # Create the parallel GPU workers. world_size = self.parallel_config.world_size @@ -55,15 +85,9 @@ def _init_executor(self) -> None: else: result_handler = ResultHandler() for rank in range(1, world_size): - worker = ProcessWorkerWrapper( - result_handler, - partial( - create_worker, - **self._get_worker_kwargs( - rank=rank, - local_rank=rank, - distributed_init_method=distributed_init_method, - ))) + worker = ProcessWorkerWrapper(result_handler, + WorkerWrapperBase, + self.vllm_config, rank) self.workers.append(worker) if rank % tensor_parallel_size == 0: self.tp_driver_workers.append(worker) @@ -77,32 +101,30 @@ def _init_executor(self) -> None: # Set up signal handlers to shutdown the executor cleanly # sometimes gc does not work well - self.driver_worker = self._create_worker( - distributed_init_method=distributed_init_method) + self.driver_worker = WorkerWrapperBase(self.vllm_config, 0) + + all_kwargs = [] + distributed_init_method = get_distributed_init_method( + get_ip(), get_open_port()) + for i in range(world_size): + local_rank = i + rank = i + kwargs = dict( + vllm_config=self.vllm_config, + local_rank=local_rank, + rank=rank, + distributed_init_method=distributed_init_method, + is_driver_worker=(not self.parallel_config) + or (rank % self.parallel_config.tensor_parallel_size == 0), + ) + all_kwargs.append(kwargs) + self._run_workers("init_worker", all_kwargs) self._run_workers("init_device") self._run_workers("load_model", max_concurrent_workers=self.parallel_config. max_parallel_loading_workers) - - def _check_executor_parameters(self): - world_size = self.parallel_config.world_size - tensor_parallel_size = self.parallel_config.tensor_parallel_size - - # Set CUDA_VISIBLE_DEVICES for the driver, inherited by workers - if "CUDA_VISIBLE_DEVICES" not in os.environ: - update_environment_variables({ - "CUDA_VISIBLE_DEVICES": (",".join(map(str, range(world_size)))) - }) - - cuda_device_count = cuda_device_count_stateless() - # Use confusing message for more common TP-only case. - assert tensor_parallel_size <= cuda_device_count, ( - f"please set tensor_parallel_size ({tensor_parallel_size}) " - f"to less than max local gpu count ({cuda_device_count})") - - assert world_size <= cuda_device_count, ( - f"please ensure that world_size ({world_size}) " - f"is less than than max local gpu count ({cuda_device_count})") + self.driver_exec_model = make_async(self.driver_worker.execute_model) + self.pp_locks: Optional[List[asyncio.Lock]] = None def shutdown(self): if (worker_monitor := getattr(self, "worker_monitor", @@ -121,12 +143,12 @@ def _driver_execute_model( def _run_workers( self, - method: str, + method: Union[str, Callable], *args, async_run_tensor_parallel_workers_only: bool = False, max_concurrent_workers: Optional[int] = None, **kwargs, - ) -> Any: + ) -> List[Any]: """Runs the given method on all workers. Args: @@ -135,6 +157,11 @@ def _run_workers( It will also be run asynchronously and return a list of futures rather than blocking on the results. """ + if isinstance(method, str): + sent_method = method + else: + sent_method = cloudpickle.dumps(method) + del method if max_concurrent_workers: raise NotImplementedError( @@ -143,18 +170,18 @@ def _run_workers( if async_run_tensor_parallel_workers_only: # Run only non-driver workers and just return futures. return [ - worker.execute_method(method, *args, **kwargs) + worker.execute_method(sent_method, *args, **kwargs) for worker in self.non_driver_workers ] # Start all remote workers first. worker_outputs = [ - worker.execute_method(method, *args, **kwargs) + worker.execute_method(sent_method, *args, **kwargs) for worker in self.workers ] - driver_worker_method = getattr(self.driver_worker, method) - driver_worker_output = driver_worker_method(*args, **kwargs) + driver_worker_output = run_method(self.driver_worker, sent_method, + args, kwargs) # Get the results of the workers. return [driver_worker_output @@ -172,15 +199,6 @@ def _wait_for_tasks_completion(self, parallel_worker_tasks: Any) -> None: for result in parallel_worker_tasks: result.get() - -class MultiprocessingGPUExecutorAsync(MultiprocessingGPUExecutor, - DistributedGPUExecutorAsync): - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - self.driver_exec_model = make_async(self.driver_worker.execute_model) - self.pp_locks: Optional[List[asyncio.Lock]] = None - async def _driver_execute_model_async( self, execute_model_req: Optional[ExecuteModelRequest] = None diff --git a/vllm/executor/multiproc_worker_utils.py b/vllm/executor/multiproc_worker_utils.py index bc32826529eef..539b6ae2d3572 100644 --- a/vllm/executor/multiproc_worker_utils.py +++ b/vllm/executor/multiproc_worker_utils.py @@ -12,9 +12,10 @@ import torch +from vllm.config import VllmConfig from vllm.logger import init_logger from vllm.triton_utils.importing import HAS_TRITON -from vllm.utils import _check_multiproc_method, get_mp_context +from vllm.utils import _check_multiproc_method, get_mp_context, run_method if HAS_TRITON: from vllm.triton_utils import maybe_set_triton_cache_manager @@ -147,7 +148,8 @@ class ProcessWorkerWrapper: for handling single-node multi-GPU tensor parallel.""" def __init__(self, result_handler: ResultHandler, - worker_factory: Callable[[], Any]) -> None: + worker_factory: Callable[[VllmConfig, int], Any], + vllm_config: VllmConfig, rank: int) -> None: self.mp = get_mp_context() self._task_queue = self.mp.Queue() self.result_queue = result_handler.result_queue @@ -159,13 +161,15 @@ def __init__(self, result_handler: ResultHandler, worker_factory=worker_factory, task_queue=self._task_queue, result_queue=self.result_queue, + vllm_config=vllm_config, + rank=rank, ), daemon=True) self.process.start() def _enqueue_task(self, future: Union[ResultFuture, asyncio.Future], - method: str, args, kwargs): + method: Union[str, bytes], args, kwargs): task_id = uuid.uuid4() self.tasks[task_id] = future try: @@ -176,12 +180,13 @@ def _enqueue_task(self, future: Union[ResultFuture, asyncio.Future], del self.tasks[task_id] raise ChildProcessError("worker died") from e - def execute_method(self, method: str, *args, **kwargs): + def execute_method(self, method: Union[str, bytes], *args, **kwargs): future: ResultFuture = ResultFuture() self._enqueue_task(future, method, args, kwargs) return future - async def execute_method_async(self, method: str, *args, **kwargs): + async def execute_method_async(self, method: Union[str, bytes], *args, + **kwargs): future = asyncio.get_running_loop().create_future() self._enqueue_task(future, method, args, kwargs) return await future @@ -199,9 +204,11 @@ def kill_worker(self): def _run_worker_process( - worker_factory: Callable[[], Any], + worker_factory: Callable[[VllmConfig, int], Any], task_queue: Queue, result_queue: Queue, + vllm_config: VllmConfig, + rank: int, ) -> None: """Worker process event loop""" @@ -212,7 +219,7 @@ def _run_worker_process( _add_prefix(sys.stderr, process_name, pid) # Initialize worker - worker = worker_factory() + worker = worker_factory(vllm_config, rank) del worker_factory # Accept tasks from the engine in task_queue @@ -224,8 +231,7 @@ def _run_worker_process( exception = None task_id, method, args, kwargs = items try: - executor = getattr(worker, method) - output = executor(*args, **kwargs) + output = run_method(worker, method, args, kwargs) except SystemExit: raise except KeyboardInterrupt: diff --git a/vllm/executor/multiproc_xpu_executor.py b/vllm/executor/multiproc_xpu_executor.py deleted file mode 100644 index a66afbf939ef0..0000000000000 --- a/vllm/executor/multiproc_xpu_executor.py +++ /dev/null @@ -1,26 +0,0 @@ -import vllm.envs as envs -from vllm.executor.multiproc_gpu_executor import ( - MultiprocessingGPUExecutor, MultiprocessingGPUExecutorAsync) -from vllm.executor.xpu_executor import XPUExecutor -from vllm.logger import init_logger -from vllm.utils import make_async - -logger = init_logger(__name__) - - -class MultiprocessingXPUExecutor(MultiprocessingGPUExecutor, XPUExecutor): - """Python multiprocessing-based multi-XPU executor""" - - def _check_executor_parameters(self): - mp_method = envs.VLLM_WORKER_MULTIPROC_METHOD - if mp_method != "spawn": - raise RuntimeError( - "XPU multiprocess executor only support spawn as mp method") - - -class MultiprocessingXPUExecutorAsync(MultiprocessingXPUExecutor, - MultiprocessingGPUExecutorAsync): - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - self.driver_exec_model = make_async(self.driver_worker.execute_model) diff --git a/vllm/executor/neuron_executor.py b/vllm/executor/neuron_executor.py deleted file mode 100644 index a9efc4f9a801c..0000000000000 --- a/vllm/executor/neuron_executor.py +++ /dev/null @@ -1,114 +0,0 @@ -from typing import List, Set, Tuple - -from vllm.executor.executor_base import ExecutorAsyncBase, ExecutorBase -from vllm.logger import init_logger -from vllm.lora.request import LoRARequest -from vllm.model_executor.layers.sampler import SamplerOutput -from vllm.sequence import ExecuteModelRequest -from vllm.utils import (get_distributed_init_method, get_ip, get_open_port, - make_async) -from vllm.worker.worker_base import WorkerWrapperBase - -logger = init_logger(__name__) - - -class NeuronExecutor(ExecutorBase): - - uses_ray: bool = False - - def _init_executor(self) -> None: - assert (self.lora_config is - None), "LoRA is not supported for Neuron backend." - assert (not self.speculative_config - ), "Speculative decoding not yet supported for Neuron backend." - - # Instantiate the worker and load the model to the device. - self._init_worker() - - def _init_worker(self): - wrapper = WorkerWrapperBase(vllm_config=self.vllm_config) - distributed_init_method = get_distributed_init_method( - get_ip(), get_open_port()) - wrapper.init_worker( - vllm_config=self.vllm_config, - local_rank=0, - rank=0, - distributed_init_method=distributed_init_method, - ) - self.driver_worker = wrapper.worker - self.driver_worker.init_device() - self.driver_worker.load_model() - - def determine_num_available_blocks(self) -> Tuple[int, int]: - """Determine the number of available KV blocks by invoking the - underlying worker. - """ - return self.driver_worker.determine_num_available_blocks() - - def initialize_cache(self, num_gpu_blocks: int, - num_cpu_blocks: int) -> None: - """Initialize the KV cache by invoking the underlying worker. - """ - self.driver_worker.initialize_cache(num_gpu_blocks, num_cpu_blocks) - - def execute_model( - self, - execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: - assert (not execute_model_req.blocks_to_swap_in - and not execute_model_req.blocks_to_swap_out - and not execute_model_req.blocks_to_copy), ( - "Cache operations are not supported for Neuron backend.") - assert execute_model_req.num_lookahead_slots == 0, ( - "lookahead not supported for Neuron backend.") - - output = self.driver_worker.execute_model(execute_model_req) - return output - - def add_lora(self, lora_request: LoRARequest) -> bool: - return self.driver_worker.add_lora(lora_request) - - def remove_lora(self, lora_id: int) -> bool: - return self.driver_worker.remove_lora(lora_id) - - def pin_lora(self, lora_id: int) -> bool: - return self.driver_worker.pin_lora(lora_id) - - def list_loras(self) -> Set[int]: - return self.driver_worker.list_loras() - - def add_prompt_adapter(self, prompt_adapter_request) -> bool: - raise NotImplementedError( - "Soft prompt is currently not supported by the Neuron backend.") - - def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool: - raise NotImplementedError( - "Soft prompt is currently not supported by the Neuron backend.") - - def pin_prompt_adapter(self, prompt_adapter_id: int) -> bool: - raise NotImplementedError( - "Soft prompt is currently not supported by the Neuron backend.") - - def list_prompt_adapters(self) -> Set[int]: - raise NotImplementedError( - "Soft prompt is currently not supported by the Neuron backend.") - - def check_health(self) -> None: - # NeuronExecutor will always be healthy as long as - # it's running. - return - - -class NeuronExecutorAsync(NeuronExecutor, ExecutorAsyncBase): - - async def execute_model_async( - self, - execute_model_req: ExecuteModelRequest, - ) -> List[SamplerOutput]: - output = await make_async(self.driver_worker.execute_model - )(execute_model_req=execute_model_req, ) - return output - - async def check_health_async(self) -> None: - # NeuronExecutor will always be healthy as long as - # it's running. - return diff --git a/vllm/executor/openvino_executor.py b/vllm/executor/openvino_executor.py deleted file mode 100644 index 057a32364e512..0000000000000 --- a/vllm/executor/openvino_executor.py +++ /dev/null @@ -1,125 +0,0 @@ -from typing import List, Set, Tuple - -import openvino as ov - -import vllm.envs as envs -from vllm.executor.executor_base import ExecutorAsyncBase, ExecutorBase -from vllm.logger import init_logger -from vllm.lora.request import LoRARequest -from vllm.model_executor.layers.sampler import SamplerOutput -from vllm.platforms import current_platform -from vllm.sequence import ExecuteModelRequest -from vllm.utils import (get_distributed_init_method, get_ip, get_open_port, - make_async) -from vllm.worker.worker_base import WorkerWrapperBase - -logger = init_logger(__name__) - - -class OpenVINOExecutor(ExecutorBase): - - uses_ray: bool = False - - def _init_executor(self) -> None: - assert self.device_config.device_type == "openvino" - assert self.lora_config is None, "OpenVINO backend doesn't support LoRA" - assert current_platform.is_openvino_cpu() or \ - current_platform.is_openvino_gpu(), \ - "OpenVINO backend supports only CPU and GPU devices" - - # Instantiate the worker and load the model to CPU. - self._init_worker() - - def _init_worker(self): - - wrapper = WorkerWrapperBase(vllm_config=self.vllm_config) - - distributed_init_method = get_distributed_init_method( - get_ip(), get_open_port()) - wrapper.init_worker( - ov_core=ov.Core(), - vllm_config=self.vllm_config, - local_rank=0, - rank=0, - distributed_init_method=distributed_init_method, - kv_cache_dtype=self.cache_config.cache_dtype, - is_driver_worker=True, - ) - self.driver_worker = wrapper.worker - self.driver_worker.init_device() - self.driver_worker.load_model() - - def determine_num_available_blocks(self) -> Tuple[int, int]: - """Determine the number of available KV blocks by invoking the - underlying worker. - """ - return self.driver_worker.determine_num_available_blocks() - - def initialize_cache(self, num_gpu_blocks: int, - num_cpu_blocks: int) -> None: - """Initialize the KV cache by invoking the underlying worker.""" - # NOTE: We log here to avoid multiple logs when number of workers is - # greater than one. We could log in the engine, but not all executors - # have GPUs. - # NOTE: In case of a CPU device, `cpu block` for OpenVINO backend - # is located on CPU memory but is referred as `gpu block`. - # Because we want to reuse the existing block management procedure. - device_blocks = num_gpu_blocks - swap_blocks = num_cpu_blocks - logger.info("OpenVINO %s: # device blocks: %d; # swap blocks: %d", - envs.VLLM_OPENVINO_DEVICE, device_blocks, swap_blocks) - self.driver_worker.initialize_cache(num_gpu_blocks, num_cpu_blocks) - - def execute_model( - self, - execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: - output = self.driver_worker.execute_model(execute_model_req) - return output - - def add_lora(self, lora_request: LoRARequest) -> bool: - return self.driver_worker.add_lora(lora_request) - - def remove_lora(self, lora_id: int) -> bool: - return self.driver_worker.remove_lora(lora_id) - - def pin_lora(self, lora_id: int) -> bool: - return self.driver_worker.pin_lora(lora_id) - - def list_loras(self) -> Set[int]: - return self.driver_worker.list_loras() - - def add_prompt_adapter(self, prompt_adapter_request) -> bool: - raise NotImplementedError( - "Soft prompt is currently not supported by the OPENVINO backend.") - - def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool: - raise NotImplementedError( - "Soft prompt is currently not supported by the OPENVINO backend.") - - def pin_prompt_adapter(self, prompt_adapter_id: int) -> bool: - raise NotImplementedError( - "Soft prompt is currently not supported by the OPENVINO backend.") - - def list_prompt_adapters(self) -> Set[int]: - raise NotImplementedError( - "Soft prompt is currently not supported by the OPENVINO backend.") - - def check_health(self) -> None: - # OpenVINOExecutor will always be healthy as long as - # it's running. - return - - -class OpenVINOExecutorAsync(OpenVINOExecutor, ExecutorAsyncBase): - - async def execute_model_async( - self, - execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: - output = await make_async(self.driver_worker.execute_model - )(execute_model_req=execute_model_req, ) - return output - - async def check_health_async(self) -> None: - # OpenVINOExecutor will always be healthy as long as - # it's running. - return diff --git a/vllm/executor/ray_gpu_executor.py b/vllm/executor/ray_distributed_executor.py similarity index 77% rename from vllm/executor/ray_gpu_executor.py rename to vllm/executor/ray_distributed_executor.py index e2c549cbd5331..2afd99f99b353 100644 --- a/vllm/executor/ray_gpu_executor.py +++ b/vllm/executor/ray_distributed_executor.py @@ -1,24 +1,30 @@ import asyncio import os from collections import defaultdict -from itertools import islice, repeat -from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple +from dataclasses import dataclass +from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Union +import cloudpickle import msgspec import vllm.envs as envs -from vllm.executor.distributed_gpu_executor import ( # yapf: disable - DistributedGPUExecutor, DistributedGPUExecutorAsync) +from vllm.executor.executor_base import ( + DistributedExecutorBase) # yapf: disable from vllm.executor.msgspec_utils import encode_hook -from vllm.executor.ray_utils import RayWorkerWrapper, ray +from vllm.executor.ray_utils import (RayWorkerWrapper, initialize_ray_cluster, + ray) from vllm.logger import init_logger from vllm.model_executor.layers.sampler import SamplerOutput +from vllm.platforms import current_platform from vllm.sequence import ExecuteModelRequest from vllm.utils import (_run_task_with_lock, get_distributed_init_method, get_ip, get_open_port, make_async) if ray is not None: + from ray.actor import ActorHandle from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy +else: + ActorHandle = None if TYPE_CHECKING: from ray.util.placement_group import PlacementGroup @@ -26,12 +32,29 @@ logger = init_logger(__name__) -class RayGPUExecutor(DistributedGPUExecutor): +@dataclass +class RayWorkerMetaData: + """ + Metadata for a Ray worker. + The order of ray worker creation can be random, + and we need to reset the rank after creating all workers. + """ + worker: ActorHandle + created_rank: int + adjusted_rank: int = -1 + ip: str = "" + + +class RayDistributedExecutor(DistributedExecutorBase): uses_ray: bool = True def _init_executor(self) -> None: self.forward_dag: Optional[ray.dag.CompiledDAG] = None + if envs.VLLM_USE_V1: + # v1 always uses the compiled DAG and SPMD worker. + os.environ["VLLM_USE_RAY_SPMD_WORKER"] = "1" + os.environ["VLLM_USE_RAY_COMPILED_DAG"] = "1" # If the env var is set, it uses the Ray's compiled DAG API # which optimizes the control plane overhead. # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it. @@ -53,6 +76,7 @@ def _init_executor(self) -> None: "VLLM_USE_RAY_COMPILED_DAG=1") assert self.uses_ray + initialize_ray_cluster(self.parallel_config) placement_group = self.parallel_config.placement_group # Disable Ray usage stats collection. @@ -66,6 +90,13 @@ def _init_executor(self) -> None: self.input_encoder = msgspec.msgpack.Encoder(enc_hook=encode_hook) self.output_decoder = msgspec.msgpack.Decoder( Optional[List[SamplerOutput]]) + self.use_v1 = envs.VLLM_USE_V1 + + self.pp_locks: Optional[List[asyncio.Lock]] = None + self.use_ray_spmd_worker = envs.VLLM_USE_RAY_SPMD_WORKER + if not self.use_ray_compiled_dag: + self.driver_exec_method = make_async( + self.driver_worker.execute_method) def shutdown(self) -> None: if hasattr(self, "forward_dag") and self.forward_dag is not None: @@ -123,9 +154,10 @@ def _init_workers_ray(self, placement_group: "PlacementGroup", # Create the workers. driver_ip = get_ip() - workers = [] + rank = 0 + worker_metadata: List[RayWorkerMetaData] = [] for bundle_id, bundle in enumerate(placement_group.bundle_specs): - if not bundle.get("GPU", 0): + if not bundle.get(current_platform.ray_device_key, 0): continue scheduling_strategy = PlacementGroupSchedulingStrategy( placement_group=placement_group, @@ -133,38 +165,51 @@ def _init_workers_ray(self, placement_group: "PlacementGroup", placement_group_bundle_index=bundle_id, ) - worker = ray.remote( - num_cpus=0, - num_gpus=num_gpus, - scheduling_strategy=scheduling_strategy, - **ray_remote_kwargs, - )(RayWorkerWrapper).remote(vllm_config=self.vllm_config) - workers.append(worker) - - worker_ip_refs = [ - worker.get_node_ip.remote() # type: ignore[attr-defined] - for worker in workers - ] - worker_ips = ray.get(worker_ip_refs) + if current_platform.ray_device_key == "GPU": + # NV+AMD GPUs, and Intel XPUs + worker = ray.remote( + num_cpus=0, + num_gpus=num_gpus, + scheduling_strategy=scheduling_strategy, + **ray_remote_kwargs, + )(RayWorkerWrapper).remote(vllm_config=self.vllm_config, + rpc_rank=rank) + else: + worker = ray.remote( + num_cpus=0, + num_gpus=0, + resources={current_platform.ray_device_key: num_gpus}, + scheduling_strategy=scheduling_strategy, + **ray_remote_kwargs, + )(RayWorkerWrapper).remote(vllm_config=self.vllm_config, + rpc_rank=rank) + worker_metadata.append( + RayWorkerMetaData(worker=worker, created_rank=rank)) + rank += 1 + + worker_ips = ray.get([ + each.worker.get_node_ip.remote() # type: ignore[attr-defined] + for each in worker_metadata + ]) + + for each, ip in zip(worker_metadata, worker_ips): + each.ip = ip if not self.use_ray_spmd_worker: - for i in range(len(workers)): - worker = workers[i] - worker_ip = worker_ips[i] + for i, each in enumerate(worker_metadata): + # find and remove the dummy worker from the list + worker = each.worker + worker_ip = each.ip if self.driver_dummy_worker is None and worker_ip == driver_ip: # If the worker is on the same node as the driver, we use it # as the resource holder for the driver process. self.driver_dummy_worker = worker self.driver_worker = RayWorkerWrapper( - vllm_config=self.vllm_config) - workers.pop(i) - worker_ips.pop(i) - self.workers = workers + vllm_config=self.vllm_config, rpc_rank=0) + worker_metadata.pop(i) break - else: - self.workers = workers - logger.debug("workers: %s", self.workers) + logger.debug("workers: %s", worker_metadata) logger.debug("driver_dummy_worker: %s", self.driver_dummy_worker) if not self.use_ray_spmd_worker and self.driver_dummy_worker is None: raise ValueError( @@ -176,9 +221,7 @@ def _init_workers_ray(self, placement_group: "PlacementGroup", for ip in worker_ips: ip_counts[ip] = ip_counts.get(ip, 0) + 1 - worker_to_ip = dict(zip(self.workers, worker_ips)) - - def sort_by_driver_then_worker_ip(worker): + def sort_by_driver_then_worker_ip(item: RayWorkerMetaData): """ Sort the workers based on 3 properties: 1. If the worker is on the same node as the driver (vllm engine), @@ -188,13 +231,23 @@ def sort_by_driver_then_worker_ip(worker): 3. Finally, if the work is on a node with smaller IP address, it should be placed first. """ - ip = worker_to_ip[worker] - return (ip != driver_ip, ip_counts[ip], ip) + ip = item.ip + return (0 if ip == driver_ip else 1, ip_counts[ip], ip) # After sorting, the workers on the same node will be # close to each other, and the workers on the driver # node will be placed first. - self.workers = sorted(self.workers, key=sort_by_driver_then_worker_ip) + sorted_worker_metadata = sorted(worker_metadata, + key=sort_by_driver_then_worker_ip) + start_rank = 0 if self.use_ray_spmd_worker else 1 + for i, item in enumerate(sorted_worker_metadata): + item.adjusted_rank = i + start_rank + self.workers = [item.worker for item in sorted_worker_metadata] + rerank_mapping = { + item.created_rank: item.adjusted_rank + for item in sorted_worker_metadata + } + self._run_workers("adjust_rank", rerank_mapping) # Get the set of GPU IDs used on each node. worker_node_and_gpu_ids = [] @@ -235,21 +288,29 @@ def sort_by_driver_then_worker_ip(worker): " each node.") # Set environment variables for the driver and workers. - all_args_to_update_environment_variables = [({ - "CUDA_VISIBLE_DEVICES": + all_args_to_update_environment_variables = [{ + current_platform.device_control_env_var: ",".join(map(str, node_gpus[node_id])), - "VLLM_TRACE_FUNCTION": - str(envs.VLLM_TRACE_FUNCTION), - **({ - "VLLM_ATTENTION_BACKEND": envs.VLLM_ATTENTION_BACKEND - } if envs.VLLM_ATTENTION_BACKEND is not None else {}) - }, ) for (node_id, _) in worker_node_and_gpu_ids] + } for (node_id, _) in worker_node_and_gpu_ids] + + for args in all_args_to_update_environment_variables: + # some carry-over env vars from the driver + # TODO: refactor platform-specific env vars + for name in [ + "VLLM_ATTENTION_BACKEND", + "TPU_CHIPS_PER_HOST_BOUNDS", + "TPU_HOST_BOUNDS", + "VLLM_USE_V1", + "VLLM_TRACE_FUNCTION", + ]: + if name in os.environ: + args[name] = os.environ[name] self._env_vars_for_all_workers = ( all_args_to_update_environment_variables) self._run_workers("update_environment_variables", - all_args=self._get_env_vars_to_be_updated()) + self._get_env_vars_to_be_updated()) if len(node_gpus) == 1: # in single node case, we don't need to get the IP address. @@ -265,14 +326,19 @@ def sort_by_driver_then_worker_ip(worker): driver_ip, get_open_port()) # Initialize the actual workers inside worker wrapper. - init_worker_all_kwargs = [ - self._get_worker_kwargs( - local_rank=node_workers[node_id].index(rank), + all_kwargs = [] + for rank, (node_id, _) in enumerate(worker_node_and_gpu_ids): + local_rank = node_workers[node_id].index(rank) + kwargs = dict( + vllm_config=self.vllm_config, + local_rank=local_rank, rank=rank, distributed_init_method=distributed_init_method, - ) for rank, (node_id, _) in enumerate(worker_node_and_gpu_ids) - ] - self._run_workers("init_worker", all_kwargs=init_worker_all_kwargs) + is_driver_worker=(not self.parallel_config) + or (rank % self.parallel_config.tensor_parallel_size == 0), + ) + all_kwargs.append(kwargs) + self._run_workers("init_worker", all_kwargs) self._run_workers("init_device") self._run_workers("load_model", @@ -332,18 +398,22 @@ def execute_model( if self.forward_dag is None: self.forward_dag = self._compiled_ray_dag(enable_asyncio=False) - serialized_data = self.input_encoder.encode(execute_model_req) + if self.use_v1: + serialized_data = execute_model_req + else: + serialized_data = self.input_encoder.encode(execute_model_req) outputs = ray.get(self.forward_dag.execute(serialized_data)) - output = self.output_decoder.decode(outputs[0]) + if self.use_v1: + output = outputs[0] + else: + output = self.output_decoder.decode(outputs[0]) return output def _run_workers( self, - method: str, + method: Union[str, Callable], *args, async_run_tensor_parallel_workers_only: bool = False, - all_args: Optional[List[Tuple[Any, ...]]] = None, - all_kwargs: Optional[List[Dict[str, Any]]] = None, max_concurrent_workers: Optional[int] = None, **kwargs, ) -> Any: @@ -356,9 +426,12 @@ def _run_workers( It will also be run asynchronously and return a list of futures rather than blocking on the results. - args/kwargs: All workers share the same args/kwargs - - all_args/all_kwargs: args/kwargs for each worker are specified - individually """ + if isinstance(method, str): + sent_method = method + else: + sent_method = cloudpickle.dumps(method) + del method if self.use_ray_spmd_worker: assert not async_run_tensor_parallel_workers_only, ( "async_run_tensor_parallel_workers_only is not supported for " @@ -368,26 +441,13 @@ def _run_workers( raise NotImplementedError( "max_concurrent_workers is not supported yet.") - count = len(self.workers) if not \ - async_run_tensor_parallel_workers_only \ - else len(self.non_driver_workers) - # If using SPMD worker, all workers are the same, so we should execute - # the args on all workers. Otherwise, we skip the first worker's args - # because those args will go to the driver worker. - first_worker_args_index: int = 0 if self.use_ray_spmd_worker else 1 - all_worker_args = repeat(args, count) if all_args is None \ - else islice(all_args, first_worker_args_index, None) - all_worker_kwargs = repeat(kwargs, count) if all_kwargs is None \ - else islice(all_kwargs, first_worker_args_index, None) - # Start the ray workers first. ray_workers = self.workers if async_run_tensor_parallel_workers_only: ray_workers = self.non_driver_workers ray_worker_outputs = [ - worker.execute_method.remote(method, *worker_args, **worker_kwargs) - for (worker, worker_args, worker_kwargs - ) in zip(ray_workers, all_worker_args, all_worker_kwargs) + worker.execute_method.remote(sent_method, *args, **kwargs) + for worker in ray_workers ] if async_run_tensor_parallel_workers_only: @@ -399,13 +459,9 @@ def _run_workers( # so we only explicitly execute on the driver worker if using a # non-SPMD worker class. if not self.use_ray_spmd_worker: - driver_args = args if all_args is None else all_args[0] - driver_kwargs = kwargs if all_kwargs is None else all_kwargs[0] - # Start the driver worker after all the ray workers. driver_worker_output = [ - self.driver_worker.execute_method(method, *driver_args, - **driver_kwargs) + self.driver_worker.execute_method(sent_method, *args, **kwargs) ] # Get the results of the ray workers. @@ -467,11 +523,18 @@ def _compiled_ray_dag(self, enable_asyncio: bool): for pp_rank, tp_group in enumerate(self.pp_tp_workers): # Each PP worker takes in the output of the previous PP worker, # and the TP group executes in SPMD fashion. - outputs = [ - worker.execute_model_spmd. - bind( # type: ignore[attr-defined] - outputs[i]) for i, worker in enumerate(tp_group) - ] + if self.use_v1: + outputs = [ + worker.execute_model. + bind( # type: ignore[attr-defined] + outputs[i]) for i, worker in enumerate(tp_group) + ] + else: + outputs = [ + worker.execute_model_spmd. + bind( # type: ignore[attr-defined] + outputs[i]) for i, worker in enumerate(tp_group) + ] last_pp_rank = len(self.pp_tp_workers) - 1 if pp_rank < last_pp_rank: @@ -497,17 +560,6 @@ def _compiled_ray_dag(self, enable_asyncio: bool): def __del__(self): self.shutdown() - -class RayGPUExecutorAsync(RayGPUExecutor, DistributedGPUExecutorAsync): - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - self.pp_locks: Optional[List[asyncio.Lock]] = None - self.use_ray_spmd_worker = envs.VLLM_USE_RAY_SPMD_WORKER - if not self.use_ray_compiled_dag: - self.driver_exec_method = make_async( - self.driver_worker.execute_method) - async def execute_model_async( self, execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: @@ -568,5 +620,7 @@ async def _start_worker_execution_loop(self): ] return await asyncio.gather(*coros) - def __del__(self): - self.shutdown() + def check_health(self) -> None: + # Assume that the Ray workers are healthy. + # TODO: check the health of the Ray workers + return diff --git a/vllm/executor/ray_hpu_executor.py b/vllm/executor/ray_hpu_executor.py deleted file mode 100644 index f3025cb537ab8..0000000000000 --- a/vllm/executor/ray_hpu_executor.py +++ /dev/null @@ -1,515 +0,0 @@ -import asyncio -import os -from collections import defaultdict -from itertools import islice, repeat -from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple - -import msgspec - -import vllm.envs as envs -from vllm.executor.distributed_gpu_executor import ( # yapf: disable - DistributedGPUExecutor, DistributedGPUExecutorAsync) -from vllm.executor.msgspec_utils import encode_hook -from vllm.executor.ray_utils import RayWorkerWrapper, ray -from vllm.logger import init_logger -from vllm.model_executor.layers.sampler import SamplerOutput -from vllm.sequence import ExecuteModelRequest -from vllm.utils import (_run_task_with_lock, get_distributed_init_method, - get_ip, get_open_port, make_async) - -if ray is not None: - from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy - -if TYPE_CHECKING: - from ray.util.placement_group import PlacementGroup - -logger = init_logger(__name__) - - -class RayHPUExecutor(DistributedGPUExecutor): - - uses_ray: bool = True - - def _init_executor(self) -> None: - self.forward_dag: Optional[ray.dag.CompiledDAG] = None - # If the env var is set, it uses the Ray's compiled DAG API - # which optimizes the control plane overhead. - # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it. - # Currently, this requires USE_RAY_SPMD_WORKER=True. - self.use_ray_compiled_dag = envs.VLLM_USE_RAY_COMPILED_DAG - # If the env var is set, then we do not distinguish between the - # "driver worker" vs other workers. Also, the rank 0 worker will - # be executed in a remote Ray worker. Currently this requires - # USE_RAY_COMPILED_DAG=True. - self.use_ray_spmd_worker = envs.VLLM_USE_RAY_SPMD_WORKER - if self.use_ray_compiled_dag: - assert self.use_ray_spmd_worker, ( - "VLLM_USE_RAY_COMPILED_DAG=1 requires " - "VLLM_USE_RAY_SPMD_WORKER=1") - if self.use_ray_spmd_worker: - # TODO: Support SPMD worker for non-DAG Ray executor. - assert self.use_ray_compiled_dag, ( - "VLLM_USE_RAY_SPMD_WORKER=1 requires " - "VLLM_USE_RAY_COMPILED_DAG=1") - - assert self.uses_ray - placement_group = self.parallel_config.placement_group - - # Disable Ray usage stats collection. - ray_usage = os.environ.get("RAY_USAGE_STATS_ENABLED", "0") - if ray_usage != "1": - os.environ["RAY_USAGE_STATS_ENABLED"] = "0" - - # Create the parallel GPU workers. - self._init_workers_ray(placement_group) - - self.input_encoder = msgspec.msgpack.Encoder(enc_hook=encode_hook) - self.output_decoder = msgspec.msgpack.Decoder( - Optional[List[SamplerOutput]]) - - def shutdown(self) -> None: - if hasattr(self, "forward_dag") and self.forward_dag is not None: - self.forward_dag.teardown() - import ray - for worker in self.workers: - ray.kill(worker) - self.forward_dag = None - - def finish_measurements(self): - self._run_workers("finish_measurements") - - def _init_workers_ray(self, placement_group: "PlacementGroup", - **ray_remote_kwargs): - # Otherwise, the ray workers are allocated with a full GPU. - num_gpus = 1 - - # The driver dummy worker does not actually use any resources. - # It holds the resource for the driver worker. - self.driver_dummy_worker: Optional[RayWorkerWrapper] = None - # The remaining workers are the actual ray actors. - self.workers: List[RayWorkerWrapper] = [] - - # Used in ray compiled DAG: indexed first by PP rank, - # and then TP rank. In other words, the inner list is - # the TP group of workers for a PP rank. - self.pp_tp_workers: List[List[RayWorkerWrapper]] = [] - - logger.info("use_ray_spmd_worker: %s", self.use_ray_spmd_worker) - - # Create the workers. - driver_ip = get_ip() - for bundle_id, bundle in enumerate(placement_group.bundle_specs): - if not bundle.get("HPU", 0): - continue - scheduling_strategy = PlacementGroupSchedulingStrategy( - placement_group=placement_group, - placement_group_capture_child_tasks=True, - placement_group_bundle_index=bundle_id, - ) - - worker = ray.remote( - num_cpus=0, - num_gpus=0, - resources={'HPU': num_gpus}, - scheduling_strategy=scheduling_strategy, - **ray_remote_kwargs, - )(RayWorkerWrapper).remote(vllm_config=self.vllm_config) - - if self.use_ray_spmd_worker: - self.workers.append(worker) - else: - worker_ip = ray.get(worker.get_node_ip.remote()) - if worker_ip == driver_ip and self.driver_dummy_worker is None: - # If the worker is on the same node as the driver, we use it - # as the resource holder for the driver process. - self.driver_dummy_worker = worker - self.driver_worker = RayWorkerWrapper( - vllm_config=self.vllm_config) - else: - # Else, added to the list of workers. - self.workers.append(worker) - - logger.debug("workers: %s", self.workers) - logger.debug("driver_dummy_worker: %s", self.driver_dummy_worker) - if not self.use_ray_spmd_worker and self.driver_dummy_worker is None: - raise ValueError( - "Ray does not allocate any GPUs on the driver node. Consider " - "adjusting the Ray placement group or running the driver on a " - "GPU node.") - - worker_ips = [ - ray.get(worker.get_node_ip.remote()) # type: ignore[attr-defined] - for worker in self.workers - ] - ip_counts: Dict[str, int] = {} - for ip in worker_ips: - ip_counts[ip] = ip_counts.get(ip, 0) + 1 - - def sort_by_driver_then_worker_ip(worker): - """ - Sort the workers based on 3 properties: - 1. If the worker is on the same node as the driver (vllm engine), - it should be placed first. - 2. Then, if the worker is on a node with fewer workers, it should - be placed first. - 3. Finally, if the work is on a node with smaller IP address, it - should be placed first. - """ - ip = ray.get(worker.get_node_ip.remote()) - return (ip != driver_ip, ip_counts[ip], ip) - - # After sorting, the workers on the same node will be - # close to each other, and the workers on the driver - # node will be placed first. - self.workers = sorted(self.workers, key=sort_by_driver_then_worker_ip) - - worker_node_and_gpu_ids = [] - for worker in [self.driver_dummy_worker] + self.workers: - if worker is None: - # driver_dummy_worker can be None when using ray spmd worker. - continue - worker_node_and_gpu_ids.append( - ray.get(worker.get_node_and_gpu_ids.remote()) \ - ) # type: ignore - - node_workers = defaultdict(list) # node id -> list of worker ranks - node_gpus = defaultdict(list) # node id -> list of gpu ids - - for i, (node_id, gpu_ids) in enumerate(worker_node_and_gpu_ids): - node_workers[node_id].append(i) - # `gpu_ids` can be a list of strings or integers. - # convert them to integers for consistency. - # NOTE: gpu_ids can be larger than 9 (e.g. 16 GPUs), - # string sorting is not sufficient. - # see https://github.com/vllm-project/vllm/issues/5590 - gpu_ids = [int(x) for x in gpu_ids] - node_gpus[node_id].extend(gpu_ids) - for node_id, gpu_ids in node_gpus.items(): - node_gpus[node_id] = sorted(gpu_ids) - - all_ips = set(worker_ips + [driver_ip]) - n_ips = len(all_ips) - n_nodes = len(node_workers) - - if n_nodes != n_ips: - raise RuntimeError( - f"Every node should have a unique IP address. Got {n_nodes}" - f" nodes with node ids {list(node_workers.keys())} and " - f"{n_ips} unique IP addresses {all_ips}. Please check your" - " network configuration. If you set `VLLM_HOST_IP` " - "environment variable, make sure it is unique for" - " each node.") - - # Set environment variables for the driver and workers. - all_args_to_update_environment_variables = [({ - "VLLM_TRACE_FUNCTION": - str(envs.VLLM_TRACE_FUNCTION), - }, ) for (node_id, _) in worker_node_and_gpu_ids] - self._run_workers("update_environment_variables", - all_args=all_args_to_update_environment_variables) - - if len(node_gpus) == 1: - # in single node case, we don't need to get the IP address. - # the loopback address is sufficient - # NOTE: a node may have several IP addresses, one for each - # network interface. `get_ip()` might return any of them, - # while they might not work for communication inside the node - # if the network setup is complicated. Using the loopback address - # solves this issue, as it always works for communication inside - # the node. - driver_ip = "127.0.0.1" - distributed_init_method = get_distributed_init_method( - driver_ip, get_open_port()) - - # Initialize the actual workers inside worker wrapper. - init_worker_all_kwargs = [ - self._get_worker_kwargs( - local_rank=node_workers[node_id].index(rank), - rank=rank, - distributed_init_method=distributed_init_method, - ) for rank, (node_id, _) in enumerate(worker_node_and_gpu_ids) - ] - self._run_workers("init_worker", all_kwargs=init_worker_all_kwargs) - - self._run_workers("init_device") - self._run_workers("load_model", - max_concurrent_workers=self.parallel_config. - max_parallel_loading_workers) - - if self.use_ray_spmd_worker: - for pp_rank in range(self.parallel_config.pipeline_parallel_size): - self.pp_tp_workers.append([]) - for tp_rank in range( - self.parallel_config.tensor_parallel_size): - # PP=2, TP=4 - # pp_tp_workers = [[0, 1, 2, 3], [4, 5, 6, 7]] - rank = (pp_rank * self.parallel_config.tensor_parallel_size - ) + tp_rank - assert len(self.pp_tp_workers[pp_rank]) == tp_rank - assert pp_rank < len(self.pp_tp_workers) - self.pp_tp_workers[pp_rank].append(self.workers[rank]) - - # This is the list of workers that are rank 0 of each TP group EXCEPT - # global rank 0. These are the workers that will broadcast to the - # rest of the workers. - self.tp_driver_workers: List[RayWorkerWrapper] = [] - # This is the list of workers that are not drivers and not the first - # worker in a TP group. These are the workers that will be - # broadcasted to. - self.non_driver_workers: List[RayWorkerWrapper] = [] - - # Enforce rank order for correct rank to return final output. - for index, worker in enumerate(self.workers): - # The driver worker is rank 0 and not in self.workers. - rank = index + 1 - if rank % self.parallel_config.tensor_parallel_size == 0: - self.tp_driver_workers.append(worker) - else: - self.non_driver_workers.append(worker) - - def _driver_execute_model( - self, execute_model_req: Optional[ExecuteModelRequest] - ) -> Optional[List[SamplerOutput]]: - """Run execute_model in the driver worker. - - Passing None will cause the driver to stop the model execution - loop running in each of the remote workers. - """ - assert not self.use_ray_spmd_worker, ( - "driver_worker does not exist for VLLM_USE_RAY_SPMD_WORKER=1") - return self.driver_worker.execute_method("execute_model", - execute_model_req) - - def execute_model( - self, - execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: - if not self.use_ray_spmd_worker: - return super().execute_model(execute_model_req) - - if self.forward_dag is None: - self.forward_dag = self._compiled_ray_dag(enable_asyncio=False) - - serialized_data = self.input_encoder.encode(execute_model_req) - outputs = ray.get(self.forward_dag.execute(serialized_data)) - output = self.output_decoder.decode(outputs[0]) - return output - - def _run_workers( - self, - method: str, - *args, - async_run_tensor_parallel_workers_only: bool = False, - all_args: Optional[List[Tuple[Any, ...]]] = None, - all_kwargs: Optional[List[Dict[str, Any]]] = None, - max_concurrent_workers: Optional[int] = None, - **kwargs, - ) -> Any: - """Runs the given method on all workers. Can be used in the following - ways: - - Args: - - async_run_tensor_parallel_workers_only: If True the method will be - run only in the remote TP workers, not the driver worker. - It will also be run asynchronously and return a list of futures - rather than blocking on the results. - - args/kwargs: All workers share the same args/kwargs - - all_args/all_kwargs: args/kwargs for each worker are specified - individually - """ - if self.use_ray_spmd_worker: - assert not async_run_tensor_parallel_workers_only, ( - "async_run_tensor_parallel_workers_only is not supported for " - "spmd mode.") - - if max_concurrent_workers: - raise NotImplementedError( - "max_concurrent_workers is not supported yet.") - - count = len(self.workers) if not \ - async_run_tensor_parallel_workers_only \ - else len(self.non_driver_workers) - # If using SPMD worker, all workers are the same, so we should execute - # the args on all workers. Otherwise, we skip the first worker's args - # because those args will go to the driver worker. - first_worker_args_index: int = 0 if self.use_ray_spmd_worker else 1 - all_worker_args = repeat(args, count) if all_args is None \ - else islice(all_args, first_worker_args_index, None) - all_worker_kwargs = repeat(kwargs, count) if all_kwargs is None \ - else islice(all_kwargs, first_worker_args_index, None) - - # Start the ray workers first. - ray_workers = self.workers - if async_run_tensor_parallel_workers_only: - ray_workers = self.non_driver_workers - ray_worker_outputs = [ - worker.execute_method.remote(method, *worker_args, **worker_kwargs) - for (worker, worker_args, worker_kwargs - ) in zip(ray_workers, all_worker_args, all_worker_kwargs) - ] - - if async_run_tensor_parallel_workers_only: - # Just return futures - return ray_worker_outputs - - driver_worker_output = [] - # In SPMD mode, the driver worker is the same as any other worker, - # so we only explicitly execute on the driver worker if using a - # non-SPMD worker class. - if not self.use_ray_spmd_worker: - driver_args = args if all_args is None else all_args[0] - driver_kwargs = kwargs if all_kwargs is None else all_kwargs[0] - - # Start the driver worker after all the ray workers. - driver_worker_output = [ - self.driver_worker.execute_method(method, *driver_args, - **driver_kwargs) - ] - - # Get the results of the ray workers. - if self.workers: - ray_worker_outputs = ray.get(ray_worker_outputs) - - return driver_worker_output + ray_worker_outputs - - def _wait_for_tasks_completion(self, parallel_worker_tasks: Any) -> None: - """Wait for futures returned from _run_workers() with - async_run_remote_workers_only to complete.""" - ray.get(parallel_worker_tasks) - - def _check_ray_adag_installation(self): - import pkg_resources - from packaging import version - - required_version = version.parse("2.35") - current_version = version.parse( - pkg_resources.get_distribution("ray").version) - # TODO: update the constraint once we adapt to the backward - # incompatible API change from ray 2.36 - if current_version != required_version: - raise ValueError(f"Ray version {required_version} is " - f"required, but found {current_version}") - - import importlib.util - adag_spec = importlib.util.find_spec( - "ray.experimental.compiled_dag_ref") - if adag_spec is None: - raise ValueError("Ray accelerated DAG is not installed. " - "Run `pip install ray[adag]` to install it.") - - def _compiled_ray_dag(self, enable_asyncio: bool): - assert self.parallel_config.use_ray - self._check_ray_adag_installation() - from ray.dag import InputNode, MultiOutputNode - from ray.experimental.channel.torch_tensor_type import TorchTensorType - - with InputNode() as input_data: - # Example DAG: PP=2, TP=4 - # (ExecuteModelReq, None) -> 0 -> (ExecuteModelReq, IntermediateOutput) -> 4 -> SamplerOutput # noqa: E501 - # -> 1 -> (ExecuteModelReq, IntermediateOutput) -> 5 -> SamplerOutput # noqa: E501 - # -> 2 -> (ExecuteModelReq, IntermediateOutput) -> 6 -> SamplerOutput # noqa: E501 - # -> 3 -> (ExecuteModelReq, IntermediateOutput) -> 7 -> SamplerOutput # noqa: E501 - - # All workers in the first TP group will take in the - # ExecuteModelRequest as input. - outputs = [input_data for _ in self.pp_tp_workers[0]] - for pp_rank, tp_group in enumerate(self.pp_tp_workers): - # Each PP worker takes in the output of the previous PP worker, - # and the TP group executes in SPMD fashion. - outputs = [ - worker.execute_model_spmd. - bind( # type: ignore[attr-defined] - outputs[i]) for i, worker in enumerate(tp_group) - ] - - last_pp_rank = len(self.pp_tp_workers) - 1 - if pp_rank < last_pp_rank: - # Specify how intermediate tensors should be passed - # between pp stages, no need to specify for the last - # pp stage. - transport = "auto" - outputs = [ - output.with_type_hint( - TorchTensorType(transport=transport)) - for output in outputs - ] - - forward_dag = MultiOutputNode(outputs) - - return forward_dag.experimental_compile(enable_asyncio=enable_asyncio) - - def __del__(self): - self.shutdown() - - -class RayHPUExecutorAsync(RayHPUExecutor, DistributedGPUExecutorAsync): - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - self.pp_locks: Optional[List[asyncio.Lock]] = None - self.use_ray_spmd_worker = envs.VLLM_USE_RAY_SPMD_WORKER - if not self.use_ray_compiled_dag: - self.driver_exec_method = make_async( - self.driver_worker.execute_method) - - async def execute_model_async( - self, - execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: - if not self.use_ray_spmd_worker: - return await super().execute_model_async(execute_model_req) - - if self.forward_dag is None: - self.forward_dag = self._compiled_ray_dag(enable_asyncio=True) - - serialized_data = self.input_encoder.encode(execute_model_req) - dag_future = await self.forward_dag.execute_async(serialized_data) - outputs = await dag_future - return self.output_decoder.decode(outputs[0]) - - async def _driver_execute_model_async( - self, - execute_model_req: Optional[ExecuteModelRequest] = None - ) -> List[SamplerOutput]: - assert not self.use_ray_spmd_worker, ( - "driver_worker does not exist for VLLM_USE_RAY_SPMD_WORKER=1") - if not self.tp_driver_workers: - return await self.driver_exec_method("execute_model", - execute_model_req) - if self.pp_locks is None: - # This locks each pipeline parallel stage so multiple virtual - # engines can't execute on the same stage at the same time - # We create the locks here to avoid creating them in the constructor - # which uses a different asyncio loop. - self.pp_locks = [ - asyncio.Lock() - for _ in range(self.parallel_config.pipeline_parallel_size) - ] - - tasks = [ - asyncio.create_task( - _run_task_with_lock(self.driver_exec_method, self.pp_locks[0], - "execute_model", execute_model_req)) - ] - for pp_rank, driver_worker in enumerate(self.tp_driver_workers, - start=1): - tasks.append( - asyncio.create_task( - _run_task_with_lock(driver_worker.execute_method.remote, - self.pp_locks[pp_rank], - "execute_model", execute_model_req))) - - results = await asyncio.gather(*tasks) - - # Only the last PP stage has the final results. - return results[-1] - - async def _start_worker_execution_loop(self): - assert not self.use_ray_spmd_worker, ( - "worker loop is disabled for VLLM_USE_RAY_SPMD_WORKER=1") - coros = [ - worker.execute_method.remote("start_worker_execution_loop") - for worker in self.non_driver_workers - ] - return await asyncio.gather(*coros) - - def __del__(self): - self.shutdown() diff --git a/vllm/executor/ray_tpu_executor.py b/vllm/executor/ray_tpu_executor.py deleted file mode 100644 index 5118c13934f0d..0000000000000 --- a/vllm/executor/ray_tpu_executor.py +++ /dev/null @@ -1,343 +0,0 @@ -import asyncio -import os -from collections import defaultdict -from itertools import islice, repeat -from typing import (TYPE_CHECKING, Any, Awaitable, Dict, List, Optional, Tuple, - Union) - -import vllm.envs as envs -from vllm.executor.executor_base import ExecutorAsyncBase -from vllm.executor.ray_utils import RayWorkerWrapper, ray -from vllm.executor.tpu_executor import TPUExecutor -from vllm.logger import init_logger -from vllm.model_executor.layers.sampler import SamplerOutput -from vllm.sequence import ExecuteModelRequest -from vllm.utils import (get_distributed_init_method, get_ip, get_open_port, - make_async) - -if ray is not None: - from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy - -if TYPE_CHECKING: - from ray.util.placement_group import PlacementGroup - -logger = init_logger(__name__) - - -class RayTPUExecutor(TPUExecutor): - - uses_ray: bool = True - - def __init__(self, *args, **kwargs): - # This is non-None when the execute model loop is running - # in the parallel workers. It's a coroutine in the AsyncLLMEngine case. - self.parallel_worker_tasks: Optional[Union[Any, Awaitable[Any]]] = None - # Updated by implementations that require additional args to be passed - # to the _run_workers execute_model call - self.extra_execute_model_run_workers_kwargs: Dict[str, Any] = {} - - super().__init__(*args, **kwargs) - - def _init_executor(self) -> None: - assert self.parallel_config.distributed_executor_backend == "ray" - placement_group = self.parallel_config.placement_group - - # Disable Ray usage stats collection. - ray_usage = os.environ.get("RAY_USAGE_STATS_ENABLED", "0") - if ray_usage != "1": - os.environ["RAY_USAGE_STATS_ENABLED"] = "0" - - # Create the parallel TPU workers. - self._init_workers_ray(placement_group) - - def _init_workers_ray(self, placement_group: "PlacementGroup", - **ray_remote_kwargs): - # The driver dummy worker does not actually use any resources. - # It holds the resource for the driver worker. - self.driver_dummy_worker: Optional[RayWorkerWrapper] = None - # The remaining workers are the actual ray actors. - self.workers: List[RayWorkerWrapper] = [] - - # Create the workers. - driver_ip = get_ip() - for bundle_id, bundle in enumerate(placement_group.bundle_specs): - if not bundle.get("TPU", 0): - continue - scheduling_strategy = PlacementGroupSchedulingStrategy( - placement_group=placement_group, - placement_group_capture_child_tasks=True, - placement_group_bundle_index=bundle_id, - ) - - # GKE does not fetch environment information from metadata server - # and instead sets these from within the Ray process. Therefore we - # need to override the Ray environment variables manually. - override_env = {} - if "TPU_CHIPS_PER_HOST_BOUNDS" in os.environ: - override_env.update({ - "TPU_CHIPS_PER_HOST_BOUNDS": - os.environ["TPU_CHIPS_PER_HOST_BOUNDS"] - }) - if "TPU_HOST_BOUNDS" in os.environ: - override_env.update( - {"TPU_HOST_BOUNDS": os.environ["TPU_HOST_BOUNDS"]}) - - worker = ray.remote( - num_cpus=0, - resources={"TPU": 1}, - scheduling_strategy=scheduling_strategy, - **ray_remote_kwargs, - )(RayWorkerWrapper).remote(vllm_config=self.vllm_config) - if override_env: - worker.override_env_vars.remote(override_env) - - worker_ip = ray.get(worker.get_node_ip.remote()) - if worker_ip == driver_ip and self.driver_dummy_worker is None: - # If the worker is on the same node as the driver, we use it - # as the resource holder for the driver process. - self.driver_dummy_worker = worker - self.driver_worker = RayWorkerWrapper( - vllm_config=self.vllm_config) - else: - # Else, added to the list of workers. - self.workers.append(worker) - - logger.debug("workers: %s", self.workers) - logger.debug("driver_dummy_worker: %s", self.driver_dummy_worker) - if self.driver_dummy_worker is None: - raise ValueError( - "Ray does not allocate any TPUs on the driver node. Consider " - "adjusting the Ray placement group or running the driver on a " - "TPU node.") - - worker_ips = [ - ray.get(worker.get_node_ip.remote()) # type: ignore[attr-defined] - for worker in self.workers - ] - ip_counts: Dict[str, int] = {} - for ip in worker_ips: - ip_counts[ip] = ip_counts.get(ip, 0) + 1 - - def sort_by_driver_then_worker_ip(worker): - """ - Sort the workers based on 3 properties: - 1. If the worker is on the same node as the driver (vllm engine), - it should be placed first. - 2. Then, if the worker is on a node with fewer workers, it should - be placed first. - 3. Finally, if the work is on a node with smaller IP address, it - should be placed first. - """ - ip = ray.get(worker.get_node_ip.remote()) - return (ip != driver_ip, ip_counts[ip], ip) - - # After sorting, the workers on the same node will be - # close to each other, and the workers on the driver - # node will be placed first. - self.workers = sorted(self.workers, key=sort_by_driver_then_worker_ip) - - # Get the set of TPU IDs used on each node. - worker_node_and_gpu_ids = [] - for worker in [self.driver_dummy_worker] + self.workers: - if worker is None: - # driver_dummy_worker can be None when using ray spmd worker. - continue - worker_node_and_gpu_ids.append( - ray.get(worker.get_node_and_gpu_ids.remote()) \ - ) # type: ignore - - node_workers = defaultdict(list) - for i, (node_id, _) in enumerate(worker_node_and_gpu_ids): - node_workers[node_id].append(i) - - # Set environment variables for the driver and workers. - all_args_to_update_environment_variables = [({ - "VLLM_TRACE_FUNCTION": - str(envs.VLLM_TRACE_FUNCTION), - }, ) for _ in worker_node_and_gpu_ids] - self._run_workers("update_environment_variables", - all_args=all_args_to_update_environment_variables) - - if len(node_workers) == 1: - # in single node case, we don't need to get the IP address. - # the loopback address is sufficient - # NOTE: a node may have several IP addresses, one for each - # network interface. `get_ip()` might return any of them, - # while they might not work for communication inside the node - # if the network setup is complicated. Using the loopback address - # solves this issue, as it always works for communication inside - # the node. - driver_ip = "127.0.0.1" - distributed_init_method = get_distributed_init_method( - driver_ip, get_open_port()) - - # Initialize the actual workers inside worker wrapper. - init_worker_all_kwargs = [ - self._get_worker_kwargs( - local_rank=node_workers[node_id].index(rank), - rank=rank, - distributed_init_method=distributed_init_method, - ) for rank, (node_id, _) in enumerate(worker_node_and_gpu_ids) - ] - self._run_workers("init_worker", all_kwargs=init_worker_all_kwargs) - - self._run_workers("init_device") - self._run_workers("load_model", - max_concurrent_workers=self.parallel_config. - max_parallel_loading_workers) - - def _driver_execute_model( - self, - execute_model_req: Optional[ExecuteModelRequest] = None - ) -> List[SamplerOutput]: - """Run execute_model in the driver worker. - - Passing None will cause the driver to stop the model execution - loop running in each of the remote workers. - """ - return self.driver_worker.execute_method("execute_model", - execute_model_req) - - def _run_workers( - self, - method: str, - *args, - async_run_remote_workers_only: bool = False, - all_args: Optional[List[Tuple[Any, ...]]] = None, - all_kwargs: Optional[List[Dict[str, Any]]] = None, - max_concurrent_workers: Optional[int] = None, - use_ray_compiled_dag: bool = False, - **kwargs, - ) -> Any: - """Runs the given method on all workers. Can be used in the following - ways: - - - async_run_remote_workers_only: If True the method will be run only - in the remote workers, not the driver worker. It will also be - run asynchronously and return a list of futures rather than blocking - on the results. - - args/kwargs: All workers share the same args/kwargs - - all_args/all_kwargs: args/kwargs for each worker are specified - individually - """ - - if max_concurrent_workers: - raise NotImplementedError( - "max_concurrent_workers is not supported yet.") - - count = len(self.workers) - all_worker_args = repeat(args, count) if all_args is None \ - else islice(all_args, 1, None) - all_worker_kwargs = repeat(kwargs, count) if all_kwargs is None \ - else islice(all_kwargs, 1, None) - - # Start the ray workers first. - ray_worker_outputs = [ - worker.execute_method.remote(method, *worker_args, **worker_kwargs) - for (worker, worker_args, worker_kwargs - ) in zip(self.workers, all_worker_args, all_worker_kwargs) - ] - - if async_run_remote_workers_only: - # Just return futures - return ray_worker_outputs - - driver_args = args if all_args is None else all_args[0] - driver_kwargs = kwargs if all_kwargs is None else all_kwargs[0] - - # Start the driver worker after all the ray workers. - driver_worker_output = self.driver_worker.execute_method( - method, *driver_args, **driver_kwargs) - # Get the results of the ray workers. - if self.workers: - ray_worker_outputs = ray.get(ray_worker_outputs) - - return [driver_worker_output] + ray_worker_outputs - - def _wait_for_tasks_completion(self, parallel_worker_tasks: Any) -> None: - """Wait for futures returned from _run_workers() with - async_run_remote_workers_only to complete.""" - ray.get(parallel_worker_tasks) - - def determine_num_available_blocks(self) -> Tuple[int, int]: - num_blocks = self._run_workers("determine_num_available_blocks", ) - num_tpu_blocks = min(b[0] for b in num_blocks) - num_cpu_blocks = min(b[1] for b in num_blocks) - return num_tpu_blocks, num_cpu_blocks - - def initialize_cache(self, num_gpu_blocks: int, - num_cpu_blocks: int) -> None: - logger.info("# TPU blocks: %d, # CPU blocks: %d", num_gpu_blocks, - num_cpu_blocks) - self.cache_config.num_gpu_blocks = num_gpu_blocks - self.cache_config.num_cpu_blocks = num_cpu_blocks - self._run_workers("initialize_cache", - num_gpu_blocks=num_gpu_blocks, - num_cpu_blocks=num_cpu_blocks) - - def execute_model( - self, - execute_model_req: ExecuteModelRequest, - ) -> List[SamplerOutput]: - if self.parallel_worker_tasks is None: - self.parallel_worker_tasks = self._run_workers( - "start_worker_execution_loop", - async_run_remote_workers_only=True, - **self.extra_execute_model_run_workers_kwargs) - - # Only the driver worker returns the sampling results. - return self._driver_execute_model(execute_model_req) - - def stop_remote_worker_execution_loop(self) -> None: - if self.parallel_worker_tasks is None: - return - - self._driver_execute_model() - parallel_worker_tasks = self.parallel_worker_tasks - self.parallel_worker_tasks = None - # Ensure that workers exit model loop cleanly - # (this will raise otherwise) - self._wait_for_tasks_completion(parallel_worker_tasks) - - -class RayTPUExecutorAsync(RayTPUExecutor, ExecutorAsyncBase): - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - self.driver_exec_method = make_async(self.driver_worker.execute_method) - - async def execute_model_async( - self, - execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]: - if self.parallel_worker_tasks is None: - # Start model execution loop running in the parallel workers - self.parallel_worker_tasks = asyncio.create_task( - self._start_worker_execution_loop()) - - # Only the driver worker returns the sampling results. - return await self._driver_execute_model_async(execute_model_req) - - async def stop_remote_worker_execution_loop_async(self) -> None: - if self.parallel_worker_tasks is None: - return - - await self._driver_execute_model_async() - parallel_worker_tasks = self.parallel_worker_tasks - self.parallel_worker_tasks = None - # Ensure that workers exit model loop cleanly - # (this will raise otherwise) - await parallel_worker_tasks - - async def _driver_execute_model_async( - self, - execute_model_req: Optional[ExecuteModelRequest] = None - ) -> List[SamplerOutput]: - return await self.driver_exec_method("execute_model", - execute_model_req) - - async def _start_worker_execution_loop(self): - coros = [ - worker.execute_method.remote("start_worker_execution_loop") - for worker in self.workers - ] - return await asyncio.gather(*coros) diff --git a/vllm/executor/ray_utils.py b/vllm/executor/ray_utils.py index 9f40f6a65dcd7..e55155ea06225 100644 --- a/vllm/executor/ray_utils.py +++ b/vllm/executor/ray_utils.py @@ -1,7 +1,7 @@ import os import time from collections import defaultdict -from typing import Dict, List, Optional, Tuple, Union +from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union import msgspec @@ -13,6 +13,10 @@ from vllm.utils import get_ip from vllm.worker.worker_base import WorkerWrapperBase +if TYPE_CHECKING: + from vllm.v1.core.scheduler import SchedulerOutput + from vllm.v1.outputs import ModelRunnerOutput + logger = init_logger(__name__) PG_WAIT_TIMEOUT = 1800 @@ -95,6 +99,26 @@ def execute_model_spmd( return output + def setup_device_if_necessary(self): + # TODO(swang): This is needed right now because Ray CG executes + # on a background thread, so we need to reset torch's current + # device. + # We can remove this API after it is fixed in compiled graph. + import torch + assert self.worker is not None, "Worker is not initialized" + if not self.compiled_dag_cuda_device_set: + torch.cuda.set_device(self.worker.device) + self.compiled_dag_cuda_device_set = True + + def execute_model( + self, + scheduler_output: "SchedulerOutput", + ) -> "ModelRunnerOutput": + self.setup_device_if_necessary() + assert self.worker is not None, "Worker is not initialized" + output = self.worker.model_runner.execute_model(scheduler_output) + return output + def override_env_vars(self, vars: Dict[str, str]): os.environ.update(vars) diff --git a/vllm/executor/ray_xpu_executor.py b/vllm/executor/ray_xpu_executor.py deleted file mode 100644 index d2086f5fef26c..0000000000000 --- a/vllm/executor/ray_xpu_executor.py +++ /dev/null @@ -1,40 +0,0 @@ -import asyncio -from typing import List, Optional - -import ray - -import vllm.envs as envs -from vllm.executor.ray_gpu_executor import RayGPUExecutor, RayGPUExecutorAsync -from vllm.executor.xpu_executor import XPUExecutor -from vllm.logger import init_logger -from vllm.utils import make_async - -logger = init_logger(__name__) - - -class RayXPUExecutor(RayGPUExecutor, XPUExecutor): - - def _get_env_vars_to_be_updated(self): - # Get the set of GPU IDs used on each node. - worker_node_and_gpu_ids = [] - for worker in [self.driver_dummy_worker] + self.workers: - if worker is None: - # driver_dummy_worker can be None when using ray spmd worker. - continue - worker_node_and_gpu_ids.append( - ray.get(worker.get_node_and_gpu_ids.remote())) # type: ignore - - # Set environment variables for the driver and workers. - all_args_to_update_environment_variables = [({ - "VLLM_TRACE_FUNCTION": - str(envs.VLLM_TRACE_FUNCTION), - }, ) for (_, _) in worker_node_and_gpu_ids] - return all_args_to_update_environment_variables - - -class RayXPUExecutorAsync(RayXPUExecutor, RayGPUExecutorAsync): - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - self.driver_exec_method = make_async(self.driver_worker.execute_method) - self.pp_locks: Optional[List[asyncio.Lock]] = None diff --git a/vllm/executor/tpu_executor.py b/vllm/executor/tpu_executor.py deleted file mode 100644 index e37e8973790db..0000000000000 --- a/vllm/executor/tpu_executor.py +++ /dev/null @@ -1,142 +0,0 @@ -from typing import Any, Dict, List, Optional, Set, Tuple - -import torch - -from vllm.executor.executor_base import ExecutorAsyncBase, ExecutorBase -from vllm.logger import init_logger -from vllm.lora.request import LoRARequest -from vllm.model_executor.layers.sampler import SamplerOutput -from vllm.sequence import ExecuteModelRequest -from vllm.utils import (get_distributed_init_method, get_ip, get_open_port, - make_async) - -logger = init_logger(__name__) - - -class TPUExecutor(ExecutorBase): - - uses_ray: bool = False - - def _init_executor(self) -> None: - assert not self.scheduler_config.chunked_prefill_enabled, ( - "Chunked prefill is not yet supported for TPU backend") - assert not self.speculative_config, ( - "Speculative decoding is not yet supported for TPU backend") - if self.model_config.dtype in (torch.float16, torch.float32): - logger.warning( - "The TPU backend currently does not support %s. " - "Using bfloat16 instead.", self.model_config.dtype) - self.model_config.dtype = torch.bfloat16 - - # Instantiate the worker and load the model to the device. - self.driver_worker = self._create_worker() - self.driver_worker.init_device() - self.driver_worker.load_model() - - def _get_worker_kwargs( - self, - local_rank: int = 0, - rank: int = 0, - distributed_init_method: Optional[str] = None, - ) -> Dict[str, Any]: - """Return worker init args for a given rank.""" - if distributed_init_method is None: - distributed_init_method = get_distributed_init_method( - get_ip(), get_open_port()) - return dict( - vllm_config=self.vllm_config, - local_rank=local_rank, - rank=rank, - distributed_init_method=distributed_init_method, - is_driver_worker=rank == 0, - ) - - def _create_worker( - self, - local_rank: int = 0, - rank: int = 0, - distributed_init_method: Optional[str] = None, - ): - if self.scheduler_config.is_multi_step: - from vllm.worker.multi_step_tpu_worker import MultiStepTPUWorker - worker = MultiStepTPUWorker(**self._get_worker_kwargs( - local_rank, rank, distributed_init_method)) - return worker - else: - from vllm.worker.tpu_worker import TPUWorker - - worker = TPUWorker(**self._get_worker_kwargs( - local_rank, rank, distributed_init_method)) - return worker - - def initialize_cache( - self, - num_gpu_blocks: int, - num_cpu_blocks: int, - ) -> None: - """Initialize the KV cache by invoking the underlying worker.""" - # NOTE: This is logged in the executor because there can be >1 worker - # with other executors. We could log in the engine level, but work - # remains to abstract away the device for non-GPU configurations. - logger.info("# TPU blocks: %d, # CPU blocks: %d", num_gpu_blocks, - num_cpu_blocks) - self.driver_worker.initialize_cache(num_gpu_blocks, num_cpu_blocks) - - def determine_num_available_blocks(self) -> Tuple[int, int]: - """Determine the number of available KV blocks by invoking the - underlying worker.""" - return self.driver_worker.determine_num_available_blocks() - - def execute_model( - self, - execute_model_req: ExecuteModelRequest, - ) -> List[SamplerOutput]: - output = self.driver_worker.execute_model(execute_model_req) - return output - - def add_lora(self, lora_request: LoRARequest) -> bool: - raise NotImplementedError( - "LoRA is currently not supported by the TPU backend.") - - def remove_lora(self, lora_id: int) -> bool: - raise NotImplementedError( - "LoRA is currently not supported by the TPU backend.") - - def pin_lora(self, lora_id: int) -> bool: - raise NotImplementedError( - "LoRA is currently not supported by the TPU backend.") - - def list_loras(self) -> Set[int]: - raise NotImplementedError( - "LoRA is currently not supported by the TPU backend.") - - def add_prompt_adapter(self, prompt_adapter_request) -> bool: - raise NotImplementedError( - "Soft prompt is currently not supported by the TPU backend.") - - def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool: - raise NotImplementedError( - "Soft prompt is currently not supported by the TPU backend.") - - def pin_prompt_adapter(self, prompt_adapter_id: int) -> bool: - raise NotImplementedError( - "Soft prompt is currently not supported by the TPU backend.") - - def list_prompt_adapters(self) -> Set[int]: - raise NotImplementedError( - "Soft prompt is currently not supported by the TPU backend.") - - def check_health(self) -> None: - # TPUExecutor will always be healthy as long as it's running. - return - - -class TPUExecutorAsync(TPUExecutor, ExecutorAsyncBase): - - async def execute_model_async( - self, - sexecute_model_req: ExecuteModelRequest, - ) -> SamplerOutput: - output = await make_async(self.driver_worker.execute_model - )(sexecute_model_req) - return output diff --git a/vllm/executor/uniproc_executor.py b/vllm/executor/uniproc_executor.py new file mode 100644 index 0000000000000..a5c4dcf0ec7f9 --- /dev/null +++ b/vllm/executor/uniproc_executor.py @@ -0,0 +1,132 @@ +import os +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import torch +import torch.distributed as dist + +import vllm.envs as envs +from vllm.executor.executor_base import ExecutorBase +from vllm.logger import init_logger +from vllm.utils import (get_distributed_init_method, get_ip, get_open_port, + run_method) +from vllm.worker.worker_base import WorkerWrapperBase + +logger = init_logger(__name__) + + +class UniProcExecutor(ExecutorBase): + + uses_ray: bool = False + + def _init_executor(self) -> None: + """Initialize the worker and load the model. + """ + self.driver_worker = WorkerWrapperBase(vllm_config=self.vllm_config, + rpc_rank=0) + distributed_init_method = get_distributed_init_method( + get_ip(), get_open_port()) + local_rank = 0 + rank = 0 + kwargs = dict( + vllm_config=self.vllm_config, + local_rank=local_rank, + rank=rank, + distributed_init_method=distributed_init_method, + is_driver_worker=(not self.parallel_config) + or (rank % self.parallel_config.tensor_parallel_size == 0), + ) + self.collective_rpc("init_worker", args=([kwargs], )) + self.collective_rpc("init_device") + self.collective_rpc("load_model") + + def collective_rpc(self, + method: Union[str, Callable], + timeout: Optional[float] = None, + args: Tuple = (), + kwargs: Optional[Dict] = None) -> List[Any]: + if kwargs is None: + kwargs = {} + answer = run_method(self.driver_worker, method, args, kwargs) + return [answer] + + def check_health(self) -> None: + # UniProcExecutor will always be healthy as long as + # it's running. + return + + +UniProcExecutorAsync = UniProcExecutor + + +class ExecutorWithExternalLauncher(UniProcExecutor): + """An executor that uses external launchers to launch engines, + specially designed for torchrun-compatible launchers, for + offline inference with tensor parallelism. + + see https://github.com/vllm-project/vllm/issues/11400 for + the motivation, and examples/offline_inference/torchrun_example.py + for the usage example. + + The key idea: although it is tensor-parallel inference, we only + create one worker per executor, users will launch multiple + engines with torchrun-compatible launchers, and all these engines + work together to process the same prompts. When scheduling is + deterministic, all the engines will generate the same outputs, + and they don't need to synchronize the states with each other. + """ + uses_ray: bool = False + + def _init_executor(self) -> None: + """Initialize the worker and load the model. + """ + assert self.vllm_config.parallel_config.pipeline_parallel_size == 1, \ + ("ExecutorWithExternalLauncher does not " + "support pipeline parallelism.") + assert self.vllm_config.scheduler_config.delay_factor == 0.0, \ + ("ExecutorWithExternalLauncher needs deterministic " + "execution, so it" + "does not support delay_factor in scheduling") + assert not envs.VLLM_USE_V1, \ + ("V1 architecture cannot guarantee deterministic execution, " + "so it is not supported in ExecutorWithExternalLauncher.") + self.driver_worker = WorkerWrapperBase(vllm_config=self.vllm_config, + rpc_rank=0) + # engines are launched in torchrun-compatible launchers + # so we can use the env:// method. + # required env vars: + # - RANK + # - MASTER_ADDR + # - MASTER_PORT + distributed_init_method = "env://" + rank = int(os.environ["RANK"]) + local_rank = rank + is_driver_worker = True + kwargs = dict( + vllm_config=self.vllm_config, + local_rank=local_rank, + rank=rank, + distributed_init_method=distributed_init_method, + is_driver_worker=is_driver_worker, + ) + self.collective_rpc("init_worker", args=([kwargs], )) + self.collective_rpc("init_device") + self.collective_rpc("load_model") + + def determine_num_available_blocks(self) -> Tuple[int, int]: + """ + Determine the number of available KV blocks. + Add an additional all_reduce to get the min across all ranks. + Note that even if we have the same `gpu_memory_utilization` and + `swap_space`, the available memory in every rank might still + differ because NCCL can take different amounts of memory in + different ranks. Therefore, it is necessary to test if all ranks + agree on the same KV cache configuration. + """ + a, b = super().determine_num_available_blocks() + from vllm.distributed.parallel_state import get_world_group + cpu_group = get_world_group().cpu_group + a_tensor = torch.tensor([a], device="cpu", dtype=torch.int64) + b_tensor = torch.tensor([b], device="cpu", dtype=torch.int64) + dist.all_reduce(a_tensor, group=cpu_group, op=dist.ReduceOp.MIN) + dist.all_reduce(b_tensor, group=cpu_group, op=dist.ReduceOp.MIN) + return a_tensor.item(), b_tensor.item() diff --git a/vllm/executor/xpu_executor.py b/vllm/executor/xpu_executor.py deleted file mode 100644 index 722b86a95ff8a..0000000000000 --- a/vllm/executor/xpu_executor.py +++ /dev/null @@ -1,39 +0,0 @@ -from typing import List, Optional, Union - -from vllm.executor.executor_base import ExecutorAsyncBase -from vllm.executor.gpu_executor import GPUExecutor -from vllm.logger import init_logger -from vllm.model_executor.layers.sampler import SamplerOutput -from vllm.sequence import ExecuteModelRequest, PoolerOutput -from vllm.utils import make_async - -logger = init_logger(__name__) - - -class XPUExecutor(GPUExecutor): - - uses_ray: bool = False - - def _init_executor(self) -> None: - assert self.device_config.device_type == "xpu" - assert self.speculative_config is None, ( - "Speculative decoding not yet supported for XPU backend") - - GPUExecutor._init_executor(self) - - def execute_model( - self, execute_model_req: ExecuteModelRequest - ) -> Optional[List[Union[SamplerOutput, PoolerOutput]]]: - output = self.driver_worker.execute_model(execute_model_req) - return output - - -class XPUExecutorAsync(XPUExecutor, ExecutorAsyncBase): - - async def execute_model_async( - self, - execute_model_req: ExecuteModelRequest, - ) -> List[SamplerOutput]: - output = await make_async(self.driver_worker.execute_model - )(execute_model_req=execute_model_req) - return output diff --git a/vllm/inputs/data.py b/vllm/inputs/data.py index b8163a7acde1d..57e85779dd587 100644 --- a/vllm/inputs/data.py +++ b/vllm/inputs/data.py @@ -9,7 +9,7 @@ if TYPE_CHECKING: from vllm.multimodal import (MultiModalDataDict, MultiModalKwargs, MultiModalPlaceholderDict) - from vllm.multimodal.inputs import MultiModalInputsV2 + from vllm.multimodal.inputs import MultiModalInputs class TextPrompt(TypedDict): @@ -207,7 +207,7 @@ def token_inputs( return inputs -DecoderOnlyInputs = Union[TokenInputs, "MultiModalInputsV2"] +DecoderOnlyInputs = Union[TokenInputs, "MultiModalInputs"] """ The inputs in :class:`~vllm.LLMEngine` before they are passed to the model executor. @@ -222,14 +222,14 @@ class EncoderDecoderInputs(TypedDict): This specifies the required data for encoder-decoder models. """ - encoder: Union[TokenInputs, "MultiModalInputsV2"] + encoder: Union[TokenInputs, "MultiModalInputs"] """The inputs for the encoder portion.""" - decoder: Union[TokenInputs, "MultiModalInputsV2"] + decoder: Union[TokenInputs, "MultiModalInputs"] """The inputs for the decoder portion.""" -SingletonInputs = Union[TokenInputs, "MultiModalInputsV2"] +SingletonInputs = Union[TokenInputs, "MultiModalInputs"] """ A processed :class:`SingletonPrompt` which can be passed to :class:`vllm.sequence.Sequence`. @@ -311,7 +311,7 @@ def multi_modal_hashes(self) -> List[str]: return inputs.get("multi_modal_hashes", []) if inputs["type"] == "multimodal": - # only the case when we use MultiModalInputsV2 + # only the case when we use MultiModalInputs return inputs.get("mm_hashes", []) # type: ignore[return-value] assert_never(inputs) # type: ignore[arg-type] diff --git a/vllm/inputs/preprocess.py b/vllm/inputs/preprocess.py index 0890883cc984f..70372e0cad22d 100644 --- a/vllm/inputs/preprocess.py +++ b/vllm/inputs/preprocess.py @@ -7,7 +7,7 @@ from vllm.logger import init_logger from vllm.lora.request import LoRARequest from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry -from vllm.multimodal.inputs import MultiModalDataDict, MultiModalInputsV2 +from vllm.multimodal.inputs import MultiModalDataDict, MultiModalInputs from vllm.prompt_adapter.request import PromptAdapterRequest from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup @@ -247,7 +247,7 @@ def _process_multimodal( mm_data: MultiModalDataDict, mm_processor_kwargs: Optional[Mapping[str, object]], lora_request: Optional[LoRARequest], - ) -> MultiModalInputsV2: + ) -> MultiModalInputs: """ Apply the model's multi-modal processor to a multi-modal prompt, returning the corresponding token IDs and metadata. @@ -271,7 +271,7 @@ async def _process_multimodal_async( mm_data: MultiModalDataDict, mm_processor_kwargs: Optional[Mapping[str, object]], lora_request: Optional[LoRARequest], - ) -> MultiModalInputsV2: + ) -> MultiModalInputs: """Async version of :meth:`_process_multimodal`.""" tokenizer_group = self.get_tokenizer_group() tokenizer = await tokenizer_group.get_lora_tokenizer_async(lora_request diff --git a/vllm/lora/layers.py b/vllm/lora/layers.py index a933ccaecf15e..e6f26d2b74b2f 100644 --- a/vllm/lora/layers.py +++ b/vllm/lora/layers.py @@ -51,6 +51,9 @@ def _get_lora_device(base_layer: nn.Module) -> torch.device: # marlin elif hasattr(base_layer, "B"): return base_layer.B.device + # HQQ marlin + elif hasattr(base_layer, "W_q"): + return base_layer.W_q.device else: raise ValueError(f"Unsupported base layer: {base_layer}") @@ -937,8 +940,8 @@ def soft_cap(self): return self.base_layer.soft_cap @property - def use_gather(self): - return self.base_layer.use_gather + def use_all_gather(self): + return self.base_layer.use_all_gather @property def org_vocab_size(self): diff --git a/vllm/lora/models.py b/vllm/lora/models.py index 5b7225bdc8f37..7dd61d27e4e4c 100644 --- a/vllm/lora/models.py +++ b/vllm/lora/models.py @@ -1,5 +1,4 @@ import copy -import json import math import os import re @@ -180,8 +179,8 @@ def from_local_checkpoint( cls, lora_dir: str, expected_lora_modules: List[str], + peft_helper: PEFTHelper, *, - max_position_embeddings: Optional[int] = None, lora_model_id: Optional[int] = None, device: str = "cuda", dtype: Optional[torch.dtype] = None, @@ -196,9 +195,7 @@ def from_local_checkpoint( lora_dir: The local path that has lora data. expected_lora_modules: Name of modules that are expected to be replaced by lora. - max_position_embeddings: Max position embedding length. Used to - scaling the largest context length. If None, the lora model's - context length is not scaled. + peft_helper: Loaded lora configuration information. lora_model_id: Lora model id. If not given, automatically set by a global counter. device: Device where the lora model is loaded. @@ -207,18 +204,12 @@ def from_local_checkpoint( Returns: Loaded LoRA Model. """ - lora_config_path = os.path.join(lora_dir, "adapter_config.json") lora_tensor_path = os.path.join(lora_dir, "adapter_model.safetensors") lora_bin_file_path = os.path.join(lora_dir, "adapter_model.bin") new_embeddings_tensor_path = os.path.join( lora_dir, "new_embeddings.safetensors") new_embeddings_bin_file_path = os.path.join(lora_dir, "new_embeddings.bin") - with open(lora_config_path) as f: - config = json.load(f) - - config["vllm_max_position_embeddings"] = max_position_embeddings - peft_helper = PEFTHelper.from_dict(config) unexpected_modules: List[Union[list[str], str]] if os.path.isfile(lora_tensor_path): tensors: Dict[str, torch.Tensor] = {} @@ -281,7 +272,8 @@ def from_local_checkpoint( new_embeddings_tensor_path) elif os.path.isfile(new_embeddings_bin_file_path): embeddings = torch.load(new_embeddings_bin_file_path, - map_location=device) + map_location=device, + weights_only=True) return cls.from_lora_tensors( lora_model_id=get_lora_id() diff --git a/vllm/lora/peft_helper.py b/vllm/lora/peft_helper.py index dacfb9ebd1480..b9c506f6e0bfd 100644 --- a/vllm/lora/peft_helper.py +++ b/vllm/lora/peft_helper.py @@ -1,9 +1,12 @@ # Adapted from: https://github.com/huggingface/peft/blob/main/src/peft/tuners/lora/config.py +import json import math +import os from dataclasses import MISSING, dataclass, field, fields -from typing import Literal, Optional, Union +from typing import List, Literal, Optional, Union +from vllm.config import LoRAConfig from vllm.logger import init_logger logger = init_logger(__name__) @@ -11,6 +14,12 @@ @dataclass class PEFTHelper: + """ + A helper class for PEFT configurations, specifically designed for LoRA. + This class handles configuration validation, compatibility checks for + various LoRA implementations. + """ + # Required fields r: int lora_alpha: int @@ -29,20 +38,18 @@ class PEFTHelper: vllm_max_position_embeddings: Optional[int] = field(default=False) vllm_long_context_scaling_factor: Optional[float] = field(default=None) - def _validate_features(self): + def _validate_features(self) -> List[str]: + """ + Check if there are any unsupported Lora features. + """ error_msg = [] - if self.modules_to_save: error_msg.append("vLLM only supports modules_to_save being None.") - if self.use_dora: error_msg.append("vLLM does not yet support DoRA.") - - if error_msg: - raise ValueError(f"{', '.join(error_msg)}") + return error_msg def __post_init__(self): - self._validate_features() if self.use_rslora: logger.info_once("Loading LoRA weights trained with rsLoRA.") self.vllm_lora_scaling_factor = self.lora_alpha / math.sqrt(self.r) @@ -78,3 +85,29 @@ def from_dict(cls, config_dict: dict) -> "PEFTHelper": for k, v in config_dict.items() if k in class_fields } return cls(**filtered_dict) + + @classmethod + def from_local_dir(cls, lora_path: str, + max_position_embeddings: Optional[int]) -> "PEFTHelper": + lora_config_path = os.path.join(lora_path, "adapter_config.json") + + with open(lora_config_path) as f: + config = json.load(f) + config["vllm_max_position_embeddings"] = max_position_embeddings + return cls.from_dict(config) + + def validate_legal(self, lora_config: LoRAConfig) -> None: + """ + Validates the LoRA configuration settings against application + constraints and requirements. + """ + error_msg = self._validate_features() + if self.r > lora_config.max_lora_rank: + error_msg.append( + f"LoRA rank {self.r} is greater than max_lora_rank" + f" {lora_config.max_lora_rank}.") + if self.bias != "none" and not lora_config.bias_enabled: + error_msg.append( + "Adapter bias cannot be used without bias_enabled.") + if error_msg: + raise ValueError(f"{' '.join(error_msg)}") diff --git a/vllm/lora/worker_manager.py b/vllm/lora/worker_manager.py index eec462743fe9d..a64296f7fd902 100644 --- a/vllm/lora/worker_manager.py +++ b/vllm/lora/worker_manager.py @@ -12,6 +12,7 @@ from vllm.logger import init_logger from vllm.lora.models import (LoRAModel, LoRAModelManager, LRUCacheLoRAModelManager, create_lora_manager) +from vllm.lora.peft_helper import PEFTHelper from vllm.lora.request import LoRARequest from vllm.lora.utils import get_adapter_absolute_path @@ -95,6 +96,13 @@ def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel: expected_lora_modules = list(set(expected_lora_modules)) lora_path = get_adapter_absolute_path(lora_request.lora_path) + peft_helper = PEFTHelper.from_local_dir( + lora_path, self.max_position_embeddings) + + # Validates the LoRA configuration against requirements before + # loading weights, throwing an exception if validation fails. + peft_helper.validate_legal(self.lora_config) + # For some models like Qwen2VL, we need to use hf_to_vllm_mapper # to ensure correct loading of lora weights. hf_to_vllm_mapper = None @@ -105,7 +113,7 @@ def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel: lora = self._lora_model_cls.from_local_checkpoint( lora_path, expected_lora_modules, - max_position_embeddings=self.max_position_embeddings, + peft_helper=peft_helper, lora_model_id=lora_request.lora_int_id, device="cpu", dtype=self.lora_config.lora_dtype, @@ -120,15 +128,14 @@ def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel: # - No adapter found to download from huggingface (or in # offline mode) # - No local adapter files found at `lora_request.lora_path` + # For NotFoundError raise ValueError( f"Loading lora {lora_request.lora_name} failed: No adapter " f"found for {lora_path}") from e except Exception as e: - raise RuntimeError(f"Loading lora {lora_path} failed") from e - if lora.rank > self.lora_config.max_lora_rank: - raise ValueError( - f"LoRA rank {lora.rank} is greater than max_lora_rank " - f"{self.lora_config.max_lora_rank}.") + # For BadRequestError + raise e + if lora.extra_vocab_size > self.lora_config.lora_extra_vocab_size: raise ValueError(f"LoRA added vocab size {lora.extra_vocab_size} " f"is greater than lora_extra_vocab_size " diff --git a/vllm/model_executor/guided_decoding/utils.py b/vllm/model_executor/guided_decoding/utils.py index 20abaefbacc51..90dfa62ec4670 100644 --- a/vllm/model_executor/guided_decoding/utils.py +++ b/vllm/model_executor/guided_decoding/utils.py @@ -20,6 +20,13 @@ def check_object(obj: dict) -> bool: ]): return True + # Check for array unsupported keywords + if obj.get("type") == "array" and any(key in obj for key in [ + "uniqueItems", "contains", "minContains", "maxContains", + "minItems", "maxItems" + ]): + return True + # Recursively check all nested objects and arrays for value in obj.values(): if isinstance(value, dict): diff --git a/vllm/model_executor/guided_decoding/xgrammar_decoding.py b/vllm/model_executor/guided_decoding/xgrammar_decoding.py index f10a8fb8e03cf..2d8594cb8aafa 100644 --- a/vllm/model_executor/guided_decoding/xgrammar_decoding.py +++ b/vllm/model_executor/guided_decoding/xgrammar_decoding.py @@ -298,8 +298,11 @@ def __call__(self, input_ids: list[int], # token_bitmask is a CPU tensor for use with accept_token and # fill_next_token_bitmask so we move it to the device of scores device_type = scores.device.type + dtype = scores.dtype if device_type != "cuda": - scores = scores.to("cpu").unsqueeze(0) + # xgrammar on cpu only supports float32 scores + # see: https://github.com/mlc-ai/xgrammar/blob/c1b64920cad24f44f235778c1c00bb52d57da01a/python/xgrammar/kernels/apply_token_bitmask_inplace_cpu.py#L22 + scores = scores.to("cpu").float().unsqueeze(0) # Note: In this method, if the tensors have different dimensions # on CPU device fails, but on GPU it runs without error. Hence the @@ -307,7 +310,7 @@ def __call__(self, input_ids: list[int], xgr.apply_token_bitmask_inplace(scores, self.token_bitmask.to(scores.device)) if device_type != "cuda": - scores = scores.to(device_type).squeeze() + scores = scores.to(dtype).to(device_type).squeeze() return scores diff --git a/vllm/model_executor/layers/activation.py b/vllm/model_executor/layers/activation.py index 1f2d9f6047780..a5a4571616a57 100644 --- a/vllm/model_executor/layers/activation.py +++ b/vllm/model_executor/layers/activation.py @@ -30,8 +30,10 @@ class FatreluAndMul(CustomOp): def __init__(self, threshold: float = 0.): super().__init__() self.threshold = threshold - if current_platform.is_cuda_alike() or current_platform.is_cpu(): + if current_platform.is_cuda_alike(): self.op = torch.ops._C.fatrelu_and_mul + elif current_platform.is_cpu(): + self._forward_method = self.forward_native def forward_native(self, x: torch.Tensor) -> torch.Tensor: d = x.shape[-1] // 2 @@ -97,6 +99,43 @@ def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: return out +@CustomOp.register("mul_and_silu") +class MulAndSilu(CustomOp): + """An activation function for SwiGLU. + + The function computes x -> x[:d] * silu(x[d:]) where d = x.shape[-1] // 2. + + Shapes: + x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d) + return: (num_tokens, d) or (batch_size, seq_len, d) + """ + + def __init__(self): + super().__init__() + if current_platform.is_cuda_alike(): + self.op = torch.ops._C.mul_and_silu + elif current_platform.is_xpu(): + from vllm._ipex_ops import ipex_ops + self.op = ipex_ops.silu_and_mul + elif current_platform.is_cpu(): + self._forward_method = self.forward_native + + def forward_native(self, x: torch.Tensor) -> torch.Tensor: + """PyTorch-native implementation equivalent to forward().""" + d = x.shape[-1] // 2 + return x[..., :d] * F.silu(x[..., d:]) + + def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: + d = x.shape[-1] // 2 + output_shape = (x.shape[:-1] + (d, )) + out = torch.empty(output_shape, dtype=x.dtype, device=x.device) + self.op(out, x) + return out + + # TODO implement forward_xpu for MulAndSilu + # def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: + + @CustomOp.register("gelu_and_mul") class GeluAndMul(CustomOp): """An activation function for GeGLU. diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json index 66aa2600226d4..b6f1d01f88652 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -10,25 +10,25 @@ }, "2": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 16, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, "4": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 32, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, "8": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 2, @@ -49,24 +49,24 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, "32": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 4, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, "48": { "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 256, - "GROUP_SIZE_M": 1, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 @@ -91,10 +91,10 @@ }, "128": { "BLOCK_SIZE_M": 64, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 4, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0 }, @@ -119,9 +119,9 @@ "1024": { "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 256, + "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, - "num_warps": 8, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json index 83be69c7e61fe..022d5ece7f87a 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json @@ -34,8 +34,8 @@ }, "8": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 32, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, "num_warps": 2, "num_stages": 2, @@ -56,8 +56,8 @@ }, "24": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 2, "num_stages": 2, @@ -67,10 +67,10 @@ }, "32": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 4, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json index 3ee1a5c267dce..d6ad63509f157 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json @@ -1,10 +1,10 @@ { "1": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 16, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -24,9 +24,9 @@ "4": { "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 256, + "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -35,7 +35,7 @@ "8": { "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, "num_warps": 2, "num_stages": 2, @@ -45,14 +45,14 @@ }, "16": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_N": 16, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 1, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 1 + "kpack": 2 }, "24": { "BLOCK_SIZE_M": 16, @@ -78,10 +78,10 @@ }, "48": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -112,7 +112,7 @@ "128": { "BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 256, + "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 4, "num_warps": 8, "num_stages": 2, @@ -190,7 +190,7 @@ "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, - "GROUP_SIZE_M": 8, + "GROUP_SIZE_M": 16, "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json index 90d0e6f6ba3fa..8323f512db015 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -28,10 +28,10 @@ }, "8": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0 }, @@ -64,18 +64,18 @@ }, "48": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0 }, "64": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 1, + "GROUP_SIZE_M": 4, "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 @@ -92,7 +92,7 @@ "128": { "BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 256, + "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 4, "num_warps": 8, "num_stages": 2, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json index 193782905a729..de2320e4b28cb 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json @@ -69,19 +69,19 @@ "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 1, + "GROUP_SIZE_M": 4, "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "48": { "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 4, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -100,9 +100,9 @@ }, "96": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 4, + "GROUP_SIZE_M": 1, "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, @@ -113,7 +113,7 @@ "BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 1, + "GROUP_SIZE_M": 4, "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, @@ -136,7 +136,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -147,7 +147,7 @@ "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json index 9d1b36acd64db..dbc1c189256c8 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -1,7 +1,7 @@ { "1": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_N": 16, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 4, @@ -20,7 +20,7 @@ "4": { "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 256, + "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 8, "num_stages": 2, @@ -31,7 +31,7 @@ "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, @@ -49,7 +49,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, @@ -58,7 +58,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, @@ -81,8 +81,8 @@ "waves_per_eu": 0 }, "96": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 4, @@ -90,18 +90,18 @@ "waves_per_eu": 0 }, "128": { - "BLOCK_SIZE_M": 64, - "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, - "GROUP_SIZE_M": 4, - "num_warps": 8, + "GROUP_SIZE_M": 1, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, "256": { "BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 2, @@ -126,11 +126,11 @@ "waves_per_eu": 0 }, "1536": { - "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0 }, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json index 2daaea099d092..811c77ab41093 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json @@ -8,7 +8,7 @@ "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "2": { "BLOCK_SIZE_M": 16, @@ -23,10 +23,10 @@ }, "4": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -44,11 +44,11 @@ "kpack": 2 }, "16": { - "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -63,12 +63,12 @@ "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "32": { "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 4, "num_warps": 4, "num_stages": 2, @@ -81,7 +81,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -92,7 +92,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 4, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -100,25 +100,25 @@ }, "96": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 1, - "num_warps": 8, + "GROUP_SIZE_M": 4, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "128": { "BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 4, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 1 + "kpack": 2 }, "256": { "BLOCK_SIZE_M": 64, @@ -133,13 +133,13 @@ }, "512": { "BLOCK_SIZE_M": 64, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, - "matrix_instr_nonkdim": 16, + "matrix_instr_nonkdim": 32, "kpack": 2 }, "1024": { diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json index 9e28dade2cee4..379ca107a9469 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -22,7 +22,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, @@ -31,7 +31,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, @@ -49,7 +49,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, @@ -58,7 +58,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, @@ -82,10 +82,10 @@ }, "96": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, @@ -121,23 +121,23 @@ "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 8, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, "1536": { - "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0 }, "2048": { "BLOCK_SIZE_M": 128, - "BLOCK_SIZE_N": 256, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 8, "num_stages": 2, @@ -154,10 +154,10 @@ }, "4096": { "BLOCK_SIZE_M": 128, - "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 8, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 } diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json index f885cd13a4adf..1df54c0052d6e 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json @@ -1,14 +1,14 @@ { "1": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 1 + "kpack": 2 }, "2": { "BLOCK_SIZE_M": 16, @@ -34,10 +34,10 @@ }, "8": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -56,32 +56,32 @@ }, "24": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "32": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 4, "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "48": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 8, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -107,18 +107,18 @@ "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 1 + "kpack": 2 }, "128": { "BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 1 + "kpack": 2 }, "256": { "BLOCK_SIZE_M": 64, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json index a971953062cf3..6a7d5df4d4623 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -13,7 +13,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0 }, @@ -57,35 +57,35 @@ "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 4, - "num_warps": 4, + "GROUP_SIZE_M": 1, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, "48": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0 }, "64": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, "96": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, @@ -117,11 +117,11 @@ "waves_per_eu": 0 }, "1024": { - "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0 }, @@ -137,9 +137,9 @@ "2048": { "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 256, + "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 8, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json index 4edf28f3e7c2e..a64d06c6d1724 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json @@ -23,14 +23,14 @@ }, "4": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "8": { "BLOCK_SIZE_M": 16, @@ -45,8 +45,8 @@ }, "16": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 2, @@ -55,15 +55,15 @@ "kpack": 2 }, "24": { - "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "32": { "BLOCK_SIZE_M": 16, @@ -85,13 +85,13 @@ "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 1 + "kpack": 2 }, "64": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, - "GROUP_SIZE_M": 1, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, @@ -121,14 +121,14 @@ "kpack": 1 }, "256": { - "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 64, - "GROUP_SIZE_M": 4, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, - "matrix_instr_nonkdim": 16, + "matrix_instr_nonkdim": 32, "kpack": 2 }, "512": { diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json index 1bad9550f0601..bd2c6fbc1b941 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -28,44 +28,44 @@ }, "8": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, "16": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, "24": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, "32": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 4, - "num_warps": 4, + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, "48": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 256, + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 2, @@ -73,25 +73,25 @@ }, "64": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 4, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, "96": { - "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, "128": { "BLOCK_SIZE_M": 64, - "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 4, "num_warps": 4, @@ -119,16 +119,16 @@ "1024": { "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 256, + "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 8, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, "1536": { - "BLOCK_SIZE_M": 256, - "BLOCK_SIZE_N": 256, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 8, "num_stages": 2, @@ -153,9 +153,9 @@ "waves_per_eu": 0 }, "4096": { - "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 8, "num_stages": 2, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json index f6d70ae78eab1..ac9a5593f9820 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json @@ -1,7 +1,7 @@ { "1": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 2, @@ -12,36 +12,36 @@ }, "2": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 1 + "kpack": 2 }, "4": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 1 + "kpack": 2 }, "8": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 1, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 1 + "kpack": 2 }, "16": { "BLOCK_SIZE_M": 16, @@ -56,10 +56,10 @@ }, "24": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 1, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -67,8 +67,8 @@ }, "32": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 4, "num_warps": 2, "num_stages": 2, @@ -77,11 +77,11 @@ "kpack": 2 }, "48": { - "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 64, + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -96,7 +96,7 @@ "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "96": { "BLOCK_SIZE_M": 32, @@ -114,7 +114,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 4, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -187,7 +187,7 @@ "kpack": 2 }, "4096": { - "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json index 91260051a5332..fc82cb162d726 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -13,7 +13,7 @@ "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 8, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, @@ -40,7 +40,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 1, "num_stages": 2, "waves_per_eu": 0 }, @@ -73,37 +73,37 @@ }, "64": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 1, - "num_warps": 2, + "GROUP_SIZE_M": 4, + "num_warps": 1, "num_stages": 2, "waves_per_eu": 0 }, "96": { - "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 256, - "GROUP_SIZE_M": 4, + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, "128": { "BLOCK_SIZE_M": 64, - "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 4, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0 }, "256": { "BLOCK_SIZE_M": 64, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 256, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 8, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0 }, @@ -121,14 +121,14 @@ "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0 }, "1536": { "BLOCK_SIZE_M": 128, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 8, "num_stages": 2, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json index d6220f55015d5..cf66868e9d57a 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json @@ -1,14 +1,14 @@ { "1": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 1, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "2": { "BLOCK_SIZE_M": 16, @@ -34,19 +34,19 @@ }, "8": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 1, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, - "kpack": 2 + "kpack": 1 }, "16": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, "num_warps": 2, "num_stages": 2, @@ -77,11 +77,11 @@ "kpack": 2 }, "48": { - "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, - "num_warps": 2, + "num_warps": 8, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -117,15 +117,15 @@ "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, - "matrix_instr_nonkdim": 32, + "matrix_instr_nonkdim": 16, "kpack": 2 }, "256": { "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, - "GROUP_SIZE_M": 4, - "num_warps": 8, + "GROUP_SIZE_M": 1, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, diff --git a/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Radeon_Graphics.json b/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Radeon_Graphics.json index d6220f55015d5..c07d5fe8e8ae6 100644 --- a/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Radeon_Graphics.json +++ b/vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Radeon_Graphics.json @@ -4,7 +4,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -26,7 +26,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -37,7 +37,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 2, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -70,7 +70,7 @@ "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 4, - "num_warps": 2, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -110,6 +110,17 @@ "kpack": 2 }, "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { "BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, @@ -136,7 +147,7 @@ "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, - "num_warps": 8, + "num_warps": 4, "num_stages": 2, "waves_per_eu": 0, "matrix_instr_nonkdim": 16, @@ -196,5 +207,38 @@ "waves_per_eu": 0, "matrix_instr_nonkdim": 16, "kpack": 2 + }, + "16384": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "18432": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "20480": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 } } diff --git a/vllm/model_executor/layers/fused_moe/fused_moe.py b/vllm/model_executor/layers/fused_moe/fused_moe.py index adca2864707d4..36529a94e2292 100644 --- a/vllm/model_executor/layers/fused_moe/fused_moe.py +++ b/vllm/model_executor/layers/fused_moe/fused_moe.py @@ -498,7 +498,10 @@ def grouped_topk(hidden_states: torch.Tensor, raise ValueError(f"Unsupported scoring function: {scoring_func}") if e_score_correction_bias is not None: - scores.add_(e_score_correction_bias.unsqueeze(0)) + # Store original scores before applying correction bias. We use biased + # scores for expert selection but original scores for routing weights + original_scores = scores + scores = scores + e_score_correction_bias.unsqueeze(0) num_token = scores.shape[0] group_scores = scores.view(num_token, num_expert_group, @@ -511,10 +514,16 @@ def grouped_topk(hidden_states: torch.Tensor, num_token, num_expert_group, scores.shape[-1] // num_expert_group).reshape(num_token, -1) # [n, e] tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e] - topk_weights, topk_ids = torch.topk(tmp_scores, - k=topk, - dim=-1, - sorted=False) + + if e_score_correction_bias is not None: + topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1] + # Use original unbiased scores for the routing weights + topk_weights = original_scores.gather(1, topk_ids) + else: + topk_weights, topk_ids = torch.topk(tmp_scores, + k=topk, + dim=-1, + sorted=False) if renormalize: topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True) @@ -870,4 +879,4 @@ def fused_moe( w2_scale=w2_scale, a1_scale=a1_scale, a2_scale=a2_scale, - block_shape=block_shape) \ No newline at end of file + block_shape=block_shape) diff --git a/vllm/model_executor/layers/fused_moe/layer.py b/vllm/model_executor/layers/fused_moe/layer.py index ce09a99dfa58d..e280b859f0391 100644 --- a/vllm/model_executor/layers/fused_moe/layer.py +++ b/vllm/model_executor/layers/fused_moe/layer.py @@ -40,7 +40,7 @@ class FusedMoEMethodBase(QuantizeMethodBase): @abstractmethod def create_weights(self, layer: torch.nn.Module, num_experts: int, - hidden_size: int, intermediate_size: int, + hidden_size: int, intermediate_size_per_partition: int, params_dtype: torch.dtype, **extra_weight_attrs): raise NotImplementedError @@ -80,22 +80,24 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp): """MoE method without quantization.""" def create_weights(self, layer: torch.nn.Module, num_experts: int, - hidden_size: int, intermediate_size: int, + hidden_size: int, intermediate_size_per_partition: int, params_dtype: torch.dtype, **extra_weight_attrs): # Fused gate_up_proj (column parallel) - w13_weight = torch.nn.Parameter(torch.empty(num_experts, - 2 * intermediate_size, - hidden_size, - dtype=params_dtype), + w13_weight = torch.nn.Parameter(torch.empty( + num_experts, + 2 * intermediate_size_per_partition, + hidden_size, + dtype=params_dtype), requires_grad=False) layer.register_parameter("w13_weight", w13_weight) set_weight_attrs(w13_weight, extra_weight_attrs) # down_proj (row parallel) - w2_weight = torch.nn.Parameter(torch.empty(num_experts, - hidden_size, - intermediate_size, - dtype=params_dtype), + w2_weight = torch.nn.Parameter(torch.empty( + num_experts, + hidden_size, + intermediate_size_per_partition, + dtype=params_dtype), requires_grad=False) layer.register_parameter("w2_weight", w2_weight) set_weight_attrs(w2_weight, extra_weight_attrs) @@ -315,13 +317,20 @@ def __init__( self.quant_method = quant_config.get_quant_method(self, prefix) assert self.quant_method is not None - self.quant_method.create_weights( - layer=self, - num_experts=num_experts, - hidden_size=hidden_size, - intermediate_size=self.intermediate_size_per_partition, - params_dtype=params_dtype, - weight_loader=self.weight_loader) + moe_quant_params = { + "num_experts": num_experts, + "hidden_size": hidden_size, + "intermediate_size_per_partition": + self.intermediate_size_per_partition, + "params_dtype": params_dtype, + "weight_loader": self.weight_loader, + } + # need full intermediate size pre-sharding for WNA16 act order + if (self.quant_method.__class__.__name__ == + "CompressedTensorsWNA16MoEMethod"): + moe_quant_params["intermediate_size_full"] = intermediate_size + + self.quant_method.create_weights(layer=self, **moe_quant_params) def _load_per_tensor_weight_scale(self, shard_id: str, param: torch.nn.Parameter, @@ -338,19 +347,30 @@ def _load_per_tensor_weight_scale(self, shard_id: str, elif shard_id == "w2": param_data[expert_id] = loaded_weight - def _load_model_weight_or_group_weight_scale(self, shard_dim: int, + def _load_model_weight_or_group_weight_scale(self, + shard_dim: int, expert_data: torch.Tensor, shard_id: str, loaded_weight: torch.Tensor, - tp_rank: int): - # Load grouped weight scales for group quantization - # or model weights + tp_rank: int, + load_full_w2: bool = False): + """ + Load grouped weight scales for group quantization or model weights + :param shard_dim: dimension to shard + :param expert_data: parameter for a particular expert + :param shard_id: either w1, w2, or w3 + :param loaded_weight: checkpoint weight to load into the param + :param tp_rank: tensor parallel rank + :param load_full_w2: whether or not the w2 loaded should be sharded. + """ if shard_id == "w2": - self._load_w2(shard_id=shard_id, - shard_dim=shard_dim, + # In the case where we have actorder/g_idx, we do not partition the + # w2 scales, as indicated by `load_full` argument, for all tp cases + self._load_w2(shard_dim=shard_dim, loaded_weight=loaded_weight, expert_data=expert_data, - tp_rank=tp_rank) + tp_rank=tp_rank, + load_full=load_full_w2) elif shard_id in ("w1", "w3"): self._load_w13(shard_id=shard_id, shard_dim=shard_dim, @@ -390,15 +410,21 @@ def _load_w13(self, expert_data: torch.Tensor, shard_dim: int, expert_data = expert_data.narrow(shard_dim, shard_size, shard_size) expert_data.copy_(loaded_weight) - def _load_w2(self, expert_data: torch.Tensor, shard_dim: int, - shard_id: str, loaded_weight: torch.Tensor, tp_rank: int): + def _load_w2(self, + expert_data: torch.Tensor, + shard_dim: int, + loaded_weight: torch.Tensor, + tp_rank: int, + load_full: bool = False): # Index the loaded weight for tp sharding. # down_proj: "RowParallel" so tp sharding on input_dim # Narrow parameter and load. shard_size = expert_data.shape[shard_dim] - loaded_weight = loaded_weight.narrow(shard_dim, shard_size * tp_rank, - shard_size) + if not load_full: + loaded_weight = loaded_weight.narrow(shard_dim, + shard_size * tp_rank, + shard_size) # w2, down_proj: Load into only logical weight of w2. expert_data.copy_(loaded_weight) @@ -413,8 +439,7 @@ def _load_g_idx(self, shard_id: str, expert_data: torch.Tensor, shard_dim: int, loaded_weight: torch.Tensor, tp_rank: int): if shard_id == "w2": - self._load_w2(shard_id=shard_id, - shard_dim=shard_dim, + self._load_w2(shard_dim=shard_dim, loaded_weight=loaded_weight, expert_data=expert_data, tp_rank=tp_rank) @@ -442,7 +467,7 @@ def weight_loader(self, param: torch.nn.Parameter, ] # Fetch the dim to shard the parameter/loaded weight # based on the shard id. This will be whatever - # dimension intermediate_size is used. + # dimension intermediate_size_per_partition is used. SHARD_ID_TO_SHARDED_DIM = {"w1": 0, "w2": 1, "w3": 0} expert_data = param.data[expert_id] @@ -450,11 +475,11 @@ def weight_loader(self, param: torch.nn.Parameter, # is_transposed: if the dim to shard the weight # should be flipped. Required by GPTQ, compressed-tensors - # should be whatever dimension intermediate_size is + # should be whatever dimension intermediate_size_per_partition is is_transposed = getattr(param, "is_transposed", False) shard_dim = SHARD_ID_TO_SHARDED_DIM[shard_id] if is_transposed: - shard_dim = ~shard_dim + shard_dim = int(not shard_dim) # Case input scale: input_scale loading is only supported for fp8 if "input_scale" in weight_name: @@ -506,7 +531,8 @@ def weight_loader(self, param: torch.nn.Parameter, shard_dim=shard_dim, loaded_weight=loaded_weight, expert_data=expert_data, - tp_rank=tp_rank) + tp_rank=tp_rank, + load_full_w2=getattr(param, "load_full_w2", False)) elif quant_method == FusedMoeWeightScaleSupported.TENSOR.value: self._load_per_tensor_weight_scale(shard_id=shard_id, param=param, diff --git a/vllm/model_executor/layers/linear.py b/vllm/model_executor/layers/linear.py index fdba34f760a9d..e8ced7a14ac7a 100644 --- a/vllm/model_executor/layers/linear.py +++ b/vllm/model_executor/layers/linear.py @@ -32,7 +32,7 @@ "MarlinLinearMethod", "QQQLinearMethod", "GPTQMarlin24LinearMethod", "TPUInt8LinearMethod", "GPTQLinearMethod", "FBGEMMFp8LinearMethod", "ModelOptFp8LinearMethod", "IPEXAWQLinearMethod", "IPEXGPTQLinearMethod", - "HQQMarlinMethod" + "HQQMarlinMethod", "QuarkLinearMethod" ] @@ -343,11 +343,13 @@ def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor): param.materialize(loaded_weight.shape, dtype=loaded_weight.dtype) use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False) + is_sharded_weight = getattr(param, "is_sharded_weight", False) + # bitsandbytes loads the weights of the specific portion + # no need to narrow + is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit param_data = param.data - # bitsandbytes loads the weights of the specific portion - # no need to narrow here - if output_dim is not None and not use_bitsandbytes_4bit: + if output_dim is not None and not is_sharded_weight: shard_size = param_data.shape[output_dim] start_idx = tp_rank * shard_size loaded_weight = loaded_weight.narrow(output_dim, start_idx, @@ -545,6 +547,11 @@ def weight_loader(self, use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False) + is_sharded_weight = getattr(param, "is_sharded_weight", False) + # bitsandbytes loads the weights of the specific portion + # no need to narrow + is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit + if use_bitsandbytes_4bit: shard_size = loaded_weight.shape[output_dim] shard_offset = loaded_weight.shape[output_dim] * \ @@ -553,9 +560,7 @@ def weight_loader(self, param_data = param_data.narrow(output_dim, shard_offset, shard_size) start_idx = tp_rank * shard_size - # bitsandbytes loads the weights of the specific portion - # no need to narrow here - if not use_bitsandbytes_4bit: + if not is_sharded_weight: loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size) # Special case for AQLM codebooks. @@ -940,6 +945,11 @@ def weight_loader(self, use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False) + is_sharded_weight = getattr(param, "is_sharded_weight", False) + # bitsandbytes loads the weights of the specific portion + # no need to narrow + is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit + if use_bitsandbytes_4bit: orig_qkv_offsets = { "q": (0, self.num_heads * self.head_size), @@ -963,9 +973,7 @@ def weight_loader(self, shard_id = tp_rank // self.num_kv_head_replicas start_idx = shard_id * shard_size - # bitsandbytes loads the weights of the specific portion - # no need to narrow here - if not use_bitsandbytes_4bit: + if not is_sharded_weight: loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size) @@ -1069,6 +1077,10 @@ def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor): tp_size = get_tensor_model_parallel_world_size() input_dim = getattr(param, "input_dim", None) use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False) + is_sharded_weight = getattr(param, "is_sharded_weight", False) + # bitsandbytes loads the weights of the specific portion + # no need to narrow + is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit # Special case for GGUF is_gguf_weight = getattr(param, "is_gguf_weight", False) @@ -1084,9 +1096,7 @@ def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor): param.materialize(tuple(weight_shape), dtype=loaded_weight.dtype) param_data = param.data - # bitsandbytes loads the weights of the specific portion - # no need to narrow here - if input_dim is not None and not use_bitsandbytes_4bit: + if input_dim is not None and not is_sharded_weight: shard_size = param_data.shape[input_dim] start_idx = tp_rank * shard_size loaded_weight = loaded_weight.narrow(input_dim, start_idx, diff --git a/vllm/model_executor/layers/logits_processor.py b/vllm/model_executor/layers/logits_processor.py index 2bc7e458494f7..42decde1d0f79 100644 --- a/vllm/model_executor/layers/logits_processor.py +++ b/vllm/model_executor/layers/logits_processor.py @@ -6,6 +6,7 @@ import torch.nn as nn import vllm.envs as envs +from vllm.config import get_current_vllm_config from vllm.distributed import (tensor_model_parallel_all_gather, tensor_model_parallel_gather) from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -44,8 +45,10 @@ def __init__(self, self.soft_cap = soft_cap # Whether to use gather or all-gather to gather the logits. - self.use_gather = not current_platform.is_tpu( - ) and not envs.VLLM_USE_V1 + parallel_config = get_current_vllm_config().parallel_config + self.use_all_gather = current_platform.is_tpu() \ + or envs.VLLM_USE_V1 \ + or parallel_config.distributed_executor_backend == "external_launcher" # noqa def forward( self, @@ -88,16 +91,17 @@ def _get_logits( logits = lm_head.linear_method.apply(lm_head, hidden_states, bias=embedding_bias) - if self.use_gather: - # None may be returned for rank > 0 - logits = tensor_model_parallel_gather(logits) - else: + + if self.use_all_gather: # Gather is not supported for some devices such as TPUs. # Use all-gather instead. # NOTE(woosuk): Here, the outputs of every device should not be None # because XLA requires strict SPMD among all devices. Every device # should execute the same operations after gathering the logits. logits = tensor_model_parallel_all_gather(logits) + else: + # None may be returned for rank > 0 + logits = tensor_model_parallel_gather(logits) # Remove paddings in vocab (if any). if logits is not None: logits = logits[..., :self.org_vocab_size] diff --git a/vllm/model_executor/layers/quantization/__init__.py b/vllm/model_executor/layers/quantization/__init__.py index dd10c434f0752..d2bde13fcf546 100644 --- a/vllm/model_executor/layers/quantization/__init__.py +++ b/vllm/model_executor/layers/quantization/__init__.py @@ -26,14 +26,56 @@ "experts_int8", "neuron_quant", "ipex", + "quark" ] +# The customized quantization methods which will be added to this dict. +_CUSTOMIZED_METHOD_TO_QUANT_CONFIG = {} + + +def register_quantization_config(quantization: str): + """Register a customized vllm quantization config. + + When a quantization method is not supported by vllm, you can register a customized + quantization config to support it. + + Args: + quantization (str): The quantization method name. + + Examples: + >>> from vllm.model_executor.layers.quantization import register_quantization_config + >>> from vllm.model_executor.layers.quantization import get_quantization_config + >>> from vllm.model_executor.layers.quantization.base_config import QuantizationConfig + >>> + >>> @register_quantization_config("my_quant") + ... class MyQuantConfig(QuantizationConfig): + ... pass + >>> + >>> get_quantization_config("my_quant") + + """ # noqa: E501 + + def _wrapper(quant_config_cls): + if quantization in QUANTIZATION_METHODS: + raise ValueError( + f"The quantization method `{quantization}` is already exists.") + if not issubclass(quant_config_cls, QuantizationConfig): + raise ValueError("The quantization config must be a subclass of " + "`QuantizationConfig`.") + _CUSTOMIZED_METHOD_TO_QUANT_CONFIG[quantization] = quant_config_cls + QUANTIZATION_METHODS.append(quantization) + return quant_config_cls + + return _wrapper + def get_quantization_config(quantization: str) -> Type[QuantizationConfig]: if quantization not in QUANTIZATION_METHODS: raise ValueError(f"Invalid quantization method: {quantization}") # lazy import to avoid triggering `torch.compile` too early + from vllm.model_executor.layers.quantization.quark.quark import QuarkConfig + from .aqlm import AQLMConfig from .awq import AWQConfig from .awq_marlin import AWQMarlinConfig @@ -79,7 +121,10 @@ def get_quantization_config(quantization: str) -> Type[QuantizationConfig]: "experts_int8": ExpertsInt8Config, "neuron_quant": NeuronQuantConfig, "ipex": IPEXConfig, + "quark": QuarkConfig } + # Update the `method_to_config` with customized quantization methods. + method_to_config.update(_CUSTOMIZED_METHOD_TO_QUANT_CONFIG) return method_to_config[quantization] diff --git a/vllm/model_executor/layers/quantization/awq_marlin.py b/vllm/model_executor/layers/quantization/awq_marlin.py index c28fd0c6737e0..0c3c9816878e9 100644 --- a/vllm/model_executor/layers/quantization/awq_marlin.py +++ b/vllm/model_executor/layers/quantization/awq_marlin.py @@ -303,7 +303,7 @@ def __init__(self, quant_config: AWQMarlinConfig): self.quant_config = quant_config def create_weights(self, layer: torch.nn.Module, num_experts: int, - hidden_size: int, intermediate_size: int, + hidden_size: int, intermediate_size_per_partition: int, params_dtype: torch.dtype, **extra_weight_attrs): extra_weight_attrs.update({ "is_transposed": @@ -312,17 +312,18 @@ def create_weights(self, layer: torch.nn.Module, num_experts: int, FusedMoeWeightScaleSupported.GROUP.value, }) - w13_qweight = Parameter(torch.empty(num_experts, - hidden_size, - 2 * intermediate_size // - self.quant_config.pack_factor, - dtype=torch.int32), - requires_grad=False) + w13_qweight = Parameter( + torch.empty(num_experts, + hidden_size, + 2 * intermediate_size_per_partition // + self.quant_config.pack_factor, + dtype=torch.int32), + requires_grad=False) layer.register_parameter("w13_qweight", w13_qweight) set_weight_attrs(w13_qweight, extra_weight_attrs) w2_qweight = Parameter(torch.empty(num_experts, - intermediate_size, + intermediate_size_per_partition, hidden_size // self.quant_config.pack_factor, dtype=torch.int32), @@ -331,13 +332,14 @@ def create_weights(self, layer: torch.nn.Module, num_experts: int, set_weight_attrs(w2_qweight, extra_weight_attrs) num_groups_w13 = hidden_size // self.quant_config.group_size - num_groups_w2 = intermediate_size // self.quant_config.group_size + num_groups_w2 = (intermediate_size_per_partition // + self.quant_config.group_size) # WEIGHT_SCALES # Allocate 2 scales for w1 and w3 respectively. w13_scales = Parameter(torch.empty(num_experts, num_groups_w13, - intermediate_size * 2, + intermediate_size_per_partition * 2, dtype=params_dtype), requires_grad=False) layer.register_parameter("w13_scales", w13_scales) @@ -353,12 +355,13 @@ def create_weights(self, layer: torch.nn.Module, num_experts: int, # WEIGHT_ZERO_POINT # Allocate 2 zero points for w1 and w3 respectively. - w13_qzeros = Parameter(torch.empty(num_experts, - num_groups_w13, - 2 * intermediate_size // - self.quant_config.pack_factor, - dtype=torch.int32), - requires_grad=False) + w13_qzeros = Parameter( + torch.empty(num_experts, + num_groups_w13, + 2 * intermediate_size_per_partition // + self.quant_config.pack_factor, + dtype=torch.int32), + requires_grad=False) layer.register_parameter("w13_qzeros", w13_qzeros) set_weight_attrs(w13_qzeros, extra_weight_attrs) diff --git a/vllm/model_executor/layers/quantization/base_config.py b/vllm/model_executor/layers/quantization/base_config.py index 6dfac8aad5358..2fb2642dd5156 100644 --- a/vllm/model_executor/layers/quantization/base_config.py +++ b/vllm/model_executor/layers/quantization/base_config.py @@ -133,3 +133,6 @@ def get_quant_method(self, layer: torch.nn.Module, method. """ raise NotImplementedError + + def get_cache_scale(self, name: str) -> Optional[str]: + return None diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py index 0c1fc18228f5c..48214393381f8 100644 --- a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py +++ b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py @@ -9,6 +9,7 @@ QuantizationType) from pydantic import BaseModel +from vllm.logger import init_logger from vllm.model_executor.layers.fused_moe import FusedMoE from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase, UnquantizedLinearMethod) @@ -27,6 +28,8 @@ from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod from vllm.platforms import current_platform +logger = init_logger(__name__) + __all__ = ["CompressedTensorsLinearMethod"] SPARSITY_CONFIG_NAME: Literal["sparsity_config"] = "sparsity_config" @@ -79,6 +82,8 @@ def get_quant_method( return UnquantizedLinearMethod() if isinstance(layer, LinearBase): scheme = self.get_scheme(layer=layer, layer_name=prefix) + if scheme is None: + return UnquantizedLinearMethod() layer.scheme = scheme return CompressedTensorsLinearMethod(self) if isinstance(layer, Attention): @@ -113,7 +118,7 @@ def _sparsity_scheme_map_from_config( :return: A dictionary mapping target layer names to their corresponding sparsity compression configurations """ - if (sparsity_config := config.get(SPARSITY_CONFIG_NAME)) is None: + if not (sparsity_config := config.get(SPARSITY_CONFIG_NAME)): return dict() sparsity_config = SparsityCompressionConfig.model_validate( @@ -340,10 +345,10 @@ def _get_scheme_from_parts( raise NotImplementedError( "No compressed-tensors compatible scheme was found.") - def get_scheme( - self, - layer: torch.nn.Module, - layer_name: Optional[str] = None) -> "CompressedTensorsScheme": + def get_scheme(self, + layer: torch.nn.Module, + layer_name: Optional[str] = None + ) -> Optional["CompressedTensorsScheme"]: """ compressed-tensors supports non uniform in the following way: @@ -353,10 +358,7 @@ def get_scheme( which can be a full layer_name, a regex for a layer_name, or an nn.Module name. - We first check whether a layer is in the ignore group and use - CompressedTensorsUnquantized (i.e. fp16/bf16) scheme for the layer - - We then detect whether a layer_name is found in any target and + Detect whether a layer_name is found in any target and use the quantization scheme corresponding to the matched target to select the CompressedTensorsScheme used for infernece. """ @@ -394,6 +396,13 @@ def get_scheme( if self.supports_cutlass_24(weight_quant=weight_quant, input_quant=input_quant, sparsity_scheme=sparsity_scheme): + # FIXME(tlrmchlsmth): layers using W16A16 CUTLASS 2:4 sparse kernels + # currently produce bad output in some cases + if weight_quant is None: + logger.warning_once( + "CompressedTensors24 scheme is disabled for the w16a16 " + "case. Falling back to UnquantizedLinearMethod") + return None # Have a valid sparsity scheme # Validate layer is supported by Cutlass 2:4 Kernel scheme = CompressedTensors24(quantized=weight_quant is not None @@ -412,6 +421,26 @@ def get_scheme( self._check_scheme_supported(scheme.get_min_capability()) return scheme + def get_cache_scale(self, name: str) -> Optional[str]: + """ + Check whether the param name matches the format for k/v cache scales + in compressed-tensors. If this is the case, return its equivalent + param name expected by vLLM + + :param name: param name + :return: matching param name for KV cache scale in vLLM + """ + if name.endswith(".output_scale") and ".k_proj" in name: + return name.replace(".k_proj.output_scale", ".attn.k_scale") + if name.endswith(".output_scale") and ".v_proj" in name: + return name.replace(".v_proj.output_scale", ".attn.v_scale") + if name.endswith(".output_scale") and ".q_proj" in name: + return name.replace(".q_proj.output_scale", ".attn.q_scale") + if name.endswith("self_attn.prob_output_scale"): + return name.replace(".prob_output_scale", ".attn.prob_scale") + # If no matches, return None + return None + @staticmethod def supports_cutlass_24( weight_quant: Optional[QuantizationArgs], diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py index 4fb8fd84e92d4..e1c45f4e42e41 100644 --- a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py +++ b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py @@ -13,6 +13,7 @@ FusedMoeWeightScaleSupported) from vllm.model_executor.layers.quantization.compressed_tensors.schemes import ( WNA16_SUPPORTED_BITS) +from vllm.model_executor.layers.quantization.utils import replace_parameter from vllm.model_executor.layers.quantization.utils.w8a8_utils import ( all_close_1d, normalize_e4m3fn_to_e4m3fnuz, per_tensor_dequantize) from vllm.model_executor.utils import set_weight_attrs @@ -75,24 +76,26 @@ def __init__( self.static_input_scales = not self.input_quant.dynamic def create_weights(self, layer: torch.nn.Module, num_experts: int, - hidden_size: int, intermediate_size: int, + hidden_size: int, intermediate_size_per_partition: int, params_dtype: torch.dtype, **extra_weight_attrs): params_dtype = torch.float8_e4m3fn # WEIGHTS - w13_weight = torch.nn.Parameter(torch.empty(num_experts, - 2 * intermediate_size, - hidden_size, - dtype=params_dtype), + w13_weight = torch.nn.Parameter(torch.empty( + num_experts, + 2 * intermediate_size_per_partition, + hidden_size, + dtype=params_dtype), requires_grad=False) layer.register_parameter("w13_weight", w13_weight) set_weight_attrs(w13_weight, extra_weight_attrs) - w2_weight = torch.nn.Parameter(torch.empty(num_experts, - hidden_size, - intermediate_size, - dtype=params_dtype), + w2_weight = torch.nn.Parameter(torch.empty( + num_experts, + hidden_size, + intermediate_size_per_partition, + dtype=params_dtype), requires_grad=False) layer.register_parameter("w2_weight", w2_weight) set_weight_attrs(w2_weight, extra_weight_attrs) @@ -254,6 +257,7 @@ def __init__( self.packed_factor = 32 // config.num_bits self.strategy = config.strategy self.group_size = config.group_size + self.actorder = config.actorder assert config.symmetric, ( "Only symmetric quantization is supported for MoE") @@ -266,9 +270,16 @@ def __init__( f"{WNA16_SUPPORTED_BITS}") def create_weights(self, layer: torch.nn.Module, num_experts: int, - hidden_size: int, intermediate_size: int, + hidden_size: int, intermediate_size_per_partition: int, params_dtype: torch.dtype, **extra_weight_attrs): + assert params_dtype == torch.float16, ( + "float16 is required for MoE compressed models. Set dtype=torch.float16" # noqa: E501 + ) + + intermediate_size_full = extra_weight_attrs.pop( + "intermediate_size_full") + # Will transpose the loaded weight along the # intermediate and hidden dim sizes. Will # shard for TP along the transposed dims @@ -276,35 +287,45 @@ def create_weights(self, layer: torch.nn.Module, num_experts: int, "is_transposed": True, "quant_method": self.strategy }) - w13_weight = torch.nn.Parameter(torch.empty(num_experts, - hidden_size // - self.packed_factor, - 2 * intermediate_size, - dtype=torch.int32), + w13_weight = torch.nn.Parameter(torch.empty( + num_experts, + hidden_size // self.packed_factor, + 2 * intermediate_size_per_partition, + dtype=torch.int32), requires_grad=False) layer.register_parameter("w13_weight_packed", w13_weight) set_weight_attrs(w13_weight, extra_weight_attrs) - w2_weight = torch.nn.Parameter(torch.empty(num_experts, - intermediate_size // - self.packed_factor, - hidden_size, - dtype=torch.int32), + w2_weight = torch.nn.Parameter(torch.empty( + num_experts, + intermediate_size_per_partition // self.packed_factor, + hidden_size, + dtype=torch.int32), requires_grad=False) layer.register_parameter("w2_weight_packed", w2_weight) set_weight_attrs(w2_weight, extra_weight_attrs) + # In the case where we have actorder/g_idx, + # we do not partition the w2 scales + load_full_w2 = self.actorder and self.group_size != -1 + w2_scales_size = (intermediate_size_full + if load_full_w2 else intermediate_size_per_partition) + + self.is_k_full = (not self.actorder) or ( + intermediate_size_per_partition == intermediate_size_full) + if self.strategy == "channel": num_groups_w2 = num_groups_w13 = 1 self.group_size = -1 else: - num_groups_w2 = intermediate_size // self.group_size + num_groups_w2 = w2_scales_size // self.group_size num_groups_w13 = hidden_size // self.group_size - w13_scale = torch.nn.Parameter(torch.ones(num_experts, - num_groups_w13, - 2 * intermediate_size, - dtype=params_dtype), + w13_scale = torch.nn.Parameter(torch.ones( + num_experts, + num_groups_w13, + 2 * intermediate_size_per_partition, + dtype=params_dtype), requires_grad=False) layer.register_parameter("w13_weight_scale", w13_scale) set_weight_attrs(w13_scale, extra_weight_attrs) @@ -316,6 +337,7 @@ def create_weights(self, layer: torch.nn.Module, num_experts: int, requires_grad=False) layer.register_parameter("w2_weight_scale", w2_scale) set_weight_attrs(w2_scale, extra_weight_attrs) + set_weight_attrs(w2_scale, {"load_full_w2": load_full_w2}) w2_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2), requires_grad=False) @@ -335,18 +357,18 @@ def create_weights(self, layer: torch.nn.Module, num_experts: int, ), requires_grad=False, ) - layer.register_parameter("w13_g_idx", w13_g_idx) + layer.register_parameter("w13_weight_g_idx", w13_g_idx) set_weight_attrs(w13_g_idx, extra_weight_attrs) w2_g_idx = torch.nn.Parameter( torch.empty( num_experts, - intermediate_size, + intermediate_size_per_partition, dtype=torch.int32, ), requires_grad=False, ) - layer.register_parameter("w2_g_idx", w2_g_idx) + layer.register_parameter("w2_weight_g_idx", w2_g_idx) set_weight_attrs(w2_g_idx, extra_weight_attrs) w13_g_idx_sort_indices = torch.nn.Parameter( @@ -364,7 +386,7 @@ def create_weights(self, layer: torch.nn.Module, num_experts: int, w2_g_idx_sort_indices = torch.nn.Parameter( torch.empty( num_experts, - intermediate_size, + intermediate_size_per_partition, dtype=torch.int32, ), requires_grad=False, @@ -422,24 +444,55 @@ def marlin_moe_permute_scales(s: torch.Tensor, size_k: int, size_k2 = layer.w2_weight_packed.shape[2] size_k13 = layer.w13_weight_packed.shape[2] - num_experts = layer.w13_g_idx.shape[0] - device = layer.w13_g_idx.device - layer.w13_g_idx = torch.nn.Parameter( - torch.empty((num_experts, 0), dtype=torch.int32, device=device), - requires_grad=False, - ) - layer.w2_g_idx = torch.nn.Parameter( - torch.empty((num_experts, 0), dtype=torch.int32, device=device), - requires_grad=False, - ) - layer.w13_g_idx_sort_indices = torch.nn.Parameter( - torch.empty((num_experts, 0), dtype=torch.int32, device=device), - requires_grad=False, - ) - layer.w2_g_idx_sort_indices = torch.nn.Parameter( - torch.empty((num_experts, 0), dtype=torch.int32, device=device), - requires_grad=False, - ) + num_experts = layer.w13_weight_g_idx.shape[0] + device = layer.w13_weight_g_idx.device + + # when running models with grouped act order, + # resort to g_idx values provided in checkpoint + if self.actorder == "group": + w13_g_idx_sort_indices = torch.empty_like(layer.w13_weight_g_idx) + w2_g_idx_sort_indices = torch.empty_like(layer.w2_weight_g_idx) + w13_sorted_g_idx = torch.empty_like(layer.w13_weight_g_idx) + w2_sorted_g_idx = torch.empty_like(layer.w2_weight_g_idx) + + for e in range(num_experts): + w13_g_idx_sort_indices[e] = torch.argsort( + layer.w13_weight_g_idx[e]).to(torch.int32) + w2_g_idx_sort_indices[e] = torch.argsort( + layer.w2_weight_g_idx[e]).to(torch.int32) + w13_sorted_g_idx[e] = layer.w13_weight_g_idx[e][ + w13_g_idx_sort_indices[e]] + w2_sorted_g_idx[e] = layer.w2_weight_g_idx[e][ + w2_g_idx_sort_indices[e]] + + replace_parameter(layer, "w13_weight_g_idx", w13_sorted_g_idx) + replace_parameter(layer, "w2_weight_g_idx", w2_sorted_g_idx) + replace_parameter(layer, "w13_g_idx_sort_indices", + w13_g_idx_sort_indices) + replace_parameter(layer, "w2_g_idx_sort_indices", + w2_g_idx_sort_indices) + + else: + layer.w13_weight_g_idx = torch.nn.Parameter( + torch.empty((num_experts, 0), dtype=torch.int32, + device=device), + requires_grad=False, + ) + layer.w2_weight_g_idx = torch.nn.Parameter( + torch.empty((num_experts, 0), dtype=torch.int32, + device=device), + requires_grad=False, + ) + layer.w13_g_idx_sort_indices = torch.nn.Parameter( + torch.empty((num_experts, 0), dtype=torch.int32, + device=device), + requires_grad=False, + ) + layer.w2_g_idx_sort_indices = torch.nn.Parameter( + torch.empty((num_experts, 0), dtype=torch.int32, + device=device), + requires_grad=False, + ) marlin_w13_qweight = ops.gptq_marlin_moe_repack( layer.w13_weight_packed, @@ -511,9 +564,9 @@ def apply( router_logits, topk_weights, topk_ids, - g_idx1=layer.w13_g_idx, - g_idx2=layer.w2_g_idx, + g_idx1=layer.w13_weight_g_idx, + g_idx2=layer.w2_weight_g_idx, sort_indices1=layer.w13_g_idx_sort_indices, sort_indices2=layer.w2_g_idx_sort_indices, num_bits=self.num_bits, - ) + is_k_full=self.is_k_full) diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_24.py b/vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_24.py index bc697ef93b34b..21e6fe7a22616 100644 --- a/vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_24.py +++ b/vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_24.py @@ -42,7 +42,7 @@ def create_weights(self, layer: torch.nn.Module, input_size: int, if not sparse_cutlass_supported(): raise ValueError( - "Sparse CUTLASS not supported. vLLM must be built with" + "Sparse CUTLASS not supported. vLLM must be built with " "CUDA 12.2 or later to use this feature") self.output_dtype = params_dtype diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_24.py b/vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_24.py index 61d1c911cd1ad..2e1b5e3c2d3b1 100644 --- a/vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_24.py +++ b/vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_24.py @@ -62,7 +62,7 @@ def create_weights(self, layer: torch.nn.Module, input_size: int, **kwargs): assert params_dtype == torch.float16, ( - "float16 is required for marlin24 compressd models. Set dtype=torch.float16" # noqa: E501 + "float16 is required for marlin24 compressed models. Set dtype=torch.float16" # noqa: E501 ) pack_factor = 32 // self.quant_type.size_bits diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/triton_scaled_mm.py b/vllm/model_executor/layers/quantization/compressed_tensors/triton_scaled_mm.py index 2659afcdc74a9..f4c1dbc0361c6 100644 --- a/vllm/model_executor/layers/quantization/compressed_tensors/triton_scaled_mm.py +++ b/vllm/model_executor/layers/quantization/compressed_tensors/triton_scaled_mm.py @@ -136,6 +136,10 @@ def triton_scaled_mm(input: torch.Tensor, assert N > 0 and K > 0 and M > 0 assert weight.shape[0] == K assert input.dtype == weight.dtype + + scale_a = scale_a.reshape(-1, 1) if scale_a.dim() <= 1 else scale_a + scale_b = scale_b.reshape(-1, 1) if scale_b.dim() <= 1 else scale_b + assert scale_a.dtype == scale_b.dtype and scale_a.is_floating_point() assert scale_a.shape == torch.Size([1, 1]) or scale_a.shape == torch.Size( [M, 1]) diff --git a/vllm/model_executor/layers/quantization/experts_int8.py b/vllm/model_executor/layers/quantization/experts_int8.py index 209f12c6dfec9..100cbfa4c9598 100644 --- a/vllm/model_executor/layers/quantization/experts_int8.py +++ b/vllm/model_executor/layers/quantization/experts_int8.py @@ -52,7 +52,7 @@ def __init__(self, quant_config: ExpertsInt8Config): self.quant_config = quant_config def create_weights(self, layer: torch.nn.Module, num_experts: int, - hidden_size: int, intermediate_size: int, + hidden_size: int, intermediate_size_per_partition: int, params_dtype: torch.dtype, **extra_weight_attrs): int8_dtype = torch.int8 @@ -64,26 +64,29 @@ def create_weights(self, layer: torch.nn.Module, num_experts: int, extra_weight_attrs['weight_loader'] = wrapped_weight_loader # Fused gate_up_proj (column parallel) - w13_weight = torch.nn.Parameter(torch.empty(num_experts, - 2 * intermediate_size, - hidden_size, - dtype=int8_dtype), + w13_weight = torch.nn.Parameter(torch.empty( + num_experts, + 2 * intermediate_size_per_partition, + hidden_size, + dtype=int8_dtype), requires_grad=False) layer.register_parameter("w13_weight", w13_weight) set_weight_attrs(w13_weight, extra_weight_attrs) # down_proj (row parallel) - w2_weight = torch.nn.Parameter(torch.empty(num_experts, - hidden_size, - intermediate_size, - dtype=int8_dtype), + w2_weight = torch.nn.Parameter(torch.empty( + num_experts, + hidden_size, + intermediate_size_per_partition, + dtype=int8_dtype), requires_grad=False) layer.register_parameter("w2_weight", w2_weight) set_weight_attrs(w2_weight, extra_weight_attrs) - w13_scale = torch.nn.Parameter(torch.zeros(num_experts, - 2 * intermediate_size, - dtype=torch.float32), + w13_scale = torch.nn.Parameter(torch.zeros( + num_experts, + 2 * intermediate_size_per_partition, + dtype=torch.float32), requires_grad=False) layer.register_parameter("w13_scale", w13_scale) diff --git a/vllm/model_executor/layers/quantization/fp8.py b/vllm/model_executor/layers/quantization/fp8.py index 9b0c3b06649f1..a29239b2badde 100644 --- a/vllm/model_executor/layers/quantization/fp8.py +++ b/vllm/model_executor/layers/quantization/fp8.py @@ -113,6 +113,26 @@ def get_quant_method(self, layer: torch.nn.Module, return Fp8KVCacheMethod(self) return None + def get_cache_scale(self, name: str) -> Optional[str]: + """ + Check whether the param name matches the format for k/v cache scales + in compressed-tensors. If this is the case, return its equivalent + param name expected by vLLM + + :param name: param name + :return: matching param name for KV cache scale in vLLM + """ + if name.endswith(".output_scale") and ".k_proj" in name: + return name.replace(".k_proj.output_scale", ".attn.k_scale") + if name.endswith(".output_scale") and ".v_proj" in name: + return name.replace(".v_proj.output_scale", ".attn.v_scale") + if name.endswith(".output_scale") and ".q_proj" in name: + return name.replace(".q_proj.output_scale", ".attn.q_scale") + if name.endswith("self_attn.prob_output_scale"): + return name.replace(".prob_output_scale", ".attn.prob_scale") + # If no matches, return None + return None + class Fp8LinearMethod(LinearMethodBase): """Linear method for FP8. @@ -145,7 +165,10 @@ def __init__(self, quant_config: Fp8Config): if current_platform.is_rocm(): self.use_marlin = False - self.default_scale = torch.finfo(torch.float32).min + self.block_quant = self.quant_config.weight_block_size is not None + if self.block_quant: + # Marlin doesn't support block-wise fp8 + self.use_marlin = False self.block_quant = self.quant_config.weight_block_size is not None if self.block_quant: @@ -166,6 +189,8 @@ def create_weights( weight_loader = extra_weight_attrs.get("weight_loader") if self.block_quant: + assert not envs.VLLM_FP8_PADDING, ( + "FP8 weight padding is not supported in block quantization.") tp_size = get_tensor_model_parallel_world_size() assert self.quant_config.weight_block_size is not None block_n, block_k = ( @@ -220,7 +245,7 @@ def create_weights( dtype=torch.float32), weight_loader=weight_loader, ) - scale[:] = self.default_scale + scale[:] = torch.finfo(torch.float32).min layer.register_parameter("weight_scale", scale) else: assert self.quant_config.activation_scheme == "dynamic" @@ -234,7 +259,7 @@ def create_weights( output_dim=0, weight_loader=weight_loader, ) - scale[:] = self.default_scale + scale[:] = torch.finfo(torch.float32).min # The weight_scale_inv name is intentional for deepseekv3 layer.register_parameter("weight_scale_inv", scale) @@ -244,7 +269,7 @@ def create_weights( len(output_partition_sizes), dtype=torch.float32), weight_loader=weight_loader) - scale[:] = self.default_scale + scale[:] = torch.finfo(torch.float32).min layer.register_parameter("input_scale", scale) else: layer.register_parameter("input_scale", None) @@ -252,6 +277,15 @@ def create_weights( def process_weights_after_loading(self, layer: Module) -> None: # Block quant doesn't need to process weights after loading if self.block_quant: + if current_platform.is_rocm() and not is_navi(): + weight, weight_scale, _ = \ + normalize_e4m3fn_to_e4m3fnuz( + weight=layer.weight, + weight_scale=layer.weight_scale_inv, + input_scale=layer.input_scale) + layer.weight = Parameter(weight, requires_grad=False) + layer.weight_scale_inv = Parameter(weight_scale, + requires_grad=False) return layer.weight = torch.nn.Parameter(layer.weight.data, requires_grad=False) @@ -276,13 +310,13 @@ def process_weights_after_loading(self, layer: Module) -> None: # If checkpoint is fp8, handle that there are N scales for N # shards in a fused module else: - layer.weight_scale.data[layer.weight_scale.data == - self.default_scale] = 1 + layer.weight_scale.data[layer.weight_scale.data == torch.finfo( + torch.float32).min] = 1 layer.weight_scale = torch.nn.Parameter(layer.weight_scale.data, requires_grad=False) if self.quant_config.activation_scheme == "static": - layer.input_scale.data[layer.input_scale.data == - self.default_scale] = 1 + layer.input_scale.data[layer.input_scale.data == torch.finfo( + torch.float32).min] = 1 layer.input_scale = torch.nn.Parameter(layer.input_scale.data, requires_grad=False) # If using marlin (w8a16), kernel uses channelwise weights, @@ -373,7 +407,8 @@ def apply(self, input_scale=layer.input_scale, bias=bias, cutlass_fp8_supported=self.cutlass_fp8_supported, - use_per_token_if_dynamic=False) + # Default to using per_token quantization if cutlass is supported + use_per_token_if_dynamic=self.cutlass_fp8_supported) def permute_weight_fp8(x: torch.Tensor) -> torch.Tensor: @@ -409,8 +444,8 @@ def __init__(self, quant_config: Fp8Config): self.block_quant = self.quant_config.weight_block_size is not None def create_weights(self, layer: Module, num_experts: int, hidden_size: int, - intermediate_size: int, params_dtype: torch.dtype, - **extra_weight_attrs): + intermediate_size_per_partition: int, + params_dtype: torch.dtype, **extra_weight_attrs): if self.quant_config.is_checkpoint_fp8_serialized: params_dtype = torch.float8_e4m3fn @@ -425,30 +460,34 @@ def create_weights(self, layer: Module, num_experts: int, hidden_size: int, # scales, the output_size of the weights for both the gate and up # layers must be divisible by block_n. # Required by column parallel or enabling merged weights - if intermediate_size % block_n != 0: + if intermediate_size_per_partition % block_n != 0: raise ValueError( f"The output_size of gate's and up's weight = " - f"{intermediate_size} is not divisible by " + f"{intermediate_size_per_partition} is not divisible by " f"weight quantization block_n = {block_n}.") - if (tp_size > 1 and intermediate_size % block_k != 0): + if (tp_size > 1 + and intermediate_size_per_partition % block_k != 0): # Required by row parallel - raise ValueError(f"The input_size of down's weight = " - f"{intermediate_size} is not divisible by " - f"weight quantization block_k = {block_k}.") + raise ValueError( + f"The input_size of down's weight = " + f"{intermediate_size_per_partition} is not divisible by " + f"weight quantization block_k = {block_k}.") # WEIGHTS - w13_weight = torch.nn.Parameter(torch.empty(num_experts, - 2 * intermediate_size, - hidden_size, - dtype=params_dtype), + w13_weight = torch.nn.Parameter(torch.empty( + num_experts, + 2 * intermediate_size_per_partition, + hidden_size, + dtype=params_dtype), requires_grad=False) layer.register_parameter("w13_weight", w13_weight) set_weight_attrs(w13_weight, extra_weight_attrs) - w2_weight = torch.nn.Parameter(torch.empty(num_experts, - hidden_size, - intermediate_size, - dtype=params_dtype), + w2_weight = torch.nn.Parameter(torch.empty( + num_experts, + hidden_size, + intermediate_size_per_partition, + dtype=params_dtype), requires_grad=False) layer.register_parameter("w2_weight", w2_weight) set_weight_attrs(w2_weight, extra_weight_attrs) @@ -469,7 +508,8 @@ def create_weights(self, layer: Module, num_experts: int, hidden_size: int, w13_weight_scale = torch.nn.Parameter( torch.ones( num_experts, - 2 * ((intermediate_size + block_n - 1) // block_n), + 2 * ((intermediate_size_per_partition + block_n - 1) // + block_n), (hidden_size + block_k - 1) // block_k, dtype=torch.float32, ), @@ -479,7 +519,7 @@ def create_weights(self, layer: Module, num_experts: int, hidden_size: int, torch.ones( num_experts, (hidden_size + block_n - 1) // block_n, - (intermediate_size + block_k - 1) // block_k, + (intermediate_size_per_partition + block_k - 1) // block_k, dtype=torch.float32, ), requires_grad=False, @@ -527,6 +567,30 @@ def create_weights(self, layer: Module, num_experts: int, hidden_size: int, def process_weights_after_loading(self, layer: Module) -> None: # Block quant doesn't need to process weights after loading if self.block_quant: + if current_platform.is_rocm() and not is_navi(): + w13_weight, w13_weight_scale_inv, w13_input_scale = \ + normalize_e4m3fn_to_e4m3fnuz( + layer.w13_weight, layer.w13_weight_scale_inv, + layer.w13_input_scale) + w2_weight, w2_weight_scale_inv, w2_input_scale = \ + normalize_e4m3fn_to_e4m3fnuz( + layer.w2_weight, layer.w2_weight_scale_inv, + layer.w2_input_scale) + # Reset the parameter + layer.w13_weight = torch.nn.Parameter(w13_weight, + requires_grad=False) + layer.w13_weight_scale_inv = torch.nn.Parameter( + w13_weight_scale_inv, requires_grad=False) + if w13_input_scale is not None: + layer.w13_input_scale = torch.nn.Parameter( + w13_input_scale, requires_grad=False) + layer.w2_weight = torch.nn.Parameter(w2_weight, + requires_grad=False) + layer.w2_weight_scale_inv = torch.nn.Parameter( + w2_weight_scale_inv, requires_grad=False) + if w2_input_scale is not None: + layer.w2_input_scale = torch.nn.Parameter( + w2_input_scale, requires_grad=False) return # If checkpoint is fp16, quantize in place. if not self.quant_config.is_checkpoint_fp8_serialized: diff --git a/vllm/model_executor/layers/quantization/gptq_marlin.py b/vllm/model_executor/layers/quantization/gptq_marlin.py index 2dbfca9b07690..4dc4b052b0410 100644 --- a/vllm/model_executor/layers/quantization/gptq_marlin.py +++ b/vllm/model_executor/layers/quantization/gptq_marlin.py @@ -317,7 +317,7 @@ def create_weights( layer: torch.nn.Module, num_experts: int, hidden_size: int, - intermediate_size: int, + intermediate_size_per_partition: int, params_dtype: torch.dtype, **extra_weight_attrs, ): @@ -326,7 +326,8 @@ def create_weights( # Supports only sym for now (no zp) if self.quant_config.group_size != -1: scales_size13 = hidden_size // self.quant_config.group_size - scales_size2 = intermediate_size // self.quant_config.group_size + scales_size2 = (intermediate_size_per_partition // + self.quant_config.group_size) strategy = FusedMoeWeightScaleSupported.GROUP.value else: scales_size13 = 1 @@ -342,7 +343,7 @@ def create_weights( torch.empty( num_experts, hidden_size // self.quant_config.pack_factor, - 2 * intermediate_size, + 2 * intermediate_size_per_partition, dtype=torch.int32, ), requires_grad=False, @@ -353,7 +354,8 @@ def create_weights( w2_qweight = torch.nn.Parameter( torch.empty( num_experts, - intermediate_size // self.quant_config.pack_factor, + intermediate_size_per_partition // + self.quant_config.pack_factor, hidden_size, dtype=torch.int32, ), @@ -365,7 +367,7 @@ def create_weights( w13_scales = torch.nn.Parameter( torch.empty(num_experts, scales_size13, - 2 * intermediate_size, + 2 * intermediate_size_per_partition, dtype=torch.half), requires_grad=False, ) @@ -385,7 +387,8 @@ def create_weights( w13_qzeros = torch.nn.Parameter( torch.empty(num_experts, scales_size13, - 2 * intermediate_size // self.quant_config.pack_factor, + 2 * intermediate_size_per_partition // + self.quant_config.pack_factor, dtype=params_dtype), requires_grad=False, ) @@ -414,7 +417,7 @@ def create_weights( w2_g_idx = torch.nn.Parameter( torch.empty( num_experts, - intermediate_size, + intermediate_size_per_partition, dtype=torch.int32, ), requires_grad=False, @@ -435,7 +438,7 @@ def create_weights( w2_g_idx_sort_indices = torch.nn.Parameter( torch.empty( num_experts, - intermediate_size, + intermediate_size_per_partition, dtype=torch.int32, ), requires_grad=False, diff --git a/vllm/model_executor/layers/quantization/kernels/scaled_mm/__init__.py b/vllm/model_executor/layers/quantization/kernels/scaled_mm/__init__.py index 586752d3d34e3..4824a11804163 100644 --- a/vllm/model_executor/layers/quantization/kernels/scaled_mm/__init__.py +++ b/vllm/model_executor/layers/quantization/kernels/scaled_mm/__init__.py @@ -5,8 +5,8 @@ CutlassScaledMMLinearKernel) from vllm.model_executor.layers.quantization.kernels.scaled_mm.ScaledMMLinearKernel import ( # noqa: E501 ScaledMMLinearKernel, ScaledMMLinearLayerConfig) -# from vllm.model_executor.layers.quantization.kernels.scaled_mm.triton import ( -# TritonScaledMMLinear) +from vllm.model_executor.layers.quantization.kernels.scaled_mm.triton import ( + TritonScaledMMLinearKernel) from vllm.model_executor.layers.quantization.kernels.scaled_mm.xla import ( XLAScaledMMLinearKernel) from vllm.platforms import PlatformEnum, current_platform @@ -15,9 +15,7 @@ _POSSIBLE_KERNELS: Dict[PlatformEnum, List[Type[ScaledMMLinearKernel]]] = { PlatformEnum.CPU: [CutlassScaledMMLinearKernel], PlatformEnum.CUDA: [CutlassScaledMMLinearKernel], - # TODO(rob): Create TritonScaledMMLinear kernel. ROCM will - # incorrectly attempt to run AZP models if prompted to. - PlatformEnum.ROCM: [CutlassScaledMMLinearKernel], + PlatformEnum.ROCM: [TritonScaledMMLinearKernel], PlatformEnum.TPU: [XLAScaledMMLinearKernel], } diff --git a/vllm/model_executor/layers/quantization/kernels/scaled_mm/triton.py b/vllm/model_executor/layers/quantization/kernels/scaled_mm/triton.py new file mode 100644 index 0000000000000..97ec8cb0500d7 --- /dev/null +++ b/vllm/model_executor/layers/quantization/kernels/scaled_mm/triton.py @@ -0,0 +1,38 @@ +from typing import Optional, Tuple + +import torch + +from vllm.platforms import current_platform + +from .cutlass import CutlassScaledMMLinearKernel +from .ScaledMMLinearKernel import ScaledMMLinearLayerConfig + + +class TritonScaledMMLinearKernel(CutlassScaledMMLinearKernel): + + @classmethod + def get_min_capability(cls) -> int: + return 75 + + @classmethod + def can_implement( + cls, c: ScaledMMLinearLayerConfig) -> Tuple[bool, Optional[str]]: + if current_platform.is_cpu(): + return ( + False, + "TritonScaledMMLinearKernel requires Triton which is not " + + "currently supported on CPU.") + if not c.input_symmetric: + return (False, + "TritonScaledMMLinearKernel only supports symmetric " + + "quantization.") + return True, None + + def process_weights_after_loading(self, layer: torch.nn.Module) -> None: + super().process_weights_after_loading(layer) + + def apply_weights(self, + layer: torch.nn.Module, + x: torch.Tensor, + bias: Optional[torch.Tensor] = None) -> torch.Tensor: + return super().apply_weights(layer, x, bias) diff --git a/vllm/model_executor/layers/quantization/kv_cache.py b/vllm/model_executor/layers/quantization/kv_cache.py index 6a3b02da23504..ef54c25a05573 100644 --- a/vllm/model_executor/layers/quantization/kv_cache.py +++ b/vllm/model_executor/layers/quantization/kv_cache.py @@ -79,6 +79,8 @@ def process_weights_after_loading(self, layer: torch.nn.Module) -> None: # These are used in the final Attention.forward() layer._k_scale.copy_(k_scale) layer._v_scale.copy_(v_scale) + layer._k_scale_float = k_scale + layer._v_scale_float = v_scale if (k_scale == 1.0 and v_scale == 1.0 and (layer.kv_cache_dtype != "auto" or envs.VLLM_USE_ROCM_FP8_FLASH_ATTN) diff --git a/vllm/model_executor/layers/quantization/quark/__init__.py b/vllm/model_executor/layers/quantization/quark/__init__.py new file mode 100644 index 0000000000000..e69de29bb2d1d diff --git a/vllm/model_executor/layers/quantization/quark/quark.py b/vllm/model_executor/layers/quantization/quark/quark.py new file mode 100644 index 0000000000000..144036814fafa --- /dev/null +++ b/vllm/model_executor/layers/quantization/quark/quark.py @@ -0,0 +1,394 @@ +import fnmatch +from typing import Any, Dict, List, Optional, cast + +import torch + +from vllm.model_executor.layers.fused_moe import FusedMoE +from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase, + UnquantizedLinearMethod) +from vllm.model_executor.layers.quantization.base_config import ( # noqa: E501 + QuantizationConfig, QuantizeMethodBase) +from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod +from vllm.model_executor.layers.quantization.quark.quark_moe import ( # noqa: E501 + QuarkMoEMethod) +from vllm.model_executor.layers.quantization.quark.schemes import ( + QuarkScheme, QuarkW8A8Fp8, QuarkW8A8Int8) +from vllm.model_executor.layers.quantization.quark.utils import ( + deep_compare, should_ignore_layer) +from vllm.model_executor.layers.quantization.utils.quant_utils import ( + FUSED_LAYER_NAME_MAPPING) +from vllm.platforms import current_platform + +__all__ = ["QuarkLinearMethod"] + + +class QuarkConfig(QuantizationConfig): + + def __init__(self, + quant_config: Dict[str, Any], + kv_cache_group: Optional[List[str]] = None, + kv_cache_config: Optional[Dict[str, Any]] = None, + pack_method: str = "reorder"): + if kv_cache_group is None: + kv_cache_group = [] + self.quant_config = quant_config + self.kv_cache_group = kv_cache_group + self.kv_cache_config = kv_cache_config + self.pack_method = pack_method + + def get_linear_method(self) -> "QuarkLinearMethod": + return QuarkLinearMethod(self) + + def get_supported_act_dtypes(cls) -> List[torch.dtype]: + return [torch.float16, torch.bfloat16] + + @classmethod + def get_min_capability(cls) -> int: + return 70 + + def get_name(self) -> str: + return "quark" + + def get_quant_method(self, layer: torch.nn.Module, + prefix: str) -> Optional["QuantizeMethodBase"]: + from vllm.attention.layer import Attention # Avoid circular import + + # Check if the layer is skipped for quantization. + exclude_layers = cast(List[str], self.quant_config.get("exclude")) + if should_ignore_layer(prefix, ignore=exclude_layers): + return UnquantizedLinearMethod() + if isinstance(layer, LinearBase): + scheme = self.get_scheme(layer=layer, layer_name=prefix) + layer.scheme = scheme + return QuarkLinearMethod(self) + if isinstance(layer, Attention): + return QuarkKVCacheMethod(self) + if isinstance(layer, FusedMoE): + return QuarkMoEMethod.get_moe_method(self, + module=layer, + layer_name=prefix) + return None + + @classmethod + def from_config(cls, config: Dict[str, Any]) -> "QuarkConfig": + export_config = config.get("export") + if export_config is None: + raise ValueError("The export key should be included in " + "the configurations of Quark quantized model") + kv_cache_group = cast(List[str], export_config.get("kv_cache_group")) + pack_method = cast(str, export_config.get("pack_method")) + + # In the export model of quark, the quantization configuration + # of kv_cache is stored in layer_quant_config. First, it is + # judged whether kv_cache_group exists, and then it is judged + # whether layer_quant_config has a quantization configuration + # that matches kv_cache. + if len(kv_cache_group) == 0: + kv_cache_config = None + else: + kv_cache_set = set(kv_cache_group) + layer_quant_config = cast(Dict[str, Any], + config.get("layer_quant_config")) + layer_quant_names = list(layer_quant_config.keys()) + layer_quant_set = set(layer_quant_names) + + if not kv_cache_set.issubset(layer_quant_set): + raise ValueError("The Quark quantized model has the " + "kv_cache_group parameter setting, " + "but no kv_cache quantization settings " + "were found in the quantization " + "configuration.") + + q_configs = [ + cast(Dict[str, Any], layer_quant_config.get(name)) + for name in kv_cache_group + ] + if not all( + deep_compare(q_config, q_configs[0]) + for q_config in q_configs): + raise ValueError( + "The quantization method used for kv_cache should " + "be the same, but the quantization method for the " + "kv_cache layer in the config is different.") + kv_cache_config = q_configs[0].get("output_tensors") + if kv_cache_config is None: + raise ValueError( + "The kv_cache quantization configuration is empty.") + + # Since we have already set kv_cache quantization configurations, + # we will remove the quantization configuration for the + # output_tensors corresponding to the kv_cache layer. + for q_config in q_configs: + q_config["output_tensors"] = None + + # In case q_proj output is also quantized, remove the configuration + # to keep qkv consistency. + q_proj_q_config = cast(Dict[str, Any], + layer_quant_config.get("*q_proj")) + q_proj_q_config["output_tensors"] = None + + return cls(quant_config=config, + kv_cache_group=kv_cache_group, + kv_cache_config=kv_cache_config, + pack_method=pack_method) + + @classmethod + def get_config_filenames(cls) -> List[str]: + return [] + + def _check_scheme_supported(self, + min_capability: int, + error: bool = True) -> bool: + capability_tuple = current_platform.get_device_capability() + + if capability_tuple is not None: + capability = capability_tuple.to_int() + supported = capability >= min_capability + if error and not supported: + raise RuntimeError( + "Quantization scheme is not supported for ", + f"the current GPU. Min capability: {min_capability}. ", + f"Current capability: {capability}.") + return supported + else: + return False + + def is_fp8_w8a8(self) -> bool: + # Returns True if all quantized layers in model are fp8 w8a8 + global_quant_config = cast( + Dict[str, Any], self.quant_config.get("global_quant_config")) + layer_quant_configs = cast(Dict[str, Any], + self.quant_config.get("layer_quant_config")) + for config in (global_quant_config, *layer_quant_configs.values()): + weight_config = cast(Dict[str, Any], config.get("weight")) + input_config = cast(Dict[str, Any], config.get("input_tensors")) + if not self._is_fp8_w8a8(weight_config, input_config): + return False + return True + + def _is_fp8_w8a8(self, weight_quant: Optional[Dict[str, Any]], + input_quant: Optional[Dict[str, Any]]) -> bool: + # Confirm weights and input quantized. + if weight_quant is None or input_quant is None: + return False + + # Confirm weight scheme is supported + is_fp8_dtype = (weight_quant.get("dtype") == "fp8_e4m3" + and input_quant.get("dtype") == "fp8_e4m3") + is_static_weight = not weight_quant.get("is_dynamic") + is_per_tensor_or_channel_weight = (weight_quant.get("qscheme") + in ["per_tensor", "per_channel"]) + + if not (is_fp8_dtype and is_static_weight + and is_per_tensor_or_channel_weight): + return False + + # Dynamic quantization is always supported if weights supported. + if input_quant.get("is_dynamic"): + return True + + # Confirm activation scheme is supported. + is_per_tensor_activation = (input_quant.get("qscheme") == "per_tensor") + return is_per_tensor_activation + + def _is_static_tensor_w8a8(self, weight_quant: Optional[Dict[str, Any]], + input_quant: Optional[Dict[str, Any]]) -> bool: + # Confirm weights and input quantized. + if weight_quant is None or input_quant is None: + return False + + is_int8_dtype = (weight_quant.get("dtype") == "int8" + and input_quant.get("dtype") == "int8") + + is_tensor = (weight_quant.get("qscheme") + in ["per_tensor", "per_channel"] + and input_quant.get("qscheme") == "per_tensor") + + is_static = (not weight_quant.get("is_dynamic") + and not input_quant.get("is_dynamic")) + + is_weight_symmetric = (weight_quant.get("symmetric") is True) + + # Both symmetric and asymmetric input quantization supported. + # Only symmetric weight quantization supported. + return is_int8_dtype and is_tensor and is_weight_symmetric and is_static + + def _find_matched_config(self, layer_name: str, + module: torch.nn.Module) -> Dict[str, Any]: + + proj_name = layer_name.split(".")[-1] + if proj_name in FUSED_LAYER_NAME_MAPPING: + shard_proj_names = FUSED_LAYER_NAME_MAPPING[proj_name] + + # Convert fused_name --> [shard_names] + shard_names = [ + layer_name.replace(proj_name, shard_proj_name) + for shard_proj_name in shard_proj_names + ] + shard_configs = [ + self._find_matched_config(shard_name, module) + for shard_name in shard_names + ] + if not all( + deep_compare(q_config, shard_configs[0]) + for q_config in shard_configs): + raise ValueError( + f"Found a different quantization configuration for " + f"{shard_proj_names} in {layer_name}. vLLM " + "requires all to use the same scheme.") + return shard_configs[0] + else: + layer_quant_config = cast( + Dict[str, Any], self.quant_config.get("layer_quant_config")) + for name_pattern in layer_quant_config: + if fnmatch.fnmatch(layer_name, name_pattern): + return layer_quant_config[name_pattern] + + layer_type = cast(str, type(module)) + layer_type_quant_config = cast( + Dict[str, Any], + self.quant_config.get("layer_type_quant_config")) + if layer_type in layer_type_quant_config: + return layer_type_quant_config[layer_type] + + global_quant_config = cast( + Dict[str, Any], self.quant_config.get("global_quant_config")) + return global_quant_config + + def _get_scheme_from_config(self, config: Dict[str, Any]) -> "QuarkScheme": + if config.get("output_tensors") or config.get("bias"): + raise NotImplementedError( + "Currently, Quark models with output_tensors " + "and bias quantized are not supported") + weight_config = cast(Dict[str, Any], config.get("weight")) + input_config = cast(Dict[str, Any], config.get("input_tensors")) + + if self._is_fp8_w8a8(weight_config, input_config): + is_fp8_w8a8_supported = self._check_scheme_supported( + QuarkW8A8Fp8.get_min_capability(), error=False) + if is_fp8_w8a8_supported: + weight_qscheme = cast(str, weight_config.get("qscheme")) + input_static = (input_config is not None and + not cast(bool, input_config.get("is_dynamic"))) + return QuarkW8A8Fp8(qscheme=weight_qscheme, + is_static_input_scheme=input_static) + elif self._is_static_tensor_w8a8(weight_config, input_config): + weight_qscheme = cast(str, weight_config.get("qscheme")) + return QuarkW8A8Int8(qscheme=weight_qscheme, + is_static_input_scheme=True, + input_symmetric=input_config.get("symmetric")) + + raise NotImplementedError("No quark compatible scheme was found. " + f"Weight config: {weight_config}, " + f"Input config: {input_config}") + + def get_scheme(self, layer: torch.nn.Module, + layer_name: str) -> "QuarkScheme": + + layer_quant_config = self._find_matched_config(layer_name, layer) + + # Find the quant_scheme + scheme = self._get_scheme_from_config(layer_quant_config) + # Raise error if device does not support the scheme + # (e.g. fp8 needs ada lovelace) + self._check_scheme_supported(scheme.get_min_capability()) + + return scheme + + def get_cache_scale(self, name: str) -> Optional[str]: + """ + Check whether the param name matches the format for k/v cache scales + in quark. If this is the case, return its equivalent param name + expected by vLLM + + :param name: param name + :return: matching param name for KV cache scale in vLLM + """ + if name.endswith(".output_scale") and ".k_proj" in name: + return name.replace(".k_proj.output_scale", ".attn.k_scale") + if name.endswith(".output_scale") and ".v_proj" in name: + return name.replace(".v_proj.output_scale", ".attn.v_scale") + if name.endswith(".output_scale") and ".q_proj" in name: + return name.replace(".q_proj.output_scale", ".attn.q_scale") + if name.endswith("self_attn.prob_output_scale"): + return name.replace(".prob_output_scale", ".attn.prob_scale") + + # If no matches, return None + return None + + +class QuarkLinearMethod(LinearMethodBase): + + def __init__(self, quantization_config: QuarkConfig): + self.quantization_config = quantization_config + + def process_weights_after_loading(self, layer: torch.nn.Module) -> None: + layer.scheme.process_weights_after_loading(layer) + + def create_weights(self, layer: torch.nn.Module, + input_size_per_partition: int, + output_partition_sizes: List[int], input_size: int, + output_size: int, params_dtype: torch.dtype, + **extra_weight_attrs): + """ + Use the CompressedTensorsScheme associated with each layer to create + the necessary parameters for the layer. See LinearMethodBase for param + details + """ + weight_loader = extra_weight_attrs.get("weight_loader") + layer.scheme.create_weights( + layer=layer, + input_size=input_size, + input_size_per_partition=input_size_per_partition, + output_partition_sizes=output_partition_sizes, + output_size=output_size, + params_dtype=params_dtype, + weight_loader=weight_loader) + + def apply(self, + layer: torch.nn.Module, + x: torch.Tensor, + bias: Optional[torch.Tensor] = None): + """ + Use the output of create_weights and the CompressedTensorsScheme + associated with the layer to apply the forward pass with the + layer input. See LinearMethodBase for param details + + """ + scheme = layer.scheme + if scheme is None: + raise ValueError("A scheme must be defined for each layer") + return scheme.apply_weights(layer, x, bias=bias) + + +class QuarkKVCacheMethod(BaseKVCacheMethod): + """ + Supports loading kv-cache scaling factors from quark checkpoints. + """ + + def __init__(self, quant_config: QuarkConfig): + self.validate_kv_cache_config(quant_config.kv_cache_config) + super().__init__(quant_config) + + @staticmethod + def validate_kv_cache_config(kv_cache_config: Optional[Dict[str, Any]]): + """ + Validator for the kv cache configuration. Useful for controlling the + kv cache quantization schemes, that are being supported in vLLM + :param kv_cache_config: the quark kv cache scheme + """ + if kv_cache_config is None: + return + + dtype = kv_cache_config.get("dtype") + if dtype != "fp8_e4m3": + raise NotImplementedError( + "Currently supported kv cache quantization is " + f"dtype=fp8_e4m3, however received {dtype}") + + qscheme = kv_cache_config.get("qscheme") + if qscheme != "per_tensor": + raise NotImplementedError( + "Only support per-tensor scaling factor " + "for quark KV cache. " + f"Expected qscheme: per_tensor, found qscheme: {qscheme}") diff --git a/vllm/model_executor/layers/quantization/quark/quark_moe.py b/vllm/model_executor/layers/quantization/quark/quark_moe.py new file mode 100644 index 0000000000000..68a3954540763 --- /dev/null +++ b/vllm/model_executor/layers/quantization/quark/quark_moe.py @@ -0,0 +1,227 @@ +from typing import Any, Callable, Dict, Optional + +import torch + +import vllm.model_executor.layers.fused_moe # noqa +from vllm import _custom_ops as ops +from vllm.logger import init_logger +from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase, + FusedMoeWeightScaleSupported) +from vllm.model_executor.layers.quantization.utils.w8a8_utils import ( + all_close_1d, normalize_e4m3fn_to_e4m3fnuz, per_tensor_dequantize) +from vllm.model_executor.utils import set_weight_attrs +from vllm.platforms import current_platform + +logger = init_logger(__name__) + +__all__ = ["QuarkMoEMethod", "QuarkW8A8Fp8MoEMethod"] + + +class QuarkMoEMethod(FusedMoEMethodBase): + + @staticmethod + def get_moe_method( + quant_config: "QuarkConfig", # type: ignore # noqa E501 # noqa F821 + module: torch.nn.Module, + layer_name: str) -> "QuarkMoEMethod": + layer_quant_config = quant_config._find_matched_config( + layer_name, module) + + if (layer_quant_config.get("output_tensors") + or layer_quant_config.get("bias")): + raise NotImplementedError("Currently, Quark models with " + "output_tensors and bias " + "quantized are not supported") + weight_config = layer_quant_config.get("weight") + input_config = layer_quant_config.get("input_tensors") + + if quant_config._is_fp8_w8a8(weight_config, input_config): + return QuarkW8A8Fp8MoEMethod(weight_config, input_config) + else: + raise RuntimeError("Unsupported FusedMoe scheme") + + +class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod): + + def __init__(self, weight_config: Dict[str, Any], input_config: Dict[str, + Any]): + self.weight_quant = weight_config + self.input_quant = input_config + + weight_qscheme = self.weight_quant.get("qscheme") + input_qscheme = self.input_quant.get("qscheme") + if not (weight_qscheme == "per_tensor" + and input_qscheme == "per_tensor"): + raise ValueError( + "For FP8 Fused MoE layers, only per-tensor scales" + "for weights and activations are supported. Found " + f"{weight_qscheme}, {input_qscheme}") # noqa E501 + + self.static_input_scales = not self.input_quant.get("is_dynamic") + + def create_weights(self, layer: torch.nn.Module, num_experts: int, + hidden_size: int, intermediate_size_per_partition: int, + params_dtype: torch.dtype, **extra_weight_attrs): + + params_dtype = torch.float8_e4m3fn + + # WEIGHTS + w13_weight = torch.nn.Parameter(torch.empty( + num_experts, + 2 * intermediate_size_per_partition, + hidden_size, + dtype=params_dtype), + requires_grad=False) + layer.register_parameter("w13_weight", w13_weight) + set_weight_attrs(w13_weight, extra_weight_attrs) + + w2_weight = torch.nn.Parameter(torch.empty( + num_experts, + hidden_size, + intermediate_size_per_partition, + dtype=params_dtype), + requires_grad=False) + layer.register_parameter("w2_weight", w2_weight) + set_weight_attrs(w2_weight, extra_weight_attrs) + + # WEIGHT_SCALES + # Allocate 2 scales for w1 and w3 respectively. + # They will be combined to a single scale after weight loading. + w13_weight_scale = torch.nn.Parameter(torch.ones(num_experts, + 2, + dtype=torch.float32), + requires_grad=False) + layer.register_parameter("w13_weight_scale", w13_weight_scale) + + w2_weight_scale = torch.nn.Parameter(torch.ones(num_experts, + dtype=torch.float32), + requires_grad=False) + layer.register_parameter("w2_weight_scale", w2_weight_scale) + # Add the quantization method used (per tensor/grouped/channel) + # to ensure the weight scales are loaded in properly + extra_weight_attrs.update( + {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}) + set_weight_attrs(w13_weight_scale, extra_weight_attrs) + set_weight_attrs(w2_weight_scale, extra_weight_attrs) + + # INPUT_SCALES + if self.static_input_scales: + w13_input_scale = torch.nn.Parameter(torch.ones( + num_experts, dtype=torch.float32), + requires_grad=False) + layer.register_parameter("w13_input_scale", w13_input_scale) + set_weight_attrs(w13_input_scale, extra_weight_attrs) + + w2_input_scale = torch.nn.Parameter(torch.ones( + num_experts, dtype=torch.float32), + requires_grad=False) + layer.register_parameter("w2_input_scale", w2_input_scale) + set_weight_attrs(w2_input_scale, extra_weight_attrs) + else: + layer.w13_input_scale = None + layer.w2_input_scale = None + + def process_weights_after_loading(self, layer: torch.nn.Module) -> None: + # Fp8 moe kernels require a single activation scale. + # We take the max of all the scales in case they differ. + if self.static_input_scales: + if (layer.w13_input_scale is None or layer.w2_input_scale is None): + raise ValueError( + "QuantConfig has static quantization, but found " + "activation scales are None.") + if (not all_close_1d(layer.w13_input_scale) + or not all_close_1d(layer.w2_input_scale)): + logger.warning_once( + "Found input_scales that are not equal for " + "fp8 MoE layer. Using the maximum across experts " + "for each layer. ") + layer.w13_input_scale = torch.nn.Parameter( + layer.w13_input_scale.max(), requires_grad=False) + layer.w2_input_scale = torch.nn.Parameter( + layer.w2_input_scale.max(), requires_grad=False) + + # If rocm, normalize the weights and scales to e4m3fnuz + if current_platform.is_rocm(): + # Normalize the weights and scales + w13_weight, w13_weight_scale, w13_input_scale = \ + normalize_e4m3fn_to_e4m3fnuz( + layer.w13_weight, layer.w13_weight_scale, + layer.w13_input_scale) + w2_weight, w2_weight_scale, w2_input_scale = \ + normalize_e4m3fn_to_e4m3fnuz( + layer.w2_weight, layer.w2_weight_scale, + layer.w2_input_scale) + # Reset the parameter + layer.w13_weight = torch.nn.Parameter(w13_weight, + requires_grad=False) + layer.w13_weight_scale = torch.nn.Parameter(w13_weight_scale, + requires_grad=False) + if w13_input_scale is not None: + layer.w13_input_scale = torch.nn.Parameter(w13_input_scale, + requires_grad=False) + layer.w2_weight = torch.nn.Parameter(w2_weight, + requires_grad=False) + layer.w2_weight_scale = torch.nn.Parameter(w2_weight_scale, + requires_grad=False) + if w2_input_scale is not None: + layer.w2_input_scale = torch.nn.Parameter(w2_input_scale, + requires_grad=False) + + # Fp8 moe kernel needs single weight scale for w13 per expert. + # We take the max then dequant and requant each expert. + assert layer.w13_weight_scale is not None + shard_size = layer.intermediate_size_per_partition + max_w13_scales = layer.w13_weight_scale.max(dim=1).values + for expert_id in range(layer.num_experts): + start = 0 + for shard_id in range(2): + dq_weight = per_tensor_dequantize( + layer.w13_weight[expert_id][start:start + shard_size, :], + layer.w13_weight_scale[expert_id][shard_id]) + layer.w13_weight[expert_id][ + start:start + shard_size, :], _ = ops.scaled_fp8_quant( + dq_weight, max_w13_scales[expert_id]) + start += shard_size + + layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales, + requires_grad=False) + + def apply( + self, + layer: torch.nn.Module, + x: torch.Tensor, + router_logits: torch.Tensor, + top_k: int, + renormalize: bool, + use_grouped_topk: bool = False, + topk_group: Optional[int] = None, + num_expert_group: Optional[int] = None, + custom_routing_function: Optional[Callable] = None, + scoring_func: str = "softmax", + e_score_correction_bias: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + from vllm.model_executor.layers.fused_moe import fused_experts + + topk_weights, topk_ids = FusedMoE.select_experts( + hidden_states=x, + router_logits=router_logits, + use_grouped_topk=use_grouped_topk, + top_k=top_k, + renormalize=renormalize, + topk_group=topk_group, + num_expert_group=num_expert_group, + custom_routing_function=custom_routing_function, + scoring_func=scoring_func, + e_score_correction_bias=e_score_correction_bias) + + return fused_experts(x, + layer.w13_weight, + layer.w2_weight, + topk_weights=topk_weights, + topk_ids=topk_ids, + inplace=True, + use_fp8_w8a8=True, + w1_scale=layer.w13_weight_scale, + w2_scale=layer.w2_weight_scale, + a1_scale=layer.w13_input_scale, + a2_scale=layer.w2_input_scale) diff --git a/vllm/model_executor/layers/quantization/quark/schemes/__init__.py b/vllm/model_executor/layers/quantization/quark/schemes/__init__.py new file mode 100644 index 0000000000000..fb0ba9bd5220c --- /dev/null +++ b/vllm/model_executor/layers/quantization/quark/schemes/__init__.py @@ -0,0 +1,5 @@ +from .quark_scheme import QuarkScheme +from .quark_w8a8_fp8 import QuarkW8A8Fp8 +from .quark_w8a8_int8 import QuarkW8A8Int8 + +__all__ = ["QuarkScheme", "QuarkW8A8Fp8", "QuarkW8A8Int8"] diff --git a/vllm/model_executor/layers/quantization/quark/schemes/quark_scheme.py b/vllm/model_executor/layers/quantization/quark/schemes/quark_scheme.py new file mode 100644 index 0000000000000..239597fa4be0e --- /dev/null +++ b/vllm/model_executor/layers/quantization/quark/schemes/quark_scheme.py @@ -0,0 +1,52 @@ +from abc import ABC, abstractmethod +from typing import Optional + +import torch + +__all__ = ["QuarkScheme"] + + +class QuarkScheme(ABC): + """ + Abstract class used to describe the weight creation and forward pass + of different quantization schemes supported by Quark. + """ + + @classmethod + @abstractmethod + def get_min_capability(cls) -> int: + """ + Get minimum device capability. + """ + raise NotImplementedError + + @abstractmethod + def create_weights(self, *args, **kwargs): + """ + Weight creation for the particular scheme. Inputs to this function + + """ + raise NotImplementedError + + @abstractmethod + def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor, + bias: Optional[torch.Tensor]): + """ + Run the forward pass for the particular scheme. This is where + scheme-specific dequant/quant steps/kernels should be applied. + + :param layer: torch.nn.Module with the registered weights and + other parameters relevant to the particular scheme. + :param x: input to the layer + :param bias: bias parameter + + """ + raise NotImplementedError + + @abstractmethod + def process_weights_after_loading(self, layer: torch.nn.Module): + """ + Called after weight loading is complete for any cleanup that + needs to occur. + """ + raise NotImplementedError diff --git a/vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_fp8.py b/vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_fp8.py new file mode 100644 index 0000000000000..447911a648639 --- /dev/null +++ b/vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_fp8.py @@ -0,0 +1,142 @@ +from typing import Callable, List, Optional + +import torch +from torch.nn import Parameter + +from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme +from vllm.model_executor.layers.quantization.utils.w8a8_utils import ( + apply_fp8_linear, cutlass_fp8_supported, normalize_e4m3fn_to_e4m3fnuz, + requantize_with_max_scale) +from vllm.model_executor.parameter import (ChannelQuantScaleParameter, + ModelWeightParameter, + PerTensorScaleParameter) +from vllm.platforms import current_platform + +__all__ = ["QuarkW8A8Fp8"] + + +class QuarkW8A8Fp8(QuarkScheme): + + def __init__(self, qscheme: str, is_static_input_scheme: Optional[bool]): + self.qscheme = qscheme + self.is_static_input_scheme = is_static_input_scheme + self.cutlass_fp8_supported = cutlass_fp8_supported() + self.out_dtype = torch.get_default_dtype() + + @classmethod + def get_min_capability(cls) -> int: + # lovelace and up + return 89 + + def process_weights_after_loading(self, layer) -> None: + # If per tensor, when we have a fused module (e.g. QKV) with per + # tensor scales (thus N scales being passed to the kernel), + # requantize so we can always run per tensor + if self.qscheme == "per_tensor": + max_w_scale, weight = requantize_with_max_scale( + weight=layer.weight, + weight_scale=layer.weight_scale, + logical_widths=layer.logical_widths, + ) + + if current_platform.is_rocm(): + weight, max_w_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz( + weight=weight, + weight_scale=max_w_scale, + input_scale=layer.input_scale) + if input_scale is not None: + layer.input_scale = Parameter(input_scale, + requires_grad=False) + + layer.weight = Parameter(weight.t(), requires_grad=False) + layer.weight_scale = Parameter(max_w_scale, requires_grad=False) + + # If channelwise, scales are already lined up, so just transpose. + elif self.qscheme == "per_channel": + weight = layer.weight + + if current_platform.is_rocm(): + weight, weight_scale, input_scale = \ + normalize_e4m3fn_to_e4m3fnuz( + weight=weight, + weight_scale=layer.weight_scale, + input_scale=layer.input_scale) + if input_scale is not None: + layer.input_scale = Parameter(input_scale, + requires_grad=False) + else: + weight_scale = layer.weight_scale.data + + layer.weight = Parameter(weight.t(), requires_grad=False) + # required by torch.compile to be torch.nn.Parameter + layer.weight_scale = Parameter(weight_scale, requires_grad=False) + + else: + raise ValueError(f"Unknown quantization scheme {self.qscheme}") + + # INPUT SCALE + if self.is_static_input_scheme: + layer.input_scale = Parameter(layer.input_scale.max(), + requires_grad=False) + else: + layer.input_scale = None + + def create_weights(self, layer: torch.nn.Module, + output_partition_sizes: List[int], + input_size_per_partition: int, + params_dtype: torch.dtype, weight_loader: Callable, + **kwargs): + output_size_per_partition = sum(output_partition_sizes) + layer.logical_widths = output_partition_sizes + + # WEIGHT + weight = ModelWeightParameter(data=torch.empty( + output_size_per_partition, + input_size_per_partition, + dtype=torch.float8_e4m3fn), + input_dim=1, + output_dim=0, + weight_loader=weight_loader) + layer.register_parameter("weight", weight) + + # WEIGHT SCALE + # TODO: update create_xxx_parameter functions to return + # the newly added parameters + if self.qscheme == "per_channel": + weight_scale = ChannelQuantScaleParameter( + data=torch.empty((sum(output_partition_sizes), 1), + dtype=torch.float32), + output_dim=0, + weight_loader=weight_loader) + else: + assert self.qscheme == "per_tensor" + weight_scale = PerTensorScaleParameter(data=torch.empty( + len(output_partition_sizes), dtype=torch.float32), + weight_loader=weight_loader) + + # min requirement for fp8 kernels + weight_scale[:] = torch.finfo(torch.float32).min + layer.register_parameter("weight_scale", weight_scale) + + # INPUT SCALE + if self.is_static_input_scheme: + input_scale = PerTensorScaleParameter(data=torch.empty( + len(output_partition_sizes), dtype=torch.float32), + weight_loader=weight_loader) + input_scale[:] = torch.finfo(torch.float32).min + layer.register_parameter("input_scale", input_scale) + + def apply_weights(self, + layer: torch.nn.Module, + x: torch.Tensor, + bias: Optional[torch.Tensor] = None) -> torch.Tensor: + + return apply_fp8_linear( + input=x, + weight=layer.weight, + weight_scale=layer.weight_scale, + out_dtype=self.out_dtype, + input_scale=layer.input_scale, + bias=bias, + cutlass_fp8_supported=self.cutlass_fp8_supported, + use_per_token_if_dynamic=True) diff --git a/vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_int8.py b/vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_int8.py new file mode 100644 index 0000000000000..8cb47e9c37e56 --- /dev/null +++ b/vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_int8.py @@ -0,0 +1,105 @@ +from typing import Callable, List, Optional, Set + +import torch + +from vllm.logger import init_logger +from vllm.model_executor.layers.quantization.kernels.scaled_mm import ( + ScaledMMLinearLayerConfig, choose_scaled_mm_linear_kernel) +from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme +from vllm.model_executor.parameter import (BasevLLMParameter, + ChannelQuantScaleParameter, + ModelWeightParameter, + PerTensorScaleParameter) + +logger = init_logger(__name__) + + +class QuarkW8A8Int8(QuarkScheme): + _kernel_backends_being_used: Set[str] = set() + + def __init__(self, qscheme: str, is_static_input_scheme: Optional[bool], + input_symmetric: Optional[bool]): + self.qscheme = qscheme + self.is_static_input_scheme = is_static_input_scheme + self.input_symmetric = input_symmetric + + @classmethod + def get_min_capability(cls) -> int: + # turing and up + return 75 + + def create_weights(self, layer: torch.nn.Module, + output_partition_sizes: List[int], + input_size_per_partition: int, + params_dtype: torch.dtype, weight_loader: Callable, + **kwargs): + self.logical_widths = output_partition_sizes + + scaled_mm_linear_kernel_config = ScaledMMLinearLayerConfig( + is_channelwise=(self.qscheme == "per_channel"), + is_static_input_scheme=(self.is_static_input_scheme is True), + input_symmetric=(self.input_symmetric is True)) + + kernel_type = choose_scaled_mm_linear_kernel( + scaled_mm_linear_kernel_config) + + if kernel_type.__name__ not in self._kernel_backends_being_used: + logger.info("Using %s for QuarkW8A8Int8", kernel_type.__name__) + self._kernel_backends_being_used.add(kernel_type.__name__) + + # WEIGHT + weight = ModelWeightParameter(data=torch.empty( + sum(output_partition_sizes), + input_size_per_partition, + dtype=torch.int8), + input_dim=1, + output_dim=0, + weight_loader=weight_loader) + + layer.register_parameter("weight", weight) + + # WEIGHT SCALE + if self.qscheme == "per_channel": + weight_scale = ChannelQuantScaleParameter( + data=torch.empty((sum(output_partition_sizes), 1), + dtype=torch.float32), + output_dim=0, + weight_loader=weight_loader) + else: + assert self.qscheme == "per_tensor" + weight_scale = PerTensorScaleParameter(data=torch.empty( + len(output_partition_sizes), dtype=torch.float32), + weight_loader=weight_loader) + layer.register_parameter("weight_scale", weight_scale) + + # INPUT SCALE + if self.is_static_input_scheme: + input_scale = BasevLLMParameter(data=torch.empty( + 1, dtype=torch.float32), + weight_loader=weight_loader) + layer.register_parameter("input_scale", input_scale) + + if not self.input_symmetric: + # Note: quark stores the zp using the same dtype + # as the weights + # AZP loaded as int8 but used as int32 + input_zero_point = BasevLLMParameter( + data=torch.empty(1, dtype=torch.int8), + weight_loader=weight_loader) + layer.register_parameter("input_zero_point", input_zero_point) + + self.kernel = kernel_type(c=scaled_mm_linear_kernel_config, + w_q_param_name="weight", + w_s_param_name="weight_scale", + i_s_param_name="input_scale", + i_zp_param_name="input_zero_point", + azp_adj_param_name="azp_adj") + + # Checkpoints are serialized in quark format, which is + # different from the format the kernel may want. Handle repacking here. + def process_weights_after_loading(self, layer: torch.nn.Module) -> None: + self.kernel.process_weights_after_loading(layer) + + def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor, + bias: Optional[torch.Tensor]) -> torch.Tensor: + return self.kernel.apply_weights(layer, x, bias) diff --git a/vllm/model_executor/layers/quantization/quark/utils.py b/vllm/model_executor/layers/quantization/quark/utils.py new file mode 100644 index 0000000000000..742a629bdb1c5 --- /dev/null +++ b/vllm/model_executor/layers/quantization/quark/utils.py @@ -0,0 +1,99 @@ +import re +from typing import Any, Iterable, Optional + +from vllm.model_executor.layers.quantization.utils.quant_utils import ( + FUSED_LAYER_NAME_MAPPING) + + +def deep_compare(dict1: Any, dict2: Any) -> bool: + if type(dict1) is not type(dict2): + return False + if isinstance(dict1, dict): + if dict1.keys() != dict2.keys(): + return False + return all(deep_compare(dict1[k], dict2[k]) for k in dict1) + elif isinstance(dict1, list): + return set(dict1) == set(dict2) + else: + return dict1 == dict2 + + +def should_ignore_layer(layer_name: Optional[str], + ignore: Iterable[str]) -> bool: + if layer_name is None: + return False + + # layer_name = model.layers.0.self_attn.qkv_proj + # proj_name = qkv_proj + proj_name = layer_name.split(".")[-1] + + # Fused layers like gate_up_proj or qkv_proj will not be fused + # in the safetensors checkpoint. So, we convert the name + # from the fused version to unfused + check to make sure that + # each shard of the fused layer has the same scheme. + if proj_name in FUSED_LAYER_NAME_MAPPING: + shard_proj_names = FUSED_LAYER_NAME_MAPPING[proj_name] + + # Convert fused_name --> [shard_names] + shard_names = [ + layer_name.replace(proj_name, shard_proj_name) + for shard_proj_name in shard_proj_names + ] + + # Layer should be ignored if shards are ignored. + should_ignore_layer = None + for shard_name in shard_names: + should_ignore_shard = check_equal_or_regex_match( + layer_name=shard_name, targets=ignore) + + # If shard_idx=0, set layer ignore to match shard. + if should_ignore_layer is None: + should_ignore_layer = should_ignore_shard + + # If shard_idx=1+ confirm scheme matches prior shards. + elif should_ignore_shard != should_ignore_layer: + raise ValueError(f"Found a different quantization schemes for " + f"{shard_proj_names} in {layer_name}. vLLM " + "requires all to use the same scheme.") + + # Unfused layers like down_proj and o_proj will match + # the safetensors checkpoint already. + else: + should_ignore_layer = check_equal_or_regex_match(layer_name=layer_name, + targets=ignore) + + assert should_ignore_layer is not None + return should_ignore_layer + + +def check_equal_or_regex_match(layer_name: str, + targets: Iterable[str]) -> bool: + """ + Checks whether a layer_name is exactly equal or a regex match for + if target starts with 're:' to any target in list. + """ + for target in targets: + if _is_equal_or_regex_match(layer_name, target): + return True + return False + + +def _is_equal_or_regex_match(value: str, + target: str, + check_contains: bool = False) -> bool: + """ + Checks whether a value is exactly equal or a regex match for target + if target starts with 're:'. If check_contains is set to True, + additionally checks if the target string is contained within the value. + """ + + if target.startswith("re:"): + pattern = target[3:] + if re.match(pattern, value): + return True + elif check_contains: + if target.lower() in value.lower(): + return True + elif target == value: + return True + return False diff --git a/vllm/model_executor/layers/quantization/utils/fp8_utils.py b/vllm/model_executor/layers/quantization/utils/fp8_utils.py index f3c3e130e4161..b2ce1f9e29650 100644 --- a/vllm/model_executor/layers/quantization/utils/fp8_utils.py +++ b/vllm/model_executor/layers/quantization/utils/fp8_utils.py @@ -5,6 +5,9 @@ import triton import triton.language as tl +from vllm.platforms import current_platform +from vllm.utils import is_navi + def apply_w8a8_block_fp8_linear( input: torch.Tensor, @@ -33,11 +36,14 @@ def apply_w8a8_block_fp8_linear( def input_to_float8( - x: torch.Tensor, - dtype: torch.dtype = torch.float8_e4m3fn + x: torch.Tensor, + dtype: Optional[torch.dtype] = None ) -> Tuple[torch.Tensor, torch.Tensor]: """This function quantizes input values to float8 values " "with tensor-wise quantization.""" + if dtype is None: + dtype = (torch.float8_e4m3fnuz if current_platform.is_rocm() + and not is_navi() else torch.float8_e4m3fn) finfo = torch.finfo(dtype) min_val, max_val = x.aminmax() amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12) @@ -125,7 +131,7 @@ def per_token_group_quant_fp8( x: torch.Tensor, group_size: int, eps: float = 1e-10, - dtype: torch.dtype = torch.float8_e4m3fn, + dtype: Optional[torch.dtype] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: """Function to perform per-token-group quantization on an input tensor `x`. It converts the tensor values into signed float8 values and returns the @@ -140,6 +146,9 @@ def per_token_group_quant_fp8( Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the scaling factor for quantization. """ + if dtype is None: + dtype = (torch.float8_e4m3fnuz if current_platform.is_rocm() + and not is_navi() else torch.float8_e4m3fn) assert (x.shape[-1] % group_size == 0), ( f"the last dimension of `x` {x.shape[-1]} must be divisible " f"by `group_size` {group_size}") diff --git a/vllm/model_executor/layers/rotary_embedding.py b/vllm/model_executor/layers/rotary_embedding.py index 3fcd81a3c4213..d071cfe888f05 100644 --- a/vllm/model_executor/layers/rotary_embedding.py +++ b/vllm/model_executor/layers/rotary_embedding.py @@ -841,6 +841,37 @@ def get_input_positions( ) -> Tuple[List[List[int]], int]: """Get mrope input positions and delta value.""" + llm_positions, mrope_position_delta = \ + MRotaryEmbedding.get_input_positions_tensor( + input_tokens, + image_grid_thw, + video_grid_thw, + image_token_id, + video_token_id, + vision_start_token_id, + vision_end_token_id, + spatial_merge_size, + context_len, + seq_len, + ) + + return llm_positions.tolist(), mrope_position_delta + + @staticmethod + def get_input_positions_tensor( + input_tokens: List[int], + image_grid_thw: Union[List[List[int]], torch.Tensor], + video_grid_thw: Union[List[List[int]], torch.Tensor], + image_token_id: int, + video_token_id: int, + vision_start_token_id: int, + vision_end_token_id: int, + spatial_merge_size: int, + context_len: int = 0, + seq_len: Optional[int] = None, + ) -> Tuple[torch.Tensor, int]: + """Get mrope input positions and delta value.""" + if isinstance(image_grid_thw, torch.Tensor): image_grid_thw = image_grid_thw.tolist() if isinstance(video_grid_thw, torch.Tensor): @@ -916,7 +947,7 @@ def get_input_positions( len(input_tokens)).item() llm_positions = llm_positions[:, context_len:seq_len] - return llm_positions.tolist(), mrope_position_delta + return llm_positions, mrope_position_delta @staticmethod def get_next_input_positions( @@ -930,6 +961,17 @@ def get_next_input_positions( seq_len + mrope_position_delta)) for _ in range(3) ] + @staticmethod + def get_next_input_positions_tensor( + mrope_position_delta: int, + context_len: int, + seq_len: int, + ) -> torch.Tensor: + return torch.arange( + mrope_position_delta + context_len, + mrope_position_delta + seq_len, + ).expand(3, -1) + _ROPE_DICT: Dict[Tuple, RotaryEmbedding] = {} diff --git a/vllm/model_executor/layers/vocab_parallel_embedding.py b/vllm/model_executor/layers/vocab_parallel_embedding.py index 4afa367b387da..996d9b2505d3a 100644 --- a/vllm/model_executor/layers/vocab_parallel_embedding.py +++ b/vllm/model_executor/layers/vocab_parallel_embedding.py @@ -356,7 +356,7 @@ def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor): elif isinstance(param, UninitializedParameter): shape = list(loaded_weight.shape) if output_dim is not None: - shape[output_dim] = shape[output_dim] // self.tp_size + shape[output_dim] = self.num_embeddings_per_partition param.materialize(tuple(shape), dtype=loaded_weight.dtype) # If parameter does not have output dim, then it should @@ -382,7 +382,7 @@ def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor): else: assert loaded_weight.shape[output_dim] == self.org_vocab_size - # Copy the data. + # Copy the data. Select chunk corresponding to current shard. loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size) if current_platform.is_hpu(): diff --git a/vllm/model_executor/model_loader/loader.py b/vllm/model_executor/model_loader/loader.py index 0033fbff0e9ac..91ea1db624eb9 100644 --- a/vllm/model_executor/model_loader/loader.py +++ b/vllm/model_executor/model_loader/loader.py @@ -39,7 +39,8 @@ from vllm.model_executor.model_loader.tensorizer import ( TensorizerConfig, is_vllm_tensorized, load_with_tensorizer, serialize_vllm_model, tensorizer_weights_iterator) -from vllm.model_executor.model_loader.utils import (get_model_architecture, +from vllm.model_executor.model_loader.utils import (ParamMapping, + get_model_architecture, set_default_torch_dtype) from vllm.model_executor.model_loader.weight_utils import ( download_safetensors_index_file_from_hf, download_weights_from_hf, @@ -113,7 +114,7 @@ def _initialize_model( all_params = [param.name for param in signatures.parameters.values()] if "vllm_config" in all_params and "prefix" in all_params: # new-style model class - with set_current_vllm_config(vllm_config): + with set_current_vllm_config(vllm_config, check_compile=True): return model_class(vllm_config=vllm_config, prefix=prefix) msg = ("vLLM model class should accept `vllm_config` and `prefix` as " @@ -141,7 +142,7 @@ def _initialize_model( kwargs["lora_config"] = vllm_config.lora_config if "scheduler_config" in all_params: kwargs["scheduler_config"] = vllm_config.scheduler_config - with set_current_vllm_config(vllm_config): + with set_current_vllm_config(vllm_config, check_compile=True): return model_class(**kwargs) @@ -181,6 +182,9 @@ class Source: fall_back_to_pt: bool = True """Whether .pt weights can be used.""" + allow_patterns_overrides: Optional[list[str]] = None + """If defined, weights will load exclusively using these patterns.""" + def __init__(self, load_config: LoadConfig): super().__init__(load_config) if load_config.model_loader_extra_config: @@ -217,6 +221,7 @@ def _prepare_weights( model_name_or_path: str, revision: Optional[str], fall_back_to_pt: bool, + allow_patterns_overrides: Optional[list[str]], ) -> Tuple[str, List[str], bool]: """Prepare weights for the model. @@ -248,6 +253,9 @@ def _prepare_weights( if fall_back_to_pt: allow_patterns += ["*.pt"] + if allow_patterns_overrides is not None: + allow_patterns = allow_patterns_overrides + if not is_local: hf_folder = download_weights_from_hf( model_name_or_path, @@ -297,7 +305,8 @@ def _get_weights_iterator( ) -> Generator[Tuple[str, torch.Tensor], None, None]: """Get an iterator for the model weights based on the load format.""" hf_folder, hf_weights_files, use_safetensors = self._prepare_weights( - source.model_or_path, source.revision, source.fall_back_to_pt) + source.model_or_path, source.revision, source.fall_back_to_pt, + source.allow_patterns_overrides) if self.load_config.load_format == LoadFormat.NPCACHE: # Currently np_cache only support *.bin checkpoints assert use_safetensors is False @@ -339,6 +348,8 @@ def _get_all_weights( prefix="", fall_back_to_pt=getattr(model, "fall_back_to_pt_during_load", True), + allow_patterns_overrides=getattr(model, "allow_patterns_overrides", + None), ) yield from self._get_weights_iterator(primary_weights) @@ -352,7 +363,8 @@ def _get_all_weights( def download_model(self, model_config: ModelConfig) -> None: self._prepare_weights(model_config.model, model_config.revision, - fall_back_to_pt=True) + fall_back_to_pt=True, + allow_patterns_overrides=None) def load_model(self, vllm_config: VllmConfig) -> nn.Module: device_config = vllm_config.device_config @@ -983,21 +995,11 @@ def _unquantized_generator(self, hf_weights_files, use_safetensors, def _get_bnb_target_modules(self, model: nn.Module) -> None: - # TODO: Maybe we can replace bitsandbytes_stacked_params_mapping with - # packed_modules_mapping. - inverse_stacked_mapping: Dict[str, List[str]] = {} - for orig, ( - packed, - idx, - ) in model.bitsandbytes_stacked_params_mapping.items(): - if packed not in inverse_stacked_mapping: - inverse_stacked_mapping[packed] = [] - inverse_stacked_mapping[packed].insert(idx, orig) - for name, module in model.named_modules(): if isinstance(module, (LinearBase, )): last_name = name.split(".")[-1] - if sub_modules := inverse_stacked_mapping.get(last_name, []): + if sub_modules := self.modules_mapping.packed_mapping.get( + last_name, []): # Map vllm's names to transformers's names. for sub_name in sub_modules: self.target_modules.append( @@ -1018,10 +1020,18 @@ def _load_weights(self, model_config: ModelConfig, "The required method 'load_weights' is not defined in class" f" {type(model).__name__}.") - if not hasattr(model, "bitsandbytes_stacked_params_mapping"): + if not hasattr(model, "packed_modules_mapping"): raise AttributeError( f"Model {type(model).__name__} does not support BitsAndBytes " - "quantization yet.") + "quantization yet. No 'packed_modules_mapping' found.") + + self.modules_mapping = ParamMapping( + copy.deepcopy(model.packed_modules_mapping)) + + # For some models like Molmo, we need to use hf_to_vllm_mapper + # to ensure correct loading of weights. + if hf_to_vllm_mapper := getattr(model, "hf_to_vllm_mapper", None): + self.weight_mapper = lambda name: hf_to_vllm_mapper._map_name(name) # For some models like Molmo, we need to use hf_to_vllm_mapper # to ensure correct loading of weights. @@ -1070,8 +1080,8 @@ def _load_weights(self, model_config: ModelConfig, # weight tensor. So TP does not work with pre_quantized bnb models. if pre_quant and get_tensor_model_parallel_world_size() > 1: raise ValueError( - "Prequant BitsAndBytes models with TP is not supported." - "Please try with PP.") + "Prequant BitsAndBytes models with tensor parallelism is not " + "supported. Please try with pipeline parallelism.") load_8bit = False if pre_quant: @@ -1109,16 +1119,23 @@ def _load_weights(self, model_config: ModelConfig, for shard_name, ( weight_name, index, - ) in model.bitsandbytes_stacked_params_mapping.items(): - shard_pos = quant_param_name.find(shard_name) + ) in self.modules_mapping.inverse_packed_mapping.items(): # Some models, such as MiniCPM V2.5/2.6, contain both # module names 'kv_proj' and 'qkv_proj'. To prevent 'kv_proj' # from being incorrectly identified as being present in # 'vpm.encoder.layers.0.self_attn.qkv_proj.weight - if shard_pos > 0 and quant_param_name[shard_pos - 1] == ".": + shard_pos = quant_param_name.find(shard_name) + can_correct_rename = (shard_pos > 0) and ( + quant_param_name[shard_pos - 1] == ".") + # If the quant_param_name is packed, it won't occur in the + # param_dict before renaming. + new_quant_param_name = quant_param_name.replace( + shard_name, weight_name) + need_rename = (quant_param_name not in param_dict) \ + and (new_quant_param_name in param_dict) + if can_correct_rename and need_rename: shard_index = index - quant_param_name = quant_param_name.replace( - shard_name, weight_name) + quant_param_name = new_quant_param_name break # Models like Clip/Siglip may skip some layers in initialization, diff --git a/vllm/model_executor/model_loader/tensorizer.py b/vllm/model_executor/model_loader/tensorizer.py index fbd4937112e11..e359aef9dcb7f 100644 --- a/vllm/model_executor/model_loader/tensorizer.py +++ b/vllm/model_executor/model_loader/tensorizer.py @@ -288,7 +288,8 @@ def _init_model(self): model_args.torch_dtype = self.tensorizer_config.dtype assert self.tensorizer_config.model_class is not None # TODO: Do we need to consider old-style model class? - with no_init_or_tensor(), set_current_vllm_config(self.vllm_config): + with no_init_or_tensor(), set_current_vllm_config(self.vllm_config, + check_compile=True): return self.tensorizer_config.model_class( vllm_config=self.vllm_config, ) @@ -459,16 +460,7 @@ def tensorize_vllm_model(engine_args: EngineArgs, stream.write(encryption_params.key) engine = LLMEngine.from_engine_args(engine_args) - if tensorizer_config._is_sharded: - # if the engine is a distributed engine (for tensor parallel) then each - # worker shard needs to serialize its part of the model. - engine.model_executor._run_workers( - "save_tensorized_model", - tensorizer_config=tensorizer_config, - ) - else: - # with a single worker, we can get to the underlying model directly - serialize_vllm_model( - engine.model_executor.driver_worker.model_runner.model, - tensorizer_config, - ) + engine.model_executor.collective_rpc( + "save_tensorized_model", + kwargs=dict(tensorizer_config=tensorizer_config), + ) diff --git a/vllm/model_executor/model_loader/utils.py b/vllm/model_executor/model_loader/utils.py index 44978a55e072d..3f923d2f6632a 100644 --- a/vllm/model_executor/model_loader/utils.py +++ b/vllm/model_executor/model_loader/utils.py @@ -1,6 +1,7 @@ """Utilities for selecting and loading models.""" import contextlib -from typing import Tuple, Type +from dataclasses import dataclass, field +from typing import Dict, List, Tuple, Type import torch from torch import nn @@ -49,3 +50,26 @@ def get_model_architecture( def get_architecture_class_name(model_config: ModelConfig) -> str: return get_model_architecture(model_config)[1] + + +@dataclass +class ParamMapping: + """ + A class to handle parameter mapping for model weight loading. + It creates a bidirectional mapping between packed parameters and their + constituent parts. + """ + packed_mapping: Dict[str, List[str]] + inverse_packed_mapping: Dict[str, Tuple[str, + int]] = field(default_factory=dict) + + def __post_init__(self): + for packed_name, sub_params in self.packed_mapping.items(): + # Skip self-contained cases (e.g., {"W_pack": ["W_pack"]}) + if len(sub_params) == 1 and sub_params[0] == packed_name: + continue + for index, param_name in enumerate(sub_params): + self.inverse_packed_mapping[param_name] = ( + packed_name, + index, + ) diff --git a/vllm/model_executor/model_loader/weight_utils.py b/vllm/model_executor/model_loader/weight_utils.py index b70407221312a..b764a940b1742 100644 --- a/vllm/model_executor/model_loader/weight_utils.py +++ b/vllm/model_executor/model_loader/weight_utils.py @@ -93,7 +93,7 @@ def convert_bin_to_safetensor_file( pt_filename: str, sf_filename: str, ) -> None: - loaded = torch.load(pt_filename, map_location="cpu") + loaded = torch.load(pt_filename, map_location="cpu", weights_only=True) if "state_dict" in loaded: loaded = loaded["state_dict"] shared = _shared_pointers(loaded) @@ -381,7 +381,9 @@ def np_cache_weights_iterator( disable=not enable_tqdm, bar_format=_BAR_FORMAT, ): - state = torch.load(bin_file, map_location="cpu") + state = torch.load(bin_file, + map_location="cpu", + weights_only=True) for name, param in state.items(): param_path = os.path.join(np_folder, name) with open(param_path, "wb") as f: @@ -447,7 +449,7 @@ def pt_weights_iterator( disable=not enable_tqdm, bar_format=_BAR_FORMAT, ): - state = torch.load(bin_file, map_location="cpu") + state = torch.load(bin_file, map_location="cpu", weights_only=True) yield from state.items() del state torch.cuda.empty_cache() diff --git a/vllm/model_executor/models/aria.py b/vllm/model_executor/models/aria.py index 0ad403ec49220..8c6873de13627 100644 --- a/vllm/model_executor/models/aria.py +++ b/vllm/model_executor/models/aria.py @@ -1,9 +1,11 @@ -from typing import (Callable, Iterable, List, Mapping, Optional, Set, Tuple, - TypedDict, Union) +from typing import (Iterable, List, Mapping, Optional, Set, Tuple, TypedDict, + Union) import torch import torch.nn as nn -from transformers import BatchFeature, PretrainedConfig +from transformers import AriaConfig, AriaTextConfig, BatchFeature +from transformers.models.aria.modeling_aria import AriaCrossAttention +from transformers.models.aria.processing_aria import AriaProcessor from vllm.attention import AttentionMetadata from vllm.config import CacheConfig, QuantizationConfig, VllmConfig @@ -13,8 +15,6 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear, RowParallelLinear) from vllm.model_executor.layers.logits_processor import LogitsProcessor -from vllm.model_executor.layers.quantization.compressed_tensors.utils import ( - get_compressed_tensors_cache_scale) from vllm.model_executor.layers.sampler import (SamplerOutput, SamplingMetadata, get_sampler) from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead @@ -28,10 +28,12 @@ BaseProcessingInfo, PromptReplacement) from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs from vllm.sequence import IntermediateTensors -from vllm.transformers_utils.configs.aria import (AriaMoELMConfig, - AriaVisionConfig) -from .idefics2_vision_model import Idefics2VisionTransformer +# yapf: disable +from .idefics2_vision_model import Idefics2VisionConfig +from .idefics2_vision_model import ( + Idefics2VisionTransformer as Idefics3VisionTransformer) +# yapf: enable from .interfaces import SupportsMultiModal from .llama import LlamaDecoderLayer, LlamaMLP, LlamaModel from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn, @@ -49,87 +51,69 @@ class AriaImagePixelInputs(TypedDict): """ -class AriaVisionTransformer(Idefics2VisionTransformer): - """ - AriaVisionTransformer is a modified version of Idefics2VisionTransformer - that replaces the post-layernorm with an identity layer. - """ +class AriaVisionTransformer(Idefics3VisionTransformer): def __init__( self, - config: AriaVisionConfig, + config: Idefics2VisionConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: super().__init__(config, quant_config, prefix) + # Unlike Idefics3VisionTransformer which uses LayerNorm after the + # final layer, Aria omits this normalization, so we replace it with an + # Identity layer self.post_layernorm = nn.Identity() + def load_weights(self, weights: Iterable[Tuple[str, + torch.Tensor]]) -> Set[str]: + stacked_params_mapping = [ + # (param_name, shard_name, shard_id) + ("qkv_proj", "q_proj", "q"), + ("qkv_proj", "k_proj", "k"), + ("qkv_proj", "v_proj", "v"), + ] + params_dict = dict(self.named_parameters()) + loaded_params: Set[str] = set() + for name, loaded_weight in weights: + + # NOTE: post_layernorm is not used in Aria + if "post_layernorm" in name: + continue + + for param_name, weight_name, shard_id in stacked_params_mapping: + if weight_name not in name: + continue + name = name.replace(weight_name, param_name) + param = params_dict[name] + weight_loader = param.weight_loader + weight_loader(param, loaded_weight, shard_id) + break + else: + param = params_dict[name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + weight_loader(param, loaded_weight) + loaded_params.add(name) + return loaded_params + -class AriaVisionModel(nn.Module): - config_class = AriaVisionConfig +class AriaProjectorMLP(nn.Module): def __init__( self, - config: AriaVisionConfig, - quant_config: Optional[QuantizationConfig] = None, - *, - prefix: str = "", + in_features: int, + hidden_features: int, + output_dim: int, ) -> None: super().__init__() - self.vision_model = AriaVisionTransformer( - config, - quant_config, - prefix=f"{prefix}.vision_model", - ) - - def forward( - self, - pixel_values: torch.Tensor, - pixel_mask: Optional[torch.Tensor] = None, - ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]: - patch_attention_mask = self._create_patch_attention_mask(pixel_mask) - - vit_oup = self.vision_model( - pixel_values=pixel_values, - patch_attention_mask=patch_attention_mask, - ) - - image_atts = self._create_image_attention_mask(patch_attention_mask) - - return vit_oup, image_atts - - def _create_patch_attention_mask( - self, pixel_mask: Optional[torch.Tensor]) -> torch.Tensor: - if pixel_mask is None: - return None - - patches_subgrid = pixel_mask.unfold( - dimension=1, - size=self.vision_model.config.patch_size, - step=self.vision_model.config.patch_size, - ).unfold( - dimension=2, - size=self.vision_model.config.patch_size, - step=self.vision_model.config.patch_size, - ) - return (patches_subgrid.sum(dim=(-1, -2)) > 0).bool() - - def _create_image_attention_mask( - self, patch_attention_mask: torch.Tensor) -> torch.Tensor: - if patch_attention_mask is None: - return None - - flattened_mask = patch_attention_mask.flatten(1) - return torch.logical_not(flattened_mask) - - -class FFN(nn.Module): - - def __init__(self, embed_dim: int, ff_dim: int, output_dim: int) -> None: - super().__init__() - self.linear_in = ColumnParallelLinear(embed_dim, ff_dim, bias=False) - self.linear_out = RowParallelLinear(ff_dim, output_dim, bias=False) + self.linear_in = ColumnParallelLinear(in_features, + hidden_features, + bias=False) + self.linear_out = RowParallelLinear(hidden_features, + output_dim, + bias=False) self.act = get_act_fn("gelu_new") def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: @@ -139,46 +123,6 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: return hidden_states -class CrossAttention(nn.Module): - - def __init__(self, kv_dim: int, embed_dim: int, num_heads: int) -> None: - super().__init__() - self.num_heads = num_heads - self.q_proj = nn.Linear(embed_dim, embed_dim, bias=False) - self.k_proj = nn.Linear(kv_dim, embed_dim, bias=False) - self.v_proj = nn.Linear(kv_dim, embed_dim, bias=False) - - self.multihead_attn = nn.MultiheadAttention(embed_dim, num_heads) - self.linear = nn.Linear(embed_dim, embed_dim) - - self.layer_norm = nn.LayerNorm(embed_dim) - self.ln_kv = nn.LayerNorm(kv_dim) - - def forward( - self, - x: torch.Tensor, - hidden_states: torch.Tensor, - attn_mask: Optional[torch.Tensor] = None, - ) -> torch.Tensor: - normed_hidden_states = self.layer_norm(hidden_states) - query = self.q_proj(normed_hidden_states).permute(1, 0, 2) - - x = self.ln_kv(x) - key = self.k_proj(x).permute(1, 0, 2) - value = self.v_proj(x).permute(1, 0, 2) - - attn_output, _ = self.multihead_attn(query, - key, - value, - attn_mask=attn_mask) - - attn_output = attn_output.permute(1, 0, 2) - - attn_output = self.linear(attn_output) - - return attn_output - - class AriaProjector(nn.Module): """ A projection module with one cross attention layer and one FFN layer, which @@ -200,42 +144,42 @@ class AriaProjector(nn.Module): A tensor with the shape of (batch_size, query_number, output_dim) """ - def __init__( - self, - patch_to_query_dict: dict[int, int], - embed_dim: int, - num_heads: int, - kv_dim: int, - ff_dim: int, - output_dim: int, - norm_layer: Callable[[int], nn.Module] = nn.LayerNorm, - ) -> None: + def __init__(self, config: AriaConfig) -> None: super().__init__() - self.patch_to_query_dict = patch_to_query_dict - self.embed_dim = embed_dim - self.num_heads = num_heads + + self.patch_to_query_dict = config.projector_patch_to_query_dict + self.in_features = config.vision_config.hidden_size + self.num_heads = config.vision_config.num_attention_heads + self.kv_dim = config.vision_config.hidden_size + self.hidden_features = config.text_config.hidden_size + self.output_dim = config.text_config.hidden_size self.query = nn.Parameter( - torch.empty(max(patch_to_query_dict.values()), self.embed_dim)) + torch.empty(config.max_value_projector_patch_to_query_dict, + self.in_features)) - self.cross_attn = CrossAttention(kv_dim, embed_dim, num_heads) + self.cross_attn = AriaCrossAttention(config) - self.ln_ffn = norm_layer(embed_dim) - self.ffn = FFN(embed_dim, ff_dim, output_dim) + self.layer_norm = nn.LayerNorm(self.in_features) + self.feed_forward = AriaProjectorMLP(self.in_features, + self.hidden_features, + self.output_dim) def forward( self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None, ) -> torch.Tensor: - bs = x.shape[0] - queries = self.query.unsqueeze(0).repeat(bs, 1, 1) + batch_size, num_patches = x.shape[0], x.shape[1] + + if num_patches not in self.patch_to_query_dict: + raise KeyError(f"Number of patches {num_patches} not found in " + "patch_to_query_dict amongst possible values " + f"{self.patch_to_query_dict.keys()}.") - query_num = self.patch_to_query_dict.get(x.shape[1], None) - assert (query_num is not None - ), f"Query number for {x.shape[1]} patches is not provided" + query_num = self.patch_to_query_dict[num_patches] - queries = queries[:, :query_num, :] + queries = self.query[:query_num].unsqueeze(0).repeat(batch_size, 1, 1) if attn_mask is not None: attn_mask = attn_mask.repeat_interleave(self.num_heads, 0) @@ -243,7 +187,7 @@ def forward( attention_out = self.cross_attn(x, queries, attn_mask=attn_mask) - out = self.ffn(self.ln_ffn(attention_out)) + out = self.feed_forward(self.layer_norm(attention_out)) return out @@ -280,7 +224,7 @@ def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor, param.data.copy_(loaded_weight.transpose(1, 2)) -class MoELayer(nn.Module): +class AriaTextMoELayer(nn.Module): """ Mixture of Experts (MoE) Layer for the AriaMoE model. @@ -291,7 +235,7 @@ class MoELayer(nn.Module): def __init__( self, - config: AriaMoELMConfig, + config: AriaTextConfig, quant_config: Optional[QuantizationConfig], ) -> None: super().__init__() @@ -305,15 +249,16 @@ def __init__( num_experts=config.moe_num_experts, top_k=config.moe_topk, hidden_size=config.hidden_size, - intermediate_size=config.moe_intermediate_size, + intermediate_size=config.intermediate_size, quant_config=quant_config, reduce_results=True, ) self.shared_experts = LlamaMLP( config.hidden_size, - config.moe_intermediate_size * config.moe_num_shared_experts, + config.intermediate_size * config.moe_num_shared_experts, "silu", quant_config=quant_config, + bias=config.mlp_bias, ) def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: @@ -331,13 +276,15 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: router_output = torch.nn.functional.linear(hidden_states, self.router_weight) - shared_expert_output = self.shared_experts(hidden_states) + hidden_states_copy = hidden_states.clone() + # NOTE: hidden_states will be modified inplace by `FusedMoE` sparse_expert_output = self.experts(hidden_states, router_output) + shared_expert_output = self.shared_experts(hidden_states_copy) return sparse_expert_output + shared_expert_output -class MoEDecoderLayer(LlamaDecoderLayer): +class AriaTextDecoderLayer(LlamaDecoderLayer): """ Custom Decoder Layer for the AriaMoE model which modifies the standard `LlamaDecoderLayer` by replacing the traditional MLP with a Mixture of @@ -346,16 +293,16 @@ class MoEDecoderLayer(LlamaDecoderLayer): def __init__( self, - config: AriaMoELMConfig, + config: AriaTextConfig, cache_config: Optional[CacheConfig] = None, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: super().__init__(config, cache_config, quant_config, prefix) - self.mlp = MoELayer(config, quant_config=quant_config) + self.mlp = AriaTextMoELayer(config, quant_config=quant_config) -class AriaMoELMModel(LlamaModel): +class AriaTextModel(LlamaModel): """ Custom LlamaModel for the AriaMoE model which modifies the standard LlamaModel by replacing the `LlamaDecoderLayer` with `MoEDecoderLayer`. @@ -364,7 +311,7 @@ class AriaMoELMModel(LlamaModel): def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__(vllm_config=vllm_config, prefix=prefix, - layer_type=MoEDecoderLayer) + layer_type=AriaTextDecoderLayer) # Adapted from LlamaModel.load_weights with the modification of adding # the expert weights mapping to `stacked_params_mapping` @@ -390,14 +337,14 @@ def load_weights(self, weights: Iterable[Tuple[str, # Models trained using ColossalAI may include these tensors in # the checkpoint. Skip them. continue - if scale_name := get_compressed_tensors_cache_scale(name): - # Loading kv cache scales for compressed-tensors quantization + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales param = params_dict[scale_name] weight_loader = getattr(param, "weight_loader", default_weight_loader) - if loaded_weight.shape: - # scalar shape is torch.Size([1]), not torch.Size([]) - loaded_weight = loaded_weight[0] + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) weight_loader(param, loaded_weight) loaded_params.add(scale_name) continue @@ -436,25 +383,17 @@ def load_weights(self, weights: Iterable[Tuple[str, return loaded_params -def build_mm_projector(config: PretrainedConfig): - return AriaProjector( - patch_to_query_dict=config.projector_patch_to_query_dict, - embed_dim=config.vision_config.hidden_size, - num_heads=config.vision_config.num_attention_heads, - kv_dim=config.vision_config.hidden_size, - ff_dim=config.text_config.hidden_size, - output_dim=config.text_config.hidden_size, - ) - - class AriaProcessingInfo(BaseProcessingInfo): def get_hf_config(self): - return self.ctx.get_hf_config() + return self.ctx.get_hf_config(AriaConfig) - def get_vision_config(self) -> AriaVisionConfig: + def get_vision_config(self): return self.get_hf_config().vision_config + def get_hf_processor(self): + return self.ctx.get_hf_processor(AriaProcessor) + def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]: return {"image": None} @@ -486,7 +425,7 @@ def get_dummy_processor_inputs( } hf_processor = self.info.get_hf_processor() - image_token: str = hf_processor.image_token # type: ignore + image_token: str = hf_processor.tokenizer.image_token # type: ignore return ProcessorInputs( prompt_text=image_token * num_images, @@ -556,10 +495,14 @@ def __init__( quant_config = vllm_config.quant_config self.config = config - self.vision_tower = AriaVisionModel(config.vision_config) - self.multi_modal_projector = build_mm_projector(config) + self.vision_tower = AriaVisionTransformer( + config.vision_config, + quant_config, + prefix=f"{prefix}.vision_tower", + ) + self.multi_modal_projector = AriaProjector(config) self.vocab_size = config.text_config.vocab_size - self.language_model = AriaMoELMModel( + self.language_model = AriaTextModel( vllm_config=vllm_config.with_hf_config(config.text_config), prefix=maybe_prefix(prefix, "language_model.model"), ) @@ -610,6 +553,22 @@ def _parse_and_validate_image_input( pixel_mask=pixel_mask, ) + def _create_patch_attention_mask( + self, pixel_mask: Optional[torch.Tensor]) -> torch.Tensor: + if pixel_mask is None: + return None + + patches_subgrid = pixel_mask.unfold( + dimension=1, + size=self.vision_tower.config.patch_size, + step=self.vision_tower.config.patch_size, + ).unfold( + dimension=2, + size=self.vision_tower.config.patch_size, + step=self.vision_tower.config.patch_size, + ) + return (patches_subgrid.sum(dim=(-1, -2)) > 0).bool() + def _process_image_input( self, image_input: AriaImagePixelInputs ) -> Tuple[torch.Tensor, torch.Tensor]: @@ -618,9 +577,18 @@ def _process_image_input( pixel_values = image_input['pixel_values'] pixel_mask = image_input['pixel_mask'] - image_feature, image_attn_mask = self.vision_tower( - pixel_values, pixel_mask=pixel_mask) - return self.multi_modal_projector(image_feature, image_attn_mask) + patch_attention_mask = self._create_patch_attention_mask(pixel_mask) + + image_outputs = self.vision_tower( + pixel_values=pixel_values, + patch_attention_mask=patch_attention_mask, + ) + image_attn_mask = None + if patch_attention_mask is not None: + flattened_mask = patch_attention_mask.flatten(1) + image_attn_mask = torch.logical_not(flattened_mask) + + return self.multi_modal_projector(image_outputs, image_attn_mask) def get_multimodal_embeddings(self, **kwargs) -> Optional[NestedTensors]: image_input = self._parse_and_validate_image_input(**kwargs) @@ -685,6 +653,5 @@ def sample( return next_tokens def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): - loader = AutoWeightsLoader(self) loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) diff --git a/vllm/model_executor/models/baichuan.py b/vllm/model_executor/models/baichuan.py index 5e68b7f165bf4..a923ed36a9db2 100644 --- a/vllm/model_executor/models/baichuan.py +++ b/vllm/model_executor/models/baichuan.py @@ -350,13 +350,6 @@ class BaiChuanBaseForCausalLM(nn.Module, SupportsLoRA, SupportsPP): embedding_modules = {} embedding_padding_modules = [] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } - def __init__( self, *, diff --git a/vllm/model_executor/models/blip2.py b/vllm/model_executor/models/blip2.py index 917b88e802071..b559ac677a740 100644 --- a/vllm/model_executor/models/blip2.py +++ b/vllm/model_executor/models/blip2.py @@ -14,12 +14,12 @@ from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY -from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, - MultiModalInputsV2, MultiModalKwargs, - NestedTensors, PlaceholderRange) +from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs, + NestedTensors) from vllm.multimodal.parse import MultiModalDataItems from vllm.multimodal.processing import (BaseMultiModalProcessor, - BaseProcessingInfo, PromptReplacement) + BaseProcessingInfo, PromptReplacement, + PromptReplacementDetails) from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs from vllm.sequence import IntermediateTensors @@ -475,36 +475,27 @@ def _get_prompt_replacements( hf_processor_mm_kwargs: Mapping[str, object], out_mm_kwargs: MultiModalKwargs, ) -> list[PromptReplacement]: + tokenizer = self.info.get_tokenizer() + vocab = tokenizer.get_vocab() + + bos_token_id = tokenizer.bos_token_id + assert isinstance(bos_token_id, int) + + image_token_id = vocab[""] num_image_tokens = self.info.get_num_image_tokens() + image_tokens = [image_token_id] * num_image_tokens return [ PromptReplacement( modality="image", - target="", - replacement="" * num_image_tokens + "", + target=[bos_token_id], + replacement=PromptReplacementDetails( + full=image_tokens + [bos_token_id], + features=image_tokens, + ), ) ] - def apply( - self, - prompt: Union[str, list[int]], - mm_data: MultiModalDataDict, - hf_processor_mm_kwargs: Mapping[str, object], - ) -> MultiModalInputsV2: - result = super().apply(prompt, mm_data, hf_processor_mm_kwargs) - - # Only tokens should be considered as placeholders, - # so we ignore the trailing bos_token - result["mm_placeholders"] = { - modality: [ - PlaceholderRange(offset=p["offset"], length=p["length"] - 1) - for p in ps - ] - for modality, ps in result["mm_placeholders"].items() - } - - return result - @MULTIMODAL_REGISTRY.register_processor(Blip2MultiModalProcessor, info=Blip2ProcessingInfo, diff --git a/vllm/model_executor/models/chameleon.py b/vllm/model_executor/models/chameleon.py index a6634204699c9..e834c9004f140 100644 --- a/vllm/model_executor/models/chameleon.py +++ b/vllm/model_executor/models/chameleon.py @@ -28,12 +28,12 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.model_executor.utils import set_weight_attrs from vllm.multimodal import MULTIMODAL_REGISTRY -from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, - MultiModalInputsV2, MultiModalKwargs, - NestedTensors, PlaceholderRange) +from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs, + NestedTensors) from vllm.multimodal.parse import MultiModalDataItems from vllm.multimodal.processing import (BaseMultiModalProcessor, - BaseProcessingInfo, PromptReplacement) + BaseProcessingInfo, PromptReplacement, + PromptReplacementDetails) from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs from vllm.sequence import IntermediateTensors @@ -122,8 +122,9 @@ def _apply_hf_processor_tokens_only( ) -> list[int]: # HF processor adds sep token for chat mode tokenizer = self.info.get_tokenizer() - sep_token_id: int = \ - tokenizer.vocab[tokenizer.sep_token] # type: ignore + vocab = tokenizer.get_vocab() + + sep_token_id = vocab[tokenizer.sep_token] # type: ignore return prompt_tokens + [sep_token_id] @@ -141,39 +142,27 @@ def _get_prompt_replacements( out_mm_kwargs: MultiModalKwargs, ) -> list[PromptReplacement]: processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) + tokenizer = self.info.get_tokenizer() + vocab = tokenizer.get_vocab() + + image_start_id = vocab[processor.image_start_token] + image_token_id = vocab[processor.image_token] + image_end_id = vocab[processor.image_end_token] + + num_image_tokens = self.info.get_num_image_tokens() + image_tokens = [image_token_id] * num_image_tokens return [ PromptReplacement( modality="image", - target="", - replacement="".join([ - processor.image_start_token, - processor.image_token * self.info.get_num_image_tokens(), - processor.image_end_token, - ]), + target=[image_token_id], + replacement=PromptReplacementDetails( + full=([image_start_id] + image_tokens + [image_end_id]), + features=image_tokens, + ), ) ] - def apply( - self, - prompt: Union[str, list[int]], - mm_data: MultiModalDataDict, - hf_processor_mm_kwargs: Mapping[str, object], - ) -> MultiModalInputsV2: - result = super().apply(prompt, mm_data, hf_processor_mm_kwargs) - - # Only tokens should be considered as placeholders, - # so we ignore the image_start_token and image_end_token - result["mm_placeholders"] = { - modality: [ - PlaceholderRange(offset=p["offset"] + 1, - length=p["length"] - 2) for p in ps - ] - for modality, ps in result["mm_placeholders"].items() - } - - return result - class ChameleonLayerNorm(nn.LayerNorm): diff --git a/vllm/model_executor/models/chatglm.py b/vllm/model_executor/models/chatglm.py index 7e37ce3086e6b..d5f9b4d19e5ca 100644 --- a/vllm/model_executor/models/chatglm.py +++ b/vllm/model_executor/models/chatglm.py @@ -41,7 +41,7 @@ from vllm.transformers_utils.configs import ChatGLMConfig from .interfaces import SupportsLoRA, SupportsMultiModal, SupportsPP -from .utils import (is_pp_missing_parameter, +from .utils import (AutoWeightsLoader, WeightsMapper, is_pp_missing_parameter, make_empty_intermediate_tensors_factory, make_layers, maybe_prefix) @@ -605,9 +605,50 @@ def forward( return IntermediateTensors({"hidden_states": hidden_states}) return hidden_states + def load_weights(self, weights: Iterable[Tuple[str, + torch.Tensor]]) -> Set[str]: + stacked_params_mapping = [ + # (param_name, shard_name, shard_id) + ("linear_proj.merged_proj", "linear_proj.gate_proj", 0), + ("linear_proj.merged_proj", "linear_proj.dense_h_to_4h", 1), + ] + params_dict = dict(self.named_parameters()) + loaded_params: Set[str] = set() + + for name, loaded_weight in weights: + for (param_name, weight_name, shard_id) in stacked_params_mapping: + if weight_name not in name: + continue + name = name.replace(weight_name, param_name) + # Skip loading extra bias for GPTQ models. + if name.endswith(".bias") and name not in params_dict: + continue + if is_pp_missing_parameter(name, self): + continue + param = params_dict[name] + weight_loader = param.weight_loader + weight_loader(param, loaded_weight, shard_id) + break + else: + if "rotary_pos_emb.inv_freq" in name: + continue + if name.endswith(".bias") and name not in params_dict: + continue + if is_pp_missing_parameter(name, self): + continue + param = params_dict[name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + weight_loader(param, loaded_weight) + loaded_params.add(name) + return loaded_params + class ChatGLMBaseModel(nn.Module, SupportsLoRA, SupportsPP): + hf_to_vllm_mapper = WeightsMapper( + orig_to_new_substr={".word_embeddings": ""}, ) + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_config @@ -660,52 +701,9 @@ def sample( next_tokens = self.sampler(logits, sampling_metadata) return next_tokens - def load_weights(self, weights: Iterable[Tuple[str, - torch.Tensor]]) -> Set[str]: - # Merge two ColumnParallelLinear into one MergedColumnParallelLinear - merged_weights_dict: Dict[str, Dict[str, Optional[torch.Tensor]]] = { - "transformer.vision.linear_proj.merged_proj.weight": { - "transformer.vision.linear_proj.gate_proj.weight": None, - "transformer.vision.linear_proj.dense_h_to_4h.weight": None, - } - } - - params_dict = dict(self.named_parameters(remove_duplicate=False)) - loaded_params: Set[str] = set() - for name, loaded_weight in weights: - is_weight_to_be_merge = False - for _, merged_weight_dict in merged_weights_dict.items(): - if name in merged_weight_dict: - assert merged_weight_dict[name] is None - merged_weight_dict[name] = loaded_weight - is_weight_to_be_merge = True - if is_weight_to_be_merge: - continue - if "rotary_pos_emb.inv_freq" in name: - continue - if "word_embeddings" in name: - name = name.replace(".word_embeddings", "") - # Skip loading extra bias for GPTQ models. - if name.endswith(".bias") and name not in params_dict: - continue - if is_pp_missing_parameter(name, self): - continue - param = params_dict[name] - weight_loader = getattr(param, "weight_loader", - default_weight_loader) - weight_loader(param, loaded_weight) - loaded_params.add(name) - - for combined_name, merged_weight_dict in merged_weights_dict.items(): - if combined_name in params_dict: - param = params_dict[combined_name] - combined_weight = torch.cat(list(merged_weight_dict.values()), - dim=0) - weight_loader = getattr(param, "weight_loader", - default_weight_loader) - weight_loader(param, combined_weight) - loaded_params.add(combined_name) - return loaded_params + def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): + loader = AutoWeightsLoader(self) + return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) class ChatGLM(ChatGLMBaseModel): @@ -726,6 +724,7 @@ class ChatGLM(ChatGLMBaseModel): class ChatGLMV(ChatGLMBaseModel, SupportsMultiModal): + packed_modules_mapping = { "query_key_value": ["query_key_value"], "dense_h_to_4h": ["dense_h_to_4h"], @@ -777,7 +776,7 @@ def __new__( ) -> None: config = vllm_config.model_config.hf_config # Initialize VL - if hasattr(config, "visual"): + if hasattr(config, "vision_config"): return ChatGLMV(vllm_config=vllm_config, prefix=prefix) # Initialize LLM else: diff --git a/vllm/model_executor/models/commandr.py b/vllm/model_executor/models/commandr.py index 570fc796d4b2e..1d39f9b7527f5 100644 --- a/vllm/model_executor/models/commandr.py +++ b/vllm/model_executor/models/commandr.py @@ -435,6 +435,19 @@ def load_weights(self, weights: Iterable[Tuple[str, params_dict = dict(self.named_parameters()) loaded_params: Set[str] = set() for name, loaded_weight in weights: + + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales + param = params_dict[scale_name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) + weight_loader(param, loaded_weight) + loaded_params.add(scale_name) + continue + for param_name, shard_name, shard_id in stacked_params_mapping: if shard_name not in name: continue diff --git a/vllm/model_executor/models/dbrx.py b/vllm/model_executor/models/dbrx.py index 2c64ba67d877b..b2aa3c0709bd4 100644 --- a/vllm/model_executor/models/dbrx.py +++ b/vllm/model_executor/models/dbrx.py @@ -442,14 +442,26 @@ def sample( def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]) -> Set[str]: - expert_params_mapping = [( "w13" if weight_name in ["w1", "v1"] else "w2", f"mlp.{weight_name}", ) for weight_name in ["w1", "v1", "w2"]] params_dict = dict(self.named_parameters(remove_duplicate=False)) loaded_params: Set[str] = set() + for name, loaded_weight in weights: + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales + param = params_dict[scale_name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) + weight_loader(param, loaded_weight) + loaded_params.add(scale_name) + continue + if name.endswith(("w1", "w2", "v1")): name = name + "_weight" for param_name, weight_name in expert_params_mapping: @@ -462,6 +474,7 @@ def load_weights(self, weights: Iterable[Tuple[str, weight_loader = param.weight_loader weight_loader(param, loaded_weight, weight_name, name) break + else: # Remapping the name of FP8 kv-scale. name = maybe_remap_kv_scale_name(name, params_dict) diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py index 5968716a5c06c..7636360b143e5 100644 --- a/vllm/model_executor/models/deepseek_v2.py +++ b/vllm/model_executor/models/deepseek_v2.py @@ -262,14 +262,8 @@ def __init__( mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim)) self.scaling = self.scaling * mscale * mscale - # self.attn = Attention(self.num_heads, - # self.qk_head_dim, - # self.scaling, - # num_kv_heads=self.num_heads) - - # TODO, support head_size 192 self.attn = Attention(self.num_local_heads, - 256, + self.qk_head_dim, self.scaling, num_kv_heads=self.num_local_heads, cache_config=cache_config, @@ -319,18 +313,14 @@ def forward( k = torch.empty_like(q) k[..., :self.qk_nope_head_dim] = k_nope k[..., self.qk_nope_head_dim:] = k_pe - q = torch.nn.functional.pad(q, [0, 256 - self.qk_head_dim], - value=0).view(-1, - self.num_local_heads * 256) - k = torch.nn.functional.pad(k, [0, 256 - self.qk_head_dim], - value=0).view(-1, - self.num_local_heads * 256) - v = torch.nn.functional.pad(v, [0, 256 - self.v_head_dim], - value=0).view(-1, - self.num_local_heads * 256) + # padding value to qk_head_dim for alignment + v = torch.nn.functional.pad( + v, [0, self.qk_head_dim - self.v_head_dim], + value=0).view(-1, self.num_local_heads * self.qk_head_dim) attn_output = self.attn(q, k, v, kv_cache, attn_metadata) attn_output = attn_output.view( - -1, self.num_local_heads, 256)[..., :self.v_head_dim].reshape( + -1, self.num_local_heads, + self.qk_head_dim)[..., :self.v_head_dim].reshape( -1, self.num_local_heads * self.v_head_dim) output, _ = self.o_proj(attn_output) return output diff --git a/vllm/model_executor/models/deepseek_v3.py b/vllm/model_executor/models/deepseek_v3.py index ca79b14c55fea..0b44f0d062c40 100644 --- a/vllm/model_executor/models/deepseek_v3.py +++ b/vllm/model_executor/models/deepseek_v3.py @@ -269,14 +269,8 @@ def __init__( mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim)) self.scaling = self.scaling * mscale * mscale - # self.attn = Attention(self.num_heads, - # self.qk_head_dim, - # self.scaling, - # num_kv_heads=self.num_heads) - - # TODO, support head_size 192 self.attn = Attention(self.num_local_heads, - 256, + self.qk_head_dim, self.scaling, num_kv_heads=self.num_local_heads, cache_config=cache_config, @@ -326,18 +320,14 @@ def forward( k = torch.empty_like(q) k[..., :self.qk_nope_head_dim] = k_nope k[..., self.qk_nope_head_dim:] = k_pe - q = torch.nn.functional.pad(q, [0, 256 - self.qk_head_dim], - value=0).view(-1, - self.num_local_heads * 256) - k = torch.nn.functional.pad(k, [0, 256 - self.qk_head_dim], - value=0).view(-1, - self.num_local_heads * 256) - v = torch.nn.functional.pad(v, [0, 256 - self.v_head_dim], - value=0).view(-1, - self.num_local_heads * 256) + # padding value to qk_head_dim for alignment + v = torch.nn.functional.pad( + v, [0, self.qk_head_dim - self.v_head_dim], + value=0).view(-1, self.num_local_heads * self.qk_head_dim) attn_output = self.attn(q, k, v, kv_cache, attn_metadata) attn_output = attn_output.view( - -1, self.num_local_heads, 256)[..., :self.v_head_dim].reshape( + -1, self.num_local_heads, + self.qk_head_dim)[..., :self.v_head_dim].reshape( -1, self.num_local_heads * self.v_head_dim) output, _ = self.o_proj(attn_output) return output diff --git a/vllm/model_executor/models/deepseek_vl2.py b/vllm/model_executor/models/deepseek_vl2.py index 99fa941c055d2..344832d8b33e6 100644 --- a/vllm/model_executor/models/deepseek_vl2.py +++ b/vllm/model_executor/models/deepseek_vl2.py @@ -1,7 +1,7 @@ # adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/faf18023f24b962b32d9f0a2d89e402a8d383a78/deepseek_vl2/models/modeling_deepseek_vl_v2.py """Inference-only Deepseek-VL2 model compatible with HuggingFace weights.""" import math -from functools import cached_property, partial +from functools import cached_property from typing import (Iterable, List, Literal, Mapping, Optional, Set, Tuple, TypedDict, Union) @@ -9,7 +9,7 @@ import torch.nn as nn import torch.nn.functional as F from einops import rearrange, repeat -from transformers import AutoProcessor, BatchFeature, ProcessorMixin +from transformers import BatchFeature from vllm.attention import AttentionMetadata from vllm.config import VllmConfig @@ -31,6 +31,8 @@ from vllm.transformers_utils.configs.deepseek_vl2 import (DeepseekVLV2Config, MlpProjectorConfig, VisionEncoderConfig) +from vllm.transformers_utils.processors.deepseek_vl2 import ( + DeepseekVLV2Processor) from vllm.utils import is_list_of from .interfaces import SupportsMultiModal, SupportsPP @@ -129,25 +131,8 @@ class DeepseekVL2ProcessingInfo(BaseProcessingInfo): def get_hf_config(self): return self.ctx.get_hf_config(DeepseekVLV2Config) - def get_hf_processor(self) -> ProcessorMixin: - # TODO(Isotr0py): we should get rid of dependency on deepseek_vl2 - # in the future, because it's flasky and lack of maintenance. - try: - from deepseek_vl2.models.processing_deepseek_vl_v2 import ( - DeepseekVLV2Processor, select_best_resolution) - AutoProcessor.register("DeepseekVLV2Processor", - DeepseekVLV2Processor) - except ModuleNotFoundError as exc: - raise ModuleNotFoundError( - "You need to `pip install " - "git+https://github.com/deepseek-ai/DeepSeek-VL2.git` " - "to use this model") from exc - - processor = self.ctx.get_hf_processor(DeepseekVLV2Processor) - processor.select_best_resolution = partial( - select_best_resolution, - candidate_resolutions=processor.candidate_resolutions) - return processor + def get_hf_processor(self) -> DeepseekVLV2Processor: + return self.ctx.get_hf_processor(DeepseekVLV2Processor) def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]: return {"image": None} @@ -224,31 +209,21 @@ def _call_hf_processor( mm_kwargs: Mapping[str, object], ) -> BatchFeature: if mm_data: - outputs = self.info.ctx.call_hf_processor( + processed_outputs = self.info.ctx.call_hf_processor( self.info.get_hf_processor(**mm_kwargs), dict(prompt=prompt, **mm_data), mm_kwargs, ) - - # Deepseek-vl2 processor don't return BatchFeature, - # we need to manually create it - processed_outputs = dict(input_ids=outputs["input_ids"]) - processed_outputs = BatchFeature(data=dict(processed_outputs), - tensor_type="pt") - - # Remove batch dimension from processor outputs, - # because we will try batch to create NestedTensors target_dtype = self.info.ctx.model_config.dtype - pixel_values = outputs["images"].to(target_dtype).squeeze(0) - images_spatial_crop = outputs["images_spatial_crop"].squeeze(0) + pixel_values = processed_outputs.pop("pixel_values").to( + target_dtype) + # split pixel values into patches corresponding to each image + images_spatial_crop = processed_outputs["images_spatial_crop"] patches_per_image = [ x.prod().item() + 1 for x in images_spatial_crop ] - - # Rename `images` -> `pixel_values` to avoid confusion - processed_outputs["pixel_values"] = list( - pixel_values.split(patches_per_image)) - processed_outputs["images_spatial_crop"] = images_spatial_crop + pixel_values = pixel_values.split(patches_per_image) + processed_outputs["pixel_values"] = pixel_values else: tokenizer = self.info.get_tokenizer() processed_outputs = tokenizer(prompt, @@ -274,8 +249,10 @@ def _get_prompt_replacements( hf_processor_mm_kwargs: Mapping[str, object], out_mm_kwargs: MultiModalKwargs, ) -> list[PromptReplacement]: - hf_processor = self.info.get_hf_processor() - image_token_id: int = hf_processor.image_token_id + hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) + + image_token_id = hf_processor.image_token_id + assert isinstance(image_token_id, int) def get_replacement_deepseek_vl2(item_idx: int): images = mm_items.get_items( @@ -356,13 +333,18 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): f"Only 2D tile_tag is supported currently, got: {self.tile_tag}" ) + if self.text_config.topk_method == "noaux_tc": + architectures = ["DeepseekV3ForCausalLM"] + elif not self.text_config.use_mla: + architectures = ["DeepseekForCausalLM"] + else: + architectures = ["DeepseekV2ForCausalLM"] + self.language_model = init_vllm_registered_model( vllm_config=vllm_config, hf_config=self.text_config, prefix=maybe_prefix(prefix, "language"), - architectures=["DeepseekV3ForCausalLM"] - if self.text_config.topk_method == "noaux_tc" else - ["DeepseekV2ForCausalLM"], + architectures=architectures, ) self.make_empty_intermediate_tensors = ( diff --git a/vllm/model_executor/models/exaone.py b/vllm/model_executor/models/exaone.py index ca85f418d5762..bc3295da7b60a 100644 --- a/vllm/model_executor/models/exaone.py +++ b/vllm/model_executor/models/exaone.py @@ -38,8 +38,6 @@ RowParallelLinear) from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig -from vllm.model_executor.layers.quantization.compressed_tensors.utils import ( - get_compressed_tensors_cache_scale) from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -428,14 +426,6 @@ class ExaoneForCausalLM(nn.Module, SupportsLoRA, SupportsPP): "lm_head": "output_embeddings", } embedding_padding_modules = ["lm_head"] - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "c_fc_0": ("gate_up_proj", 0), - "c_fc_1": ("gate_up_proj", 1), - } def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() @@ -445,6 +435,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): self.config = config self.lora_config = lora_config + self.quant_config = quant_config self.transformer = ExaoneModel( vllm_config=vllm_config, @@ -538,14 +529,14 @@ def load_weights(self, weights: Iterable[Tuple[str, # processed with quantization, LoRA, fine-tuning, etc. if self.config.tie_word_embeddings and "lm_head.weight" in name: continue - if scale_name := get_compressed_tensors_cache_scale(name): - # Loading kv cache scales for compressed-tensors quantization + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales param = params_dict[scale_name] weight_loader = getattr(param, "weight_loader", default_weight_loader) - if loaded_weight.shape: - # scalar shape is torch.Size([1]), not torch.Size([]) - loaded_weight = loaded_weight[0] + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) weight_loader(param, loaded_weight) loaded_params.add(scale_name) continue diff --git a/vllm/model_executor/models/fairseq2_llama.py b/vllm/model_executor/models/fairseq2_llama.py new file mode 100644 index 0000000000000..b93a68680375d --- /dev/null +++ b/vllm/model_executor/models/fairseq2_llama.py @@ -0,0 +1,151 @@ +# Copyright 2024 The vLLM team. +# Copyright 2024 Meta Platforms, Inc. and affiliates. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Llama model for fairseq2 weights.""" + +from typing import Iterable, Set, Tuple + +import torch +from torch.nn import Parameter + +from vllm.config import VllmConfig +from vllm.distributed import (get_tensor_model_parallel_rank, + get_tensor_model_parallel_world_size) +from vllm.model_executor.layers.linear import set_weight_attrs +from vllm.model_executor.models.llama import LlamaForCausalLM + +from .utils import AutoWeightsLoader, WeightsMapper + + +class Fairseq2LlamaForCausalLM(LlamaForCausalLM): + + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): + super().__init__(vllm_config=vllm_config, prefix=prefix) + self.tp_rank = get_tensor_model_parallel_rank() + self.tp_size = get_tensor_model_parallel_world_size() + # For the model loader to read only the relevant checkpoint files + self.allow_patterns_overrides = [ + # either the full checkpoint + "model.pt", + # or the tp-sharded checkpoint of the current rank + f"model.{self.tp_rank}.pt", + ] + + def load_weights(self, weights: Iterable[Tuple[str, + torch.Tensor]]) -> Set[str]: + # fairseq2's serialization adds a wrapper to usual .pt state_dict's: + # { "model_key": my_model_name, "my_model_name": state_dict } + # which we first need to unpack + weights_wrapped = dict(weights) + weights = weights_wrapped[ + weights_wrapped["model_key"]].items() # type: ignore + + # remap keys + fs2_to_vllm_mapper = WeightsMapper( + orig_to_new_prefix={ + "decoder_frontend.embed.": "model.embed_tokens.", + "decoder.": "model.", + "final_proj.": "lm_head.", + }, + orig_to_new_substr={ + ".self_attn_layer_norm.": ".input_layernorm.", + ".ffn_layer_norm.": ".post_attention_layernorm.", + ".self_attn.output_proj.": ".self_attn.o_proj.", + ".ffn.gate_proj.": ".mlp.gate_proj.", + ".ffn.inner_proj.": ".mlp.up_proj.", + ".ffn.output_proj.": ".mlp.down_proj.", + ".layer_norm.": ".norm.", + }, + ) + weights = fs2_to_vllm_mapper.apply(weights) + + params = dict(self.named_parameters()) + + loader = AutoWeightsLoader( + self, + skip_prefixes=(["lm_head."] + if self.config.tie_word_embeddings else None), + ) + return loader.load_weights( + (self.reshape_fairseq2_weights(name, loaded_weight, params) + for name, loaded_weight in weights)) + + def flag_sharded_weights(self, params: dict[str, Parameter]): + """Sets the `is_sharded_weight` flag to True for all sharded weights""" + for name, param in params.items(): + modules = name.split(".") + if "norm" in name and len(param.size()) < 2: + # layer norms are not sharded + continue + elif any(emb in modules for emb in ["embed_tokens", "lm_head"]): + # for now we repeat embedding layers for compatibility + continue + else: + # all other layers are sharded + set_weight_attrs(param, {"is_sharded_weight": True}) + + def reshape_fairseq2_weights( + self, + name: str, + loaded_weight: torch.Tensor, + params: dict[str, Parameter], + ) -> Tuple[str, torch.Tensor]: + """Reshape fairseq2's weights.""" + + def permute(w: torch.Tensor, n_heads: int) -> torch.Tensor: + attn_in = self.config.head_dim * n_heads + # check for a sharded weight on dim 0 + if attn_in // self.tp_size == w.size()[0]: + attn_in //= self.tp_size + n_heads //= self.tp_size + attn_out = self.config.hidden_size + return (w.view(n_heads, attn_in // n_heads // 2, 2, + attn_out).transpose(1, + 2).reshape(attn_in, attn_out)) + + modules = name.split(".") + + # rotary embeds should be sliced + if "k_proj" in modules: + loaded_weight = permute(loaded_weight, + self.config.num_key_value_heads) + + elif "q_proj" in modules: + loaded_weight = permute(loaded_weight, + self.config.num_attention_heads) + + # We make the loaded weights compatible with both + # full checkpoints and tp sharded checkpoints. + # Embeddings are repeated to fit the vocab size. + # Other weights are flagged for the weight_loader calls. + if any(emb in modules for emb in ["embed_tokens", "lm_head"]): + # Embeddings are sharded on dim 0 + dim = 0 + # In fairseq2, vocab size has to be divisible by tp_size + # so we don't worry about padding + if self.tp_size > 1 and loaded_weight.shape[ + dim] < self.config.vocab_size: + assert loaded_weight.shape[ + dim] * self.tp_size == self.config.vocab_size, \ + "vocab_size should be divisible by tp_size." + repeats = [1] * len(loaded_weight.size()) + repeats[dim] = self.tp_size + # repeat to match vocab size and to be easily 'narrow'able + loaded_weight = loaded_weight.repeat(repeats) + set_weight_attrs(params[name], {"is_sharded_weight": False}) + # if embeddings are sharded, the rest is too + if "embed_tokens" in modules: + self.flag_sharded_weights(params) + + return name, loaded_weight diff --git a/vllm/model_executor/models/falcon.py b/vllm/model_executor/models/falcon.py index 8660cf79b9cdb..c503a368e8244 100644 --- a/vllm/model_executor/models/falcon.py +++ b/vllm/model_executor/models/falcon.py @@ -409,9 +409,9 @@ def forward( class FalconForCausalLM(nn.Module, SupportsPP): - - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = {} + packed_modules_mapping = { + "query_key_value": ["query_key_value"], + } def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() diff --git a/vllm/model_executor/models/fuyu.py b/vllm/model_executor/models/fuyu.py index 63e7147f84e03..dbf9da50cc9de 100644 --- a/vllm/model_executor/models/fuyu.py +++ b/vllm/model_executor/models/fuyu.py @@ -16,7 +16,7 @@ """ PyTorch Fuyu model.""" import math from typing import (Iterable, List, Literal, Mapping, Optional, Set, Tuple, - TypedDict, Union) + TypedDict) import torch import torch.nn as nn @@ -30,13 +30,13 @@ from vllm.model_executor.models.persimmon import PersimmonForCausalLM from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY -from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, - MultiModalInputsV2, MultiModalKwargs, - NestedTensors, PlaceholderRange) +from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs, + NestedTensors) from vllm.multimodal.parse import (ImageProcessorItems, ImageSize, MultiModalDataItems) from vllm.multimodal.processing import (BaseMultiModalProcessor, - BaseProcessingInfo, PromptReplacement) + BaseProcessingInfo, PromptReplacement, + PromptReplacementDetails) from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs from vllm.sequence import IntermediateTensors @@ -183,7 +183,9 @@ def _apply_hf_processor_tokens_only( ) -> list[int]: # HF processor adds boa_token_id tokenizer = self.info.get_tokenizer() - boa_token_id: int = tokenizer.vocab["<0x04>"] # type: ignore + vocab = tokenizer.get_vocab() + + boa_token_id = vocab["<0x04>"] return prompt_tokens + [boa_token_id] @@ -202,6 +204,7 @@ def _get_prompt_replacements( ) -> list[PromptReplacement]: hf_config = self.info.get_hf_config() bos_token_id = hf_config.bos_token_id + assert isinstance(bos_token_id, int) tokenizer = self.info.get_tokenizer() eot_token_id = tokenizer.bos_token_id @@ -215,9 +218,13 @@ def get_replacement_fuyu(item_idx: int): image_width=image_size.width, image_height=image_size.height, ) + image_tokens = ([_IMAGE_TOKEN_ID] * ncols + + [_NEWLINE_TOKEN_ID]) * nrows - return (([_IMAGE_TOKEN_ID] * ncols + [_NEWLINE_TOKEN_ID]) * nrows + - [bos_token_id]) + return PromptReplacementDetails( + full=image_tokens + [bos_token_id], + features=image_tokens, + ) return [ PromptReplacement( @@ -227,26 +234,6 @@ def get_replacement_fuyu(item_idx: int): ) ] - def apply( - self, - prompt: Union[str, list[int]], - mm_data: MultiModalDataDict, - hf_processor_mm_kwargs: Mapping[str, object], - ) -> MultiModalInputsV2: - result = super().apply(prompt, mm_data, hf_processor_mm_kwargs) - - # Only |SPEAKER| (image) tokens should be considered as placeholders, - # so we ignore the trailing bos_token_id - result["mm_placeholders"] = { - modality: [ - PlaceholderRange(offset=p["offset"], length=p["length"] - 1) - for p in ps - ] - for modality, ps in result["mm_placeholders"].items() - } - - return result - @MULTIMODAL_REGISTRY.register_processor(FuyuMultiModalProcessor, info=FuyuProcessingInfo, diff --git a/vllm/model_executor/models/gemma.py b/vllm/model_executor/models/gemma.py index b28715c48adfb..6de0c866bc2f0 100644 --- a/vllm/model_executor/models/gemma.py +++ b/vllm/model_executor/models/gemma.py @@ -349,15 +349,6 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA, SupportsPP): "gate_up_proj", "down_proj", ] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } # Gemma does not apply LoRA to the embedding layer. embedding_modules = {} diff --git a/vllm/model_executor/models/gemma2.py b/vllm/model_executor/models/gemma2.py index f4530e4771960..f0dc7693974be 100644 --- a/vllm/model_executor/models/gemma2.py +++ b/vllm/model_executor/models/gemma2.py @@ -31,8 +31,6 @@ RowParallelLinear) from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig -from vllm.model_executor.layers.quantization.compressed_tensors.utils import ( - get_compressed_tensors_cache_scale) from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -254,6 +252,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): cache_config = vllm_config.cache_config quant_config = vllm_config.quant_config self.config = config + self.quant_config = quant_config self.embed_tokens = VocabParallelEmbedding( config.vocab_size, @@ -329,7 +328,8 @@ def load_weights(self, weights: Iterable[Tuple[str, params_dict = dict(self.named_parameters()) loaded_params: Set[str] = set() for name, loaded_weight in weights: - if scale_name := get_compressed_tensors_cache_scale(name): + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): # Loading kv cache scales for compressed-tensors quantization param = params_dict[scale_name] weight_loader = getattr(param, "weight_loader", @@ -399,16 +399,6 @@ class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): embedding_modules = {} embedding_padding_modules = [] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } - def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): config = vllm_config.model_config.hf_config quant_config = vllm_config.quant_config diff --git a/vllm/model_executor/models/glm4_vision_encoder.py b/vllm/model_executor/models/glm4_vision_encoder.py index 5db9786ea05b7..7de7c622f4fc7 100644 --- a/vllm/model_executor/models/glm4_vision_encoder.py +++ b/vllm/model_executor/models/glm4_vision_encoder.py @@ -43,7 +43,8 @@ def forward(self, images: torch.Tensor) -> torch.Tensor: torch.Tensor Transformed tensor with shape (B, L, D) """ - images = images.to(self.proj.weight.device) + images = images.to(device=self.proj.weight.device, + dtype=self.proj.weight.dtype) x = self.proj(images) x = x.flatten(2).transpose(1, 2) cls_token = self.cls_embedding.expand(x.shape[0], -1, -1) diff --git a/vllm/model_executor/models/gpt2.py b/vllm/model_executor/models/gpt2.py index 1656a3cc9e46d..2f1aa2d68653c 100644 --- a/vllm/model_executor/models/gpt2.py +++ b/vllm/model_executor/models/gpt2.py @@ -258,13 +258,13 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): self.transformer = GPT2Model(vllm_config=vllm_config, prefix=maybe_prefix( prefix, "transformer")) + self.lm_head = ParallelLMHead(self.config.vocab_size, + self.config.hidden_size, + quant_config=quant_config, + prefix=f"{prefix}.lm_head") if self.config.tie_word_embeddings: - self.lm_head = self.transformer.wte - else: - self.lm_head = ParallelLMHead(self.config.vocab_size, - self.config.hidden_size, - quant_config=quant_config, - prefix=f"{prefix}.lm_head") + self.lm_head = self.lm_head.tie_weights(self.transformer.wte) + self.logits_processor = LogitsProcessor(config.vocab_size) self.sampler = get_sampler() self.make_empty_intermediate_tensors = ( @@ -309,15 +309,12 @@ def load_weights(self, weights: Iterable[Tuple[str, params_dict = dict(self.named_parameters(remove_duplicate=False)) loaded_params: Set[str] = set() for name, loaded_weight in weights: - if name.startswith("lm_head"): - # GPT-2 ties the weights of the embedding layer and the final - # linear layer. - continue if ".attn.bias" in name or ".attn.masked_bias" in name: # Skip attention mask. # NOTE: "c_attn.bias" should not be skipped. continue - if not name.startswith("transformer."): + if not name.startswith("transformer.") and not name.startswith( + "lm_head"): name = "transformer." + name if is_pp_missing_parameter(name, self): diff --git a/vllm/model_executor/models/gpt_j.py b/vllm/model_executor/models/gpt_j.py index 4829578a56959..08298cc0db36f 100644 --- a/vllm/model_executor/models/gpt_j.py +++ b/vllm/model_executor/models/gpt_j.py @@ -313,6 +313,19 @@ def load_weights(self, weights: Iterable[Tuple[str, for name, loaded_weight in weights: if "attn.bias" in name or "attn.masked_bias" in name: continue + + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales + param = params_dict[scale_name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) + weight_loader(param, loaded_weight) + loaded_params.add(scale_name) + continue + for (param_name, weight_name, shard_id) in stacked_params_mapping: if weight_name not in name: continue diff --git a/vllm/model_executor/models/granite.py b/vllm/model_executor/models/granite.py index 3c5dfc2e7f45d..543b4e2f5e286 100644 --- a/vllm/model_executor/models/granite.py +++ b/vllm/model_executor/models/granite.py @@ -38,8 +38,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization.base_config import ( QuantizationConfig) -from vllm.model_executor.layers.quantization.compressed_tensors.utils import ( - get_compressed_tensors_cache_scale) from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -360,14 +358,6 @@ class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP): "lm_head": "output_embeddings", } embedding_padding_modules = ["lm_head"] - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() @@ -377,6 +367,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): self.config = config self.lora_config = lora_config + self.quant_config = quant_config self.model = GraniteModel(vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")) @@ -480,14 +471,14 @@ def load_weights(self, weights: Iterable[Tuple[str, # processed with quantization, LoRA, fine-tuning, etc. if self.config.tie_word_embeddings and "lm_head.weight" in name: continue - if scale_name := get_compressed_tensors_cache_scale(name): - # Loading kv cache scales for compressed-tensors quantization + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales param = params_dict[scale_name] weight_loader = getattr(param, "weight_loader", default_weight_loader) - if loaded_weight.shape: - # scalar shape is torch.Size([1]), not torch.Size([]) - loaded_weight = loaded_weight[0] + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) weight_loader(param, loaded_weight) loaded_params.add(scale_name) continue diff --git a/vllm/model_executor/models/granitemoe.py b/vllm/model_executor/models/granitemoe.py index 51296ef0cc08e..b518a0a6cbdee 100644 --- a/vllm/model_executor/models/granitemoe.py +++ b/vllm/model_executor/models/granitemoe.py @@ -348,6 +348,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): self.config = config self.lora_config = lora_config + self.quant_config = quant_config # Required by MixtralForCausalLM self.model = GraniteMoeModel(vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")) diff --git a/vllm/model_executor/models/grok1.py b/vllm/model_executor/models/grok1.py index 99940c547f3d5..fb082fbf29b16 100644 --- a/vllm/model_executor/models/grok1.py +++ b/vllm/model_executor/models/grok1.py @@ -37,6 +37,7 @@ from vllm.model_executor.layers.quantization.base_config import ( QuantizationConfig) from vllm.model_executor.layers.quantization.fp8 import Fp8Config +from vllm.model_executor.layers.quantization.quark.quark import QuarkConfig from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.sampler import Sampler, SamplerOutput from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -197,7 +198,9 @@ def __init__( ) -> None: super().__init__() self.hidden_size = config.hidden_size - self.use_fp8 = isinstance(quant_config, Fp8Config) + self.use_fp8 = isinstance( + quant_config, Fp8Config) or (isinstance(quant_config, QuarkConfig) + and quant_config.is_fp8_w8a8()) # Requires transformers > 4.32.0 rope_theta = getattr(config, "rope_theta", 10000) self.attn = Grok1Attention(hidden_size=self.hidden_size, diff --git a/vllm/model_executor/models/idefics3.py b/vllm/model_executor/models/idefics3.py index 17e772e7faa32..d16a77f862d98 100644 --- a/vllm/model_executor/models/idefics3.py +++ b/vllm/model_executor/models/idefics3.py @@ -662,16 +662,6 @@ class Idefics3ForConditionalGeneration(nn.Module, SupportsMultiModal, "down_proj", ] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } - embedding_modules = {} embedding_padding_modules = [] diff --git a/vllm/model_executor/models/interfaces_base.py b/vllm/model_executor/models/interfaces_base.py index 4c353ae6ffc13..37b91a803d71e 100644 --- a/vllm/model_executor/models/interfaces_base.py +++ b/vllm/model_executor/models/interfaces_base.py @@ -3,7 +3,6 @@ import torch import torch.nn as nn -from transformers import PretrainedConfig from typing_extensions import TypeIs, TypeVar from vllm.logger import init_logger @@ -19,9 +18,6 @@ logger = init_logger(__name__) -# The type of HF config -C_co = TypeVar("C_co", bound=PretrainedConfig, covariant=True) - # The type of hidden states # Currently, T = torch.Tensor for all models except for Medusa # which has T = List[torch.Tensor] @@ -34,7 +30,7 @@ @runtime_checkable -class VllmModel(Protocol[C_co, T_co]): +class VllmModel(Protocol[T_co]): """The interface required for all models in vLLM.""" def __init__( @@ -97,7 +93,7 @@ def is_vllm_model( @runtime_checkable -class VllmModelForTextGeneration(VllmModel[C_co, T], Protocol[C_co, T]): +class VllmModelForTextGeneration(VllmModel[T], Protocol[T]): """The interface required for all generative models in vLLM.""" def compute_logits( @@ -143,7 +139,7 @@ def is_text_generation_model( @runtime_checkable -class VllmModelForPooling(VllmModel[C_co, T], Protocol[C_co, T]): +class VllmModelForPooling(VllmModel[T], Protocol[T]): """The interface required for all pooling models in vLLM.""" def pooler( diff --git a/vllm/model_executor/models/llama.py b/vllm/model_executor/models/llama.py index d991fabc92b08..a82fb4398ebbc 100644 --- a/vllm/model_executor/models/llama.py +++ b/vllm/model_executor/models/llama.py @@ -39,9 +39,8 @@ RowParallelLinear) from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig -from vllm.model_executor.layers.quantization.compressed_tensors.utils import ( - get_compressed_tensors_cache_scale) from vllm.model_executor.layers.quantization.fp8 import Fp8Config +from vllm.model_executor.layers.quantization.quark.quark import QuarkConfig from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -86,7 +85,9 @@ def __init__( quant_config=quant_config, prefix=f"{prefix}.down_proj", ) - self.use_fp8 = (isinstance(quant_config, Fp8Config) + self.use_fp8 = (isinstance(quant_config, Fp8Config) or + (isinstance(quant_config, QuarkConfig) + and quant_config.is_fp8_w8a8()) if current_platform.is_rocm() and not is_navi() else False) if hidden_act != "silu": @@ -113,20 +114,19 @@ def forward(self, x): class LlamaAttention(nn.Module): - def __init__( - self, - config: LlamaConfig, - hidden_size: int, - num_heads: int, - num_kv_heads: int, - rope_theta: float = 10000, - rope_scaling: Optional[Dict[str, Any]] = None, - max_position_embeddings: int = 8192, - quant_config: Optional[QuantizationConfig] = None, - bias: bool = False, - cache_config: Optional[CacheConfig] = None, - prefix: str = "", - ) -> None: + def __init__(self, + config: LlamaConfig, + hidden_size: int, + num_heads: int, + num_kv_heads: int, + rope_theta: float = 10000, + rope_scaling: Optional[Dict[str, Any]] = None, + max_position_embeddings: int = 8192, + quant_config: Optional[QuantizationConfig] = None, + bias: bool = False, + bias_o_proj: bool = False, + cache_config: Optional[CacheConfig] = None, + prefix: str = "") -> None: super().__init__() layer_idx = extract_layer_index(prefix) self.hidden_size = hidden_size @@ -166,7 +166,7 @@ def __init__( self.o_proj = RowParallelLinear( input_size=self.total_num_heads * self.head_dim, output_size=hidden_size, - bias=bias, + bias=bias_o_proj, quant_config=quant_config, prefix=f"{prefix}.o_proj", ) @@ -199,10 +199,13 @@ def __init__( sliding_window = None # For CUDA devices and Navi4x, attn_fp8 will be set to false. + use_fp8 = isinstance( + quant_config, Fp8Config) or (isinstance(quant_config, QuarkConfig) + and quant_config.is_fp8_w8a8()) self.attn_fp8_out = envs.VLLM_USE_ROCM_CUSTOM_PAGED_ATTN_FP8_OUT \ and current_platform.is_rocm() \ and not is_navi() \ - and isinstance(quant_config, Fp8Config) + and use_fp8 self.attn = Attention( self.num_heads, @@ -243,7 +246,9 @@ def __init__( ) -> None: super().__init__() self.hidden_size = config.hidden_size - self.use_fp8 = (isinstance(quant_config, Fp8Config) + self.use_fp8 = (isinstance(quant_config, Fp8Config) or + (isinstance(quant_config, QuarkConfig) + and quant_config.is_fp8_w8a8()) if current_platform.is_rocm() and not is_navi() else False) rope_theta = getattr(config, "rope_theta", 10000) @@ -258,6 +263,11 @@ def __init__( # Support internlm/internlm-7b with bias attention_bias = getattr(config, "attention_bias", False) or getattr( config, "bias", False) + bias_o_proj = attention_bias + # support internlm/internlm3-8b with qkv_bias + if hasattr(config, 'qkv_bias'): + attention_bias = config.qkv_bias + self.self_attn = LlamaAttention( config=config, hidden_size=self.hidden_size, @@ -269,6 +279,7 @@ def __init__( max_position_embeddings=max_position_embeddings, quant_config=quant_config, bias=attention_bias, + bias_o_proj=bias_o_proj, cache_config=cache_config, prefix=f"{prefix}.self_attn", ) @@ -331,6 +342,7 @@ def __init__(self, lora_config = vllm_config.lora_config self.config = config + self.quant_config = quant_config self.padding_idx = config.pad_token_id lora_vocab = (lora_config.lora_extra_vocab_size * (lora_config.max_loras or 1)) if lora_config else 0 @@ -421,14 +433,14 @@ def load_weights(self, weights: Iterable[Tuple[str, # Models trained using ColossalAI may include these tensors in # the checkpoint. Skip them. continue - if scale_name := get_compressed_tensors_cache_scale(name): - # Loading kv cache scales for compressed-tensors quantization + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales param = params_dict[scale_name] weight_loader = getattr(param, "weight_loader", default_weight_loader) - if loaded_weight.shape: - # scalar shape is torch.Size([1]), not torch.Size([]) - loaded_weight = loaded_weight[0] + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) weight_loader(param, loaded_weight) loaded_params.add(scale_name) continue @@ -484,16 +496,6 @@ class LlamaForCausalLM(nn.Module, SupportsLoRA, SupportsPP): } embedding_padding_modules = ["lm_head"] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } - # Mistral/Llama models can also be loaded with --load-format mistral # from consolidated.safetensors checkpoints mistral_mapping = { diff --git a/vllm/model_executor/models/llava.py b/vllm/model_executor/models/llava.py index bb3db60c7d8ed..296af2aac5660 100644 --- a/vllm/model_executor/models/llava.py +++ b/vllm/model_executor/models/llava.py @@ -5,9 +5,11 @@ import torch import torch.nn as nn +from packaging.version import Version from transformers import (BatchFeature, CLIPVisionConfig, LlavaConfig, PixtralVisionConfig, PretrainedConfig, SiglipVisionConfig) +from transformers import __version__ as TRANSFORMERS_VERSION from transformers.models.llava import LlavaProcessor from transformers.models.pixtral import PixtralProcessor @@ -22,7 +24,7 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, - MultiModalInputsV2, MultiModalKwargs, + MultiModalInputs, MultiModalKwargs, NestedTensors) from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems, ImageSize, MultiModalDataItems) @@ -313,13 +315,14 @@ def _get_prompt_replacements( hf_processor_mm_kwargs: Mapping[str, object], out_mm_kwargs: MultiModalKwargs, ) -> list[PromptReplacement]: + processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) hf_config = self.info.get_hf_config() - image_token_id = hf_config.image_token_index + tokenizer = self.info.get_tokenizer() + vocab = tokenizer.get_vocab() - processor = self.info.get_hf_processor() - image_token = processor.image_token - image_break_token = processor.image_break_token - image_end_token = processor.image_end_token + image_break_id = vocab[processor.image_break_token] + image_token_id = hf_config.image_token_index + image_end_id = vocab[processor.image_end_token] vision_config = hf_config.vision_config assert isinstance(vision_config, PixtralVisionConfig) @@ -334,10 +337,10 @@ def get_replacement(item_idx: int): image_height=image_size.height, ) - tokens = ([image_token] * ncols + [image_break_token]) * nrows - tokens[-1] = image_end_token + tokens = ([image_token_id] * ncols + [image_break_id]) * nrows + tokens[-1] = image_end_id - return "".join(tokens) + return tokens return [ PromptReplacement( @@ -463,14 +466,10 @@ def init_vision_tower_for_llava( info=_build_llava_or_pixtral_hf_info, dummy_inputs=LlavaDummyInputsBuilder) class LlavaForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP): - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), + + packed_modules_mapping = { + "qkv_proj": ["q_proj", "k_proj", "v_proj"], + "gate_up_proj": ["gate_proj", "up_proj"] } def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: @@ -720,6 +719,27 @@ def load_weights(self, weights: Iterable[Tuple[str, return loader.load_weights(weights) +class MantisProcessingInfo(LlavaProcessingInfo): + + def get_hf_processor(self): + hf_config = self.get_hf_config() + vision_info = self.get_vision_encoder_info() + + if Version(TRANSFORMERS_VERSION) < Version("4.48"): + # BUG: num_additional_image_tokens = 0 but treated as 1, + # so we set vision_feature_select_strategy to None to offset this + vision_feature_select_strategy = None + else: + # FIXED: https://github.com/huggingface/transformers/pull/33424/files#diff-6a37acc21efcadaae622b079b2712a131131448ff64262bd219aa346aeec38faL150 + vision_feature_select_strategy = hf_config.vision_feature_select_strategy # noqa: E501 + + return self.ctx.get_hf_processor( + LlavaProcessor, + patch_size=vision_info.get_patch_size(), + vision_feature_select_strategy=vision_feature_select_strategy, + ) + + class MantisMultiModalProcessor(LlavaMultiModalProcessor): def apply( @@ -727,7 +747,7 @@ def apply( prompt: Union[str, list[int]], mm_data: MultiModalDataDict, hf_processor_mm_kwargs: Mapping[str, object], - ) -> MultiModalInputsV2: + ) -> MultiModalInputs: hf_config = self.info.get_hf_config() image_token_id = hf_config.image_token_index @@ -786,7 +806,7 @@ def get_replacement_mantis(item_idx: int): for modality, placeholders in mm_placeholders.items() } - return MultiModalInputsV2( + return MultiModalInputs( type="multimodal", prompt=prompt, prompt_token_ids=prompt_ids, @@ -798,7 +818,7 @@ def get_replacement_mantis(item_idx: int): # To use this model, please use # `--hf_overrides '{"architectures": ["MantisForConditionalGeneration"]}'` @MULTIMODAL_REGISTRY.register_processor(MantisMultiModalProcessor, - info=LlavaProcessingInfo, + info=MantisProcessingInfo, dummy_inputs=LlavaDummyInputsBuilder) class MantisForConditionalGeneration(LlavaForConditionalGeneration): pass diff --git a/vllm/model_executor/models/llava_onevision.py b/vllm/model_executor/models/llava_onevision.py index 78a47e64d9afc..5b0f35b08646b 100644 --- a/vllm/model_executor/models/llava_onevision.py +++ b/vllm/model_executor/models/llava_onevision.py @@ -19,8 +19,8 @@ from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs, NestedTensors) -from vllm.multimodal.parse import (MultiModalDataItems, VideoEmbeddingItems, - VideoProcessorItems) +from vllm.multimodal.parse import (ImageSize, MultiModalDataItems, + VideoEmbeddingItems, VideoProcessorItems) from vllm.multimodal.processing import PromptReplacement from vllm.multimodal.profiling import ProcessorInputs from vllm.sequence import IntermediateTensors @@ -145,6 +145,10 @@ def _get_num_unpadded_features( return (unpadded_features, newline_features) + def get_image_size_with_most_features(self) -> ImageSize: + # NOTE: This hardcoded value is found via processor tests + return ImageSize(width=1153, height=944) + def _get_num_frame_tokens( self, *, @@ -550,10 +554,12 @@ def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict: # Preserve the order of modalities if there are multiple of them # from the order of kwargs. for input_key in kwargs: - if input_key == "pixel_values" and "images" not in modalities: + if input_key in ("pixel_values", + "image_embeds") and "images" not in modalities: modalities["images"] = self._parse_and_validate_image_input( **kwargs) - if input_key == "pixel_values_videos" and "videos" not in modalities: # noqa E501 + if input_key in ("pixel_values_videos", + "video_embeds") and "videos" not in modalities: modalities["videos"] = self._parse_and_validate_video_input( **kwargs) @@ -810,7 +816,7 @@ def apply_pooling(self, image_features, stride=2): return image_feature def get_multimodal_embeddings( - self, **kwargs) -> Optional[List[Tuple[NestedTensors, str]]]: + self, **kwargs) -> Optional[tuple[torch.Tensor, ...]]: modalities = self._parse_and_validate_multimodal_inputs(**kwargs) if not modalities: return None @@ -836,8 +842,7 @@ def get_multimodal_embeddings( def get_input_embeddings( self, input_ids: torch.Tensor, - multimodal_embeddings: Optional[List[Tuple[NestedTensors, - str]]] = None, + multimodal_embeddings: Optional[tuple[torch.Tensor, ...]] = None, ) -> torch.Tensor: inputs_embeds = self.language_model.get_input_embeddings(input_ids) if multimodal_embeddings is not None: @@ -846,6 +851,34 @@ def get_input_embeddings( [self.config.image_token_index, self.config.video_token_index]) return inputs_embeds + def get_input_embeddings_v0( + self, + input_ids: torch.Tensor, + image_input: Optional[NestedTensors] = None, + video_input: Optional[NestedTensors] = None, + ) -> torch.Tensor: + + inputs_embeds = self.get_input_embeddings(input_ids) + if image_input is not None: + image_embeds = self._process_image_input(image_input) + inputs_embeds = merge_multimodal_embeddings( + input_ids, + inputs_embeds, + image_embeds, + placeholder_token_id=self.config.image_token_index, + ) + + if video_input is not None: + video_embeds = self._process_video_pixels(video_input) + inputs_embeds = merge_multimodal_embeddings( + input_ids, + inputs_embeds, + video_embeds, + placeholder_token_id=self.config.video_token_index, + ) + + return inputs_embeds + def forward( self, input_ids: torch.Tensor, @@ -865,13 +898,21 @@ def forward( if intermediate_tensors is not None: inputs_embeds = None - # NOTE: In v1, inputs_embeds is always generated at model runner, this - # condition is for v0 compatibility. + # NOTE: In v1, inputs_embeds is always generated at model runner from + # `get_multimodal_embeddings` and `get_input_embeddings`, this + # condition is only for v0 compatibility. elif inputs_embeds is None: - multimodal_embeddings = self.get_multimodal_embeddings(**kwargs) - inputs_embeds = self.get_input_embeddings(input_ids, - multimodal_embeddings) - input_ids = None + image_input = self._parse_and_validate_image_input(**kwargs) + video_input = self._parse_and_validate_video_input(**kwargs) + + if image_input is None and video_input is None: + inputs_embeds = None + else: + inputs_embeds = self.get_input_embeddings_v0( + input_ids, + image_input=image_input, + video_input=video_input) + input_ids = None hidden_states = self.language_model.model(input_ids, positions, diff --git a/vllm/model_executor/models/minicpm3.py b/vllm/model_executor/models/minicpm3.py index e9d7eada1d16c..5e1e6c6fa6141 100644 --- a/vllm/model_executor/models/minicpm3.py +++ b/vllm/model_executor/models/minicpm3.py @@ -241,11 +241,5 @@ class MiniCPM3ForCausalLM(MiniCPMForCausalLM): # `embedding_modules` and `embedding_padding_modules` # are inherited from MiniCPMForCausalLM - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } - def _init_model(self, *, vllm_config: VllmConfig, prefix: str = ""): return MiniCPM3Model(vllm_config=vllm_config, prefix=prefix) diff --git a/vllm/model_executor/models/minicpmv.py b/vllm/model_executor/models/minicpmv.py index ff7dab89e4da8..1aa529056893b 100644 --- a/vllm/model_executor/models/minicpmv.py +++ b/vllm/model_executor/models/minicpmv.py @@ -761,16 +761,6 @@ class MiniCPMV2_5(MiniCPMVBaseModel, SupportsLoRA): "kv_proj", ] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } - embedding_modules = {} embedding_padding_modules = [] @@ -881,16 +871,6 @@ class MiniCPMV2_6(MiniCPMVBaseModel, SupportsLoRA): "kv_proj", ] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } - embedding_modules = {} embedding_padding_modules = [] diff --git a/vllm/model_executor/models/mixtral.py b/vllm/model_executor/models/mixtral.py index a5b364fe5ec85..da415cdae96ed 100644 --- a/vllm/model_executor/models/mixtral.py +++ b/vllm/model_executor/models/mixtral.py @@ -347,6 +347,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): lora_config = vllm_config.lora_config self.config = config self.lora_config = lora_config + self.quant_config = quant_config self.model = MixtralModel(vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")) @@ -428,6 +429,18 @@ def load_weights(self, weights: Iterable[Tuple[str, if "rotary_emb.inv_freq" in name: continue + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales + param = params_dict[scale_name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) + weight_loader(param, loaded_weight) + loaded_params.add(scale_name) + continue + for (param_name, weight_name, shard_id) in stacked_params_mapping: if weight_name not in name: continue diff --git a/vllm/model_executor/models/mllama.py b/vllm/model_executor/models/mllama.py index 7ab7e12284bb8..0f9acf50eb9b0 100644 --- a/vllm/model_executor/models/mllama.py +++ b/vllm/model_executor/models/mllama.py @@ -1126,20 +1126,16 @@ def forward( @INPUT_REGISTRY.register_dummy_encoder_data(dummy_encoder_data_for_mllama) @INPUT_REGISTRY.register_input_processor(input_processor_for_mllama) class MllamaForConditionalGeneration(nn.Module, SupportsMultiModal): - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), + packed_modules_mapping = { + "qkv_proj": ["q_proj", "k_proj", "v_proj"], + "gate_up_proj": ["gate_proj", "up_proj"] } def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_config quant_config = vllm_config.quant_config + self.quant_config = quant_config self.vocab_size = config.text_config.vocab_size self.hidden_size = config.text_config.hidden_size self.max_num_tiles = config.vision_config.max_num_tiles @@ -1453,6 +1449,17 @@ def load_weights(self, weights: Iterable[Tuple[str, name = name.replace('patch_embedding.weight', 'patch_embedding._linear.weight') loaded_weight = loaded_weight.view(loaded_weight.shape[0], -1) + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales + param = params_dict[scale_name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) + weight_loader(param, loaded_weight) + updated_params.add(scale_name) + continue for (param_name, weight_name, shard_id) in stacked_params_mapping: if weight_name not in name: continue diff --git a/vllm/model_executor/models/molmo.py b/vllm/model_executor/models/molmo.py index c45ee9b921c9e..5c7ae0deefcd8 100644 --- a/vllm/model_executor/models/molmo.py +++ b/vllm/model_executor/models/molmo.py @@ -23,7 +23,8 @@ from vllm.inputs import (INPUT_REGISTRY, DecoderOnlyInputs, DummyData, InputContext, token_inputs) from vllm.model_executor import SamplingMetadata -from vllm.model_executor.layers.activation import QuickGELU, SiluAndMul +from vllm.model_executor.layers.activation import (MulAndSilu, QuickGELU, + SiluAndMul) from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import (ColumnParallelLinear, MergedColumnParallelLinear, @@ -462,15 +463,6 @@ def forward( return output -class SwiGLU(nn.Module): - - def forward(self, x: torch.Tensor) -> torch.Tensor: - x, gate = x.chunk(2, dim=-1) - # Note that the order is reversed compared to - # SiluAndMul. - return x * F.silu(gate) - - class LanuageModelMLP(nn.Module): """Molmo's LLM mlp.""" @@ -489,7 +481,7 @@ def __init__(self, quant_config=quant_config, ) # Activation function. - self.act_fn = SwiGLU() + self.act_fn = MulAndSilu() # Feed-forward output projection. self.down_proj = RowParallelLinear( self.intermediate_size, @@ -1193,12 +1185,6 @@ class MolmoForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, embedding_modules = {} embedding_padding_modules = [] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - "gate_proj": ("merged_linear", 0), - "up_proj": ("merged_linear", 1), - } - def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_config diff --git a/vllm/model_executor/models/nemotron.py b/vllm/model_executor/models/nemotron.py index 34cb9981c167b..2340283b69665 100644 --- a/vllm/model_executor/models/nemotron.py +++ b/vllm/model_executor/models/nemotron.py @@ -395,12 +395,6 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP): "lm_head": "output_embeddings", } embedding_padding_modules = ["lm_head"] - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - } def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() @@ -411,6 +405,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): self.config = config self.lora_config = lora_config + self.quant_config = quant_config self.model = NemotronModel(vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")) @@ -495,6 +490,17 @@ def load_weights(self, weights: Iterable[Tuple[str, # Models trained using ColossalAI may include these tensors in # the checkpoint. Skip them. continue + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales + param = params_dict[scale_name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) + weight_loader(param, loaded_weight) + loaded_params.add(scale_name) + continue for (param_name, weight_name, shard_id) in stacked_params_mapping: if weight_name not in name: continue diff --git a/vllm/model_executor/models/opt.py b/vllm/model_executor/models/opt.py index 7edafcd20b5db..ea1185aa80dc6 100644 --- a/vllm/model_executor/models/opt.py +++ b/vllm/model_executor/models/opt.py @@ -329,13 +329,9 @@ def forward( class OPTForCausalLM(nn.Module, SupportsPP): - - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), + packed_modules_mapping = { + "qkv_proj": ["q_proj", "k_proj", "v_proj"], + "gate_up_proj": ["gate_proj", "up_proj"] } def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): diff --git a/vllm/model_executor/models/paligemma.py b/vllm/model_executor/models/paligemma.py index f9ad0c67adaba..5a28b1ffbb7b4 100644 --- a/vllm/model_executor/models/paligemma.py +++ b/vllm/model_executor/models/paligemma.py @@ -136,6 +136,17 @@ def forward(self, image_features: torch.Tensor) -> torch.Tensor: @INPUT_REGISTRY.register_input_processor(input_processor_for_paligemma) class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP): + packed_modules_mapping = { + "qkv_proj": [ + "q_proj", + "k_proj", + "v_proj", + ], + "gate_up_proj": [ + "gate_proj", + "up_proj", + ], + } def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() diff --git a/vllm/model_executor/models/phi.py b/vllm/model_executor/models/phi.py index f9e972688ddd1..59b7508a370f8 100644 --- a/vllm/model_executor/models/phi.py +++ b/vllm/model_executor/models/phi.py @@ -279,14 +279,6 @@ class PhiForCausalLM(nn.Module, SupportsLoRA, SupportsPP): "fc2", ] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - } - embedding_modules = {} embedding_padding_modules = [] diff --git a/vllm/model_executor/models/phi3.py b/vllm/model_executor/models/phi3.py index 937858ee3b8c2..34141511ea791 100644 --- a/vllm/model_executor/models/phi3.py +++ b/vllm/model_executor/models/phi3.py @@ -14,7 +14,3 @@ class Phi3ForCausalLM(LlamaForCausalLM): "gate_up_proj", ], } - - # BitandBytes specific attributes - # Initialize an empty dict when there is no stacked parameter mapping. - bitsandbytes_stacked_params_mapping = {} diff --git a/vllm/model_executor/models/phi3v.py b/vllm/model_executor/models/phi3v.py index 7a230e5beb367..0fcda81da2800 100644 --- a/vllm/model_executor/models/phi3v.py +++ b/vllm/model_executor/models/phi3v.py @@ -30,15 +30,19 @@ VocabParallelEmbedding) from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY -from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, - MultiModalInputsV2, MultiModalKwargs, - NestedTensors, PlaceholderRange) +from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs, + NestedTensors) from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems, ImageSize, MultiModalDataItems) +# yapf conflicts with isort for this block +# yapf: disable from vllm.multimodal.processing import (BaseMultiModalProcessor, BaseProcessingInfo, BoundPromptReplacement, - PlaceholderInfo, PromptReplacement) + PlaceholderFeaturesInfo, + PromptReplacement, + PromptReplacementDetails) +# yapf: enable from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs from vllm.sequence import IntermediateTensors from vllm.utils import is_list_of @@ -437,7 +441,12 @@ def get_replacement_phi3v(item_idx: int): processor=hf_processor, ) - return [_IMAGE_TOKEN_ID] * num_image_tokens + [bos_token_id] + image_tokens = [_IMAGE_TOKEN_ID] * num_image_tokens + + return PromptReplacementDetails( + full=image_tokens + [bos_token_id], + features=image_tokens, + ) num_images = mm_items.get_count("image", strict=False) @@ -454,7 +463,7 @@ def _apply_prompt_replacements( token_ids: list[int], mm_prompt_repls: Mapping[str, Sequence[BoundPromptReplacement]], mm_item_counts: Mapping[str, int], - ) -> tuple[list[int], str, Mapping[str, list[PlaceholderInfo]]]: + ) -> tuple[list[int], str, Mapping[str, list[PlaceholderFeaturesInfo]]]: token_ids, text, placeholders = super()._apply_prompt_replacements( token_ids=token_ids, mm_prompt_repls=mm_prompt_repls, @@ -467,11 +476,11 @@ def _apply_prompt_replacements( token_ids = [token_ids[0], *token_ids[2:]] placeholders = { modality: [ - PlaceholderInfo( + PlaceholderFeaturesInfo( modality=p.modality, item_idx=p.item_idx, start_idx=p.start_idx - 1, - replacement=p.replacement, + tokens=p.tokens, ) for p in ps ] for modality, ps in placeholders.items() @@ -479,26 +488,6 @@ def _apply_prompt_replacements( return token_ids, text, placeholders - def apply( - self, - prompt: Union[str, list[int]], - mm_data: MultiModalDataDict, - hf_processor_mm_kwargs: Mapping[str, object], - ) -> MultiModalInputsV2: - result = super().apply(prompt, mm_data, hf_processor_mm_kwargs) - - # Only <|image|> tokens should be considered as placeholders, - # so we ignore the trailing bos_token_id - result["mm_placeholders"] = { - modality: [ - PlaceholderRange(offset=p["offset"], length=p["length"] - 1) - for p in ps - ] - for modality, ps in result["mm_placeholders"].items() - } - - return result - @MULTIMODAL_REGISTRY.register_processor(Phi3VMultiModalProcessor, info=Phi3VProcessingInfo, diff --git a/vllm/model_executor/models/phimoe.py b/vllm/model_executor/models/phimoe.py index 1febd62f2f705..881c09ea9db99 100644 --- a/vllm/model_executor/models/phimoe.py +++ b/vllm/model_executor/models/phimoe.py @@ -546,6 +546,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): lora_config = vllm_config.lora_config self.config = config self.lora_config = lora_config + self.quant_config = vllm_config.quant_config self.model = PhiMoEModel(vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")) @@ -623,6 +624,18 @@ def load_weights(self, weights: Iterable[Tuple[str, if "rotary_emb.inv_freq" in name: continue + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales + param = params_dict[scale_name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) + weight_loader(param, loaded_weight) + loaded_params.add(scale_name) + continue + for param_name, weight_name, shard_id in stacked_params_mapping: if weight_name not in name: continue diff --git a/vllm/model_executor/models/qwen.py b/vllm/model_executor/models/qwen.py index baf955f6b515d..1345b381f0a99 100644 --- a/vllm/model_executor/models/qwen.py +++ b/vllm/model_executor/models/qwen.py @@ -1028,13 +1028,6 @@ class QWenLLM(QWenBaseModel): embedding_modules = {} embedding_padding_modules = [] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "w2": ("gate_up_proj", 0), - "w1": ("gate_up_proj", 1), - } - class QWenVL(QWenBaseModel, SupportsMultiModal): packed_modules_mapping = { diff --git a/vllm/model_executor/models/qwen2.py b/vllm/model_executor/models/qwen2.py index d20fb150f7e39..82de1c3574090 100644 --- a/vllm/model_executor/models/qwen2.py +++ b/vllm/model_executor/models/qwen2.py @@ -256,7 +256,15 @@ def forward( return hidden_states, residual -@support_torch_compile +@support_torch_compile( + dynamic_arg_dims={ + "input_ids": 0, + # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl, + # otherwise (seq_len, ). + "positions": -1, + "intermediate_tensors": 0, + "inputs_embeds": 0, + }) class Qwen2Model(nn.Module): def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): @@ -279,6 +287,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): )) self.config = config + self.quant_config = quant_config self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size @@ -364,6 +373,17 @@ def load_weights(self, weights: Iterable[Tuple[str, for name, loaded_weight in weights: if "rotary_emb.inv_freq" in name: continue + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales + param = params_dict[scale_name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) + weight_loader(param, loaded_weight) + loaded_params.add(scale_name) + continue for (param_name, weight_name, shard_id) in stacked_params_mapping: if weight_name not in name: continue @@ -418,16 +438,6 @@ class Qwen2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): embedding_modules = {} embedding_padding_modules = [] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } - def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_config diff --git a/vllm/model_executor/models/qwen2_audio.py b/vllm/model_executor/models/qwen2_audio.py index 0dff9595c6c08..fc5aed5c94abb 100644 --- a/vllm/model_executor/models/qwen2_audio.py +++ b/vllm/model_executor/models/qwen2_audio.py @@ -41,7 +41,8 @@ from vllm.multimodal.parse import (AudioProcessorItems, MultiModalDataItems, MultiModalDataParser) from vllm.multimodal.processing import (BaseMultiModalProcessor, - BaseProcessingInfo, PromptReplacement) + BaseProcessingInfo, PromptReplacement, + PromptReplacementDetails) from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs from vllm.sequence import IntermediateTensors @@ -153,29 +154,24 @@ def _call_hf_processor( mm_data: Mapping[str, object], mm_kwargs: Mapping[str, Any], ) -> BatchFeature: - mm_data = dict(mm_data) - audios = mm_data.pop("audios", []) - - if audios: - mm_data["audios"] = audios - - feature_extractor = self.info.get_feature_extractor(**mm_kwargs) - mm_kwargs = dict( - **mm_kwargs, - sampling_rate=feature_extractor.sampling_rate, - ) - else: - # NOTE: WhisperFeatureExtractor cannot handle empty list of audios - pass + # Text-only input not supported in composite processor + if not mm_data or not mm_data.get("audios", []): + prompt_ids = self.info.get_tokenizer().encode(prompt) + prompt_ids = self._apply_hf_processor_tokens_only(prompt_ids) + return BatchFeature(dict(input_ids=[prompt_ids]), tensor_type="pt") + + feature_extractor = self.info.get_feature_extractor(**mm_kwargs) + mm_kwargs = dict( + **mm_kwargs, + sampling_rate=feature_extractor.sampling_rate, + ) - processed_outputs = super()._call_hf_processor( + return super()._call_hf_processor( prompt=prompt, mm_data=mm_data, mm_kwargs=mm_kwargs, ) - return processed_outputs - def _get_mm_fields_config( self, hf_inputs: BatchFeature, @@ -192,8 +188,20 @@ def _get_prompt_replacements( hf_processor_mm_kwargs: Mapping[str, object], out_mm_kwargs: MultiModalKwargs, ) -> list[PromptReplacement]: - hf_config = self.info.get_hf_config() - placeholder = hf_config.audio_token_index + processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) + tokenizer = self.info.get_tokenizer() + vocab = tokenizer.get_vocab() + + # Use getattr with default to be compatible with transformers<4.48 + audio_token = getattr(processor, "audio_token", "<|AUDIO|>") + audio_bos_token = getattr(processor, "audio_bos_token", + "<|audio_bos|>") + audio_eos_token = getattr(processor, "audio_eos_token", + "<|audio_eos|>") + + audio_token_id = vocab[audio_token] + audio_bos_id = vocab[audio_bos_token] + audio_eos_id = vocab[audio_eos_token] feature_attention_mask = out_mm_kwargs.get("feature_attention_mask") if feature_attention_mask is None: @@ -206,20 +214,25 @@ def _get_prompt_replacements( audio_output_lengths = audio_output_lens.tolist() def get_replacement_qwen2_audio(item_idx: int): - num_placeholders = audio_output_lengths[item_idx] - if num_placeholders == 0: + num_features = audio_output_lengths[item_idx] + if num_features == 0: audios = mm_items.get_items("audio", AudioProcessorItems) audio = audios.get(item_idx) raise ValueError( f"The audio {audio} (len={len(audio)}) is too short " "to be represented inside the model") - return [placeholder] * num_placeholders + audio_tokens = [audio_token_id] * num_features + + return PromptReplacementDetails( + full=[audio_bos_id] + audio_tokens + [audio_eos_id], + features=audio_tokens, + ) return [ PromptReplacement( modality="audio", - target=[placeholder], + target=audio_token, replacement=get_replacement_qwen2_audio, ) ] diff --git a/vllm/model_executor/models/qwen2_rm.py b/vllm/model_executor/models/qwen2_rm.py index 988d682d36be3..593ce4857af0f 100644 --- a/vllm/model_executor/models/qwen2_rm.py +++ b/vllm/model_executor/models/qwen2_rm.py @@ -12,7 +12,7 @@ from vllm.config import VllmConfig from vllm.model_executor.layers.linear import (ColumnParallelLinear, RowParallelLinear) -from vllm.model_executor.layers.pooler import Pooler, PoolingType +from vllm.model_executor.layers.pooler import Pooler, PoolingType, SimplePooler from vllm.model_executor.pooling_metadata import PoolingMetadata from vllm.sequence import IntermediateTensors, PoolerOutput @@ -32,7 +32,7 @@ def forward(self, input): return self.activation(input) -class Qwen2ForRewardModel(nn.Module, SupportsLoRA, SupportsPP): +class Qwen2RewardBaseModel(nn.Module, SupportsLoRA, SupportsPP): packed_modules_mapping = { "qkv_proj": [ "q_proj", @@ -60,7 +60,6 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): config = vllm_config.model_config.hf_config quant_config = vllm_config.quant_config lora_config = vllm_config.lora_config - pooler_config = vllm_config.model_config.pooler_config self.config = config self.lora_config = lora_config @@ -74,14 +73,11 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): config.hidden_size, quant_config=quant_config), ReLU(), - RowParallelLinear(config.hidden_size, 1, + RowParallelLinear(config.hidden_size, + config.num_labels, quant_config=quant_config), ) - self._pooler = Pooler.from_config_with_defaults( - pooler_config, - pooling_type=PoolingType.ALL, - normalize=False, - softmax=False) + self._pooler: SimplePooler self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors) @@ -115,3 +111,31 @@ def load_weights(self, weights: Iterable[Tuple[str, loader = AutoWeightsLoader(self, ignore_unexpected_prefixes=["lm_head."]) return loader.load_weights(weights) + + +class Qwen2ForRewardModel(Qwen2RewardBaseModel): + + def __init__(self, *, vllm_config, prefix=""): + vllm_config.model_config.hf_config.num_labels = 1 + super().__init__(vllm_config=vllm_config, prefix=prefix) + pooler_config = vllm_config.model_config.pooler_config + self._pooler = Pooler.from_config_with_defaults( + pooler_config, + pooling_type=PoolingType.ALL, + normalize=False, + softmax=False) + + +class Qwen2ForProcessRewardModel(Qwen2RewardBaseModel): + + def __init__(self, *, vllm_config, prefix=""): + vllm_config.model_config.hf_config.num_labels = 2 + super().__init__(vllm_config=vllm_config, prefix=prefix) + pooler_config = vllm_config.model_config.pooler_config + self._pooler = Pooler.from_config_with_defaults( + pooler_config, + pooling_type=PoolingType.STEP, + normalize=False, + softmax=True, + step_tag_id=151651, + ) diff --git a/vllm/model_executor/models/qwen2_vl.py b/vllm/model_executor/models/qwen2_vl.py index 76a810e8f0c20..a2778ee73810e 100644 --- a/vllm/model_executor/models/qwen2_vl.py +++ b/vllm/model_executor/models/qwen2_vl.py @@ -55,7 +55,7 @@ from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (ImageItem, ModalityData, MultiModalFieldConfig, MultiModalKwargs, - NestedTensors, VideoItem) + VideoItem) from vllm.multimodal.parse import (ImageSize, ModalityDataItems, MultiModalDataItems, MultiModalDataParser) from vllm.multimodal.processing import (BaseMultiModalProcessor, @@ -67,11 +67,15 @@ from .interfaces import SupportsLoRA, SupportsMultiModal, SupportsPP from .utils import (AutoWeightsLoader, WeightsMapper, - init_vllm_registered_model, maybe_prefix) + init_vllm_registered_model, maybe_prefix, + merge_multimodal_embeddings) from .vision import get_vit_attn_backend logger = init_logger(__name__) +# For profile run +_MAX_FRAMES_PER_VIDEO = 16 + # === Vision Inputs === # @@ -135,7 +139,7 @@ class Qwen2VLVideoEmbeddingInputs(TypedDict): - List[`torch.Tensor`]: A list of tensors holding all videos' features. Each tensor holds an video's features. - `torch.Tensor`: A tensor holding all videos' features - (concatenation of all videos' feature tensors). + (concatenation of all videos' feature tensors). Tensor shape: `(num_image_features, hidden_size)` - `num_image_features` varies based on @@ -611,6 +615,7 @@ def forward( # adapter x = self.merger(x) + return x def load_weights(self, weights: Iterable[Tuple[str, @@ -874,8 +879,8 @@ def get_num_frames_with_most_features(self, seq_len: int) -> int: max_image_tokens = self.get_max_image_tokens() * max_images max_total_frames = self._get_max_video_frames(seq_len - max_image_tokens) - - num_frames = max(max_total_frames // max(max_videos, 1), 1) + num_frames = min(max(max_total_frames // max(max_videos, 1), 1), + _MAX_FRAMES_PER_VIDEO) # Temporary workaround for https://github.com/huggingface/transformers/issues/35412 if num_frames > 1 and num_frames % 2 == 1: @@ -948,26 +953,29 @@ def _get_prompt_replacements( hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) image_processor = self.info.get_image_processor( **hf_processor_mm_kwargs) + tokenizer = self.info.get_tokenizer() + vocab = tokenizer.get_vocab() # NOTE: Only Qwen2VLProcessor in transformers 4.47.0 has # image_token and video_token registered placeholder = { - "image": hf_processor.image_token, - "video": hf_processor.video_token, + "image": vocab[hf_processor.image_token], + "video": vocab[hf_processor.video_token], } + merge_length = image_processor.merge_size**2 def get_replacement_qwen2vl(item_idx: int, modality: str): grid_thw = out_mm_kwargs[f"{modality}_grid_thw"][item_idx] assert isinstance(grid_thw, torch.Tensor) - num_tokens = grid_thw.prod() // merge_length - return placeholder[modality] * num_tokens + num_tokens = int(grid_thw.prod()) // merge_length + return [placeholder[modality]] * num_tokens return [ PromptReplacement( modality=modality, - target=placeholder[modality], + target=[placeholder[modality]], replacement=partial(get_replacement_qwen2vl, modality=modality), ) for modality in ("image", "video") @@ -1038,16 +1046,6 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal, embedding_modules = {} embedding_padding_modules = [] - # BitandBytes specific attributes - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } - # To ensure correct weight loading and mapping. hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={ "lm_head.": "language_model.lm_head.", @@ -1057,11 +1055,8 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal, def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config: Qwen2VLConfig = vllm_config.model_config.hf_config - cache_config = vllm_config.cache_config quant_config = vllm_config.quant_config multimodal_config = vllm_config.model_config.multimodal_config - assert not cache_config.enable_prefix_caching, \ - "Qwen2-VL currently does not support prefix caching" self.config = config self.multimodal_config = multimodal_config @@ -1183,85 +1178,122 @@ def _parse_and_validate_video_input( video_embeds=video_embeds, video_grid_thw=video_grid_thw) - def _process_image_input(self, - image_input: Qwen2VLImageInputs) -> torch.Tensor: + def _process_image_input( + self, image_input: Qwen2VLImageInputs) -> tuple[torch.Tensor, ...]: + + grid_thw = image_input["image_grid_thw"] + assert grid_thw.ndim == 2 + if image_input["type"] == "image_embeds": - return image_input["image_embeds"].type(self.visual.dtype) + image_embeds = image_input["image_embeds"].type(self.visual.dtype) + else: + pixel_values = image_input["pixel_values"].type(self.visual.dtype) + image_embeds = self.visual(pixel_values, grid_thw=grid_thw) + + # Split concatenated embeddings for each image item. + merge_size = self.visual.spatial_merge_size + sizes = grid_thw.prod(-1) // merge_size // merge_size - pixel_values = image_input["pixel_values"].type(self.visual.dtype) - image_embeds = self.visual(pixel_values, - grid_thw=image_input["image_grid_thw"]) - return image_embeds + return image_embeds.split(sizes.tolist()) + + def _process_video_input( + self, video_input: Qwen2VLVideoInputs) -> tuple[torch.Tensor, ...]: + + grid_thw = video_input["video_grid_thw"] + assert grid_thw.ndim == 2 - def _process_video_input(self, - video_input: Qwen2VLVideoInputs) -> torch.Tensor: if video_input["type"] == "video_embeds": - return video_input["video_embeds"].type(self.visual.dtype) + video_embeds = video_input["video_embeds"].type(self.visual.dtype) + else: + pixel_values_videos = video_input["pixel_values_videos"].type( + self.visual.dtype) + video_embeds = self.visual(pixel_values_videos, grid_thw=grid_thw) - pixel_values_videos = video_input["pixel_values_videos"].type( - self.visual.dtype) - video_embeds = self.visual(pixel_values_videos, - grid_thw=video_input["video_grid_thw"]) - return video_embeds + # Split concatenated embeddings for each video item. + merge_size = self.visual.spatial_merge_size + sizes = grid_thw.prod(-1) // merge_size // merge_size - def _merge_multimodal_embeddings( - self, - input_ids: torch.Tensor, - inputs_embeds: torch.Tensor, - multimodal_embeddings: torch.Tensor, - placeholder_token_id: int, - ) -> torch.Tensor: - mask = (input_ids == placeholder_token_id) - inputs_embeds[mask, :] = multimodal_embeddings - return inputs_embeds + return video_embeds.split(sizes.tolist()) + + def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict: + modalities = {} + + # Preserve the order of modalities if there are multiple of them + # from the order of kwargs. + for input_key in kwargs: + if input_key in ("pixel_values", + "image_embeds") and "images" not in modalities: + modalities["images"] = self._parse_and_validate_image_input( + **kwargs) + if input_key in ("pixel_values_videos", + "video_embeds") and "videos" not in modalities: + modalities["videos"] = self._parse_and_validate_video_input( + **kwargs) + + return modalities def get_multimodal_embeddings( - self, **kwargs) -> Optional[List[Tuple[NestedTensors, str]]]: + self, **kwargs) -> Optional[tuple[torch.Tensor, ...]]: - image_input = self._parse_and_validate_image_input(**kwargs) - video_input = self._parse_and_validate_video_input(**kwargs) - if image_input is None and video_input is None: + modalities = self._parse_and_validate_multimodal_inputs(**kwargs) + if not modalities: return None - # We make a tuple of each embedding with its modality string. This is a - # temporary workaround for models to handle mixed modalities when - # get_multimodal_embeddings and get_input_embeddings are called - # separately. - # TODO(ywang96): Add support for mixed-modality inference for v1. - multimodal_embeddings: List[Tuple[NestedTensors, str]] = [] - - if image_input is not None: - image_embeds = self._process_image_input(image_input) - multimodal_embeddings.append((image_embeds, "image")) - if video_input is not None: - video_embeds = self._process_video_input(video_input) - multimodal_embeddings.append((video_embeds, "video")) + # The result multimodal_embeddings is tuple of tensors, with each + # tensor correspoending to a multimodal data item (image or video). + multimodal_embeddings: tuple[torch.Tensor, ...] = () + + # NOTE: It is important to iterate over the keys in this dictionary + # to preserve the order of the modalities. + for modality in modalities: + if modality == "images": + image_input = modalities["images"] + vision_embeddings = self._process_image_input(image_input) + multimodal_embeddings += vision_embeddings + if modality == "videos": + video_input = modalities["videos"] + video_embeddings = self._process_video_input(video_input) + multimodal_embeddings += video_embeddings return multimodal_embeddings def get_input_embeddings( self, input_ids: torch.Tensor, - multimodal_embeddings: Optional[List[Tuple[NestedTensors, - str]]] = None, + multimodal_embeddings: Optional[tuple[torch.Tensor, ...]] = None, ) -> torch.Tensor: inputs_embeds = self.language_model.get_input_embeddings(input_ids) if multimodal_embeddings is not None: - for embeddings, modality in multimodal_embeddings: - if modality == "image": - inputs_embeds = self._merge_multimodal_embeddings( - input_ids, - inputs_embeds, - embeddings, - placeholder_token_id=self.config.image_token_id, - ) - if modality == "video": - inputs_embeds = self._merge_multimodal_embeddings( - input_ids, - inputs_embeds, - embeddings, - placeholder_token_id=self.config.video_token_id, - ) + inputs_embeds = merge_multimodal_embeddings( + input_ids, inputs_embeds, multimodal_embeddings, + [self.config.image_token_id, self.config.video_token_id]) + return inputs_embeds + + def get_input_embeddings_v0( + self, + input_ids: torch.Tensor, + image_input: Optional[tuple[torch.Tensor, ...]] = None, + video_input: Optional[tuple[torch.Tensor, ...]] = None, + ) -> torch.Tensor: + + inputs_embeds = self.get_input_embeddings(input_ids) + if image_input is not None: + image_embeds = self._process_image_input(image_input) + inputs_embeds = merge_multimodal_embeddings( + input_ids, + inputs_embeds, + image_embeds, + placeholder_token_id=self.config.image_token_id, + ) + + if video_input is not None: + video_embeds = self._process_video_input(video_input) + inputs_embeds = merge_multimodal_embeddings( + input_ids, + inputs_embeds, + video_embeds, + placeholder_token_id=self.config.video_token_id, + ) return inputs_embeds def forward( @@ -1297,22 +1329,25 @@ def forward( if intermediate_tensors is not None: inputs_embeds = None - # NOTE: In v1, inputs_embeds is always generated at model runner, this - # condition is for v0 compatibility. + # NOTE: In v1, inputs_embeds is always generated at model runner from + # `get_multimodal_embeddings` and `get_input_embeddings`, this + # condition is only for v0 compatibility. elif inputs_embeds is None: - multimodal_embeddings = self.get_multimodal_embeddings(**kwargs) - - # We need to check for usage of mrope here in case there is - # multimodal data. - # TODO (ywang96): move this to model runner in V1. - if multimodal_embeddings is not None and uses_mrope(self.config): - assert positions.ndim == 2 and positions.size(0) == 3, ( - "multimodal section rotary embedding requires " - f"(3, seq_len) positions, but got {positions.size()}") - - inputs_embeds = self.get_input_embeddings(input_ids, - multimodal_embeddings) - input_ids = None + image_input = self._parse_and_validate_image_input(**kwargs) + video_input = self._parse_and_validate_video_input(**kwargs) + + if image_input is None and video_input is None: + inputs_embeds = None + else: + if uses_mrope(self.config): + assert positions.ndim == 2 and positions.size(0) == 3, ( + "multimodal section rotary embedding requires " + f"(3, seq_len) positions, but got {positions.size()}") + inputs_embeds = self.get_input_embeddings_v0( + input_ids, + image_input=image_input, + video_input=video_input) + input_ids = None hidden_states = self.language_model.model( input_ids=input_ids, diff --git a/vllm/model_executor/models/registry.py b/vllm/model_executor/models/registry.py index 7a193491e02f8..9983017e645f2 100644 --- a/vllm/model_executor/models/registry.py +++ b/vllm/model_executor/models/registry.py @@ -47,6 +47,7 @@ "DeepseekV3ForCausalLM": ("deepseek_v3", "DeepseekV3ForCausalLM"), "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"), "FalconForCausalLM": ("falcon", "FalconForCausalLM"), + "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"), "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"), "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"), "GlmForCausalLM": ("glm", "GlmForCausalLM"), @@ -60,6 +61,7 @@ "InternLMForCausalLM": ("llama", "LlamaForCausalLM"), "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"), "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"), + "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"), "JAISLMHeadModel": ("jais", "JAISLMHeadModel"), "JambaForCausalLM": ("jamba", "JambaForCausalLM"), "LlamaForCausalLM": ("llama", "LlamaForCausalLM"), @@ -126,6 +128,7 @@ "Qwen2Model": ("qwen2", "Qwen2EmbeddingModel"), "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"), "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"), + "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"), "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"), # [Multimodal] "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"), # noqa: E501 diff --git a/vllm/model_executor/models/siglip.py b/vllm/model_executor/models/siglip.py index cca42842bc06e..1e51018973e8c 100644 --- a/vllm/model_executor/models/siglip.py +++ b/vllm/model_executor/models/siglip.py @@ -344,10 +344,14 @@ def __init__( self.config = config self.activation_fn = get_act_fn(config.hidden_act) - - # For quantization, we require the hidden size to be a multiple of 64 - quantizable = (config.hidden_size % 64 == 0 - and config.intermediate_size % 64 == 0) + # Special handling for BNB quantization + if quant_config and quant_config.get_name() == "bitsandbytes": + quantizable = True + else: + # For other quantization, we require the hidden size to be a + # multiple of 64 + quantizable = (config.hidden_size % 64 == 0 + and config.intermediate_size % 64 == 0) self.fc1 = ColumnParallelLinear( config.hidden_size, config.intermediate_size, diff --git a/vllm/model_executor/models/solar.py b/vllm/model_executor/models/solar.py index 4f3cdbbcee9f4..e6d919f23c85d 100644 --- a/vllm/model_executor/models/solar.py +++ b/vllm/model_executor/models/solar.py @@ -38,8 +38,6 @@ RowParallelLinear) from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig -from vllm.model_executor.layers.quantization.compressed_tensors.utils import ( - get_compressed_tensors_cache_scale) from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -399,14 +397,6 @@ class SolarForCausalLM(nn.Module, SupportsLoRA, SupportsPP): "lm_head": "output_embeddings", } embedding_padding_modules = ["lm_head"] - bitsandbytes_stacked_params_mapping = { - # shard_name, weight_name, index - "q_proj": ("qkv_proj", 0), - "k_proj": ("qkv_proj", 1), - "v_proj": ("qkv_proj", 2), - "gate_proj": ("gate_up_proj", 0), - "up_proj": ("gate_up_proj", 1), - } def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() @@ -415,6 +405,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): lora_config = vllm_config.lora_config self.config = config self.lora_config = lora_config + self.quant_config = quant_config self.model = SolarModel( vllm_config=vllm_config, @@ -497,14 +488,14 @@ def load_weights(self, weights: Iterable[Tuple[str, # Models trained using ColossalAI may include these tensors in # the checkpoint. Skip them. continue - if scale_name := get_compressed_tensors_cache_scale(name): - # Loading kv cache scales for compressed-tensors quantization + if (self.quant_config is not None and + (scale_name := self.quant_config.get_cache_scale(name))): + # Loading kv cache quantization scales param = params_dict[scale_name] weight_loader = getattr(param, "weight_loader", default_weight_loader) - if loaded_weight.shape: - # scalar shape is torch.Size([1]), not torch.Size([]) - loaded_weight = loaded_weight[0] + loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else + loaded_weight[0]) weight_loader(param, loaded_weight) loaded_params.add(scale_name) continue diff --git a/vllm/model_executor/models/ultravox.py b/vllm/model_executor/models/ultravox.py index 3edfb5107683a..d577e545a473b 100644 --- a/vllm/model_executor/models/ultravox.py +++ b/vllm/model_executor/models/ultravox.py @@ -16,7 +16,7 @@ from vllm import envs from vllm.attention import AttentionMetadata from vllm.config import VllmConfig -from vllm.model_executor.layers.activation import SiluAndMul, get_act_fn +from vllm.model_executor.layers.activation import MulAndSilu, get_act_fn from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler from vllm.model_executor.model_loader.loader import DefaultModelLoader @@ -137,7 +137,7 @@ def _call_hf_processor( mm_kwargs: Mapping[str, object], ) -> BatchFeature: # Text-only input not supported in composite processor - if not mm_data: + if not mm_data or not mm_data.get("audios", []): prompt_ids = self.info.get_tokenizer().encode(prompt) prompt_ids = self._apply_hf_processor_tokens_only(prompt_ids) return BatchFeature(dict(input_ids=[prompt_ids]), tensor_type="pt") @@ -146,13 +146,6 @@ def _call_hf_processor( audios = mm_data.pop("audios", []) assert isinstance(audios, list) - if not audios: - return super()._call_hf_processor( - prompt=prompt, - mm_data=mm_data, - mm_kwargs=mm_kwargs, - ) - feature_extractor = self.info.get_feature_extractor() mm_kwargs = dict( **mm_kwargs, @@ -212,11 +205,15 @@ def _get_prompt_replacements( out_mm_kwargs: MultiModalKwargs, ) -> list[PromptReplacement]: hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) - placeholder = hf_processor.audio_token_replacement # type: ignore + tokenizer = self.info.get_tokenizer() + vocab = tokenizer.get_vocab() + + replacement_id = vocab[ + hf_processor.audio_token_replacement] # type: ignore def get_replacement_ultravox(item_idx: int): audio_token_len = out_mm_kwargs["audio_token_len"][item_idx] - return placeholder * audio_token_len + return [replacement_id] * int(audio_token_len) # type: ignore return [ PromptReplacement( @@ -248,15 +245,6 @@ def forward(self, audio_embeds: torch.Tensor) -> torch.Tensor: return audio_embeds -class FlippedSiluAndMul(SiluAndMul): - """Ultravox is trained with SwiGLU with flipped halves.""" - - def forward(self, x: torch.Tensor): - a, b = x.chunk(2, dim=-1) - flipped = torch.cat((b, a), dim=-1) - return super().forward(flipped) - - class UltravoxProjector(nn.Module): def __init__(self, config: UltravoxConfig): @@ -269,7 +257,7 @@ def __init__(self, config: UltravoxConfig): dim = self.hidden_dim if config.projector_act == "swiglu": - self.act = FlippedSiluAndMul() + self.act = MulAndSilu() dim = dim // 2 else: self.act = get_act_fn(config.projector_act) diff --git a/vllm/model_executor/models/vision.py b/vllm/model_executor/models/vision.py index a1395982af44c..57166f05cd9bf 100644 --- a/vllm/model_executor/models/vision.py +++ b/vllm/model_executor/models/vision.py @@ -82,23 +82,25 @@ def get_vit_attn_backend(support_fa: bool = False) -> _Backend: if backend_by_env_var is not None: selected_backend = backend_name_to_enum(backend_by_env_var) if selected_backend is None: - # For Volta and Turing GPUs, use xformers instead. - device_available = current_platform.has_device_capability(80) - if device_available and support_fa: - from transformers.utils import is_flash_attn_2_available - if is_flash_attn_2_available(): - selected_backend = _Backend.FLASH_ATTN + if current_platform.is_cuda(): + device_available = current_platform.has_device_capability(80) + if device_available and support_fa: + from transformers.utils import is_flash_attn_2_available + if is_flash_attn_2_available(): + selected_backend = _Backend.FLASH_ATTN + else: + logger.warning_once( + "Current `vllm-flash-attn` has a bug inside vision " + "module, so we use xformers backend instead. You can " + "run `pip install flash-attn` to use flash-attention " + "backend.") + selected_backend = _Backend.XFORMERS else: - logger.warning_once( - "Current `vllm-flash-attn` has a bug inside vision module, " - "so we use xformers backend instead. You can run " - "`pip install flash-attn` to use flash-attention backend.") + # For Volta and Turing GPUs, use xformers instead. selected_backend = _Backend.XFORMERS - elif current_platform.is_cpu() or current_platform.is_rocm(): - # ROCM doesn't support xformers - selected_backend = _Backend.TORCH_SDPA else: - selected_backend = _Backend.XFORMERS + # Default to torch SDPA for other non-GPU platforms. + selected_backend = _Backend.TORCH_SDPA return selected_backend diff --git a/vllm/model_executor/models/whisper.py b/vllm/model_executor/models/whisper.py index c1f3bb0ca33c2..b8512b735da94 100644 --- a/vllm/model_executor/models/whisper.py +++ b/vllm/model_executor/models/whisper.py @@ -729,7 +729,22 @@ def sample( def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]) -> Set[str]: loader = AutoWeightsLoader(self, skip_prefixes=["proj_out."]) - loaded_weights = [(name, loaded_weight) - for name, loaded_weight in weights] mapper = WeightsMapper({".fc1.": ".mlp.fc1.", ".fc2.": ".mlp.fc2."}) - return loader.load_weights(loaded_weights, mapper=mapper) + # add fake zeros bias for k_proj to state_dict + weights = _create_fake_bias_for_k_proj(weights) + return loader.load_weights(weights, mapper=mapper) + + +def _create_fake_bias_for_k_proj( + weights: Iterable[Tuple[str, torch.Tensor]] +) -> Iterable[Tuple[str, torch.Tensor]]: + """ + Create full zeros bias for k_proj weight in self-attention layers. + So that the bias for k_proj in qkv_proj can be initialized with zeros. + """ + for name, weight in weights: + if ".self_attn.k_proj.weight" in name: + bias = torch.zeros(weight.size(0)) + bias_name = name.replace("weight", "bias") + yield from [(name, weight), (bias_name, bias)] + yield name, weight diff --git a/vllm/model_executor/parameter.py b/vllm/model_executor/parameter.py index fc5a3e7fba674..a9ce8af15d3bb 100644 --- a/vllm/model_executor/parameter.py +++ b/vllm/model_executor/parameter.py @@ -56,8 +56,14 @@ def __init__(self, data: torch.Tensor, weight_loader: Callable): def weight_loader(self): return self._weight_loader + def _is_1d_and_scalar(self, loaded_weight: torch.Tensor): + cond1 = self.data.ndim == 1 and self.data.numel() == 1 + cond2 = loaded_weight.ndim == 0 and loaded_weight.numel() == 1 + return (cond1 and cond2) + def _assert_and_load(self, loaded_weight: torch.Tensor): - assert self.data.shape == loaded_weight.shape + assert (self.data.shape == loaded_weight.shape + or self._is_1d_and_scalar(loaded_weight)) self.data.copy_(loaded_weight) def load_column_parallel_weight(self, loaded_weight: torch.Tensor): diff --git a/vllm/multimodal/inputs.py b/vllm/multimodal/inputs.py index 4b63703585214..b35184f6855ab 100644 --- a/vllm/multimodal/inputs.py +++ b/vllm/multimodal/inputs.py @@ -491,7 +491,7 @@ def get_items(self, modality: str) -> Sequence[MultiModalKwargsItem]: """ -class MultiModalInputsV2(TypedDict): +class MultiModalInputs(TypedDict): """ Represents the outputs of :class:`vllm.multimodal.processing.BaseMultiModalProcessor`, diff --git a/vllm/multimodal/processing.py b/vllm/multimodal/processing.py index 8b47dfb07387f..308e4a2e322e8 100644 --- a/vllm/multimodal/processing.py +++ b/vllm/multimodal/processing.py @@ -1,7 +1,8 @@ import re from abc import ABC, abstractmethod from collections import defaultdict -from collections.abc import Callable, ItemsView, Iterable, Mapping, Sequence +from collections.abc import (Callable, Generator, ItemsView, Iterable, Mapping, + Sequence) from dataclasses import dataclass, field from functools import lru_cache from typing import (TYPE_CHECKING, Generic, NamedTuple, Optional, Protocol, @@ -18,8 +19,8 @@ from .hasher import MultiModalHasher from .inputs import (MultiModalDataDict, MultiModalFieldConfig, - MultiModalInputsV2, MultiModalKwargs, - MultiModalKwargsItem, PlaceholderRange) + MultiModalInputs, MultiModalKwargs, MultiModalKwargsItem, + PlaceholderRange) from .parse import MultiModalDataItems, MultiModalDataParser if TYPE_CHECKING: @@ -30,21 +31,100 @@ _S = TypeVar("_S", str, list[int]) _PromptSeq = Union[str, list[int]] +PromptSeq = Union[str, list[int]] +"""A token sequence (list of token IDs) or text.""" + + +@dataclass +class PromptReplacementDetails: + """Details about the replacement token sequence or text.""" + + full: PromptSeq + """The full replacement.""" + + features: PromptSeq + """ + The part of the replacement that corresponds to feature placeholders; + this will be replaced by the output of the vision encoder during model + inference. + """ + + @staticmethod + def from_seq(seq: PromptSeq) -> "PromptReplacementDetails": + return PromptReplacementDetails(full=seq, features=seq) + + +PromptRepl = Union[PromptSeq, PromptReplacementDetails] +""" +The replacement token sequence or text. + +If only part of the replacement corresponds to feature placeholders, you can +use :class:`PromptReplacementDetails` to specify which part. +""" + @dataclass class PromptReplacement: """ Defines how to replace portions of an input prompt with placeholder tokens. + + Example: + + For each image, replace one ```` input placeholder in the prompt + with a number of ```` feature placeholders + equal to the feature size of the vision encoder: + + .. code-block:: python + + PromptReplacement( + modality="image", + target="", + replacement="" * image_feature_size, + ) + + As above, but further pad the feature placeholders with ```` + and ```, which are not supposed to be passed to the vision + encoder: + + .. code-block:: python + + PromptReplacement( + modality="image", + target="", + replacement=PromptReplacementDetails( + full="".join([ + "", + "" * image_feature_size, + "", + ]), + features="" * image_feature_size, + ), + ) + + To avoid unnecessary tokenization during prompt replacement, + we recommended passing token sequences instead of text: + + .. code-block:: python + + PromptReplacement( + modality="image", + target=[image_token_id], + replacement=PromptReplacementDetails( + full=([image_bos_id] + [image_token_id] * image_feature_size + + [image_eos_id]), + features=[image_token_id] * image_feature_size, + ), + ) """ modality: str """The modality for which the replacement is made.""" - target: _PromptSeq + target: PromptSeq """The token sequence (or text) to find and replace.""" - replacement: Union[Callable[[int], _PromptSeq], - _PromptSeq] = field(repr=False) + replacement: Union[Callable[[int], PromptRepl], + PromptRepl] = field(repr=False) """ Given the index of the processed item within :attr:`modality`, output the replacement token sequence (or text). @@ -107,11 +187,26 @@ def full_groupby_modality(values: Iterable[_M]) -> ItemsView[str, list[_M]]: @dataclass class _BoundPromptSequence: + """ + A :data:`_PromptSeq` bound to a tokenizer to automatically + convert between token sequence and text representations. + """ tokenizer: AnyTokenizer = field(repr=False) _text: Optional[str] _token_ids: Optional[list[int]] + @staticmethod + def from_seq( + tokenizer: AnyTokenizer, + seq: PromptSeq, + ) -> "_BoundPromptSequence": + return _BoundPromptSequence( + tokenizer=tokenizer, + _text=seq if isinstance(seq, str) else None, + _token_ids=seq if isinstance(seq, list) else None, + ) + def __post_init__(self) -> None: if self._text is None and self._token_ids is None: raise ValueError("At least one of 'text' and 'token_ids' must be " @@ -134,6 +229,12 @@ def token_ids(self) -> list[int]: return self._token_ids +@dataclass +class _BoundPromptReplacementGroup: + full: _BoundPromptSequence + features: _BoundPromptSequence + + @dataclass class BoundPromptReplacement: """ @@ -144,25 +245,19 @@ class BoundPromptReplacement: tokenizer: AnyTokenizer = field(repr=False) modality: str - _target: _PromptSeq - _replacement: Union[Callable[[int], _PromptSeq], - _PromptSeq] = field(repr=False) + _target: PromptSeq + _replacement: Union[Callable[[int], PromptRepl], + PromptRepl] = field(repr=False) def __post_init__(self) -> None: - self._replacement_cache = dict[int, _BoundPromptSequence]() + self._replacement_cache = dict[int, _BoundPromptReplacementGroup]() @property def target(self) -> _BoundPromptSequence: """The token sequence (or text) to find and replace.""" - target = self._target - - return _BoundPromptSequence( - tokenizer=self.tokenizer, - _text=target if isinstance(target, str) else None, - _token_ids=target if isinstance(target, list) else None, - ) + return _BoundPromptSequence.from_seq(self.tokenizer, self._target) - def get_replacement(self, item_idx: int) -> _BoundPromptSequence: + def get_replacement(self, item_idx: int) -> _BoundPromptReplacementGroup: """ Given the index of the processed item within :attr:`modality`, output the replacement token sequence (or text). @@ -177,10 +272,16 @@ def get_replacement(self, item_idx: int) -> _BoundPromptSequence: else: cache_key = None - bound_replacement = _BoundPromptSequence( - tokenizer=self.tokenizer, - _text=replacement if isinstance(replacement, str) else None, - _token_ids=replacement if isinstance(replacement, list) else None, + if not isinstance(replacement, PromptReplacementDetails): + replacement = PromptReplacementDetails.from_seq(replacement) + + bound_full = _BoundPromptSequence.from_seq(self.tokenizer, + replacement.full) + bound_features = _BoundPromptSequence.from_seq(self.tokenizer, + replacement.features) + bound_replacement = _BoundPromptReplacementGroup( + full=bound_full, + features=bound_features, ) if cache_key is not None: @@ -197,7 +298,7 @@ class _TokenMatch(NamedTuple): def iter_token_matches( token_ids: list[int], match_ids: list[int], -) -> Iterable[_TokenMatch]: +) -> Generator[_TokenMatch]: """ Yield each occurrence of :code:`match_ids` in :code:`token_ids`. @@ -272,15 +373,15 @@ def end_idx(self) -> int: @dataclass -class PlaceholderInfo: +class PlaceholderFeaturesInfo: modality: str item_idx: int start_idx: int - replacement: list[int] + tokens: list[int] @property def length(self) -> int: - return len(self.replacement) + return len(self.tokens) def to_range(self) -> PlaceholderRange: return PlaceholderRange( @@ -314,7 +415,7 @@ def find_text_matches( def _resolve_matches( - prompt: _PromptSeq, + prompt: PromptSeq, mm_matches: Mapping[str, Sequence[_PromptReplacementMatch]], ) -> list[_PromptReplacementMatch]: """ @@ -362,10 +463,10 @@ def _replace_matches( replacement = repl_info.get_replacement(item_idx) if isinstance(prompt, str): - repl_seq = replacement.text + repl_seq = replacement.full.text out_seqs.append(prompt[prev_end_idx:start_idx] + repl_seq) else: - repl_seq = replacement.token_ids + repl_seq = replacement.full.token_ids out_seqs.append(prompt[prev_end_idx:start_idx] + repl_seq) prev_end_idx = end_idx @@ -404,78 +505,77 @@ def replace_text_matches( return "".join(texts) -def _iter_modality_placeholders( +def _iter_placeholders( + mm_prompt_repls: Mapping[str, Sequence[BoundPromptReplacement]], prompt: list[int], - modality: str, - modality_repls: Sequence[BoundPromptReplacement], - modal_item_count: int, -) -> Iterable[PlaceholderInfo]: - if modal_item_count == 0: - return + mm_item_counts: Mapping[str, int], +) -> Iterable[PlaceholderFeaturesInfo]: + """ + Yield each set of placeholder tokens found in :code:`prompt`. + + Matches are exclusive even when multiple modalities share + the same placeholder tokens. In that case, the modality that + appears earlier in `mm_prompt_repls` takes priority. + Note that empty matches are ignored. + """ prompt_len = len(prompt) - item_idx = 0 + item_idx_by_modality = defaultdict[str, int](lambda: 0) start_idx = 0 while start_idx < prompt_len: found = False - for repl_info in modality_repls: - replacement = repl_info.get_replacement(item_idx) - repl_tokens = replacement.token_ids - repl_len = len(repl_tokens) - end_idx = start_idx + repl_len - - if repl_len == 0 or end_idx > prompt_len: + for modality, modality_repls in mm_prompt_repls.items(): + item_idx = item_idx_by_modality[modality] + if item_idx >= mm_item_counts.get(modality, 0): continue - if prompt[start_idx:end_idx] == repl_tokens: - yield PlaceholderInfo( - modality=modality, - item_idx=item_idx, - start_idx=start_idx, - replacement=repl_tokens, - ) - - item_idx += 1 - if item_idx >= modal_item_count: - return - - # Exclude overlapping matches - start_idx = end_idx - found = True - break + for repl_info in modality_repls: + replacement = repl_info.get_replacement(item_idx) + repl_tokens_full = replacement.full.token_ids + repl_len_full = len(repl_tokens_full) + end_idx_full = start_idx + repl_len_full + + if repl_len_full == 0 or end_idx_full > prompt_len: + continue + + if prompt[start_idx:end_idx_full] == repl_tokens_full: + repl_tokens_feat = replacement.features.token_ids + + try: + match = next( + iter_token_matches(repl_tokens_full, + repl_tokens_feat)) + yield PlaceholderFeaturesInfo( + modality=modality, + item_idx=item_idx, + start_idx=start_idx + match.start_idx, + tokens=repl_tokens_feat, + ) + except StopIteration: + raise AssertionError( + f"{repl_tokens_feat=} should be a " + f"subsequence of {repl_tokens_full=}") from None + + # Exclude overlapping matches + start_idx = end_idx_full + item_idx_by_modality[modality] += 1 + found = True + break + + if found: + break # Go back to the outer while loop if not found: start_idx += 1 -def _iter_placeholders( - mm_prompt_repls: Mapping[str, Sequence[BoundPromptReplacement]], - prompt: list[int], - mm_item_counts: Mapping[str, int], -) -> Iterable[PlaceholderInfo]: - """ - For each modality, yield each set of placeholder tokens found in - :code:`prompt`. - - Note that empty matches are ignored. - """ - for modality, modal_item_count in mm_item_counts.items(): - if modality in mm_prompt_repls: - yield from _iter_modality_placeholders( - prompt, - modality, - mm_prompt_repls[modality], - modal_item_count, - ) - - def find_mm_placeholders( mm_prompt_repls: Mapping[str, Sequence[BoundPromptReplacement]], prompt: list[int], mm_item_counts: Mapping[str, int], -) -> Mapping[str, list[PlaceholderInfo]]: +) -> Mapping[str, list[PlaceholderFeaturesInfo]]: it = _iter_placeholders(mm_prompt_repls, prompt, mm_item_counts) return dict(full_groupby_modality(it)) @@ -620,7 +720,7 @@ def __call__( prompt: str, mm_data: MultiModalDataDict, hf_processor_mm_kwargs: Mapping[str, object], - ) -> MultiModalInputsV2: + ) -> MultiModalInputs: return self.apply(prompt, mm_data, hf_processor_mm_kwargs) def _get_data_parser(self) -> MultiModalDataParser: @@ -690,7 +790,7 @@ def _find_mm_placeholders( mm_prompt_repls: Mapping[str, Sequence[BoundPromptReplacement]], new_token_ids: list[int], mm_item_counts: Mapping[str, int], - ) -> Mapping[str, list[PlaceholderInfo]]: + ) -> Mapping[str, list[PlaceholderFeaturesInfo]]: return find_mm_placeholders(mm_prompt_repls, new_token_ids, mm_item_counts) @@ -959,7 +1059,7 @@ def _apply_prompt_replacements( token_ids: list[int], mm_prompt_repls: Mapping[str, Sequence[BoundPromptReplacement]], mm_item_counts: Mapping[str, int], - ) -> tuple[list[int], str, Mapping[str, list[PlaceholderInfo]]]: + ) -> tuple[list[int], str, Mapping[str, list[PlaceholderFeaturesInfo]]]: tokenizer = self.info.get_tokenizer() mm_token_matches = { @@ -1048,7 +1148,7 @@ def _validate_mm_kwargs( def _validate_mm_placeholders( self, - mm_placeholders: Mapping[str, list[PlaceholderInfo]], + mm_placeholders: Mapping[str, list[PlaceholderFeaturesInfo]], mm_item_counts: Mapping[str, int], *, allow_missing: bool = False, @@ -1078,7 +1178,7 @@ def apply( prompt: Union[str, list[int]], mm_data: MultiModalDataDict, hf_processor_mm_kwargs: Mapping[str, object], - ) -> MultiModalInputsV2: + ) -> MultiModalInputs: """ Process multi-modal inputs to be used in vLLM. @@ -1156,7 +1256,7 @@ def apply( # If HF processor already inserts placeholder tokens, # there is no need for us to insert them - if all(len(repls) == 0 for repls in mm_missing_repls.items()): + if all(len(repls) == 0 for repls in mm_missing_repls.values()): tokenizer = self.info.get_tokenizer() prompt = decode_tokens(tokenizer, prompt_ids) mm_placeholders = hf_mm_placeholders @@ -1180,7 +1280,7 @@ def apply( for modality, placeholders in mm_placeholders.items() } - return MultiModalInputsV2( + return MultiModalInputs( type="multimodal", prompt=prompt, prompt_token_ids=prompt_ids, diff --git a/vllm/multimodal/profiling.py b/vllm/multimodal/profiling.py index ec580cd6ecddd..c68edaff80167 100644 --- a/vllm/multimodal/profiling.py +++ b/vllm/multimodal/profiling.py @@ -11,7 +11,7 @@ from vllm.inputs import DummyData from vllm.logger import init_logger -from .inputs import MultiModalDataDict, MultiModalInputsV2 +from .inputs import MultiModalDataDict, MultiModalInputs from .processing import BaseMultiModalProcessor, BaseProcessingInfo logger = init_logger(__name__) @@ -106,7 +106,7 @@ def processing_info(self) -> BaseProcessingInfo: def dummy_inputs(self) -> BaseDummyInputsBuilder[_I]: return self.processor.dummy_inputs - def _get_mm_limits(self) -> Mapping[str, int]: + def get_mm_limits(self) -> Mapping[str, int]: mm_config = self.processing_info.ctx.get_mm_config() mm_limit_per_prompt = mm_config.limit_per_prompt @@ -131,7 +131,7 @@ def _get_dummy_mm_inputs( self, seq_len: int, mm_counts: Mapping[str, int], - ) -> MultiModalInputsV2: + ) -> MultiModalInputs: factory = self.dummy_inputs processor_inputs = factory.get_dummy_processor_inputs( seq_len, mm_counts) @@ -146,7 +146,7 @@ def get_dummy_data(self, seq_len: int) -> DummyData: # Avoid circular import from vllm.sequence import SequenceData - mm_counts = self._get_mm_limits() + mm_counts = self.get_mm_limits() info = self.processing_info mm_max_tokens_per_item = info.get_mm_max_tokens_per_item(seq_len) diff --git a/vllm/multimodal/registry.py b/vllm/multimodal/registry.py index 804a91da8c889..7a4b85385cac9 100644 --- a/vllm/multimodal/registry.py +++ b/vllm/multimodal/registry.py @@ -17,7 +17,7 @@ from .inputs import MultiModalDataDict, MultiModalKwargs, NestedTensors from .processing import (BaseMultiModalProcessor, BaseProcessingInfo, ProcessingCache) -from .profiling import BaseDummyInputsBuilder +from .profiling import BaseDummyInputsBuilder, MultiModalProfiler from .utils import cached_get_tokenizer from .video import VideoPlugin @@ -252,14 +252,14 @@ def get_max_tokens_per_item_by_modality( model_config: "ModelConfig", ) -> Mapping[str, int]: """ - Get the maximum number of tokens per data item from each modality - for profiling the memory usage of a model. - - Note: - This is currently directly used only in V1. + Get the maximum number of tokens per data item from each modality based + on underlying model configuration. """ if self.has_processor(model_config): - tokenizer = cached_get_tokenizer(model_config.tokenizer) + tokenizer = cached_get_tokenizer( + model_config.tokenizer, + trust_remote_code=model_config.trust_remote_code, + ) processor = self.create_processor(model_config, tokenizer) seq_len = model_config.max_model_len return processor.info.get_mm_max_tokens_per_item(seq_len) @@ -269,6 +269,28 @@ def get_max_tokens_per_item_by_modality( for key, plugin in self._plugins.items() } + def get_max_tokens_per_item_by_nonzero_modality( + self, + model_config: "ModelConfig", + ) -> Mapping[str, int]: + """ + Get the maximum number of tokens per data item from each modality based + on underlying model configuration, excluding modalities that user + explicitly disabled via `limit_mm_per_prompt`. + + Note: + This is currently directly used only in V1 for profiling the memory + usage of a model. + """ + mm_limits = self.get_mm_limits_per_prompt(model_config) + + return { + key: max_tokens_per_mm_item + for key, max_tokens_per_mm_item in + self.get_max_tokens_per_item_by_modality(model_config).items() + if mm_limits[key] > 0 + } + def get_max_tokens_by_modality( self, model_config: "ModelConfig", @@ -282,10 +304,10 @@ def get_max_tokens_by_modality( Note: This should be called after :meth:`init_mm_limits_per_prompt`. """ - limits_per_plugin = self._limits_by_model[model_config] + mm_limits = self.get_mm_limits_per_prompt(model_config) return { - key: limits_per_plugin[key] * max_tokens_per_mm_item + key: mm_limits[key] * max_tokens_per_mm_item for key, max_tokens_per_mm_item in self.get_max_tokens_per_item_by_modality(model_config).items() } @@ -349,6 +371,15 @@ def get_mm_limits_per_prompt( Note: This should be called after :meth:`init_mm_limits_per_prompt`. """ + if self.has_processor(model_config): + tokenizer = cached_get_tokenizer( + model_config.tokenizer, + trust_remote_code=model_config.trust_remote_code, + ) + processor = self.create_processor(model_config, tokenizer) + profiler = MultiModalProfiler(processor) + return profiler.get_mm_limits() + return self._limits_by_model[model_config] def register_processor( diff --git a/vllm/multimodal/utils.py b/vllm/multimodal/utils.py index 1c6bbf77b926f..8a296ae2eec3c 100644 --- a/vllm/multimodal/utils.py +++ b/vllm/multimodal/utils.py @@ -1,4 +1,5 @@ from functools import lru_cache +from itertools import groupby from pathlib import Path from typing import TYPE_CHECKING, Optional, TypeVar, Union from urllib.parse import ParseResult, urlparse @@ -26,7 +27,7 @@ if TYPE_CHECKING: from .hasher import MultiModalHashDict - from .inputs import MultiModalPlaceholderDict + from .inputs import MultiModalKwargs, MultiModalPlaceholderDict class MediaConnector: @@ -477,3 +478,39 @@ def merge_and_sort_multimodal_metadata( merged_hashes = None return sorted_modalities, merged_placeholders, merged_hashes + + +def group_mm_inputs_by_modality( + mm_inputs: list["MultiModalKwargs"]) -> list[list["MultiModalKwargs"]]: + """Group consecutive MultiModalKwargs from mm_inputs with the same modality + together into the same list for batching purpose. For MultiModalKwargs with + multiple modalities, put them into their own list. + + Args: + mm_inputs: List of MultiModalKwargs. + + Returns: + list[list[MultiModalKwargs]]: List of list of MultiModalKwargs, each + inner list contains consecutive MultiModalKwargs with same modality, or + one with multimodal modalities. + """ + if not mm_inputs: + return [] + + def modality_group_func(mm_input: "MultiModalKwargs") -> Union[str, int]: + # If the input has multiple modalities, return a id as the unique key + # for the mm_input input. + if len(mm_input.modalities) > 1: + return id(mm_input) + + elif len(mm_input.modalities) == 1: + return list(mm_input.modalities)[0] + + # FIXME(Isotr0py): Modality of mm_input from legacy pipeline is empty, + # this is used to make InternVL with legacy pipeline still work with v1. + else: + return "" + + return [ + list(group) for _, group in groupby(mm_inputs, key=modality_group_func) + ] \ No newline at end of file diff --git a/vllm/outputs.py b/vllm/outputs.py index b519c159b1531..25b2265285d16 100644 --- a/vllm/outputs.py +++ b/vllm/outputs.py @@ -1,6 +1,6 @@ import time from dataclasses import dataclass -from typing import Dict, Generic, List, Optional +from typing import Dict, Generic, List, MutableSequence, Optional from typing import Sequence as GenericSequence from typing import Union @@ -162,6 +162,26 @@ def new( finished=finished, ) + def add(self, next_output: "RequestOutput") -> None: + """Merge subsequent RequestOutput into this one""" + + self.prompt = next_output.prompt + self.prompt_token_ids = next_output.prompt_token_ids + self.prompt_logprobs = next_output.prompt_logprobs + self.finished |= next_output.finished + + #TODO assuming n == 1 for now + completion = self.outputs[0] + next_completion = next_output.outputs[0] + completion.text += next_completion.text + if not isinstance(completion.token_ids, MutableSequence): + completion.token_ids = list(completion.token_ids) + completion.token_ids.extend(next_completion.token_ids) + if next_completion.logprobs: + assert completion.logprobs is not None + completion.logprobs.extend(next_completion.logprobs) + completion.cumulative_logprob = next_completion.cumulative_logprob + @classmethod def from_seq_group( cls, seq_group: SequenceGroup, use_cache: bool, @@ -172,9 +192,9 @@ def from_seq_group( if seq_group.request_id in seq_id_to_seq_group: group: SequenceGroupBase = seq_id_to_seq_group[ seq_group.request_id] + assembled_seq_group = group.maybe_assemble_group(seq_group) if finished: group.finish_seq(seq_group) - assembled_seq_group = group.maybe_assemble_group(seq_group) if assembled_seq_group is None: return None return cls.from_seq_group(assembled_seq_group, use_cache, diff --git a/vllm/platforms/__init__.py b/vllm/platforms/__init__.py index 6ca95b41dbb07..ddbdc43ca5710 100644 --- a/vllm/platforms/__init__.py +++ b/vllm/platforms/__init__.py @@ -101,6 +101,10 @@ def cpu_platform_plugin() -> Optional[str]: try: from importlib.metadata import version is_cpu = "cpu" in version("vllm") + if not is_cpu: + import platform + is_cpu = platform.machine().lower().startswith("arm") + except Exception: pass @@ -213,8 +217,11 @@ def __getattr__(name: str): global _init_trace _init_trace = "".join(traceback.format_stack()) return _current_platform - else: + elif name in globals(): return globals()[name] + else: + raise AttributeError( + f"No attribute named '{name}' exists in {__name__}.") __all__ = [ diff --git a/vllm/platforms/cpu.py b/vllm/platforms/cpu.py index 4d3b84fea887f..74948202cbe48 100644 --- a/vllm/platforms/cpu.py +++ b/vllm/platforms/cpu.py @@ -1,3 +1,4 @@ +import os from typing import TYPE_CHECKING, Optional import psutil @@ -105,6 +106,32 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None: else: parallel_config.worker_cls = "vllm.worker.cpu_worker.CPUWorker" + assert vllm_config.device_config.device_type == "cpu" + + # + # Environment variables for CPU executor + # + + # Disable torch async compiling which won't work with daemonic processes + os.environ["TORCHINDUCTOR_COMPILE_THREADS"] = "1" + + # Intel OpenMP setting + ld_prealod_str = os.getenv("LD_PRELOAD", "") + if "libiomp5.so" in ld_prealod_str: + # The time(milliseconds) that a thread should wait after + # completing the execution of a parallel region, before sleeping. + os.environ['KMP_BLOCKTIME'] = "1" + # Prevents the CPU to run into low performance state + os.environ['KMP_TPAUSE'] = "0" + # Provides fine granularity parallelism + os.environ['KMP_FORKJOIN_BARRIER_PATTERN'] = "dist,dist" + os.environ['KMP_PLAIN_BARRIER_PATTERN'] = "dist,dist" + os.environ['KMP_REDUCTION_BARRIER_PATTERN'] = "dist,dist" + + # To hint IPEX uses shared memory based AllReduce + os.environ["LOCAL_WORLD_SIZE"] = str( + vllm_config.parallel_config.tensor_parallel_size) + @classmethod def is_pin_memory_available(cls) -> bool: logger.warning("Pin memory is not supported on CPU.") diff --git a/vllm/platforms/cuda.py b/vllm/platforms/cuda.py index 80cefcb492531..2587e3a11dde3 100644 --- a/vllm/platforms/cuda.py +++ b/vllm/platforms/cuda.py @@ -143,6 +143,13 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None: if cache_config and cache_config.block_size is None: cache_config.block_size = 16 + @classmethod + def get_current_memory_usage(cls, + device: Optional[torch.types.Device] = None + ) -> float: + torch.cuda.reset_peak_memory_stats(device) + return torch.cuda.max_memory_allocated(device) + @classmethod def get_attn_backend_cls(cls, selected_backend, head_size, dtype, kv_cache_dtype, block_size, use_v1) -> str: diff --git a/vllm/platforms/hpu.py b/vllm/platforms/hpu.py index 242c2c127979a..a32c262c84efa 100644 --- a/vllm/platforms/hpu.py +++ b/vllm/platforms/hpu.py @@ -1,7 +1,9 @@ +import os from typing import TYPE_CHECKING, Optional import torch +from vllm import envs from vllm.logger import init_logger from .interface import Platform, PlatformEnum, _Backend @@ -58,6 +60,22 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None: cache_config = vllm_config.cache_config if cache_config and cache_config.block_size is None: cache_config.block_size = 128 + if (parallel_config.distributed_executor_backend == 'mp' + and envs.VLLM_WORKER_MULTIPROC_METHOD == 'fork'): + if os.environ.get("VLLM_WORKER_MULTIPROC_METHOD", + None) is not None: + logger.warning("On HPU, VLLM_WORKER_MULTIPROC_METHOD=fork " + "might cause application hangs on exit. Using " + "VLLM_WORKER_MULTIPROC_METHOD=fork anyway, " + "as it was explicitly requested.") + else: + logger.warning( + "On HPU, VLLM_WORKER_MULTIPROC_METHOD=fork " + "might cause application hangs on exit. Setting " + "VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. " + "To override that behavior, please set " + "VLLM_WORKER_MULTIPROC_METHOD=fork explicitly.") + os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" @classmethod def is_pin_memory_available(cls): diff --git a/vllm/platforms/interface.py b/vllm/platforms/interface.py index 3c2ec9636df91..f2ecec3203fb7 100644 --- a/vllm/platforms/interface.py +++ b/vllm/platforms/interface.py @@ -33,6 +33,7 @@ class _Backend(enum.Enum): HPU_ATTN = enum.auto() PALLAS = enum.auto() IPEX = enum.auto() + BLOCK_SPARSE_FLASH_ATTN = enum.auto() NO_ATTENTION = enum.auto() @@ -276,6 +277,15 @@ def is_pin_memory_available(cls) -> bool: return False return True + @classmethod + def get_current_memory_usage(cls, + device: Optional[torch.types.Device] = None + ) -> float: + """ + Return the memory usage in bytes. + """ + raise NotImplementedError + @classmethod def get_punica_wrapper(cls) -> str: """ diff --git a/vllm/platforms/neuron.py b/vllm/platforms/neuron.py index 0696f73cc17b4..ead3dab05a6b1 100644 --- a/vllm/platforms/neuron.py +++ b/vllm/platforms/neuron.py @@ -35,6 +35,14 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None: parallel_config.worker_cls = \ "vllm.worker.neuron_worker.NeuronWorker" + if parallel_config.world_size > 1: + parallel_config.distributed_executor_backend = "uni" + + assert (vllm_config.lora_config is + None), "LoRA is not supported for Neuron backend." + assert (not vllm_config.speculative_config + ), "Speculative decoding not yet supported for Neuron backend." + cache_config = vllm_config.cache_config if cache_config: # neuron needs block_size = max_model_len diff --git a/vllm/platforms/openvino.py b/vllm/platforms/openvino.py index 9390eda535c8f..7d414165a8188 100644 --- a/vllm/platforms/openvino.py +++ b/vllm/platforms/openvino.py @@ -66,9 +66,8 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None: from vllm.utils import GiB_bytes parallel_config = vllm_config.parallel_config - assert ( - parallel_config.world_size == 1 - ), "OpenVINOExecutor only supports single CPU socket currently." + assert (parallel_config.world_size == 1 + ), "OpenVINO only supports single CPU socket currently." if parallel_config.worker_cls == "auto": parallel_config.worker_cls = \ @@ -141,3 +140,10 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None: raise RuntimeError( "Invalid environment variable VLLM_OPENVINO_KVCACHE_SPACE" f" {kv_cache_space}, expect a positive integer value.") + + assert vllm_config.device_config.device_type == "openvino" + assert vllm_config.lora_config is None, \ + "OpenVINO backend doesn't support LoRA" + assert cls.is_openvino_cpu() or \ + cls.is_openvino_gpu(), \ + "OpenVINO backend supports only CPU and GPU devices" diff --git a/vllm/platforms/rocm.py b/vllm/platforms/rocm.py index 75b4474001b29..53e2812343156 100644 --- a/vllm/platforms/rocm.py +++ b/vllm/platforms/rocm.py @@ -108,7 +108,7 @@ class RocmPlatform(Platform): supported_quantization: list[str] = [ "awq", "gptq", "fp8", "compressed_tensors", "compressed-tensors", - "fbgemm_fp8", "gguf" + "fbgemm_fp8", "gguf", "quark" ] @classmethod @@ -224,3 +224,10 @@ def verify_quantization(cls, quant: str) -> None: @classmethod def get_punica_wrapper(cls) -> str: return "vllm.lora.punica_wrapper.punica_gpu.PunicaWrapperGPU" + + @classmethod + def get_current_memory_usage(cls, + device: Optional[torch.types.Device] = None + ) -> float: + torch.cuda.reset_peak_memory_stats(device) + return torch.cuda.mem_get_info()[1] - torch.cuda.mem_get_info()[0] diff --git a/vllm/platforms/tpu.py b/vllm/platforms/tpu.py index ff9487daac7a7..05a3aa4305cfa 100644 --- a/vllm/platforms/tpu.py +++ b/vllm/platforms/tpu.py @@ -72,6 +72,16 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None: assert vllm_config.speculative_config is None, \ "TPU does not support speculative decoding" + assert not vllm_config.scheduler_config.chunked_prefill_enabled, ( + "Chunked prefill is not yet supported for TPU backend") + assert not vllm_config.speculative_config, ( + "Speculative decoding is not yet supported for TPU backend") + if vllm_config.model_config.dtype in (torch.float16, torch.float32): + logger.warning( + "The TPU backend currently does not support %s. " + "Using bfloat16 instead.", vllm_config.model_config.dtype) + vllm_config.model_config.dtype = torch.bfloat16 + parallel_config = vllm_config.parallel_config scheduler_config = vllm_config.scheduler_config if parallel_config.worker_cls == "auto": diff --git a/vllm/platforms/xpu.py b/vllm/platforms/xpu.py index f34376b44e689..c34b5b58672e7 100644 --- a/vllm/platforms/xpu.py +++ b/vllm/platforms/xpu.py @@ -78,19 +78,40 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None: raise NotImplementedError( "XPU does not support speculative decoding") + if vllm_config.device_config is not None: + assert vllm_config.device_config.device_type == "xpu" + # check and update parallel config parallel_config = vllm_config.parallel_config - if (parallel_config.distributed_executor_backend is not None - and parallel_config.distributed_executor_backend != "ray"): + if parallel_config.worker_cls == "auto": + parallel_config.worker_cls = "vllm.worker.xpu_worker.XPUWorker" + + if parallel_config.distributed_executor_backend is None: + parallel_config.distributed_executor_backend = "ray" + elif parallel_config.distributed_executor_backend == "mp": + # FIXME(kunshang): + # spawn needs calling `if __name__ == '__main__':`` + # fork is not supported for xpu start new process. + logger.error( + "Both start methods (spawn and fork) have issue " + "on XPU if you use mp backend, setting it to ray instead.") + parallel_config.distributed_executor_backend = "ray" + + elif parallel_config.distributed_executor_backend != "ray": logger.warning( "%s is not supported on XPU, fallback to ray distributed" " executor backend.", parallel_config.distributed_executor_backend) parallel_config.distributed_executor_backend = "ray" - if parallel_config.worker_cls == "auto": - parallel_config.worker_cls = "vllm.worker.xpu_worker.XPUWorker" @classmethod def is_pin_memory_available(cls): logger.warning("Pin memory is not supported on XPU.") return False + + @classmethod + def get_current_memory_usage(cls, + device: Optional[torch.types.Device] = None + ) -> float: + torch.xpu.reset_peak_memory_stats(device) + return torch.xpu.max_memory_allocated(device) diff --git a/vllm/plugins/__init__.py b/vllm/plugins/__init__.py index e5fa4f0e4a2f6..a78a054917756 100644 --- a/vllm/plugins/__init__.py +++ b/vllm/plugins/__init__.py @@ -50,22 +50,18 @@ def load_general_plugins(): processes. They should be designed in a way that they can be loaded multiple times without causing issues. """ + global plugins_loaded + if plugins_loaded: + return + plugins_loaded = True - # all processes created by vllm will load plugins, - # and here we can inject some common environment variables - # for all processes. - - # see https://github.com/vllm-project/vllm/issues/10480 - os.environ['TORCHINDUCTOR_COMPILE_THREADS'] = '1' - # see https://github.com/vllm-project/vllm/issues/10619 - torch._inductor.config.compile_threads = 1 - + # some platform-specific configurations from vllm.platforms import current_platform if current_platform.is_xpu(): # see https://github.com/pytorch/pytorch/blob/43c5f59/torch/_dynamo/config.py#L158 torch._dynamo.config.disable = True - if current_platform.is_hpu(): + elif current_platform.is_hpu(): # NOTE(kzawora): PT HPU lazy backend (PT_HPU_LAZY_MODE = 1) # does not support torch.compile # Eager backend (PT_HPU_LAZY_MODE = 0) must be selected for @@ -78,10 +74,6 @@ def load_general_plugins(): # see https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html # noqa: E501 os.environ['PT_HPU_ENABLE_LAZY_COLLECTIVES'] = 'true' - global plugins_loaded - if plugins_loaded: - return - plugins_loaded = True plugins = load_plugins_by_group(group='vllm.general_plugins') # general plugins, we only need to execute the loaded functions for func in plugins.values(): diff --git a/vllm/profiler/layerwise_profile.py b/vllm/profiler/layerwise_profile.py index 33babfebdca1e..29c0edd0ee535 100644 --- a/vllm/profiler/layerwise_profile.py +++ b/vllm/profiler/layerwise_profile.py @@ -1,7 +1,7 @@ import copy from collections import defaultdict from dataclasses import asdict, dataclass, field -from typing import Callable, Dict, List, Optional, Tuple, TypeAlias, Union +from typing import Any, Callable, Dict, List, Optional, Tuple, TypeAlias, Union import pandas as pd from torch._C._autograd import DeviceType, _KinetoEvent, _ProfilerResult @@ -128,7 +128,7 @@ def export_summary_stats_table_csv(self, filename: str): ]) df.to_csv(filename) - def convert_stats_to_dict(self) -> str: + def convert_stats_to_dict(self) -> dict[str, Any]: return { "metadata": { "num_running_seqs": self.num_running_seqs @@ -227,7 +227,7 @@ def _total_cuda_time(self): [self._cumulative_cuda_time(root) for root in self._module_tree]) def _build_stats_trees(self): - summary_dict: Dict[str, self.StatsTreeNode] = {} + summary_dict: Dict[str, _StatsTreeNode] = {} total_cuda_time = self._total_cuda_time() def pct_cuda_time(cuda_time_us): diff --git a/vllm/prompt_adapter/utils.py b/vllm/prompt_adapter/utils.py index 473b87c89c21d..8b2732923c4e7 100644 --- a/vllm/prompt_adapter/utils.py +++ b/vllm/prompt_adapter/utils.py @@ -89,6 +89,7 @@ def load_peft_weights(model_id: str, adapters_weights = safe_load_file(filename, device=device) else: adapters_weights = torch.load(filename, - map_location=torch.device(device)) + map_location=torch.device(device), + weights_only=True) return adapters_weights diff --git a/vllm/sequence.py b/vllm/sequence.py index 5857f656dfc10..74320db709f94 100644 --- a/vllm/sequence.py +++ b/vllm/sequence.py @@ -815,7 +815,9 @@ def set_finished_time(self, time: Optional[float]) -> None: def get_max_num_running_seqs(self) -> int: """The maximum number of sequences running in parallel in the remaining lifetime of the request.""" - return 0 if self.first_seq.is_finished() else 1 + if self.is_single_seq: + return 0 if self.first_seq.is_finished() else 1 + return self.num_seqs() - self.num_finished_seqs() def get_seqs( self, @@ -824,7 +826,10 @@ def get_seqs( if status is None: return self.seqs - return self.seqs if self.first_seq.status == status else [] + if self.is_single_seq: + return self.seqs if self.first_seq.status == status else [] + + return [seq for seq in self.seqs if seq.status == status] def is_encoder_decoder(self) -> bool: return self.encoder_seq is not None @@ -833,19 +838,22 @@ def get_encoder_seq(self) -> Optional[Sequence]: return self.encoder_seq def get_finished_seqs(self) -> List[Sequence]: - return self.seqs if self.first_seq.is_finished() else [] + if self.is_single_seq: + return self.seqs if self.first_seq.is_finished() else [] + + return [seq for seq in self.seqs if seq.is_finished()] def update_num_computed_tokens(self, num_new_computed_tokens: int): """Update number of tokens computed so far.""" - seq = self.first_seq - if not seq.is_finished(): - seq.data.update_num_computed_tokens(num_new_computed_tokens) + for seq in self.seqs: + if not seq.is_finished(): + seq.data.update_num_computed_tokens(num_new_computed_tokens) def get_num_uncomputed_tokens(self) -> int: num_uncomputed_tokens = 0 - seq = self.first_seq - if not seq.is_finished(): - num_uncomputed_tokens += seq.data.get_num_uncomputed_tokens() + for seq in self.seqs: + if not seq.is_finished(): + num_uncomputed_tokens += seq.data.get_num_uncomputed_tokens() return num_uncomputed_tokens def num_seqs(self, status: Optional[SequenceStatus] = None) -> int: @@ -860,10 +868,14 @@ def num_seqs(self, status: Optional[SequenceStatus] = None) -> int: return len(self.get_seqs(status)) def num_finished_seqs(self) -> int: - return 1 if self.first_seq.is_finished() else 0 + if self.is_single_seq: + return 1 if self.seqs[0].is_finished() else 0 + return len(self.get_finished_seqs()) def is_finished(self) -> bool: - return self.first_seq.is_finished() + if self.is_single_seq: + return self.first_seq.is_finished() + return all(seq.is_finished() for seq in self.seqs) def is_prefill(self) -> bool: return self.first_seq.is_prefill() @@ -1391,13 +1403,15 @@ class ParallelSampleSequenceGroup(SequenceGroupBase): @staticmethod def add_request(request_id: str, engine, params, **kwargs): original_params = params - params = original_params.clone() - params.n = 1 group = ParallelSampleSequenceGroup(request_id) seqs = [] for i in range(original_params.n): request_id_i = f"{request_id}_parallel_sample_{i}" group.seq_id_to_index[request_id_i] = i + params = copy.deepcopy(original_params) + params.n = 1 + if params.seed is not None: + params.seed += i seq_group = engine._add_processed_request( request_id_i, params=params, @@ -1432,33 +1446,34 @@ def maybe_assemble_group( self, seq_group: SequenceGroup) -> Optional[SequenceGroup]: # in the streaming mode, we will return the assembled sequence - # for the first sequence, and then return None for the rest of - # sequences + # for the first remaining sequence, and then return None for the + # rest of sequences if self.streaming: - if self.seq_id_to_index[seq_group.request_id] == 0: + first_remaining_id = next(iter(self.to_be_finished)) + if seq_group.request_id == first_remaining_id: return self.assembled_seq_group return None # in the non-streaming mode, we will return the assembled sequence - # once after all sequences finish, and then return None for the + # when the last sequences finishes, and then return None for the # rest of the time - - if len(self.to_be_finished) > 0: - return None - - assert self.assembled_seq_group is not None - params = self.assembled_seq_group.sampling_params - assert isinstance(params, SamplingParams) - if not self.output_produced: - self.output_produced = True - if params._real_n is not None: - # Get the top-n sequences. - n = params._real_n or params.n - seqs = self.assembled_seq_group.seqs - sorting_key = lambda seq: seq.get_cumulative_logprob() - sorted_seqs = sorted(seqs, key=sorting_key, reverse=True) - top_n_seqs = sorted_seqs[:n] - self.assembled_seq_group.seqs = top_n_seqs - return self.assembled_seq_group - if self.output_produced: - return None + if (len(self.to_be_finished) == 1 + and seq_group.request_id in self.to_be_finished + and seq_group.is_finished()): + assert self.assembled_seq_group is not None + params = self.assembled_seq_group.sampling_params + assert isinstance(params, SamplingParams) + if not self.output_produced: + self.output_produced = True + if params._real_n is not None: + # Get the top-n sequences. + n = params._real_n or params.n + seqs = self.assembled_seq_group.seqs + sorting_key = lambda seq: seq.get_cumulative_logprob() + sorted_seqs = sorted(seqs, key=sorting_key, reverse=True) + top_n_seqs = sorted_seqs[:n] + self.assembled_seq_group.seqs = top_n_seqs + return self.assembled_seq_group + if self.output_produced: + return None + return None diff --git a/vllm/spec_decode/medusa_worker.py b/vllm/spec_decode/medusa_worker.py index 1ab691a7ef047..21a58fc426275 100644 --- a/vllm/spec_decode/medusa_worker.py +++ b/vllm/spec_decode/medusa_worker.py @@ -9,17 +9,15 @@ from vllm.spec_decode.interfaces import SpeculativeProposals from vllm.spec_decode.proposer_worker_base import NonLLMProposerWorkerBase from vllm.spec_decode.top1_proposer import Top1Proposer -from vllm.worker.worker_base import WorkerWrapperBase +from vllm.worker.worker_base import DelegateWorkerBase -class MedusaWorker(NonLLMProposerWorkerBase, WorkerWrapperBase): +class MedusaWorker(NonLLMProposerWorkerBase, DelegateWorkerBase): """Worker for Medusa. """ def __init__(self, *args, **kwargs): - super().__init__(kwargs.get("vllm_config")) - self.init_worker(*args, **kwargs) - + DelegateWorkerBase.__init__(self, *args, **kwargs) # Lazy initialization list. self._proposer: Top1Proposer diff --git a/vllm/spec_decode/multi_step_worker.py b/vllm/spec_decode/multi_step_worker.py index 676ac5eb3609d..32197f8cc8f2f 100644 --- a/vllm/spec_decode/multi_step_worker.py +++ b/vllm/spec_decode/multi_step_worker.py @@ -16,10 +16,10 @@ SpeculativeProposer) from vllm.spec_decode.proposer_worker_base import ProposerWorkerBase from vllm.spec_decode.top1_proposer import Top1Proposer -from vllm.worker.worker_base import WorkerWrapperBase +from vllm.worker.worker_base import DelegateWorkerBase -class MultiStepWorker(ProposerWorkerBase, WorkerWrapperBase): +class MultiStepWorker(ProposerWorkerBase, DelegateWorkerBase): """The MultiStepWorker is equivalent to a Worker except that it allows multiple forward passes in a single call, assuming the scheduler has allocated enough space to store the additional KV. This reduces overhead @@ -32,15 +32,12 @@ class MultiStepWorker(ProposerWorkerBase, WorkerWrapperBase): """ def __init__(self, *args, **kwargs): - super().__init__(kwargs.get("vllm_config")) - self.init_worker(*args, **kwargs) - + DelegateWorkerBase.__init__(self, *args, **kwargs) # Lazy initialization list. self._proposer: SpeculativeProposer def init_device(self) -> None: self.worker.init_device() - self._proposer = Top1Proposer( weakref.proxy(self), # type: ignore[arg-type] self.device, @@ -56,18 +53,6 @@ def set_should_modify_greedy_probs_inplace(self) -> None: self.model_runner.model.sampler.should_modify_greedy_probs_inplace = ( True) - def determine_num_available_blocks(self) -> Tuple[int, int]: - return self.worker.determine_num_available_blocks() - - def get_cache_block_size_bytes(self) -> int: - return self.worker.get_cache_block_size_bytes() - - def initialize_cache(self, *args, **kwargs) -> None: - self.worker.initialize_cache(*args, **kwargs) - - def execute_model(self, *args, **kwargs) -> List[SamplerOutput]: - return self.worker.execute_model(*args, **kwargs) - @torch.inference_mode() def sampler_output( self, diff --git a/vllm/spec_decode/ngram_worker.py b/vllm/spec_decode/ngram_worker.py index bb6b99135580e..e906b1789cde8 100644 --- a/vllm/spec_decode/ngram_worker.py +++ b/vllm/spec_decode/ngram_worker.py @@ -2,6 +2,7 @@ from typing import List, Optional, Set, Tuple import torch +import torch.nn as nn from vllm.model_executor.layers.sampler import SamplerOutput from vllm.sequence import ExecuteModelRequest @@ -10,6 +11,10 @@ from vllm.spec_decode.top1_proposer import Top1Proposer +class _DummyModel(nn.Module): + pass + + class NGramWorker(NonLLMProposerWorkerBase): """NGramWorker provides a light drafter without need for model. @@ -36,7 +41,6 @@ def set_ngram_window_size(self, ngram_prompt_lookup_min: int, def init_device(self): self.device = torch.device(f"{self.device_type}:{self.local_rank}") - self.load_model = lambda *args, **kwargs: None # Current NGramWorker only supports Top1Proposer self._proposer = Top1Proposer( @@ -45,6 +49,12 @@ def init_device(self): vocab_size=self.vocab_size, ) + def load_model(self) -> None: + pass # Dummy + + def get_model(self) -> nn.Module: + return _DummyModel() + def sampler_output( self, execute_model_req: ExecuteModelRequest, diff --git a/vllm/spec_decode/smaller_tp_proposer_worker.py b/vllm/spec_decode/smaller_tp_proposer_worker.py index 8896b7dbc6b8a..c6ff5e52f9388 100644 --- a/vllm/spec_decode/smaller_tp_proposer_worker.py +++ b/vllm/spec_decode/smaller_tp_proposer_worker.py @@ -1,6 +1,7 @@ from typing import List, Optional, Set, Tuple import torch +import torch.nn as nn from vllm.distributed.parallel_state import (get_tp_group, init_model_parallel_group, @@ -15,6 +16,10 @@ logger = init_logger(__name__) +class _DummyModel(nn.Module): + pass + + class SmallerTpProposerWorker(ProposerWorkerBase): """Class which allows a speculative draft model to run with smaller tensor parallel degree than target model. @@ -139,6 +144,13 @@ def get_spec_proposals( return self._worker.get_spec_proposals( execute_model_req, seq_ids_with_bonus_token_in_last_step) + def get_model(self) -> nn.Module: + if self._is_dummy: + return _DummyModel() + + with self._patch_tensor_parallel_group(): + return self._worker.get_model() + def execute_model( self, execute_model_req: Optional[ExecuteModelRequest] = None diff --git a/vllm/spec_decode/spec_decode_worker.py b/vllm/spec_decode/spec_decode_worker.py index e369da1a70c23..0d66ede3d907a 100644 --- a/vllm/spec_decode/spec_decode_worker.py +++ b/vllm/spec_decode/spec_decode_worker.py @@ -4,6 +4,7 @@ from typing import Any, Dict, List, Optional, Set, Tuple, Type import torch +import torch.nn as nn from vllm.config import ParallelConfig, SpeculativeConfig, VllmConfig from vllm.distributed.communication_op import broadcast_tensor_dict @@ -40,8 +41,8 @@ get_all_num_logprobs, get_sampled_token_logprobs, nvtx_range, split_batch_by_proposal_len) -from vllm.worker.worker_base import (LoraNotSupportedWorkerBase, WorkerBase, - WorkerWrapperBase) +from vllm.utils import resolve_obj_by_qualname +from vllm.worker.worker_base import LoraNotSupportedWorkerBase, WorkerBase logger = init_logger(__name__) @@ -64,8 +65,9 @@ def create_spec_worker(*args, **kwargs) -> "SpecDecodeWorker": target_worker_config = copy.deepcopy(vllm_config) target_worker_config.parallel_config.worker_cls =\ target_worker_config.parallel_config.sd_worker_cls - target_worker = WorkerWrapperBase(vllm_config=target_worker_config) - target_worker.init_worker(*args, **kwargs) + cls = resolve_obj_by_qualname( + target_worker_config.parallel_config.worker_cls) + target_worker = cls(*args, **kwargs) # Set the disable_logprobs variable in the TargetModelRunner instance # as per its value specified in the SpeculativeConfig. target_worker.model_runner.disable_logprobs =\ @@ -402,6 +404,9 @@ def initialize_cache(self, num_gpu_blocks: int, self.proposer_worker.initialize_cache(num_gpu_blocks=num_gpu_blocks, num_cpu_blocks=num_cpu_blocks) + def get_model(self) -> nn.Module: + return self.scorer_worker.get_model() + @torch.inference_mode() def execute_model( self, diff --git a/vllm/tracing.py b/vllm/tracing.py index 50068d8cf9c25..72a3f85118d36 100644 --- a/vllm/tracing.py +++ b/vllm/tracing.py @@ -16,7 +16,6 @@ OTEL_EXPORTER_OTLP_TRACES_PROTOCOL) from opentelemetry.sdk.trace import TracerProvider from opentelemetry.sdk.trace.export import BatchSpanProcessor - from opentelemetry.semconv_ai import SpanAttributes as BaseSpanAttributes from opentelemetry.trace import SpanKind, Tracer, set_tracer_provider from opentelemetry.trace.propagation.tracecontext import ( TraceContextTextMapPropagator) @@ -92,21 +91,30 @@ def extract_trace_headers(headers: Mapping[str, str]) -> Mapping[str, str]: return {h: headers[h] for h in TRACE_HEADERS if h in headers} -class SpanAttributes(BaseSpanAttributes): - # The following span attribute names are added here because they are missing - # from the Semantic Conventions for LLM. - LLM_REQUEST_ID = "gen_ai.request.id" - LLM_REQUEST_N = "gen_ai.request.n" - LLM_USAGE_NUM_SEQUENCES = "gen_ai.usage.num_sequences" - LLM_LATENCY_TIME_IN_QUEUE = "gen_ai.latency.time_in_queue" - LLM_LATENCY_TIME_TO_FIRST_TOKEN = "gen_ai.latency.time_to_first_token" - LLM_LATENCY_E2E = "gen_ai.latency.e2e" - LLM_LATENCY_TIME_IN_SCHEDULER = "gen_ai.latency.time_in_scheduler" +class SpanAttributes: + # Attribute names copied from here to avoid version conflicts: + # https://github.com/open-telemetry/semantic-conventions/blob/main/docs/gen-ai/gen-ai-spans.md + GEN_AI_USAGE_COMPLETION_TOKENS = "gen_ai.usage.completion_tokens" + GEN_AI_USAGE_PROMPT_TOKENS = "gen_ai.usage.prompt_tokens" + GEN_AI_REQUEST_MAX_TOKENS = "gen_ai.request.max_tokens" + GEN_AI_REQUEST_TOP_P = "gen_ai.request.top_p" + GEN_AI_REQUEST_TEMPERATURE = "gen_ai.request.temperature" + GEN_AI_RESPONSE_MODEL = "gen_ai.response.model" + # Attribute names added until they are added to the semantic conventions: + GEN_AI_REQUEST_ID = "gen_ai.request.id" + GEN_AI_REQUEST_N = "gen_ai.request.n" + GEN_AI_USAGE_NUM_SEQUENCES = "gen_ai.usage.num_sequences" + GEN_AI_LATENCY_TIME_IN_QUEUE = "gen_ai.latency.time_in_queue" + GEN_AI_LATENCY_TIME_TO_FIRST_TOKEN = "gen_ai.latency.time_to_first_token" + GEN_AI_LATENCY_E2E = "gen_ai.latency.e2e" + GEN_AI_LATENCY_TIME_IN_SCHEDULER = "gen_ai.latency.time_in_scheduler" # Time taken in the forward pass for this across all workers - LLM_LATENCY_TIME_IN_MODEL_FORWARD = "gen_ai.latency.time_in_model_forward" + GEN_AI_LATENCY_TIME_IN_MODEL_FORWARD = ( + "gen_ai.latency.time_in_model_forward") # Time taken in the model execute function. This will include model # forward, block/sync across workers, cpu-gpu sync time and sampling time. - LLM_LATENCY_TIME_IN_MODEL_EXECUTE = "gen_ai.latency.time_in_model_execute" + GEN_AI_LATENCY_TIME_IN_MODEL_EXECUTE = ( + "gen_ai.latency.time_in_model_execute") def contains_trace_headers(headers: Mapping[str, str]) -> bool: diff --git a/vllm/transformers_utils/configs/aria.py b/vllm/transformers_utils/configs/aria.py deleted file mode 100644 index d253da0d96a34..0000000000000 --- a/vllm/transformers_utils/configs/aria.py +++ /dev/null @@ -1,47 +0,0 @@ -from transformers.models.idefics2.configuration_idefics2 import ( - Idefics2VisionConfig) -from transformers.models.llama.configuration_llama import LlamaConfig - - -class AriaVisionConfig(Idefics2VisionConfig): - model_type = "aria_vision_model" - - -class AriaMoELMConfig(LlamaConfig): - """ - Configuration class for AriaMoE language model. - - This class extends the LlamaConfig to include additional parameters specific - to the Mixture of Experts (MoE) architecture. - """ - - model_type = "aria_moe_lm" - - def __init__( - self, - moe_intermediate_size: int = 4096, - moe_num_experts: int = 8, - moe_topk: int = 2, - moe_num_shared_experts: int = 2, - **kwargs, - ): - """ - Initialize the AriaMoELMConfig. - - Args: - moe_intermediate_size (int): The intermediate size for MoE layers. - Default is 4096. - moe_num_experts (int): The number of experts in the MoE layer. - Default is 8. - moe_topk (int): The number of top experts to route to for each - token. Default is 2. - moe_num_shared_experts (int): The number of shared experts. Default - is 2. - **kwargs: Additional keyword arguments to be passed to the parent - LlamaConfig. - """ - super().__init__(**kwargs) - self.moe_intermediate_size = moe_intermediate_size - self.moe_num_experts = moe_num_experts - self.moe_topk = moe_topk - self.moe_num_shared_experts = moe_num_shared_experts diff --git a/vllm/transformers_utils/processors/__init__.py b/vllm/transformers_utils/processors/__init__.py new file mode 100644 index 0000000000000..9c71b8cada32e --- /dev/null +++ b/vllm/transformers_utils/processors/__init__.py @@ -0,0 +1,4 @@ +from vllm.transformers_utils.processors.deepseek_vl2 import ( + DeepseekVLV2Processor) + +__all__ = ["DeepseekVLV2Processor"] diff --git a/vllm/transformers_utils/processors/deepseek_vl2.py b/vllm/transformers_utils/processors/deepseek_vl2.py new file mode 100644 index 0000000000000..27cdf6bc22d0e --- /dev/null +++ b/vllm/transformers_utils/processors/deepseek_vl2.py @@ -0,0 +1,361 @@ +# yapf: disable +# ruff: noqa: E501 +# coding=utf-8 +# adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/ff23960c5cf9e6874b44be38af930cfb0ccbb620/deepseek_vl2/models/processing_deepseek_vl_v2.py +# Copyright (c) 2023-2024 DeepSeek. +# +# Permission is hereby granted, free of charge, to any person obtaining a copy of +# this software and associated documentation files (the "Software"), to deal in +# the Software without restriction, including without limitation the rights to +# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of +# the Software, and to permit persons to whom the Software is furnished to do so, +# subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included in all +# copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS +# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR +# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER +# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN +# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + +import math +from typing import List, Tuple + +import torch +import torchvision.transforms as T +from PIL import Image, ImageOps +from transformers import AutoProcessor, BatchFeature, LlamaTokenizerFast +from transformers.processing_utils import ProcessorMixin + + +class ImageTransform: + + def __init__(self, + mean: Tuple[float, float, float] = (0.5, 0.5, 0.5), + std: Tuple[float, float, float] = (0.5, 0.5, 0.5), + normalize: bool = True): + self.mean = mean + self.std = std + self.normalize = normalize + + transform_pipelines = [T.ToTensor()] + + if normalize: + transform_pipelines.append(T.Normalize(mean, std)) + + self.transform = T.Compose(transform_pipelines) + + def __call__(self, pil_img: Image.Image): + x = self.transform(pil_img) + return x + + +class DeepseekVLV2Processor(ProcessorMixin): + tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast") + attributes = ["tokenizer"] + + def __init__( + self, + tokenizer: LlamaTokenizerFast, + candidate_resolutions: Tuple[Tuple[int, int]], + patch_size: int, + downsample_ratio: int, + image_mean: Tuple[float, float, float] = (0.5, 0.5, 0.5), + image_std: Tuple[float, float, float] = (0.5, 0.5, 0.5), + normalize: bool = True, + image_token: str = "", + pad_token: str = "<｜▁pad▁｜>", + add_special_token: bool = False, + sft_format: str = "deepseek", + mask_prompt: bool = True, + ignore_id: int = -100, + **kwargs, + ): + + self.candidate_resolutions = candidate_resolutions + self.image_size = candidate_resolutions[0][0] + self.patch_size = patch_size + self.image_mean = image_mean + self.image_std = image_std + self.normalize = normalize + self.downsample_ratio = downsample_ratio + + self.image_transform = ImageTransform(mean=image_mean, std=image_std, normalize=normalize) + self.tokenizer = tokenizer + self.tokenizer.padding_side = 'left' # must set this，padding side with make a difference in batch inference + + # add the pad_token as special token to use 'tokenizer.pad_token' and 'tokenizer.pad_token_id' + if tokenizer.pad_token is None: + self.tokenizer.add_special_tokens({'pad_token': pad_token}) + + # add image token + image_token_id = self.tokenizer.vocab.get(image_token) + if image_token_id is None: + special_tokens = [image_token] + special_tokens_dict = {"additional_special_tokens": special_tokens} + self.tokenizer.add_special_tokens(special_tokens_dict) + self.image_token_id = self.tokenizer.vocab.get(image_token) + + # add five special tokens for grounding-related tasks + # <|ref|>, <|/ref|>, <|det|>, <|/det|>, <|grounding|> + special_tokens = ['<|ref|>', '<|/ref|>', '<|det|>', '<|/det|>', '<|grounding|>'] + special_tokens_dict = {"additional_special_tokens": special_tokens} + self.tokenizer.add_special_tokens(special_tokens_dict) + + # add special tokens for SFT data + special_tokens = ["<|User|>", "<|Assistant|>"] + special_tokens_dict = {"additional_special_tokens": special_tokens} + self.tokenizer.add_special_tokens(special_tokens_dict) + + self.image_token = image_token + self.pad_token = pad_token + self.add_special_token = add_special_token + self.sft_format = sft_format + self.mask_prompt = mask_prompt + self.ignore_id = ignore_id + + super().__init__( + tokenizer, + **kwargs, + ) + + def select_best_resolution(self, image_size): + # used for cropping + original_width, original_height = image_size + best_fit = None + max_effective_resolution = 0 + min_wasted_resolution = float("inf") + + for width, height in self.candidate_resolutions: + scale = min(width / original_width, height / original_height) + downscaled_width, downscaled_height = int( + original_width * scale), int(original_height * scale) + effective_resolution = min(downscaled_width * downscaled_height, + original_width * original_height) + wasted_resolution = (width * height) - effective_resolution + + if effective_resolution > max_effective_resolution or ( + effective_resolution == max_effective_resolution + and wasted_resolution < min_wasted_resolution): + max_effective_resolution = effective_resolution + min_wasted_resolution = wasted_resolution + best_fit = (width, height) + + return best_fit + + @property + def bos_id(self): + return self.tokenizer.bos_token_id + + @property + def eos_id(self): + return self.tokenizer.eos_token_id + + @property + def pad_id(self): + return self.tokenizer.pad_token_id + + def encode(self, text: str, bos: bool = True, eos: bool = False): + t = self.tokenizer.encode(text, add_special_tokens=False) + + if bos: + t = [self.bos_id] + t + if eos: + t = t + [self.eos_id] + + return t + + def decode(self, t: List[int], **kwargs) -> str: + return self.tokenizer.decode(t, **kwargs) + + def process_one( + self, + prompt: str, + images: List[Image.Image], + inference_mode: bool = True, + **kwargs, + ): + """ + + Args: + prompt (str): the formatted prompt; + conversations (List[Dict]): conversations with a list of messages; + images (List[ImageType]): the list of images; + inference_mode (bool): if True, then remove the last eos token; + system_prompt (str): the system prompt; + **kwargs: + + Returns: + outputs (BaseProcessorOutput): the output of the processor, + - input_ids (torch.LongTensor): [N + image tokens] + - target_ids (torch.LongTensor): [N + image tokens] + - pixel_values (torch.FloatTensor): [n_patches, 3, H, W] + - image_id (int): the id of the image token + - num_image_tokens (List[int]): the number of image tokens + """ + + assert (prompt is not None and images is not None + ), "prompt and images must be used at the same time." + + sft_format = prompt + tokenized_str, images_list, images_seq_mask, images_spatial_crop, num_image_tokens = self.tokenize_with_images( + sft_format, images, bos=True, eos=True, cropping=len(images) <= 2) + masked_tokenized_str = [] + for token_index in tokenized_str: + if token_index != self.image_token_id: + masked_tokenized_str.append(token_index) + else: + masked_tokenized_str.append(self.ignore_id) + + assert len(tokenized_str) == len(images_seq_mask) == len(masked_tokenized_str), \ + (f"tokenized_str's length {len(tokenized_str)}, input_ids' length {len(masked_tokenized_str)}, " + f"imags_seq_mask's length {len(images_seq_mask)}, are not equal") + + input_ids = torch.LongTensor(tokenized_str) + target_ids = torch.LongTensor(masked_tokenized_str) + images_seq_mask = torch.tensor(images_seq_mask, dtype=torch.bool) + + # set input_ids < 0 | input_ids == self.image_token_id as ignore_id + target_ids[(input_ids < 0) | + (input_ids == self.image_token_id)] = self.ignore_id + input_ids[input_ids < 0] = self.pad_id + + if inference_mode: + # 去掉结尾的eos token + assert input_ids[-1] == self.eos_id + input_ids = input_ids[:-1] + target_ids = target_ids[:-1] + images_seq_mask = images_seq_mask[:-1] + + if len(images_list) == 0: + pixel_values = torch.zeros((1, 3, self.image_size, self.image_size)) + images_spatial_crop = torch.zeros((1, 2), dtype=torch.long) + else: + pixel_values = torch.stack(images_list, dim=0) + images_spatial_crop = torch.tensor(images_spatial_crop, dtype=torch.long) + + input_ids = input_ids.unsqueeze(0) + + prepare = BatchFeature( + data=dict( + input_ids=input_ids, + pixel_values=pixel_values, + images_seq_mask=images_seq_mask, + images_spatial_crop=images_spatial_crop, + num_image_tokens=num_image_tokens, + ), + tensor_type="pt", + ) + return prepare + + def __call__( + self, + *, + prompt: str, + images: List[Image.Image], + inference_mode: bool = True, + **kwargs, + ): + """ + + Args: + prompt (str): the formatted prompt; + images (List[ImageType]): the list of images; + inference_mode (bool): if True, then remove the last eos token; + **kwargs: + + Returns: + outputs (BaseProcessorOutput): the output of the processor, + - input_ids (torch.LongTensor): [N + image tokens] + - images (torch.FloatTensor): [n_images, 3, H, W] + - image_id (int): the id of the image token + - num_image_tokens (List[int]): the number of image tokens + """ + + prepare = self.process_one( + prompt=prompt, + images=images, + inference_mode=inference_mode, + ) + + return prepare + + def tokenize_with_images( + self, + conversation: str, + images: List[Image.Image], + bos: bool = True, + eos: bool = True, + cropping: bool = True, + ): + """Tokenize text with tags.""" + assert conversation.count(self.image_token) == len(images) + text_splits = conversation.split(self.image_token) + images_list, images_seq_mask, images_spatial_crop = [], [], [] + num_image_tokens = [] + tokenized_str = [] + for text_sep, image in zip(text_splits, images): + """encode text_sep""" + tokenized_sep = self.encode(text_sep, bos=False, eos=False) + tokenized_str += tokenized_sep + images_seq_mask += [False] * len(tokenized_sep) + + """select best resolution for anyres""" + if cropping: + best_width, best_height = self.select_best_resolution(image.size) + else: + best_width, best_height = self.image_size, self.image_size + + """process the global view""" + global_view = ImageOps.pad(image, (self.image_size, self.image_size), + color=tuple(int(x * 255) for x in self.image_transform.mean)) + images_list.append(self.image_transform(global_view)) + + """process the local views""" + local_view = ImageOps.pad(image, (best_width, best_height), + color=tuple(int(x * 255) for x in self.image_transform.mean)) + for i in range(0, best_height, self.image_size): + for j in range(0, best_width, self.image_size): + images_list.append( + self.image_transform(local_view.crop((j, i, j + self.image_size, i + self.image_size)))) + + """record height / width crop num""" + num_width_tiles, num_height_tiles = best_width // self.image_size, best_height // self.image_size + images_spatial_crop.append([num_width_tiles, num_height_tiles]) + + """add image tokens""" + h = w = math.ceil((self.image_size // self.patch_size) / self.downsample_ratio) + # global views tokens h * (w + 1), 1 is for line separator + tokenized_image = [self.image_token_id] * h * (w + 1) + # add a separator between global and local views + tokenized_image += [self.image_token_id] + # local views tokens, (num_height_tiles * h) * (num_width_tiles * w + 1) + tokenized_image += [self.image_token_id] * (num_height_tiles * h) * (num_width_tiles * w + 1) + + tokenized_str += tokenized_image + images_seq_mask += [True] * len(tokenized_image) + num_image_tokens.append(len(tokenized_image)) + + """process the last text split""" + tokenized_sep = self.encode(text_splits[-1], bos=False, eos=False) + tokenized_str += tokenized_sep + images_seq_mask += [False] * len(tokenized_sep) + + """add the bos and eos tokens""" + if bos: + tokenized_str = [self.bos_id] + tokenized_str + images_seq_mask = [False] + images_seq_mask + if eos: + tokenized_str = tokenized_str + [self.eos_id] + images_seq_mask = images_seq_mask + [False] + + assert len(tokenized_str) == len( + images_seq_mask), f"tokenize_with_images func: tokenized_str's length {len(tokenized_str)} is not equal to imags_seq_mask's length {len(images_seq_mask)}" + + return tokenized_str, images_list, images_seq_mask, images_spatial_crop, num_image_tokens + + +AutoProcessor.register("DeepseekVLV2Processor", DeepseekVLV2Processor) diff --git a/vllm/transformers_utils/s3_utils.py b/vllm/transformers_utils/s3_utils.py index 6ae68161bbd97..74a56cbf57ec3 100644 --- a/vllm/transformers_utils/s3_utils.py +++ b/vllm/transformers_utils/s3_utils.py @@ -145,7 +145,8 @@ def pull_files(self, return for file in files: - destination_file = self.dir + file.removeprefix(base_dir) + destination_file = os.path.join(self.dir, + file.removeprefix(base_dir)) local_dir = Path(destination_file).parent os.makedirs(local_dir, exist_ok=True) self.s3.download_file(bucket_name, file, destination_file) diff --git a/vllm/transformers_utils/tokenizer.py b/vllm/transformers_utils/tokenizer.py index 294262484f2fb..1f1d67fabb243 100644 --- a/vllm/transformers_utils/tokenizer.py +++ b/vllm/transformers_utils/tokenizer.py @@ -67,9 +67,10 @@ def get_cached_tokenizer(tokenizer: AnyTokenizer) -> AnyTokenizer: tokenizer_all_special_tokens_extended = ( tokenizer.all_special_tokens_extended) tokenizer_all_special_tokens = set(tokenizer.all_special_tokens) + tokenizer_vocab = tokenizer.get_vocab() tokenizer_len = len(tokenizer) - max_token_id = max(tokenizer.get_vocab().values()) + max_token_id = max(tokenizer_vocab.values()) # Some tokenizers (e.g., QwenTokenizer) have special tokens that # are added and included in the implementation of the vocab_size # property, but not in get_vocab(); if there is an implementation @@ -96,6 +97,9 @@ def all_special_tokens_extended(self): def max_token_id(self): return max_token_id + def get_vocab(self): + return tokenizer_vocab + def __len__(self): return tokenizer_len diff --git a/vllm/transformers_utils/tokenizer_group/__init__.py b/vllm/transformers_utils/tokenizer_group/__init__.py index d400276796996..09569c564a58d 100644 --- a/vllm/transformers_utils/tokenizer_group/__init__.py +++ b/vllm/transformers_utils/tokenizer_group/__init__.py @@ -24,7 +24,8 @@ def init_tokenizer_from_configs(model_config: ModelConfig, max_input_length=None, tokenizer_mode=model_config.tokenizer_mode, trust_remote_code=model_config.trust_remote_code, - revision=model_config.tokenizer_revision) + revision=model_config.tokenizer_revision, + truncation_side=model_config.truncation_side) return get_tokenizer_group(parallel_config.tokenizer_pool_config, **init_kwargs) diff --git a/vllm/transformers_utils/tokenizers/mistral.py b/vllm/transformers_utils/tokenizers/mistral.py index 17d722e3d88fe..d801cf4e4c7b1 100644 --- a/vllm/transformers_utils/tokenizers/mistral.py +++ b/vllm/transformers_utils/tokenizers/mistral.py @@ -18,6 +18,7 @@ Tekkenizer) from vllm.logger import init_logger +from vllm.utils import is_list_of if TYPE_CHECKING: from vllm.entrypoints.chat_utils import ChatCompletionMessageParam @@ -27,7 +28,7 @@ @dataclass class Encoding: - input_ids: List[int] + input_ids: Union[List[int], List[List[int]]] def maybe_serialize_tool_calls(request: ChatCompletionRequest): @@ -223,17 +224,25 @@ def __len__(self) -> int: def __call__( self, - prompt: str, + prompt: Union[str, List[str], List[int]], add_special_tokens: bool = False, truncation: bool = False, max_length: Optional[int] = None, ): - # Mistral Tokenizers should not add special tokens - input_ids = self.encode(prompt) - - if truncation: - input_ids = input_ids[:max_length] - + input_ids: Union[List[int], List[List[int]]] + # For List[str], original prompt text + if is_list_of(prompt, str): + input_ids_: List[List[int]] = [] + for p in prompt: + each_input_ids = self.encode_one(p, truncation, max_length) + input_ids_.append(each_input_ids) + input_ids = input_ids_ + # For List[int], apply chat template output, already tokens. + elif is_list_of(prompt, int): + input_ids = prompt + # For str, single prompt text + else: + input_ids = self.encode_one(prompt, truncation, max_length) return Encoding(input_ids=input_ids) def get_vocab(self) -> Dict[str, int]: @@ -245,6 +254,19 @@ def get_added_vocab(self) -> Dict[str, int]: # Mistral tokenizers have no added vocabulary return {} + def encode_one( + self, + prompt: str, + truncation: bool = False, + max_length: Optional[int] = None, + ) -> List[int]: + # Mistral Tokenizers should not add special tokens + input_ids = self.encode(prompt) + + if truncation: + input_ids = input_ids[:max_length] + return input_ids + def encode(self, prompt: str) -> List[int]: # `encode` should only be used for prompt completion # it should never be used for chat_completion. diff --git a/vllm/usage/usage_lib.py b/vllm/usage/usage_lib.py index a9deee881f41a..7f5cc906382af 100644 --- a/vllm/usage/usage_lib.py +++ b/vllm/usage/usage_lib.py @@ -27,6 +27,17 @@ _GLOBAL_RUNTIME_DATA: Dict[str, Union[str, int, bool]] = {} +_USAGE_ENV_VARS_TO_COLLECT = [ + "VLLM_USE_MODELSCOPE", + "VLLM_USE_TRITON_FLASH_ATTN", + "VLLM_ATTENTION_BACKEND", + "VLLM_USE_FLASHINFER_SAMPLER", + "VLLM_PP_LAYER_PARTITION", + "VLLM_USE_TRITON_AWQ", + "VLLM_USE_V1", + "VLLM_ENABLE_V1_MULTIPROCESSING", +] + def set_runtime_usage_data(key: str, value: Union[str, int, bool]) -> None: """Set global usage data that will be sent with every usage heartbeat.""" @@ -119,9 +130,11 @@ def __init__(self) -> None: self.total_memory: Optional[int] = None self.architecture: Optional[str] = None self.platform: Optional[str] = None + self.cuda_runtime: Optional[str] = None self.gpu_count: Optional[int] = None self.gpu_type: Optional[str] = None self.gpu_memory_per_device: Optional[int] = None + self.env_var_json: Optional[str] = None # vLLM Information self.model_architecture: Optional[str] = None @@ -157,6 +170,8 @@ def _report_usage_once(self, model_architecture: str, self.gpu_count = torch.cuda.device_count() self.gpu_type = device_property.name self.gpu_memory_per_device = device_property.total_memory + if current_platform.is_cuda(): + self.cuda_runtime = torch.version.cuda self.provider = _detect_cloud_provider() self.architecture = platform.machine() self.platform = platform.platform() @@ -176,6 +191,12 @@ def _report_usage_once(self, model_architecture: str, self.vllm_version = VLLM_VERSION self.model_architecture = model_architecture + # Environment variables + self.env_var_json = json.dumps({ + env_var: getattr(envs, env_var) + for env_var in _USAGE_ENV_VARS_TO_COLLECT + }) + # Metadata self.log_time = _get_current_timestamp_ns() self.source = envs.VLLM_USAGE_SOURCE diff --git a/vllm/utils.py b/vllm/utils.py index ad02c68d16276..f9bdadbf9dcd8 100644 --- a/vllm/utils.py +++ b/vllm/utils.py @@ -36,6 +36,7 @@ overload) from uuid import uuid4 +import cloudpickle import numpy as np import numpy.typing as npt import psutil @@ -282,6 +283,37 @@ def inner(*args, **kwds): return func +class rpd_user_marker: + + def __init__(self, name=None): + self.name = name + self.marker = None + + def __enter__(self): + if is_hipScopedMarker_available(): + from hipScopedMarker import hipScopedMarker + marker_name = self.name if self.name else "UserMarker Undefined" + self.marker = hipScopedMarker(f"{marker_name}") + self.marker.__enter__() + return self + + def __exit__(self, exc_type, exc_val, exc_tb): + if is_hipScopedMarker_available() and self.marker: + self.marker.__exit__(exc_type, exc_val, exc_tb) + + def start(self): + if is_hipScopedMarker_available(): + from hipScopedMarker import hipScopedMarker + marker_name = self.name if self.name else "UserMarker Undefined" + self.marker = hipScopedMarker(f"{marker_name}") + self.marker.__enter__() + return self + + def end(self, exc_type=0, exc_val=0, exc_tb=0): + if is_hipScopedMarker_available() and self.marker: + self.marker.__exit__(exc_type, exc_val, exc_tb) + + class Device(enum.Enum): GPU = enum.auto() CPU = enum.auto() @@ -828,13 +860,7 @@ def __init__(self, device: Optional[torch.types.Device] = None): def current_memory_usage(self) -> float: # Return the memory usage in bytes. from vllm.platforms import current_platform - if current_platform.is_cuda_alike(): - torch.cuda.reset_peak_memory_stats(self.device) - mem = torch.cuda.max_memory_allocated(self.device) - elif current_platform.is_xpu(): - torch.xpu.reset_peak_memory_stats(self.device) # type: ignore - mem = torch.xpu.max_memory_allocated(self.device) # type: ignore - return mem + return current_platform.get_current_memory_usage(self.device) def __enter__(self): self.initial_memory = self.current_memory_usage() @@ -2068,36 +2094,57 @@ def kill_process_tree(pid: int): @dataclass class MemorySnapshot: """Memory snapshot.""" - torch_peak_in_bytes: int = 0 - torch_memory_in_bytes: int = 0 + torch_peak: int = 0 + cuda_memory: int = 0 + torch_memory: int = 0 + non_torch_memory: int = 0 timestamp: float = 0.0 + auto_measure: bool = True + + def __post_init__(self): + if self.auto_measure: + self.measure() def measure(self): - self.torch_peak_in_bytes = torch.cuda.memory_stats( - )["allocated_bytes.all.peak"] - self.torch_memory_in_bytes = torch.cuda.memory_stats( - )["allocated_bytes.all.current"] + # we measure the torch peak memory usage via allocated_bytes, + # rather than `torch.cuda.memory_reserved()` . + # After `torch.cuda.reset_peak_memory_stats()`, + # `torch.cuda.memory_reserved()` will keep growing, and only shrink + # when we call `torch.cuda.empty_cache()` or OOM happens. + self.torch_peak = torch.cuda.memory_stats().get( + "allocated_bytes.all.peak", 0) + + self.cuda_memory = torch.cuda.mem_get_info( + )[1] - torch.cuda.mem_get_info()[0] + + # torch.cuda.memory_reserved() is how many bytes + # PyTorch gets from cuda (by calling cudaMalloc, etc.) + # this is used to measure the non-torch memory usage + self.torch_memory = torch.cuda.memory_reserved() + + self.non_torch_memory = self.cuda_memory - self.torch_memory self.timestamp = time.time() def __sub__(self, other: "MemorySnapshot") -> "MemorySnapshot": - """support a - b""" return MemorySnapshot( - torch_peak_in_bytes=self.torch_peak_in_bytes - - other.torch_peak_in_bytes, - torch_memory_in_bytes=self.torch_memory_in_bytes - - other.torch_memory_in_bytes, - timestamp=self.timestamp - other.timestamp) + torch_peak=self.torch_peak - other.torch_peak, + cuda_memory=self.cuda_memory - other.cuda_memory, + torch_memory=self.torch_memory - other.torch_memory, + non_torch_memory=self.non_torch_memory - other.non_torch_memory, + timestamp=self.timestamp - other.timestamp, + auto_measure=False, + ) @dataclass class MemoryProfilingResult: - """Memory profiling result. - """ # noqa - baseline_memory_in_bytes: int = 0 - non_kv_cache_memory_in_bytes: int = 0 - torch_peak_increase_in_bytes: int = 0 - non_torch_increase_in_bytes: int = 0 - weights_memory_in_bytes: float = 0 + """Memory profiling result. All numbers are in bytes. + """ + non_kv_cache_memory: int = 0 + torch_peak_increase: int = 0 + non_torch_increase: int = 0 + weights_memory: float = 0 + before_create: MemorySnapshot = field(default_factory=MemorySnapshot) before_profile: MemorySnapshot = field(default_factory=MemorySnapshot) after_profile: MemorySnapshot = field(default_factory=MemorySnapshot) profile_time: float = 0.0 @@ -2105,18 +2152,14 @@ class MemoryProfilingResult: @contextlib.contextmanager def memory_profiling( - baseline_memory_in_bytes: int, weights_memory_in_bytes: int -) -> Generator[MemoryProfilingResult, None, None]: + baseline_snapshot: MemorySnapshot, + weights_memory: int) -> Generator[MemoryProfilingResult, None, None]: """Memory profiling context manager. - baseline_memory_in_bytes: memory used by all the components other than - the current vLLM instance. It contains: memory used by other processes, memory - used by another vLLM instance in the same process, etc. It is usually measured - before the current vLLM instance initialize the device. And we assume it is - constant during the profiling of the current vLLM instance. - weights_memory_in_bytes: memory used by PyTorch when loading the model weights. + baseline_snapshot: the memory snapshot before the current vLLM instance. + weights_memory: memory used by PyTorch when loading the model weights. Note that, before loading the model weights, we also initialize the device and distributed environment, which may consume some memory. This part is not - included in the weights_memory_in_bytes because PyTorch does not control it. + included in the weights_memory because PyTorch does not control it. The memory in one GPU can be classified into 3 categories: 1. memory used by anything other than the current vLLM instance. @@ -2151,20 +2194,21 @@ def memory_profiling( b. 2 GiB reserved for the peak activation tensors (category 2) c. 1 GiB used by non-torch components (category 3) - The memory used for loading weights (a.) is directly given from the argument `weights_memory_in_bytes`. + The memory used for loading weights (a.) is directly given from the argument `weights_memory`. - The increase of ``torch.cuda.memory_stats()["allocated_bytes.all.peak"]` after profiling gives (b.). + The increase of `torch.cuda.memory_stats()["allocated_bytes.all.peak"]` during profiling gives (b.). - (c.) is tricky. We measure the total memory used in this GPU (`torch.cuda.mem_get_info()[1] - torch.cuda.mem_get_info()[0]`), - subtract the baseline memory, the memory used by the model weights, and diff of `torch.cuda.memory_stats()["allocated_bytes.all.current"]`. + The increase of `non_torch_memory` from creating the current vLLM instance until after profiling to get (c.). """ # noqa + gc.collect() + torch.cuda.empty_cache() torch.cuda.reset_peak_memory_stats() result = MemoryProfilingResult() - result.baseline_memory_in_bytes = baseline_memory_in_bytes + result.before_create = baseline_snapshot # the part of memory used for holding the model weights - result.weights_memory_in_bytes = weights_memory_in_bytes + result.weights_memory = weights_memory result.before_profile.measure() @@ -2175,13 +2219,12 @@ def memory_profiling( result.after_profile.measure() - diff = result.after_profile - result.before_profile - result.torch_peak_increase_in_bytes = diff.torch_peak_in_bytes - current_cuda_memory_bytes = torch.cuda.mem_get_info( - )[1] - torch.cuda.mem_get_info()[0] - result.non_torch_increase_in_bytes = current_cuda_memory_bytes - baseline_memory_in_bytes - weights_memory_in_bytes - diff.torch_memory_in_bytes # noqa - result.profile_time = diff.timestamp - result.non_kv_cache_memory_in_bytes = result.non_torch_increase_in_bytes + result.torch_peak_increase_in_bytes + result.weights_memory_in_bytes # noqa + diff_profile = result.after_profile - result.before_profile + diff_from_create = result.after_profile - result.before_create + result.torch_peak_increase = diff_profile.torch_peak + result.non_torch_increase = diff_from_create.non_torch_memory + result.profile_time = diff_profile.timestamp + result.non_kv_cache_memory = result.non_torch_increase + result.torch_peak_increase + result.weights_memory # noqa # Adapted from: https://github.com/sgl-project/sglang/blob/v0.4.1/python/sglang/srt/utils.py#L630 # noqa: E501 @@ -2312,3 +2355,25 @@ def bind_kv_cache( assert len(forward_ctx.kv_cache) == len(kv_cache) for ve, ve_kv_cache in enumerate(kv_cache): forward_ctx.kv_cache[ve] = ve_kv_cache[kv_cache_idx] + + +def run_method(obj: Any, method: Union[str, bytes, Callable], args: Tuple[Any], + kwargs: Dict[str, Any]) -> Any: + """ + Run a method of an object with the given arguments and keyword arguments. + If the method is string, it will be converted to a method using getattr. + If the method is serialized bytes and will be deserialized using + cloudpickle. + If the method is a callable, it will be called directly. + """ + if isinstance(method, bytes): + func = partial(cloudpickle.loads(method), obj) + elif isinstance(method, str): + try: + func = getattr(obj, method) + except AttributeError: + raise NotImplementedError(f"Method {method!r} is not" + " implemented.") from None + else: + func = partial(method, obj) # type: ignore + return func(*args, **kwargs) diff --git a/vllm/v1/attention/backends/flash_attn.py b/vllm/v1/attention/backends/flash_attn.py old mode 100644 new mode 100755 index b02bc9ffde538..4730ffd291e4b --- a/vllm/v1/attention/backends/flash_attn.py +++ b/vllm/v1/attention/backends/flash_attn.py @@ -9,12 +9,21 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, AttentionMetadata, AttentionType) +from vllm.envs import VLLM_FLASH_ATTN_VERSION +from vllm.logger import init_logger +from vllm.platforms import current_platform from vllm.utils import cdiv -from vllm.vllm_flash_attn import flash_attn_varlen_func +from vllm.vllm_flash_attn import (fa_version_unsupported_reason, + flash_attn_varlen_func, + is_fa_version_supported) + +logger = init_logger(__name__) class FlashAttentionBackend(AttentionBackend): + accept_output_buffer: bool = True + @staticmethod def get_supported_head_sizes() -> List[int]: return [32, 64, 96, 128, 160, 192, 224, 256] @@ -61,7 +70,7 @@ class FlashAttentionMetadata: max_query_len: int query_start_loc: torch.Tensor max_seq_len: int - seq_start_loc: torch.Tensor + seq_lens: torch.Tensor block_table: torch.Tensor slot_mapping: torch.Tensor @@ -69,8 +78,8 @@ class FlashAttentionMetadata: use_cascade: bool common_prefix_len: int cu_prefix_query_lens: Optional[torch.Tensor] - cu_prefix_kv_lens: Optional[torch.Tensor] - cu_suffix_kv_lens: Optional[torch.Tensor] + prefix_kv_lens: Optional[torch.Tensor] + suffix_kv_lens: Optional[torch.Tensor] # For logging. num_input_tokens: int = 0 # Number of tokens including padding. @@ -126,15 +135,33 @@ def __init__( "are not implemented for " "FlashAttentionImpl") + # if hopper default to FA3, otherwise stick to FA2 for now + # TODO(lucas): profile FA3 on ampere to see if it makes sense to + # use FA3 as default for both + if current_platform.get_device_capability()[0] >= 9: + self.fa_version = 3 if is_fa_version_supported(3) else 2 + else: + self.fa_version = 2 + + if VLLM_FLASH_ATTN_VERSION is not None: + assert VLLM_FLASH_ATTN_VERSION in [2, 3] + self.fa_version = VLLM_FLASH_ATTN_VERSION + + if not is_fa_version_supported(self.fa_version): + logger.error("Cannot use FA version %d is not supported due to %s", + self.fa_version, + fa_version_unsupported_reason(self.fa_version)) + + assert is_fa_version_supported(self.fa_version) + def forward( self, + layer: torch.nn.Module, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, kv_cache: torch.Tensor, attn_metadata: FlashAttentionMetadata, - k_scale: float = 1.0, - v_scale: float = 1.0, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: """Forward pass with FlashAttention. @@ -148,9 +175,7 @@ def forward( Returns: shape = [num_tokens, num_heads * head_size] """ - # NOTE(woosuk): FlashAttention does not support FP8 KV cache. - assert k_scale == 1.0 and v_scale == 1.0, ( - "key/v_scale is not supported in FlashAttention.") + assert output is not None, "Output tensor must be provided." assert output is not None, "Output tensor must be provided." @@ -181,8 +206,8 @@ def forward( value_cache, attn_metadata.slot_mapping, self.kv_cache_dtype, - k_scale, - v_scale, + layer._k_scale, + layer._v_scale, ) # Compute attention and update output up to `num_actual_tokens`. @@ -195,7 +220,7 @@ def forward( out=output[:num_actual_tokens], cu_seqlens_q=attn_metadata.query_start_loc, max_seqlen_q=attn_metadata.max_query_len, - cu_seqlens_k=attn_metadata.seq_start_loc, + seqused_k=attn_metadata.seq_lens, max_seqlen_k=attn_metadata.max_seq_len, softmax_scale=self.scale, causal=True, @@ -203,6 +228,7 @@ def forward( window_size=self.sliding_window, block_table=attn_metadata.block_table, softcap=self.logits_soft_cap, + fa_version=self.fa_version, ) return output @@ -215,8 +241,8 @@ def forward( cu_query_lens=attn_metadata.query_start_loc, max_query_len=attn_metadata.max_query_len, cu_prefix_query_lens=attn_metadata.cu_prefix_query_lens, - cu_prefix_kv_lens=attn_metadata.cu_prefix_kv_lens, - cu_suffix_kv_lens=attn_metadata.cu_suffix_kv_lens, + prefix_kv_lens=attn_metadata.prefix_kv_lens, + suffix_kv_lens=attn_metadata.suffix_kv_lens, max_kv_len=attn_metadata.max_seq_len, softmax_scale=self.scale, alibi_slopes=self.alibi_slopes, @@ -224,6 +250,7 @@ def forward( logits_soft_cap=self.logits_soft_cap, block_table=attn_metadata.block_table, common_prefix_len=attn_metadata.common_prefix_len, + fa_version=self.fa_version, ) return output @@ -304,8 +331,8 @@ def cascade_attention( cu_query_lens: torch.Tensor, max_query_len: int, cu_prefix_query_lens: torch.Tensor, - cu_prefix_kv_lens: torch.Tensor, - cu_suffix_kv_lens: torch.Tensor, + prefix_kv_lens: torch.Tensor, + suffix_kv_lens: torch.Tensor, max_kv_len: int, softmax_scale: float, alibi_slopes: Optional[torch.Tensor], @@ -313,6 +340,7 @@ def cascade_attention( logits_soft_cap: float, block_table: torch.Tensor, common_prefix_len: int, + fa_version: int, ) -> torch.Tensor: assert alibi_slopes is None, ("Cascade attention does not support ALiBi.") # TODO: Support sliding window. @@ -331,7 +359,7 @@ def cascade_attention( k=key_cache, v=value_cache, cu_seqlens_q=cu_prefix_query_lens, - cu_seqlens_k=cu_prefix_kv_lens, + seqused_k=prefix_kv_lens, max_seqlen_q=num_tokens, max_seqlen_k=common_prefix_len, softmax_scale=softmax_scale, @@ -340,6 +368,7 @@ def cascade_attention( block_table=block_table[:1], softcap=logits_soft_cap, return_softmax_lse=True, + fa_version=fa_version, ) # Process suffix per query. @@ -348,7 +377,7 @@ def cascade_attention( k=key_cache, v=value_cache, cu_seqlens_q=cu_query_lens, - cu_seqlens_k=cu_suffix_kv_lens, + seqused_k=suffix_kv_lens, max_seqlen_q=max_query_len, max_seqlen_k=max_kv_len - common_prefix_len, softmax_scale=softmax_scale, @@ -357,6 +386,7 @@ def cascade_attention( block_table=block_table[:, num_common_kv_blocks:], softcap=logits_soft_cap, return_softmax_lse=True, + fa_version=fa_version, ) # Merge prefix and suffix outputs, and store the result in output. diff --git a/vllm/v1/core/encoder_cache_manager.py b/vllm/v1/core/encoder_cache_manager.py index 845bd5ea05e3c..0cd8c806a3e47 100644 --- a/vllm/v1/core/encoder_cache_manager.py +++ b/vllm/v1/core/encoder_cache_manager.py @@ -1,7 +1,14 @@ -from typing import Dict, List, Set, Tuple +from typing import TYPE_CHECKING, Dict, List, Set, Tuple +from vllm.logger import init_logger +from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.v1.request import Request +if TYPE_CHECKING: + from vllm.config import ModelConfig, SchedulerConfig + +logger = init_logger(__name__) + class EncoderCacheManager: @@ -46,3 +53,72 @@ def get_freed_ids(self) -> List[Tuple[str, int]]: freed = self.freed self.freed = [] return freed + + +def compute_encoder_budget( + model_config: "ModelConfig", + scheduler_config: "SchedulerConfig", +) -> Tuple[int, int]: + """Compute the encoder cache budget based on the model and scheduler + configurations. + + Args: + model_config: Model configuration. + scheduler_config: Scheduler configuration. + + Returns: + - Compute budget for encoder execution, in unit of number of tokens + in the input sequence. + - Space budget for encoder cache size, in unit of number of tokens + in the input sequence. + """ + + if not model_config.is_multimodal_model: + return 0, 0 + + # TODO: handle encoder-decoder models once we support them. + ( + encoder_compute_budget, + encoder_cache_size, + ) = _compute_encoder_budget_multimodal(model_config, scheduler_config) + + return encoder_compute_budget, encoder_cache_size + + +def _compute_encoder_budget_multimodal( + model_config: "ModelConfig", + scheduler_config: "SchedulerConfig", +) -> Tuple[int, int]: + """Compute the encoder cache budget based on the model and scheduler + configurations for a multimodal model. + + Args: + model_config: Model configuration. + scheduler_config: Scheduler configuration. + + Returns: + - Compute budget for encoder execution, in unit of number of tokens + in the input sequence. + - Space budget for encoder cache size, in unit of number of tokens + in the input sequence. + """ + + max_tokens_by_modality_dict = MULTIMODAL_REGISTRY.get_max_tokens_per_item_by_nonzero_modality( # noqa: E501 + model_config) + + if not max_tokens_by_modality_dict: + logger.warning( + "All non-text modalities supported by the model have been " + "explicitly disabled via limit_mm_per_prompt. Encoder cache will " + "not be initialized.") + return 0, 0 + + _, max_tokens_per_mm_item = max(max_tokens_by_modality_dict.items(), + key=lambda item: item[1]) + + encoder_compute_budget = max(scheduler_config.max_num_encoder_input_tokens, + max_tokens_per_mm_item) + encoder_cache_size = max(scheduler_config.encoder_cache_size, + max_tokens_per_mm_item) + + return encoder_compute_budget, encoder_cache_size diff --git a/vllm/v1/core/kv_cache_manager.py b/vllm/v1/core/kv_cache_manager.py index 1cbff1e2d767e..18fdfdfe4a010 100644 --- a/vllm/v1/core/kv_cache_manager.py +++ b/vllm/v1/core/kv_cache_manager.py @@ -1,5 +1,5 @@ from collections import defaultdict -from typing import Dict, Iterable, List, Optional +from typing import Dict, Iterable, List, Optional, Tuple from vllm.logger import init_logger from vllm.utils import cdiv @@ -69,7 +69,8 @@ def __init__( # is finished. self.req_to_blocks: Dict[str, List[KVCacheBlock]] = {} - def get_computed_blocks(self, request: Request) -> List[KVCacheBlock]: + def get_computed_blocks( + self, request: Request) -> Tuple[List[KVCacheBlock], int]: """Get the computed (cached) blocks for the request. Note that the computed blocks must be full. @@ -77,11 +78,13 @@ def get_computed_blocks(self, request: Request) -> List[KVCacheBlock]: request: The request to get the computed blocks. Returns: - A list of blocks that are computed for the request. + A tuple containing: + - A list of blocks that are computed for the request. + - The number of computed tokens. """ if not self.enable_caching: # Prefix caching is disabled. - return [] + return [], 0 computed_blocks = [] @@ -101,7 +104,11 @@ def get_computed_blocks(self, request: Request) -> List[KVCacheBlock]: else: break - return computed_blocks + # NOTE(woosuk): Since incomplete blocks are not eligible for + # sharing, `num_computed_tokens` is always a multiple of + # `block_size`. + num_computed_tokens = len(computed_blocks) * self.block_size + return computed_blocks, num_computed_tokens def append_slots( self, @@ -278,6 +285,56 @@ def free(self, request: Request) -> None: if block.ref_cnt == 0: self.free_block_queue.append(block) + def uncache_blocks(self, request: Request) -> int: + """Uncache the blocks that are no longer full based on the + num_computed_tokens in the given request. This happens when + the blocks were full and cached due to speculative tokens, but the + speculative tokens are not accepted. + + Args: + request: The request. + + Returns: + The number of uncached blocks. + """ + blocks = self.req_to_blocks[request.request_id] + num_computed_tokens = request.num_computed_tokens + num_full_blocks = num_computed_tokens // self.block_size + num_uncached_blocks = 0 + for block in blocks[num_full_blocks:]: + # If the block is not cached, the following blocks are not cached. + if not self._maybe_evict_cached_block(block): + break + num_uncached_blocks += 1 + return num_uncached_blocks + + def reset_prefix_cache(self) -> bool: + """Reset prefix cache. This function may be used in RLHF + flows to invalid prefix caching after the weights are updated, + or used for resetting prefix caching status for benchmarking. + + Returns: + bool: True if the prefix cache is successfully reset, + False otherwise. + """ + num_used_blocks = (self.num_gpu_blocks - + self.free_block_queue.num_free_blocks) + if num_used_blocks > 0: + logger.warning( + "Failed to reset prefix cache because some " + "blocks (%d) are not freed yet", num_used_blocks) + return False + + # Remove all hashes so that no new blocks will hit. + self.cached_block_hash_to_block = defaultdict(dict) + + # Remove all hashes from all blocks. + for block in self.block_pool: + block.reset_hash() + + logger.info("Successfully reset prefix cache") + return True + def get_num_common_prefix_blocks( self, request: Request, @@ -352,7 +409,7 @@ def _get_new_blocks(self, num_blocks: int) -> List[KVCacheBlock]: # If the block is cached, evict it. if self.enable_caching: - self._evict_cached_block(curr_block) + self._maybe_evict_cached_block(curr_block) curr_block.incr_ref() ret.append(curr_block) @@ -360,13 +417,16 @@ def _get_new_blocks(self, num_blocks: int) -> List[KVCacheBlock]: return ret - def _evict_cached_block(self, block: KVCacheBlock) -> None: + def _maybe_evict_cached_block(self, block: KVCacheBlock) -> bool: """ If a block is cached in `cached_block_hash_to_block`, we reset its hash metadata and evict it from the cache. Args: block: The block to evict. + + Returns: + True if the block is evicted, False otherwise. """ block_hash = block.block_hash if block_hash and block_hash in self.cached_block_hash_to_block: @@ -376,6 +436,9 @@ def _evict_cached_block(self, block: KVCacheBlock) -> None: if len(self.cached_block_hash_to_block[block_hash]) == 0: del self.cached_block_hash_to_block[block_hash] + return True + return False + def _get_cached_block(self, block_hash: BlockHashType) -> Optional[KVCacheBlock]: """Get a cached block by the block hash, or None if cache miss. diff --git a/vllm/v1/core/kv_cache_utils.py b/vllm/v1/core/kv_cache_utils.py index 22a5d2fb08a48..bab99fe37caee 100644 --- a/vllm/v1/core/kv_cache_utils.py +++ b/vllm/v1/core/kv_cache_utils.py @@ -3,7 +3,10 @@ from dataclasses import dataclass from typing import Any, List, NamedTuple, Optional, Tuple +from vllm.config import VllmConfig from vllm.logger import init_logger +from vllm.v1.kv_cache_interface import (KVCacheConfig, KVCacheSpec, + KVCacheTensor) from vllm.v1.request import Request logger = init_logger(__name__) @@ -305,3 +308,124 @@ def hash_request_tokens(block_size: int, ret.append(block_hash) parent_block_hash_value = block_hash.hash_value return ret + + +def check_enough_kv_cache_memory(vllm_config: VllmConfig, + kv_cache_spec: KVCacheSpec, + available_memory: int): + """ + Checks whether `available_memory` is enough for the KV cache to hold at + least one request with the model's max_model_len. + + Args: + vllm_config: The global VllmConfig + kv_cache_spec: The kv cache spec of the model + available_memory: Memory available for KV cache in bytes. + + Raises: + ValueError: If there is not enough memory available for the KV cache. + """ + + if available_memory <= 0: + raise ValueError("No available memory for the cache blocks. " + "Try increasing `gpu_memory_utilization` when " + "initializing the engine.") + + max_model_len = vllm_config.model_config.max_model_len + needed_memory = 0 + for layer_spec in kv_cache_spec.values(): + needed_memory += layer_spec.bytes_for_tokens(max_model_len) + + if needed_memory > available_memory: + raise ValueError( + f"To serve at least one request with the models's max seq len " + f"({max_model_len}), ({needed_memory/1024/1024/1024:.2f} GB KV " + f"cache is needed, which is larger than the available KV cache " + f"memory ({available_memory/1024/1024/1024:.2f} GB). Try " + f"increasing `gpu_memory_utilization` or decreasing " + f"`max_model_len` when initializing the engine.") + + +def is_kv_cache_type_uniform(kv_cache_spec: KVCacheSpec) -> bool: + """ + Whether all layers in the given KVCacheSpec have the same type of KV cache. + + Args: + kv_cache_spec: The KVCacheSpec of the model + + Returns: + True if all layers have the same type, False otherwise. + """ + + layer_keys = set(layer.type_id for layer in kv_cache_spec.values()) + return len(layer_keys) == 1 + + +def _get_kv_cache_config_uniform_type(vllm_config: VllmConfig, + kv_cache_spec: KVCacheSpec, + available_memory: int) -> KVCacheConfig: + """ + Generates the KV cache configuration for a model with one type of KV cache. + Divide the available memory equally among all layers. + + Args: + vllm_config: The global VllmConfig + kv_cache_spec: The kv cache spec of the model + available_memory: Memory available for KV cache in bytes. + + Returns: + The generated KVCacheConfig + """ + + page_sizes = {layer.page_size_bytes for layer in kv_cache_spec.values()} + assert len(page_sizes) == 1 + page_size = page_sizes.pop() + + num_blocks = int(available_memory // page_size // len(kv_cache_spec)) + num_blocks = max(num_blocks, 0) + + if vllm_config.cache_config.num_gpu_blocks_override is not None: + num_gpu_blocks_override = \ + vllm_config.cache_config.num_gpu_blocks_override + logger.info( + "Overriding num_gpu_blocks=%d with " + "num_gpu_blocks_override=%d", num_blocks, num_gpu_blocks_override) + num_blocks = num_gpu_blocks_override + + logger.info("# GPU blocks: %d", num_blocks) + + per_layer_size = page_size * num_blocks + + kv_cache_config = KVCacheConfig( + num_blocks=num_blocks, + tensors={ + layer_name: KVCacheTensor(size=per_layer_size) + for layer_name in kv_cache_spec + }, + groups=[[layer_name for layer_name in kv_cache_spec]], + kv_cache_spec=kv_cache_spec) + return kv_cache_config + + +def get_kv_cache_config(vllm_config: VllmConfig, kv_cache_spec: KVCacheSpec, + available_memory: int) -> KVCacheConfig: + """ + Generates the KV cache configuration for a model + TODO: support hybrid models with more than one type of KV cache. + + Args: + vllm_config: The global VllmConfig + kv_cache_spec: The kv cache spec of the model + available_memory: Memory available for KV cache in bytes. + + Returns: + The generated KVCacheConfig + """ + check_enough_kv_cache_memory(vllm_config, kv_cache_spec, available_memory) + if is_kv_cache_type_uniform(kv_cache_spec): + # KV cache of all layers are the same, which is true for most models. + # Allocate the same amount of memory for each layer. + return _get_kv_cache_config_uniform_type(vllm_config, kv_cache_spec, + available_memory) + else: + raise NotImplementedError diff --git a/vllm/v1/core/scheduler.py b/vllm/v1/core/scheduler.py index f04e529891287..de7fb1a698df6 100644 --- a/vllm/v1/core/scheduler.py +++ b/vllm/v1/core/scheduler.py @@ -3,10 +3,11 @@ from typing import (TYPE_CHECKING, Deque, Dict, Iterable, List, Optional, Set, Tuple, Union) -from vllm.config import CacheConfig, LoRAConfig, SchedulerConfig +from vllm.config import CacheConfig, LoRAConfig, ModelConfig, SchedulerConfig from vllm.logger import init_logger from vllm.sampling_params import SamplingParams -from vllm.v1.core.encoder_cache_manager import EncoderCacheManager +from vllm.v1.core.encoder_cache_manager import (EncoderCacheManager, + compute_encoder_budget) from vllm.v1.core.kv_cache_manager import KVCacheManager from vllm.v1.engine import EngineCoreOutput, EngineCoreOutputs from vllm.v1.metrics.stats import SchedulerStats @@ -25,6 +26,7 @@ class Scheduler: def __init__( self, scheduler_config: SchedulerConfig, + model_config: ModelConfig, cache_config: CacheConfig, lora_config: Optional[LoRAConfig], ) -> None: @@ -69,16 +71,24 @@ def __init__( self.running_reqs_data: Dict[str, RunningRequestData] = {} # Encoder-related. + # Calculate encoder cache size if applicable + # NOTE: For now we use the same budget for both compute and space. + # This can be changed when we make encoder cache for embedding caching + # across requests. + encoder_compute_budget, encoder_cache_size = compute_encoder_budget( + model_config=model_config, + scheduler_config=scheduler_config, + ) + # NOTE(woosuk): Here, "encoder" includes the vision encoder (and # projector if needed). Currently, we assume that the encoder also # has the Transformer architecture (e.g., ViT). - self.max_num_encoder_input_tokens = self.scheduler_config.max_num_encoder_input_tokens #noqa: E501 - # NOTE(woosuk): For the models without encoder (e.g., text-only models), - # the encoder cache will not be initialized and used, regardless of - # the cache size. This is because the memory space for the encoder cache - # is preallocated in the profiling run. + self.max_num_encoder_input_tokens = encoder_compute_budget + # NOTE: For the models without encoder (e.g., text-only models), + # the encoder cache will not be initialized because cache size is 0 + # for these models. self.encoder_cache_manager = EncoderCacheManager( - cache_size=self.scheduler_config.encoder_cache_size) + cache_size=encoder_cache_size) def schedule(self) -> "SchedulerOutput": # NOTE(woosuk) on the scheduling algorithm: @@ -184,12 +194,8 @@ def schedule(self) -> "SchedulerOutput": request = self.waiting[0] # Get already-cached tokens. - computed_blocks = self.kv_cache_manager.get_computed_blocks( - request) - # NOTE(woosuk): Since incomplete blocks are not eligible for - # sharing, `num_computed_tokens` is always a multiple of - # `block_size`. - num_computed_tokens = len(computed_blocks) * self.block_size + computed_blocks, num_computed_tokens = \ + self.kv_cache_manager.get_computed_blocks(request) # Number of tokens to be scheduled. # We use `request.num_tokens` instead of # `request.num_prompt_tokens` to consider the resumed requests, @@ -373,18 +379,22 @@ def _try_schedule_encoder_inputs( if self.encoder_cache_manager.has_cache(request, i): # The encoder input is already computed and cached. continue - if not self.encoder_cache_manager.can_allocate(request, i): - # The encoder cache is full. We can only schedule the decoder - # tokens just before the encoder input. - num_new_tokens = start_pos - num_computed_tokens - break - if num_encoder_tokens > encoder_budget: - # The encoder budget is exhausted. We can only schedule the - # decoder tokens up until the encoder input. - # NOTE(woosuk): We assume that the encoder tokens should be - # processed altogether, as the encoder usually uses + if (not self.encoder_cache_manager.can_allocate(request, i) + or num_encoder_tokens > encoder_budget): + # The encoder cache is full or the encoder budget is exhausted. + # NOTE(woosuk): We assume that the encoder input tokens should + # be processed altogether, as the encoder usually uses # bidirectional attention. - num_new_tokens = start_pos - num_computed_tokens + if num_computed_tokens < start_pos: + # We only schedule the decoder tokens just before the + # encoder input. + num_new_tokens = start_pos - num_computed_tokens + else: + # Because of prefix caching, num_computed_tokens is greater + # than start_pos even though its encoder input is not + # available. In this case, we can't schedule any token for + # the request in this step. + num_new_tokens = 0 break encoder_budget -= num_encoder_tokens @@ -401,6 +411,10 @@ def update_from_output( num_scheduled_tokens = scheduler_output.num_scheduled_tokens new_running: List[Request] = [] outputs: List[EngineCoreOutput] = [] + + # NOTE(woosuk): As len(self.running) can be up to 1K or more, the below + # loop can be a performance bottleneck. We should do our best to avoid + # expensive operations inside the loop. for request in self.running: req_id = request.request_id request.num_computed_tokens += num_scheduled_tokens[req_id] @@ -411,13 +425,15 @@ def update_from_output( cached_encoder_input_ids = ( self.encoder_cache_manager.get_cached_input_ids(request)) - for input_id in list(cached_encoder_input_ids): - start_pos = request.mm_positions[input_id]["offset"] - num_tokens = request.mm_positions[input_id]["length"] - if start_pos + num_tokens <= request.num_computed_tokens: - # The encoder output is already processed and stored - # in the decoder's KV cache. - self.encoder_cache_manager.free(request, input_id) + # OPTIMIZATION: Avoid list(set) if the set is empty. + if cached_encoder_input_ids: + for input_id in list(cached_encoder_input_ids): + start_pos = request.mm_positions[input_id]["offset"] + num_tokens = request.mm_positions[input_id]["length"] + if start_pos + num_tokens <= request.num_computed_tokens: + # The encoder output is already processed and stored + # in the decoder's KV cache. + self.encoder_cache_manager.free(request, input_id) if request.num_computed_tokens == request.num_tokens: req_index = model_runner_output.req_id_to_index[req_id] @@ -519,6 +535,9 @@ def get_num_unfinished_requests(self) -> int: def has_unfinished_requests(self) -> bool: return self.get_num_unfinished_requests() > 0 + def reset_prefix_cache(self) -> bool: + return self.kv_cache_manager.reset_prefix_cache() + def make_stats(self) -> SchedulerStats: return SchedulerStats( num_running_reqs=len(self.running), diff --git a/vllm/v1/engine/__init__.py b/vllm/v1/engine/__init__.py index 6d90c38c72cf5..abe4952c4baff 100644 --- a/vllm/v1/engine/__init__.py +++ b/vllm/v1/engine/__init__.py @@ -66,6 +66,11 @@ class EngineCoreProfile: is_start: bool +@dataclass +class EngineCoreResetPrefixCache: + pass + + class EngineCoreRequestType(enum.Enum): """ Request types defined as hex byte strings, so it can be sent over sockets @@ -74,6 +79,8 @@ class EngineCoreRequestType(enum.Enum): ADD = b'\x00' ABORT = b'\x01' PROFILE = b'\x02' + RESET_PREFIX_CACHE = b'\x03' -EngineCoreRequestUnion = Union[EngineCoreRequest, EngineCoreProfile, List[str]] +EngineCoreRequestUnion = Union[EngineCoreRequest, EngineCoreProfile, + EngineCoreResetPrefixCache, List[str]] diff --git a/vllm/v1/engine/async_llm.py b/vllm/v1/engine/async_llm.py index a74699f7513e6..917d52d3220b8 100644 --- a/vllm/v1/engine/async_llm.py +++ b/vllm/v1/engine/async_llm.py @@ -2,9 +2,12 @@ import os from typing import AsyncGenerator, List, Mapping, Optional, Type, Union +import numpy as np + from vllm.config import ModelConfig, VllmConfig from vllm.engine.arg_utils import AsyncEngineArgs from vllm.engine.protocol import EngineClient +from vllm.envs import VLLM_V1_OUTPUT_PROC_CHUNK_SIZE from vllm.inputs import INPUT_REGISTRY, InputRegistry, PromptType from vllm.inputs.preprocess import InputPreprocessor from vllm.logger import init_logger @@ -12,16 +15,17 @@ from vllm.outputs import RequestOutput from vllm.pooling_params import PoolingParams from vllm.prompt_adapter.request import PromptAdapterRequest -from vllm.sampling_params import SamplingParams +from vllm.sampling_params import RequestOutputKind, SamplingParams from vllm.transformers_utils.tokenizer import AnyTokenizer from vllm.transformers_utils.tokenizer_group import init_tokenizer_from_configs from vllm.usage.usage_lib import UsageContext -from vllm.utils import kill_process_tree +from vllm.utils import cdiv, kill_process_tree from vllm.v1.engine.core_client import EngineCoreClient from vllm.v1.engine.output_processor import OutputProcessor from vllm.v1.engine.processor import Processor from vllm.v1.executor.abstract import Executor -from vllm.v1.metrics.loggers import LoggingStatLogger, StatLoggerBase +from vllm.v1.metrics.loggers import (LoggingStatLogger, PrometheusStatLogger, + StatLoggerBase) from vllm.v1.metrics.stats import IterationStats, SchedulerStats logger = init_logger(__name__) @@ -43,13 +47,15 @@ def __init__( assert start_engine_loop + self.model_config = vllm_config.model_config + self.log_requests = log_requests self.log_stats = log_stats self.stat_loggers: List[StatLoggerBase] = [ LoggingStatLogger(), - # TODO(rob): PrometheusStatLogger(), + PrometheusStatLogger(labels=dict( + model_name=self.model_config.served_model_name)), ] - self.model_config = vllm_config.model_config # Tokenizer (+ ensure liveness if running in another process). self.tokenizer = init_tokenizer_from_configs( @@ -205,17 +211,23 @@ async def generate( # The output_handler task pushes items into the queue. # This task pulls from the queue and yields to caller. - while True: + finished = False + while not finished: # Note: drain queue without await if possible (avoids # task switching under load which helps performance). - out = q.get_nowait() if q.qsize() > 0 else await q.get() + out = q.get_nowait() if not q.empty() else await q.get() + + # Coalesce any additional queued outputs + while not q.empty(): + next_out = q.get_nowait() + if sampling_params.output_kind == RequestOutputKind.DELTA: + out.add(next_out) + else: + out = next_out # Note: both OutputProcessor and EngineCore handle their # own request cleanup based on finished. - if out.finished: - yield out - break - + finished = out.finished yield out # If the request is disconnected by the client, the @@ -233,22 +245,41 @@ async def _run_output_handler(self): # 1) Pull EngineCoreOutputs from the EngineCore. outputs = await self.engine_core.get_output_async() - # 2) Process EngineCoreOutputs. - processed_outputs = self.output_processor.process_outputs( - outputs.outputs) - # NOTE: RequestOutputs are pushed to their queues. - assert len(processed_outputs.request_outputs) == 0 - - # 3) Abort any reqs that finished due to stop strings. - await self.engine_core.abort_requests_async( - processed_outputs.reqs_to_abort) + # Split outputs into chunks of at most + # VLLM_V1_OUTPUT_PROC_CHUNK_SIZE, so that we don't block the + # event loop for too long. + num_outputs = len(outputs.outputs) + if num_outputs <= VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: + slices = (outputs.outputs, ) + else: + slices = np.array_split( + outputs.outputs, + cdiv(num_outputs, VLLM_V1_OUTPUT_PROC_CHUNK_SIZE)) + + iteration_stats = None + for i, outputs_slice in enumerate(slices): + # 2) Process EngineCoreOutputs. + processed_outputs = self.output_processor.process_outputs( + outputs_slice, iteration_stats) + # NOTE: RequestOutputs are pushed to their queues. + assert not processed_outputs.request_outputs + iteration_stats = processed_outputs.iteration_stats + + # Allow other asyncio tasks to run between chunks + if i + 1 < len(slices): + await asyncio.sleep(0) + + # 3) Abort any reqs that finished due to stop strings. + await self.engine_core.abort_requests_async( + processed_outputs.reqs_to_abort) # 4) Logging. # TODO(rob): make into a coroutine and launch it in - # background thread once we add Prometheus. + # background thread once Prometheus overhead is non-trivial. + assert iteration_stats is not None self._log_stats( scheduler_stats=outputs.scheduler_stats, - iteration_stats=processed_outputs.iteration_stats, + iteration_stats=iteration_stats, ) except Exception as e: @@ -321,6 +352,9 @@ async def start_profile(self) -> None: async def stop_profile(self) -> None: await self.engine_core.profile_async(False) + async def reset_prefix_cache(self) -> None: + await self.engine_core.reset_prefix_cache_async() + @property def is_running(self) -> bool: return True diff --git a/vllm/v1/engine/core.py b/vllm/v1/engine/core.py index e7f90d3c62142..cf94033a38d96 100644 --- a/vllm/v1/engine/core.py +++ b/vllm/v1/engine/core.py @@ -11,15 +11,16 @@ import zmq.asyncio from msgspec import msgpack -from vllm.config import CacheConfig, VllmConfig +from vllm.config import VllmConfig from vllm.logger import init_logger from vllm.transformers_utils.config import ( maybe_register_config_serialize_by_value) from vllm.utils import get_exception_traceback, zmq_socket_ctx +from vllm.v1.core.kv_cache_utils import get_kv_cache_config from vllm.v1.core.scheduler import Scheduler from vllm.v1.engine import (EngineCoreOutputs, EngineCoreProfile, EngineCoreRequest, EngineCoreRequestType, - EngineCoreRequestUnion) + EngineCoreRequestUnion, EngineCoreResetPrefixCache) from vllm.v1.engine.mm_input_mapper import MMInputMapperServer from vllm.v1.executor.abstract import Executor from vllm.v1.request import Request, RequestStatus @@ -49,34 +50,41 @@ def __init__( # Setup KV Caches and update CacheConfig after profiling. num_gpu_blocks, num_cpu_blocks = self._initialize_kv_caches( - vllm_config.cache_config) + vllm_config) vllm_config.cache_config.num_gpu_blocks = num_gpu_blocks vllm_config.cache_config.num_cpu_blocks = num_cpu_blocks # Setup scheduler. - self.scheduler = Scheduler(vllm_config.scheduler_config, - vllm_config.cache_config, - vllm_config.lora_config) + self.scheduler = Scheduler( + scheduler_config=vllm_config.scheduler_config, + model_config=vllm_config.model_config, + cache_config=vllm_config.cache_config, + lora_config=vllm_config.lora_config, + ) self.mm_input_mapper_server = MMInputMapperServer( vllm_config.model_config) def _initialize_kv_caches(self, - cache_config: CacheConfig) -> Tuple[int, int]: + vllm_config: VllmConfig) -> Tuple[int, int]: start = time.time() - num_gpu_blocks, _ = self.model_executor.determine_num_available_blocks( - ) - if cache_config.num_gpu_blocks_override is not None: - num_gpu_blocks_override = cache_config.num_gpu_blocks_override - logger.info( - "Overriding num_gpu_blocks=%d with " - "num_gpu_blocks_override=%d", num_gpu_blocks, - num_gpu_blocks_override) - num_gpu_blocks = num_gpu_blocks_override + # Get all kv cache needed by the model + kv_cache_spec = self.model_executor.get_kv_cache_spec() + + # Profiles the peak memory usage of the model to determine how much + # memory can be allocated for kv cache. + availble_gpu_memory = self.model_executor.determine_available_memory() + # Get the kv cache tensor size + kv_cache_config = get_kv_cache_config(vllm_config, kv_cache_spec, + availble_gpu_memory) + num_gpu_blocks = kv_cache_config.num_blocks num_cpu_blocks = 0 - self.model_executor.initialize(num_gpu_blocks) + + # Initialize kv cache and warmup the execution + self.model_executor.initialize(kv_cache_config) + elapsed = time.time() - start logger.info(("init engine (profile, create kv cache, " "warmup model) took %.2f seconds"), elapsed) @@ -127,6 +135,9 @@ def shutdown(self): def profile(self, is_start: bool = True): self.model_executor.profile(is_start) + def reset_prefix_cache(self): + self.scheduler.reset_prefix_cache() + class EngineCoreProc(EngineCore): """ZMQ-wrapper for running EngineCore in background process.""" @@ -239,6 +250,8 @@ def _handle_client_request(self, request: EngineCoreRequestUnion) -> None: self.add_request(request) elif isinstance(request, EngineCoreProfile): self.model_executor.profile(request.is_start) + elif isinstance(request, EngineCoreResetPrefixCache): + self.reset_prefix_cache() else: # TODO: make an EngineCoreAbort wrapper assert isinstance(request, list) @@ -263,7 +276,9 @@ def process_input_socket(self, input_path: str): request = decoder_add_req.decode(request_data) elif request_type == EngineCoreRequestType.ABORT.value: request = decoder_abort_req.decode(request_data) - elif request_type == EngineCoreRequestType.PROFILE.value: + elif request_type in ( + EngineCoreRequestType.PROFILE.value, + EngineCoreRequestType.RESET_PREFIX_CACHE.value): request = pickle.loads(request_data) else: raise ValueError(f"Unknown RequestType: {request_type}") diff --git a/vllm/v1/engine/core_client.py b/vllm/v1/engine/core_client.py index ac0f0f14bf1ab..f3b992d6873e7 100644 --- a/vllm/v1/engine/core_client.py +++ b/vllm/v1/engine/core_client.py @@ -1,8 +1,9 @@ +import asyncio import os import signal import weakref from abc import ABC, abstractmethod -from typing import List, Type +from typing import List, Optional, Type import msgspec import zmq @@ -14,7 +15,7 @@ make_zmq_socket) from vllm.v1.engine import (EngineCoreOutputs, EngineCoreProfile, EngineCoreRequest, EngineCoreRequestType, - EngineCoreRequestUnion) + EngineCoreRequestUnion, EngineCoreResetPrefixCache) from vllm.v1.engine.core import EngineCore, EngineCoreProc from vllm.v1.executor.abstract import Executor from vllm.v1.serial_utils import PickleEncoder @@ -69,6 +70,9 @@ def add_request(self, request: EngineCoreRequest) -> None: def profile(self, is_start: bool = True) -> None: raise NotImplementedError + def reset_prefix_cache(self) -> None: + raise NotImplementedError + def abort_requests(self, request_ids: List[str]) -> None: raise NotImplementedError @@ -81,6 +85,9 @@ async def add_request_async(self, request: EngineCoreRequest) -> None: async def profile_async(self, is_start: bool = True) -> None: raise NotImplementedError + async def reset_prefix_cache_async(self) -> None: + raise NotImplementedError + async def abort_requests_async(self, request_ids: List[str]) -> None: raise NotImplementedError @@ -108,12 +115,15 @@ def abort_requests(self, request_ids: List[str]) -> None: if len(request_ids) > 0: self.engine_core.abort_requests(request_ids) - def shutdown(self): + def shutdown(self) -> None: self.engine_core.shutdown() def profile(self, is_start: bool = True) -> None: self.engine_core.profile(is_start) + def reset_prefix_cache(self) -> None: + self.engine_core.reset_prefix_cache() + class MPClient(EngineCoreClient): """ @@ -229,6 +239,10 @@ def profile(self, is_start: bool = True) -> None: self._send_input(EngineCoreRequestType.PROFILE, EngineCoreProfile(is_start)) + def reset_prefix_cache(self) -> None: + self._send_input(EngineCoreRequestType.RESET_PREFIX_CACHE, + EngineCoreResetPrefixCache()) + class AsyncMPClient(MPClient): """Asyncio-compatible client for multi-proc EngineCore.""" @@ -242,10 +256,24 @@ def __init__(self, vllm_config: VllmConfig, log_stats=True, ) + self.outputs_queue: Optional[asyncio.Queue[bytes]] = None + self.queue_task: Optional[asyncio.Task] = None + async def get_output_async(self) -> EngineCoreOutputs: + if self.outputs_queue is None: + # Perform IO in separate task to parallelize as much as possible + self.outputs_queue = asyncio.Queue() + + async def process_outputs_socket(): + assert self.outputs_queue is not None + while True: + (frame, ) = await self.output_socket.recv_multipart( + copy=False) + self.outputs_queue.put_nowait(frame.buffer) - frames = await self.output_socket.recv_multipart(copy=False) - return self.decoder.decode(frames[0].buffer) + self.queue_task = asyncio.create_task(process_outputs_socket()) + + return self.decoder.decode(await self.outputs_queue.get()) async def _send_input(self, request_type: EngineCoreRequestType, request: EngineCoreRequestUnion) -> None: @@ -266,3 +294,7 @@ async def abort_requests_async(self, request_ids: List[str]) -> None: async def profile_async(self, is_start: bool = True) -> None: await self._send_input(EngineCoreRequestType.PROFILE, EngineCoreProfile(is_start)) + + async def reset_prefix_cache_async(self) -> None: + await self._send_input(EngineCoreRequestType.RESET_PREFIX_CACHE, + EngineCoreResetPrefixCache()) diff --git a/vllm/v1/engine/llm_engine.py b/vllm/v1/engine/llm_engine.py index f5999ccda6447..55d314ebeb955 100644 --- a/vllm/v1/engine/llm_engine.py +++ b/vllm/v1/engine/llm_engine.py @@ -162,6 +162,9 @@ def start_profile(self): def stop_profile(self): self.engine_core.profile(False) + def reset_prefix_cache(self): + self.engine_core.reset_prefix_cache() + def get_tokenizer_group( self, group_type: Type[_G] = BaseTokenizerGroup, diff --git a/vllm/v1/engine/output_processor.py b/vllm/v1/engine/output_processor.py index 749f4f5043c97..564eab51bd3a8 100644 --- a/vllm/v1/engine/output_processor.py +++ b/vllm/v1/engine/output_processor.py @@ -101,6 +101,7 @@ def add_request( def process_outputs( self, engine_core_outputs: List[EngineCoreOutput], + iteration_stats: Optional[IterationStats] = None, ) -> OutputProcessorOutput: """ Process the EngineCoreOutputs: @@ -133,7 +134,8 @@ def process_outputs( request_outputs: List[RequestOutput] = [] reqs_to_abort: List[str] = [] - iteration_stats = IterationStats(self.log_stats) + if not iteration_stats: + iteration_stats = IterationStats(self.log_stats) for engine_core_output in engine_core_outputs: req_id = engine_core_output.request_id req_state = self.request_states.get(req_id) @@ -175,8 +177,8 @@ def process_outputs( iteration_stats=iteration_stats, ) + @staticmethod def _make_request_output( - self, request_state: RequestState, detokenizer_output: Optional[DetokenizerOutput], ) -> Optional[RequestOutput]: diff --git a/vllm/v1/executor/abstract.py b/vllm/v1/executor/abstract.py index 5d74d4b01f500..131be759842c7 100644 --- a/vllm/v1/executor/abstract.py +++ b/vllm/v1/executor/abstract.py @@ -1,57 +1,92 @@ -from abc import ABC, abstractmethod -from typing import Tuple, Type +from typing import Type from vllm.config import VllmConfig +from vllm.executor.executor_base import ExecutorBase +from vllm.executor.ray_distributed_executor import ( # noqa + RayDistributedExecutor as RayDistributedExecutorV0) +from vllm.executor.uniproc_executor import ( # noqa + ExecutorWithExternalLauncher as ExecutorWithExternalLauncherV0) +from vllm.executor.uniproc_executor import ( # noqa + UniProcExecutor as UniProcExecutorV0) +from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec from vllm.v1.outputs import ModelRunnerOutput -class Executor(ABC): - """Abstract class for executors.""" +class Executor(ExecutorBase): + """ + Abstract class for v1 executors, mainly define some methods for v1. + For methods shared by v0 and v1, define them in ExecutorBase""" @staticmethod def get_class(vllm_config: VllmConfig) -> Type["Executor"]: executor_class: Type[Executor] + parallel_config = vllm_config.parallel_config distributed_executor_backend = ( - vllm_config.parallel_config.distributed_executor_backend) + parallel_config.distributed_executor_backend) + if distributed_executor_backend is None: + # If the user does not specify the distributed executor backend, + # we will choose the backend based on the world size. + if parallel_config.world_size > 1: + distributed_executor_backend = "mp" + else: + distributed_executor_backend = "uni" + if distributed_executor_backend == "ray": - from vllm.v1.executor.ray_executor import RayExecutor - executor_class = RayExecutor + executor_class = RayDistributedExecutor elif distributed_executor_backend == "mp": from vllm.v1.executor.multiproc_executor import MultiprocExecutor executor_class = MultiprocExecutor + elif distributed_executor_backend == "uni": + executor_class = UniProcExecutor + elif distributed_executor_backend == "external_launcher": + # TODO: make v1 scheduling deterministic + # to support external launcher + executor_class = ExecutorWithExternalLauncher else: - assert (distributed_executor_backend is None) - from vllm.v1.executor.uniproc_executor import UniprocExecutor - executor_class = UniprocExecutor + raise ValueError("Unknown distributed executor backend: " + f"{distributed_executor_backend}") return executor_class - @abstractmethod - def __init__(self, vllm_config: VllmConfig) -> None: - raise NotImplementedError + def initialize(self, kv_cache_config: KVCacheConfig) -> None: + """ + Initialize the KV caches and begin the model execution loop of the + underlying workers. + """ + self.collective_rpc("initialize_cache", args=(kv_cache_config, )) + self.collective_rpc("compile_or_warm_up_model") - @abstractmethod - def initialize(self, num_gpu_blocks: int) -> None: - raise NotImplementedError + def determine_available_memory(self) -> int: # in bytes + output = self.collective_rpc("determine_available_memory") + # Since we use a shared centralized controller, we take the minimum + # memory size across all workers to make sure all the memory + # operators can be applied to all workers. + return min(output) - @abstractmethod - def determine_num_available_blocks(self) -> Tuple[int, int]: - raise NotImplementedError + def get_kv_cache_spec(self) -> KVCacheSpec: + output = self.collective_rpc("get_kv_cache_spec") + for x in output: + assert x == output[0] + return output[0] - @abstractmethod def execute_model( self, scheduler_output, ) -> ModelRunnerOutput: - raise NotImplementedError + output = self.collective_rpc("execute_model", + args=(scheduler_output, )) + return output[0] - @abstractmethod def profile(self, is_start: bool = True): - raise NotImplementedError + self.collective_rpc("profile", args=(is_start, )) + + +class UniProcExecutor(UniProcExecutorV0, Executor): + pass + + +class ExecutorWithExternalLauncher(ExecutorWithExternalLauncherV0, Executor): + pass - @abstractmethod - def shutdown(self): - pass - @abstractmethod - def check_health(self) -> None: - raise NotImplementedError +class RayDistributedExecutor(RayDistributedExecutorV0, Executor): + pass diff --git a/vllm/v1/executor/multiproc_executor.py b/vllm/v1/executor/multiproc_executor.py index 41e6abbd67956..f6cf35da0106b 100644 --- a/vllm/v1/executor/multiproc_executor.py +++ b/vllm/v1/executor/multiproc_executor.py @@ -6,9 +6,11 @@ import weakref from dataclasses import dataclass from enum import Enum, auto +from functools import partial from multiprocessing.process import BaseProcess -from typing import Any, Dict, List, Optional, Tuple +from typing import Any, Callable, Dict, List, Optional, Tuple, Union +import cloudpickle import psutil import zmq @@ -23,7 +25,6 @@ from vllm.utils import (get_distributed_init_method, get_mp_context, get_open_port, get_open_zmq_ipc_path, zmq_socket_ctx) from vllm.v1.executor.abstract import Executor -from vllm.v1.outputs import ModelRunnerOutput from vllm.worker.worker_base import WorkerWrapperBase logger = init_logger(__name__) @@ -34,7 +35,7 @@ class MultiprocExecutor(Executor): - def __init__(self, vllm_config: VllmConfig) -> None: + def _init_executor(self) -> None: # Call self.shutdown at exit to clean up # and ensure workers will be terminated. self._finalizer = weakref.finalize(self, self.shutdown) @@ -52,9 +53,6 @@ def sigusr1_handler(signum, frame): signal.signal(signal.SIGUSR1, sigusr1_handler) - self.vllm_config = vllm_config - self.parallel_config = vllm_config.parallel_config - self.world_size = self.parallel_config.world_size tensor_parallel_size = self.parallel_config.tensor_parallel_size assert self.world_size == tensor_parallel_size, ( @@ -79,7 +77,8 @@ def sigusr1_handler(signum, frame): # Create workers self.workers: List[WorkerProcHandle] = [] for rank in range(self.world_size): - worker = WorkerProc.make_worker_process(vllm_config, rank, rank, + worker = WorkerProc.make_worker_process(self.vllm_config, rank, + rank, distributed_init_method, scheduler_output_handle) self.workers.append(worker) @@ -90,53 +89,24 @@ def sigusr1_handler(signum, frame): for w in self.workers: w.worker_response_mq.wait_until_ready() - def initialize(self, num_gpu_blocks: int) -> None: - """ - Initialize the KV caches and begin the model execution loop of the - underlying workers. - """ - logger.info("# GPU blocks: %d", num_gpu_blocks) - self.collective_rpc("initialize_cache", args=(num_gpu_blocks, )) - self.collective_rpc("compile_or_warm_up_model") - - def determine_num_available_blocks(self) -> Tuple[int, int]: - """ - Determine the number of available KV blocks by invoking the - underlying worker. - """ - num_blocks = self.collective_rpc("determine_num_available_blocks") - - # Since we use a shared centralized controller, we take the minimum - # number of blocks across all workers to make sure all the memory - # operators can be applied to all workers. - num_gpu_blocks = min(b[0] for b in num_blocks) - num_cpu_blocks = min(b[1] for b in num_blocks) - - return num_gpu_blocks, num_cpu_blocks - def collective_rpc(self, - method: str, + method: Union[str, Callable], timeout: Optional[float] = None, args: Tuple = (), kwargs: Optional[Dict] = None) -> List[Any]: - """ - Execute an RPC call on workers. - - Args: - method: Name of the worker method to execute - timeout: Maximum time in seconds to wait for execution. Rases a - TimeoutError on timeout. None means wait indefinitely. - args: Positional arguments to pass to the worker method - kwargs: Keyword arguments to pass to the worker method - - Returns: - List of results from each worker - """ start_time = time.monotonic() kwargs = kwargs or {} + # NOTE: If the args are heterogeneous, then we pack them into a list, + # and unpack them in the method of every worker, because every worker + # knows their own rank. try: - self.rpc_broadcast_mq.enqueue((method, args, kwargs)) + if isinstance(method, str): + send_method = method + else: + send_method = cloudpickle.dumps( + method, protocol=pickle.HIGHEST_PROTOCOL) + self.rpc_broadcast_mq.enqueue((send_method, args, kwargs)) responses = [None] * self.world_size for w in self.workers: @@ -160,18 +130,6 @@ def collective_rpc(self, # Re-raise any other exceptions raise e - def execute_model( - self, - scheduler_output, - ) -> ModelRunnerOutput: - model_output = self.collective_rpc("execute_model", - args=(scheduler_output, ))[0] - return model_output - - def profile(self, is_start: bool = True): - self.collective_rpc("profile", args=(is_start, )) - return - def _ensure_worker_termination(self): """Ensure that all worker processes are terminated. Assumes workers have received termination requests. Waits for processing, then sends @@ -246,9 +204,18 @@ def __init__( ready_path: str, ): self.rank = rank - wrapper = WorkerWrapperBase(vllm_config=vllm_config) - wrapper.init_worker(vllm_config, local_rank, rank, - distributed_init_method) + wrapper = WorkerWrapperBase(vllm_config=vllm_config, rpc_rank=rank) + # TODO: move `init_worker` to executor level as a collective rpc call + all_kwargs: List[Dict] = [ + {} for _ in range(vllm_config.parallel_config.world_size) + ] + all_kwargs[rank] = { + "vllm_config": vllm_config, + "local_rank": local_rank, + "rank": rank, + "distributed_init_method": distributed_init_method, + } + wrapper.init_worker(all_kwargs) self.worker = wrapper.worker pid = os.getpid() @@ -270,7 +237,7 @@ def __init__( ready_socket.send_string(WorkerProc.READY_STR) ready_socket.send(payload) - self.worker.initialize() + self.worker.init_device() self.worker.load_model() @staticmethod @@ -394,7 +361,11 @@ def worker_busy_loop(self): method, args, kwargs = self.rpc_broadcast_mq.dequeue() try: - output = getattr(self.worker, method)(*args, **kwargs) + if isinstance(method, str): + func = getattr(self.worker, method) + elif isinstance(method, bytes): + func = partial(cloudpickle.loads(method), self.worker) + output = func(*args, **kwargs) except Exception as e: self.worker_response_mq.enqueue( (WorkerProc.ResponseStatus.FAILURE, e)) diff --git a/vllm/v1/executor/ray_executor.py b/vllm/v1/executor/ray_executor.py deleted file mode 100644 index 79acc60001c99..0000000000000 --- a/vllm/v1/executor/ray_executor.py +++ /dev/null @@ -1,342 +0,0 @@ -import os -from collections import defaultdict -from itertools import islice, repeat -from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple - -import vllm.envs as envs -from vllm.config import VllmConfig -from vllm.logger import init_logger -from vllm.utils import get_distributed_init_method, get_ip, get_open_port -from vllm.v1.executor.abstract import Executor -from vllm.v1.executor.ray_utils import (RayWorkerWrapper, - initialize_ray_cluster, ray) -from vllm.v1.outputs import ModelRunnerOutput - -if ray is not None: - from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy - -if TYPE_CHECKING: - from ray.util.placement_group import PlacementGroup - -logger = init_logger(__name__) - - -class RayExecutor(Executor): - - def __init__(self, vllm_config: VllmConfig) -> None: - self.vllm_config = vllm_config - self.parallel_config = vllm_config.parallel_config - self.model_config = vllm_config.model_config - self.forward_dag: Optional[ray.dag.CompiledDAG] = None - - # Disable Ray usage stats collection. - ray_usage = os.environ.get("RAY_USAGE_STATS_ENABLED", "0") - if ray_usage != "1": - os.environ["RAY_USAGE_STATS_ENABLED"] = "0" - - initialize_ray_cluster(self.parallel_config) - placement_group = self.parallel_config.placement_group - - # Create the parallel GPU workers. - self._init_workers_ray(placement_group) - - def _init_workers_ray(self, placement_group: "PlacementGroup", - **ray_remote_kwargs): - # A list of workers to run a model. - self.workers: List[RayWorkerWrapper] = [] - if self.parallel_config.ray_workers_use_nsight: - ray_remote_kwargs = self._configure_ray_workers_use_nsight( - ray_remote_kwargs) - - # Create the workers. - driver_ip = get_ip() - for bundle_id, bundle in enumerate(placement_group.bundle_specs): - if not bundle.get("GPU", 0): - # Skip bundles that don't have GPUs, - # as each worker needs one GPU. - continue - scheduling_strategy = PlacementGroupSchedulingStrategy( - placement_group=placement_group, - placement_group_capture_child_tasks=True, - placement_group_bundle_index=bundle_id, - ) - - worker = ray.remote( - num_cpus=0, - num_gpus=1, - scheduling_strategy=scheduling_strategy, - **ray_remote_kwargs, - )(RayWorkerWrapper).remote(vllm_config=self.vllm_config) - self.workers.append(worker) - - logger.debug("workers: %s", self.workers) - worker_ips = [ - ray.get(worker.get_node_ip.remote()) # type: ignore[attr-defined] - for worker in self.workers - ] - ip_counts: Dict[str, int] = {} - for ip in worker_ips: - ip_counts[ip] = ip_counts.get(ip, 0) + 1 - - worker_to_ip = dict(zip(self.workers, worker_ips)) - - def sort_by_driver_then_worker_ip(worker): - """ - Sort the workers based on 3 properties: - 1. If the worker is on the same node as the driver (vllm engine), - it should be placed first. - 2. Then, if the worker is on a node with fewer workers, it should - be placed first. - 3. Finally, if the work is on a node with smaller IP address, it - should be placed first. This is simply a tiebreaker to make - sure the workers are sorted in a deterministic way. - """ - ip = worker_to_ip[worker] - return (ip != driver_ip, ip_counts[ip], ip) - - # After sorting, the workers on the same node will be - # close to each other, and the workers on the driver - # node will be placed first. - self.workers = sorted(self.workers, key=sort_by_driver_then_worker_ip) - - # Get the set of GPU IDs used on each node. - worker_node_and_gpu_ids = self._run_workers("get_node_and_gpu_ids") - - node_workers = defaultdict(list) # node id -> list of worker ranks - node_gpus = defaultdict(list) # node id -> list of gpu ids - - for i, (node_id, gpu_ids) in enumerate(worker_node_and_gpu_ids): - node_workers[node_id].append(i) - # `gpu_ids` can be a list of strings or integers. - # convert them to integers for consistency. - # NOTE: gpu_ids can be larger than 9 (e.g. 16 GPUs), - # string sorting is not sufficient. - # see https://github.com/vllm-project/vllm/issues/5590 - gpu_ids = [int(x) for x in gpu_ids] - node_gpus[node_id].extend(gpu_ids) - - for node_id, gpu_ids in node_gpus.items(): - node_gpus[node_id] = sorted(gpu_ids) - - all_ips = set(worker_ips) - n_ips = len(all_ips) - n_nodes = len(node_workers) - - if n_nodes != n_ips: - raise RuntimeError( - f"Every node should have a unique IP address. Got {n_nodes}" - f" nodes with node ids {list(node_workers.keys())} and " - f"{n_ips} unique IP addresses {all_ips}. Please check your" - " network configuration. If you set `VLLM_HOST_IP` or " - "`HOST_IP` environment variable, make sure it is unique for" - " each node.") - - # Set environment variables for the driver and workers. - all_args_to_update_environment_variables = [({ - "CUDA_VISIBLE_DEVICES": - ",".join(map(str, node_gpus[node_id])), - "VLLM_TRACE_FUNCTION": - str(envs.VLLM_TRACE_FUNCTION), - "VLLM_USE_V1": - str(int(envs.VLLM_USE_V1)), - **({ - "VLLM_ATTENTION_BACKEND": envs.VLLM_ATTENTION_BACKEND - } if envs.VLLM_ATTENTION_BACKEND is not None else {}) - }, ) for (node_id, _) in worker_node_and_gpu_ids] - - self._env_vars_for_all_workers = ( - all_args_to_update_environment_variables) - - self._run_workers("update_environment_variables", - all_args=self._get_env_vars_to_be_updated()) - - if len(node_gpus) == 1: - # in single node case, we don't need to get the IP address. - # the loopback address is sufficient - # NOTE: a node may have several IP addresses, one for each - # network interface. `get_ip()` might return any of them, - # while they might not work for communication inside the node - # if the network setup is complicated. Using the loopback address - # solves this issue, as it always works for communication inside - # the node. - driver_ip = "127.0.0.1" - distributed_init_method = get_distributed_init_method( - driver_ip, get_open_port()) - - # Initialize the actual workers inside worker wrapper. - init_worker_all_kwargs = [ - self._get_worker_kwargs( - local_rank=node_workers[node_id].index(rank), - rank=rank, - distributed_init_method=distributed_init_method, - ) for rank, (node_id, _) in enumerate(worker_node_and_gpu_ids) - ] - self._run_workers("init_worker", all_kwargs=init_worker_all_kwargs) - self._run_workers("initialize") - self._run_workers("load_model") - - def _configure_ray_workers_use_nsight(self, - ray_remote_kwargs) -> Dict[str, Any]: - # If nsight profiling is enabled, we need to set the profiling - # configuration for the ray workers as runtime env. - runtime_env = ray_remote_kwargs.setdefault("runtime_env", {}) - runtime_env.update({ - "nsight": { - "t": "cuda,cudnn,cublas", - "o": "'worker_process_%p'", - "cuda-graph-trace": "node", - } - }) - - return ray_remote_kwargs - - def _get_env_vars_to_be_updated(self): - return self._env_vars_for_all_workers - - def _get_worker_kwargs( - self, - local_rank: int = 0, - rank: int = 0, - distributed_init_method: Optional[str] = None) -> Dict[str, Any]: - """ - Return worker init args for a given rank. - """ - if distributed_init_method is None: - distributed_init_method = get_distributed_init_method( - get_ip(), get_open_port()) - return dict( - vllm_config=self.vllm_config, - local_rank=local_rank, - rank=rank, - distributed_init_method=distributed_init_method, - ) - - def determine_num_available_blocks(self) -> Tuple[int, int]: - """ - Determine the number of available KV blocks. - - This invokes `determine_num_available_blocks` on each worker and takes - the min of the results, guaranteeing that the selected cache sizes are - compatible with all workers. - - Returns: - - tuple[num_gpu_blocks, num_cpu_blocks] - """ - # Get the maximum number of blocks that can be allocated on GPU and CPU. - num_blocks = self._run_workers("determine_num_available_blocks") - - # Since we use a shared centralized controller, we take the minimum - # number of blocks across all workers to make sure all the memory - # operators can be applied to all workers. - num_gpu_blocks = min(b[0] for b in num_blocks) - num_cpu_blocks = min(b[1] for b in num_blocks) - - return num_gpu_blocks, num_cpu_blocks - - def initialize(self, num_gpu_blocks: int) -> None: - """ - Initialize the KV cache in all workers. - """ - # NOTE: This is logged in the executor because there can be >1 worker - # with other executors. We could log in the engine level, but work - # remains to abstract away the device for non-GPU configurations. - logger.info("# GPU blocks: %d", num_gpu_blocks) - self._run_workers("initialize_cache", num_gpu_blocks) - self._run_workers("compile_or_warm_up_model") - - def _run_workers( - self, - method: str, - *args, - all_args: Optional[List[Tuple[Any, ...]]] = None, - all_kwargs: Optional[List[Dict[str, Any]]] = None, - **kwargs, - ) -> Any: - """ - Runs the given method on all workers. Can be used in the following - ways: - - Args: - - args/kwargs: All workers share the same args/kwargs - - all_args/all_kwargs: args/kwargs for each worker are specified - individually - """ - count = len(self.workers) - all_worker_args = repeat(args, count) if all_args is None \ - else islice(all_args, 0, None) - all_worker_kwargs = repeat(kwargs, count) if all_kwargs is None \ - else islice(all_kwargs, 0, None) - - ray_worker_refs = [ - worker.execute_method.remote( # type: ignore[attr-defined] - method, *worker_args, **worker_kwargs) - for (worker, worker_args, worker_kwargs - ) in zip(self.workers, all_worker_args, all_worker_kwargs) - ] - return ray.get(ray_worker_refs) - - def execute_model( - self, - scheduler_output, - ) -> ModelRunnerOutput: - if self.forward_dag is None: - self.forward_dag = self._compiled_ray_dag() - # Only the first worker (with rank 0) returns the execution result. - # Others return None. - output = ray.get(self.forward_dag.execute(scheduler_output))[0] - return output - - def profile(self, is_start=True): - raise NotImplementedError - - def shutdown(self): - if hasattr(self, "forward_dag") and self.forward_dag is not None: - self.forward_dag.teardown() - import ray - for worker in self.workers: - ray.kill(worker) - self.forward_dag = None - - def check_health(self) -> None: - logger.debug("Called check_health.") - - def _check_ray_compiled_graph_installation(self): - import pkg_resources - from packaging import version - - required_version = version.parse("2.39") - current_version = version.parse( - pkg_resources.get_distribution("ray").version) - if current_version < required_version: - raise ValueError(f"Ray version {required_version} is " - f"required, but found {current_version}") - - import importlib.util - raycg = importlib.util.find_spec("ray.experimental.compiled_dag_ref") - if raycg is None: - raise ValueError("Ray Compiled Graph is not installed. " - "Run `pip install ray[adag]` to install it.") - - cupy_spec = importlib.util.find_spec("cupy") - if cupy_spec is None and envs.VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL: - raise ValueError( - "cupy is not installed but required since " - "VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL is set." - "Run `pip install ray[adag]` and check cupy installation.") - - def _compiled_ray_dag(self): - assert self.parallel_config.use_ray - self._check_ray_compiled_graph_installation() - from ray.dag import InputNode, MultiOutputNode - - with InputNode() as input_batches: - outputs = [ - worker.execute_model.bind( # type: ignore[attr-defined] - input_batches) for worker in self.workers - ] - forward_dag = MultiOutputNode(outputs) - - return forward_dag.experimental_compile() - - def __del__(self): - self.shutdown() diff --git a/vllm/v1/executor/ray_utils.py b/vllm/v1/executor/ray_utils.py deleted file mode 100644 index fc9715b7a5909..0000000000000 --- a/vllm/v1/executor/ray_utils.py +++ /dev/null @@ -1,280 +0,0 @@ -import time -from collections import defaultdict -from typing import TYPE_CHECKING, Dict, List, Optional, Tuple - -from vllm.config import ParallelConfig -from vllm.logger import init_logger -from vllm.platforms import current_platform -from vllm.utils import get_ip -from vllm.v1.outputs import ModelRunnerOutput -from vllm.worker.worker_base import WorkerWrapperBase - -if TYPE_CHECKING: - from vllm.v1.core.scheduler import SchedulerOutput - -logger = init_logger(__name__) -PG_WAIT_TIMEOUT = 60 - -try: - import ray - from ray.util import placement_group_table - from ray.util.placement_group import PlacementGroup - try: - from ray._private.state import available_resources_per_node - except ImportError: - # Ray 2.9.x doesn't expose `available_resources_per_node` - from ray._private.state import state as _state - available_resources_per_node = _state._available_resources_per_node - - class RayWorkerWrapper(WorkerWrapperBase): - - def __init__(self, *args, **kwargs) -> None: - super().__init__(*args, **kwargs) - # Since the compiled DAG runs a main execution - # in a different thread that calls cuda.set_device. - # The flag indicates is set_device is called on - # that thread. It will be removed soon. - self.compiled_dag_cuda_device_set = False - - def get_node_ip(self) -> str: - return get_ip() - - def get_node_and_gpu_ids(self) -> Tuple[str, List[int]]: - node_id = ray.get_runtime_context().get_node_id() - device_key = current_platform.ray_device_key - if not device_key: - raise RuntimeError("current platform %s does not support ray.", - current_platform.device_name) - gpu_ids = ray.get_runtime_context().get_accelerator_ids( - )[device_key] - return node_id, gpu_ids - - def setup_device_if_necessary(self): - # TODO(swang): This is needed right now because Ray CG executes - # on a background thread, so we need to reset torch's current - # device. - # We can remove this API after it is fixed in compiled graph. - import torch - assert self.worker is not None, "Worker is not initialized" - if not self.compiled_dag_cuda_device_set: - torch.cuda.set_device(self.worker.device) - self.compiled_dag_cuda_device_set = True - - def execute_model( - self, - scheduler_output: "SchedulerOutput", - ) -> ModelRunnerOutput: - self.setup_device_if_necessary() - assert self.worker is not None, "Worker is not initialized" - output = self.worker.model_runner.execute_model(scheduler_output) - return output - - ray_import_err = None - -except ImportError as e: - ray = None # type: ignore - ray_import_err = e - RayWorkerWrapper = None # type: ignore - - -def ray_is_available() -> bool: - """Returns True if Ray is available.""" - return ray is not None - - -def assert_ray_available(): - """ - Raise an exception if Ray is not available. - """ - if ray is None: - raise ValueError("Failed to import Ray, please install Ray with " - "`pip install ray`.") from ray_import_err - - -def _verify_bundles(placement_group: "PlacementGroup", - parallel_config: ParallelConfig, device_str: str): - """ - Verify a given placement group has bundles located in the right place. - - There are 2 rules. - - Warn if all tensor parallel workers cannot fit in a single node. - - Fail if driver node is not included in a placement group. - - Args: - placement_group: The placement group to verify. - parallel_config: The parallel configuration. - device_str: The required device. - """ - assert ray.is_initialized(), ( - "Ray is not initialized although distributed-executor-backend is ray.") - pg_data = placement_group_table(placement_group) - # bundle_idx -> node_id - bundle_to_node_ids = pg_data["bundles_to_node_id"] - # bundle_idx -> bundle (e.g., {"GPU": 1}) - bundles = pg_data["bundles"] - # node_id -> List of bundle (e.g., {"GPU": 1}) - node_id_to_bundle: Dict[str, List[Dict[str, float]]] = defaultdict(list) - - for bundle_idx, node_id in bundle_to_node_ids.items(): - node_id_to_bundle[node_id].append(bundles[bundle_idx]) - driver_node_id = ray.get_runtime_context().get_node_id() - - if driver_node_id not in node_id_to_bundle: - raise RuntimeError( - f"driver node id {driver_node_id} is not included in a placement " - f"group {placement_group.id}. Node id -> bundles " - f"{node_id_to_bundle}. " - "You don't have enough GPUs available in a current node. Check " - "`ray status` to see if you have available GPUs in a node " - f"{driver_node_id} before starting an vLLM engine.") - - for node_id, bundles in node_id_to_bundle.items(): - if len(bundles) < parallel_config.tensor_parallel_size: - logger.warning( - "tensor_parallel_size=%d " - "is bigger than a reserved number of %ss (%d " - "%ss) in a node %s. Tensor parallel workers can be " - "spread out to 2+ nodes which can degrade the performance " - "unless you have fast interconnect across nodes, like " - "Infiniband. To resolve this issue, make sure you have more " - "than %d GPUs available at each node.", - parallel_config.tensor_parallel_size, device_str, len(bundles), - device_str, node_id, parallel_config.tensor_parallel_size) - - -def _wait_until_pg_ready(current_placement_group: "PlacementGroup"): - """Wait until a placement group is ready. - - It prints the informative log messages if the placement group is - not created within time. - - """ - # Wait until PG is ready - this will block until all - # requested resources are available, and will timeout - # if they cannot be provisioned. - placement_group_specs = current_placement_group.bundle_specs - - s = time.time() - pg_ready_ref = current_placement_group.ready() - wait_interval = 10 - while time.time() - s < PG_WAIT_TIMEOUT: - ready, _ = ray.wait([pg_ready_ref], timeout=wait_interval) - if len(ready) > 0: - break - - # Exponential backoff for warning print. - wait_interval *= 2 - logger.info( - "Waiting for creating a placement group of specs for " - "%d seconds. specs=%s. Check " - "`ray status` to see if you have enough resources.", - int(time.time() - s), placement_group_specs) - - try: - ray.get(pg_ready_ref, timeout=0) - except ray.exceptions.GetTimeoutError: - raise ValueError( - "Cannot provide a placement group of " - f"{placement_group_specs=} within {PG_WAIT_TIMEOUT} seconds. See " - "`ray status` to make sure the cluster has enough resources." - ) from None - - -def initialize_ray_cluster( - parallel_config: ParallelConfig, - ray_address: Optional[str] = None, -): - """Initialize the distributed cluster with Ray. - - it will connect to the Ray cluster and create a placement group - for the workers, which includes the specification of the resources - for each distributed worker. - - Args: - parallel_config: The configurations for parallel execution. - ray_address: The address of the Ray cluster. If None, uses - the default Ray cluster address. - """ - assert_ray_available() - - # Connect to a ray cluster. - if current_platform.is_rocm() or current_platform.is_xpu(): - # Try to connect existing ray instance and create a new one if not found - try: - ray.init("auto") - except ConnectionError: - logger.warning( - "No existing RAY instance detected. " - "A new instance will be launched with current node resources.") - ray.init(address=ray_address, - ignore_reinit_error=True, - num_gpus=parallel_config.world_size) - else: - ray.init(address=ray_address, ignore_reinit_error=True) - - if parallel_config.placement_group: - # Placement group is already set. - return - - device_str = current_platform.ray_device_key - if not device_str: - raise ValueError( - f"current platform {current_platform.device_name} does not " - "support ray.") - # Create placement group for worker processes - current_placement_group = ray.util.get_current_placement_group() - if current_placement_group: - # We are in a placement group - bundles = current_placement_group.bundle_specs - # Verify that we can use the placement group. - device_bundles = 0 - for bundle in bundles: - bundle_devices = bundle.get(device_str, 0) - if bundle_devices > 1: - raise ValueError( - "Placement group bundle cannot have more than 1 " - f"{device_str}.") - if bundle_devices: - device_bundles += 1 - if parallel_config.world_size > device_bundles: - raise ValueError( - f"The number of required {device_str}s exceeds the total " - f"number of available {device_str}s in the placement group." - f"Required number of devices: {parallel_config.world_size}. " - f"Total number of devices: {device_bundles}.") - else: - num_devices_in_cluster = ray.cluster_resources().get(device_str, 0) - if parallel_config.world_size > num_devices_in_cluster: - raise ValueError( - f"The number of required {device_str}s exceeds the total " - f"number of available {device_str}s in the placement group.") - # Create a new placement group - placement_group_specs: List[Dict[str, float]] = ([{ - device_str: 1.0 - } for _ in range(parallel_config.world_size)]) - - # vLLM engine is also a worker to execute model with an accelerator, - # so it requires to have the device in a current node. Check if - # the current node has at least one device. - current_ip = get_ip() - current_node_id = ray.get_runtime_context().get_node_id() - current_node_resource = available_resources_per_node()[current_node_id] - if current_node_resource.get(device_str, 0) < 1: - raise ValueError( - f"Current node has no {device_str} available. " - f"{current_node_resource=}. vLLM engine cannot start without " - f"{device_str}. Make sure you have at least 1 {device_str} " - f"available in a node {current_node_id=} {current_ip=}.") - # This way, at least bundle is required to be created in a current - # node. - placement_group_specs[0][f"node:{current_ip}"] = 0.001 - - # By default, Ray packs resources as much as possible. - current_placement_group = ray.util.placement_group( - placement_group_specs, strategy="PACK") - _wait_until_pg_ready(current_placement_group) - - assert current_placement_group is not None - _verify_bundles(current_placement_group, parallel_config, device_str) - # Set the placement group in the parallel config - parallel_config.placement_group = current_placement_group diff --git a/vllm/v1/executor/uniproc_executor.py b/vllm/v1/executor/uniproc_executor.py deleted file mode 100644 index be058318de58b..0000000000000 --- a/vllm/v1/executor/uniproc_executor.py +++ /dev/null @@ -1,84 +0,0 @@ -import os -from typing import Optional, Tuple - -from vllm.config import VllmConfig -from vllm.logger import init_logger -from vllm.utils import get_distributed_init_method, get_ip, get_open_port -from vllm.v1.executor.abstract import Executor -from vllm.v1.outputs import ModelRunnerOutput -from vllm.v1.worker.gpu_worker import Worker - -logger = init_logger(__name__) - - -class UniprocExecutor(Executor): - - def __init__(self, vllm_config: VllmConfig) -> None: - self.vllm_config = vllm_config - self.model_config = vllm_config.model_config - self.cache_config = vllm_config.cache_config - self.lora_config = vllm_config.lora_config - self.load_config = vllm_config.load_config - self.parallel_config = vllm_config.parallel_config - self.scheduler_config = vllm_config.scheduler_config - self.device_config = vllm_config.device_config - self.speculative_config = vllm_config.speculative_config - self.prompt_adapter_config = vllm_config.prompt_adapter_config - self.observability_config = vllm_config.observability_config - - self.worker: Worker = self._create_worker() - self.worker.initialize() - self.worker.load_model() - - def _create_worker( - self, - local_rank: int = 0, - rank: int = 0, - distributed_init_method: Optional[str] = None) -> Worker: - """Return worker init args for a given rank.""" - # see https://github.com/NVIDIA/nccl/issues/1234 - os.environ['NCCL_CUMEM_ENABLE'] = '0' - - if distributed_init_method is None: - distributed_init_method = get_distributed_init_method( - get_ip(), get_open_port()) - return Worker( - vllm_config=self.vllm_config, - local_rank=local_rank, - rank=rank, - distributed_init_method=distributed_init_method, - ) - - def determine_num_available_blocks(self) -> Tuple[int, int]: - """Determine the number of available KV blocks by invoking the - underlying worker. - """ - return self.worker.determine_num_available_blocks() - - def initialize(self, num_gpu_blocks: int) -> None: - """Initialize the KV cache by invoking the underlying worker. - """ - # NOTE: This is logged in the executor because there can be >1 worker - # with other executors. We could log in the engine level, but work - # remains to abstract away the device for non-GPU configurations. - logger.info("# GPU blocks: %d", num_gpu_blocks) - self.worker.initialize_cache(num_gpu_blocks) - self.worker.compile_or_warm_up_model() - - def execute_model( - self, - scheduler_output, - ) -> ModelRunnerOutput: - output = self.worker.execute_model(scheduler_output) - return output - - def profile(self, is_start: bool = True): - self.worker.profile(is_start) - - def shutdown(self): - pass - - def check_health(self) -> None: - # UniprocExecutor will always be healthy as long as - # it's running. - return diff --git a/vllm/v1/kv_cache_interface.py b/vllm/v1/kv_cache_interface.py new file mode 100644 index 0000000000000..6d5cc32ffc5b8 --- /dev/null +++ b/vllm/v1/kv_cache_interface.py @@ -0,0 +1,111 @@ +from dataclasses import dataclass +from typing import Dict, List + +import torch + +from vllm.logger import init_logger +from vllm.utils import cdiv, get_dtype_size + +logger = init_logger(__name__) + + +@dataclass +class KVCacheSpecBase: + """ + A base class for specifying the KV cache format of one layer. + """ + + # number of tokens in a block + block_size: int + + @property + def type_id(self) -> str: + """ + The type identifier of this KV cache. + Return different strings for layers with different KV cache type (e.g., + different number of tokens like full attention vs sliding window + attention, different KV cache size per token like layers with different + number of heads) + + Returns: + The type identifier of this KV cache. + """ + raise NotImplementedError + + @property + def page_size_bytes(self) -> int: + """ + The size of a page with `block_size` tokens in bytes. + + Returns: + The page size + """ + raise NotImplementedError + + def bytes_for_tokens(self, num_tokens: int) -> int: + """ + The KV cache size for `num_tokens` tokens in bytes. Returns the real + memory size after padding `num_tokens` to full blocks. + + Returns: + The KV cache size + """ + raise NotImplementedError + + +@dataclass +class FullAttentionSpec(KVCacheSpecBase): + num_kv_heads: int + head_size: int + dtype: torch.dtype + + @property + def type_id(self) -> str: + return f"full_attention_{self.block_size}_{self.page_size_bytes}" + + @property + def page_size_bytes(self) -> int: + return 2 * self.block_size * self.num_kv_heads * self.head_size \ + * get_dtype_size(self.dtype) + + def bytes_for_tokens(self, num_tokens: int) -> int: + return cdiv(num_tokens, self.block_size) * self.page_size_bytes + + +KVCacheSpec = Dict[str, KVCacheSpecBase] + + +@dataclass +class KVCacheTensor: + """ + A dataclass for specifying how the workers should initialize the KV cache + for a layer. Only contains the size of KV cache for that layer for now. Will + be extended to support multiple layers sharing the same memory pool. + """ + size: int # The size of KV cache Tensor in bytes + + +@dataclass +class KVCacheConfig: + """ + The KV cache configuration of a model. + """ + """The number of KV cache blocks""" + num_blocks: int + """layer_name -> how to initialize KV cache for that layer""" + tensors: Dict[str, KVCacheTensor] + """ + A list of kv-cache groups. Each group includes a set of layers with + the same kv-cache spec, and the total page_size of layers inside a group + is same across all groups (as the KVCacheManager only supports allocating + pages of the same size). For example: + 1. A model only uses full attention: one group with all layers in the model. + 2. (not implemented yet) A model with the same number of full attention + layers and sliding window attention layers: two groups, one for full + attention layers and one for sliding window attention layers. + 3. (not implemented yet) A model with 2 full attention layers and 4 sliding + window attention layers: three groups, (full * 2), (sw * 2), (sw * 2). + """ + groups: List[List[str]] + """the KVCacheSpec of the model""" + kv_cache_spec: KVCacheSpec diff --git a/vllm/v1/metrics/loggers.py b/vllm/v1/metrics/loggers.py index 8feeef17542e6..b84f03fa3267c 100644 --- a/vllm/v1/metrics/loggers.py +++ b/vllm/v1/metrics/loggers.py @@ -1,5 +1,8 @@ import time from abc import ABC, abstractmethod +from typing import Dict + +import prometheus_client from vllm.logger import init_logger from vllm.v1.metrics.stats import SchedulerStats @@ -36,3 +39,36 @@ def log(self, scheduler_stats: SchedulerStats): scheduler_stats.num_running_reqs, scheduler_stats.num_waiting_reqs, ) + + +class PrometheusStatLogger(StatLoggerBase): + + def __init__(self, labels: Dict[str, str]): + self.labels = labels + + labelnames = self.labels.keys() + labelvalues = self.labels.values() + + self._unregister_vllm_metrics() + + self.gauge_scheduler_running = prometheus_client.Gauge( + name="vllm:num_requests_running", + documentation="Number of requests in model execution batches.", + labelnames=labelnames).labels(*labelvalues) + + self.gauge_scheduler_waiting = prometheus_client.Gauge( + name="vllm:num_requests_waiting", + documentation="Number of requests waiting to be processed.", + labelnames=labelnames).labels(*labelvalues) + + def log(self, scheduler_stats: SchedulerStats): + """Log to prometheus.""" + self.gauge_scheduler_running.set(scheduler_stats.num_running_reqs) + self.gauge_scheduler_waiting.set(scheduler_stats.num_waiting_reqs) + + @staticmethod + def _unregister_vllm_metrics(): + # Unregister any existing vLLM collectors (for CI/CD + for collector in list(prometheus_client.REGISTRY._collector_to_names): + if hasattr(collector, "_name") and "vllm" in collector._name: + prometheus_client.REGISTRY.unregister(collector) diff --git a/vllm/v1/request.py b/vllm/v1/request.py index 45450165eaefe..2cfcd8b63ccb2 100644 --- a/vllm/v1/request.py +++ b/vllm/v1/request.py @@ -58,12 +58,19 @@ def __init__( # Sanity check assert len(self.mm_inputs) == len(self.mm_positions) - assert len(self.mm_inputs) == len(self.mm_hashes) + if self.mm_hashes: + assert len(self.mm_inputs) == len(self.mm_hashes) # Cache the computed kv block hashes of the request to avoid # recomputing. self._kv_block_hashes: List[BlockHashType] = [] + # Read-only views + # Prevent directly appending to the these lists since + # they should also be updated simultaneously. + self.output_token_ids = ConstantList(self._output_token_ids) + self.all_token_ids = ConstantList(self._all_token_ids) + @classmethod def from_engine_core_request(cls, request: EngineCoreRequest) -> "Request": return cls( @@ -79,18 +86,6 @@ def from_engine_core_request(cls, request: EngineCoreRequest) -> "Request": lora_request=request.lora_request, ) - @property - def output_token_ids(self) -> ConstantList[int]: - # Prevent directly appending to the output_token_ids since - # all_token_ids should also be updated simultaneously. - return ConstantList(self._output_token_ids) - - @property - def all_token_ids(self) -> ConstantList[int]: - # Prevent directly appending to the all_token_ids since - # output_token_ids should also be updated simultaneously - return ConstantList(self._all_token_ids) - def append_output_token_ids( self, token_ids: Union[int, List[int]], diff --git a/vllm/v1/sample/sampler.py b/vllm/v1/sample/sampler.py index 7cd42ca211a22..d420f7f74e109 100644 --- a/vllm/v1/sample/sampler.py +++ b/vllm/v1/sample/sampler.py @@ -50,7 +50,6 @@ def forward( # Use int32 to reduce the tensor size. sampled = sampled.to(torch.int32) - # NOTE: CPU-GPU synchronization happens here. sampler_output = SamplerOutput( sampled_token_ids=sampled.tolist(), logprob_token_ids=topk_indices, diff --git a/vllm/v1/stats/__init__.py b/vllm/v1/stats/__init__.py new file mode 100644 index 0000000000000..e69de29bb2d1d diff --git a/vllm/v1/stats/common.py b/vllm/v1/stats/common.py new file mode 100644 index 0000000000000..500bc356fc179 --- /dev/null +++ b/vllm/v1/stats/common.py @@ -0,0 +1,451 @@ +import time +from dataclasses import dataclass +from dataclasses import field as dataclass_field +from enum import IntEnum +from typing import ClassVar, Dict, List, Optional, Set + +import msgspec +from msgspec import field as msgspec_field + +from vllm.sampling_params import SamplingParams + + +class RequestStatsUpdate( + msgspec.Struct, # type: ignore + array_like=True, + omit_defaults=True, + gc=False): + """ + An update to the request stats. + + This represents a stats update at a specific timestamp with metadata + associated with the update. + + NOTE: since there might be multiple processes generating updates at + different parts of the engine (e.g. input processor, scheduler, engine core, + etc.), we use the monotonic timestamp to record the update to compute any + intervals, and explicit wall-clock timestamp should be used for timestamps. + + WARNING: This assumes stats are generated in a single machine. If there are + potentially multiple machines, one should always generate the stats updates + on one single machine or use something else. + """ + + class Type(IntEnum): + """See `RequestStats` for the lifecycle of a request.""" + + # Request arrived at the engine frontend. + ARRIVED = 0 + # Input processed by the input processor. + INPUT_PROCESSED = 1 + # Queued on the engine core. + QUEUED = 2 + # Scheduled running prefill by the scheduler. + # A request could be running a new prefill on the prompt tokens or + # a resumed prefill on the original prefill tokens + generated output + # tokens before preemption. + PREFILLING = 3 + # Preempted by the scheduler. + PREEMPTED = 4 + # Output token is generated by the engine core. + DECODING = 5 + # Token detokenized by the detokenizer. + # We will record the timestamp for each output token, as well as the + # finish reason. + DETOKENIZED = 6 + # Request finishes (or aborts). + FINISHED = 7 + + """ + Valid state updates: + ARRIVED + │ + ├──────► INPUT_PROCESSED ──────► QUEUED ──────► PREFILLING ◄────┐ + │ │ │ │ │ + │ │ │ ▼ │ + │ │ │ -──► DECODING │ + │ │ │ | │ │ + │ │ │ | ▼ │ + │ │ │ └─ DETOKENIZED │ + │ │ │ │ │ + │ │ │ ▼ │ + │ ▼ ▼ PREEMPTED ◄──────┘ + │ │ │ │ + └──────────────┴───────────────────┴──────────────┴ + │ + ▼ + FINISHED (All could go to FINISHED) + """ + _VALID_TRANSITIONS: ClassVar[Dict[Type, Set[Type]]] = { + Type.ARRIVED: { + Type.INPUT_PROCESSED, + Type.FINISHED, + }, + Type.INPUT_PROCESSED: { + Type.QUEUED, + Type.FINISHED, + }, + Type.QUEUED: { + Type.PREFILLING, + Type.FINISHED, + }, + Type.PREFILLING: { + Type.DECODING, + Type.PREEMPTED, + Type.FINISHED, + }, + Type.DECODING: { + Type.DETOKENIZED, + Type.FINISHED, + }, + Type.DETOKENIZED: { + Type.DECODING, + Type.PREEMPTED, + Type.FINISHED, + }, + Type.PREEMPTED: {Type.PREFILLING, Type.FINISHED}, + Type.FINISHED: set(), + } + + request_id: str + + type: Type + + # Timestamp when the update is recorded. This is used to record time + # intervals between events rather than wall clock time. + monotonic_ts_s: float = msgspec_field( + default_factory=lambda: time.monotonic()) + + ############################################################ + # Metadata associated with the update. + ############################################################ + # For input_processed. Metadata needed for stats logging. + num_prompt_tokens: Optional[int] = None + sampling_params: Optional[SamplingParams] = None + + # For running. + # Number of tokens computed when scheduled to run. + num_computed_tokens: Optional[int] = None + # Number of cached tokens when scheduled to run. + num_cached_tokens: Optional[int] = None + + # For decoded. + # The number of new output tokens generated. + num_new_tokens: Optional[int] = None + + # For both detokenized and decoded. + # Finished reason. + finish_reason: Optional[str] = None + + # Non-optional fields for each update type. + _REQUIRED_FIELDS: ClassVar[Dict[Type, List[str]]] = { + Type.INPUT_PROCESSED: ["num_prompt_tokens", "sampling_params"], + Type.PREFILLING: ["num_computed_tokens", "num_cached_tokens"], + Type.DETOKENIZED: ["num_new_tokens"], + Type.FINISHED: ["finish_reason"], + } + + def __post_init__(self): + required_fields = self._REQUIRED_FIELDS.get(self.type, []) + for field in required_fields: + if getattr(self, field) is None: + raise ValueError( + f"Field {field} is required for update type {self.type}.") + + @staticmethod + def check_valid_update( + update: "RequestStatsUpdate", + last_update_type: Optional[Type], + last_updated_ts_s: Optional[float], + ): + if last_update_type is None: + assert update.type == RequestStatsUpdate.Type.ARRIVED + else: + valid_cur_update_types = RequestStatsUpdate._VALID_TRANSITIONS[ + last_update_type] + assert update.type in valid_cur_update_types, ( + f"Invalid update type: {update.type} for last_update_type: " + f"{last_update_type}.") + + if last_updated_ts_s is not None: + assert update.monotonic_ts_s >= last_updated_ts_s, ( + "Update timestamp must be monotonically increasing, but " + f"last_updated_ts_s={last_updated_ts_s} and " + f"update.monotonic_ts_s={update.monotonic_ts_s}.") + + +@dataclass +class RequestStats: + """Stats associated with a request (`Request`).""" + + ############################################################ + # Metadata + ############################################################ + request_id: str + sampling_params: Optional[SamplingParams] = None + num_prompt_tokens: Optional[int] = None + + ############################################################ + # Metrics and Stats + ############################################################ + # Timestamp when the request was last updated. + last_updated_ts_s: Optional[float] = None + + # Last update stats type. + last_update_type: Optional[RequestStatsUpdate.Type] = None + + # Timestamp when the request arrived at the llm engine. + arrival_ts_s: Optional[float] = None + + # Number of tokens cached. When part of the request prefix is cached, + # this will be set. + num_cached_tokens: int = 0 + + # Number of tokens computed. + num_computed_tokens: int = 0 + + # The timestamp when the request become waiting in the queue. + queued_ts_s: Optional[float] = None + + # When the input processor is completed. + input_processor_end_ts_s: Optional[float] = None + + # A sorted list of timestamps when the request was scheduled to prefill. + # This could be when: + # 1. the request is newly scheduled, so it's a new prefill. + # 2. the request was preempted and resumed. It is equivalent to running + # a prefill of the original prefill tokens + generated output tokens + # before preemption. + prefill_start_ts_s_lst: List[float] = dataclass_field(default_factory=list) + + # A list of timestamps when a token is decoded by the engine core. + decoding_ts_s_lst: List[float] = dataclass_field(default_factory=list) + + # A sorted list of timestamps for each output token. + output_token_ts_s_lst: List[float] = dataclass_field(default_factory=list) + + # First token's timestamp. + first_token_ts_s: Optional[float] = None + + # TODO(rickyx): we need model runner to surface these. + model_forward_duration_s: float = 0.0 + # Includes model forward, block/sync across workers, cpu-gpu sync time + # and sampling time. + model_execute_duration_s: float = 0.0 + + # A sorted list of timestamps when the request was preempted at the + # scheduler. + # TODO(rickyx): right now, we don't actually have a good high-level + # metric to measure the impact of preemption other than observation of + # large P99 TPOT. Ideally we could quantify the impact of preemption by + # measuring the number of tokens re-computed due to preemption. + preempted_ts_s_lst: List[float] = dataclass_field(default_factory=list) + + # Timestamp when the request was finished at the engine core. + finished_ts_s: Optional[float] = None + + # Finish reason. + finish_reason: Optional[str] = None + + ############################################################ + # Derived properties. + ############################################################ + @property + def prefill_ts_s(self) -> Optional[float]: + """The timestamp when the request started prefilling. + Since a request could be preempted in decoding and later resumed + to prefill the decoded tokens, we use the first prefill start timestamp. + """ + return (self.prefill_start_ts_s_lst[0] + if self.prefill_start_ts_s_lst else None) + + @property + def e2e_latency_s(self) -> Optional[float]: + if self.finished_ts_s is None or self.arrival_ts_s is None: + return None + assert self.finished_ts_s >= self.arrival_ts_s + return self.finished_ts_s - self.arrival_ts_s + + @property + def queue_duration_s(self) -> Optional[float]: + """How long the request was waiting to run.""" + if self.queued_ts_s is None or self.prefill_ts_s is None: + # Either not queued or not running yet. + return None + assert self.queued_ts_s <= self.prefill_ts_s + return self.prefill_ts_s - self.queued_ts_s + + @property + def inference_latency_s(self) -> Optional[float]: + """How long the request was running inference + (prefill and decode).""" + if self.finished_ts_s is None or self.prefill_ts_s is None: + return None + assert self.finished_ts_s >= self.prefill_ts_s + return self.finished_ts_s - self.prefill_ts_s + + @property + def first_token_latency_s(self) -> Optional[float]: + if self.first_token_ts_s is None or self.arrival_ts_s is None: + return None + assert self.first_token_ts_s >= self.arrival_ts_s + return self.first_token_ts_s - self.arrival_ts_s + + @property + def prefill_latency_s(self) -> Optional[float]: + if self.first_token_ts_s is None or self.prefill_ts_s is None: + return None + assert self.first_token_ts_s >= self.prefill_ts_s + return self.first_token_ts_s - self.prefill_ts_s + + @property + def decode_latency_s(self) -> Optional[float]: + if self.e2e_latency_s is None or self.first_token_latency_s is None: + return None + assert self.e2e_latency_s >= self.first_token_latency_s + return self.e2e_latency_s - self.first_token_latency_s + + @property + def output_token_latency_s_lst(self) -> List[float]: + if len(self.output_token_ts_s_lst) == 0: + return [] + latency_s_lst = [] + for i in range(1, len(self.output_token_ts_s_lst)): + assert (self.output_token_ts_s_lst[i] >= + self.output_token_ts_s_lst[i - 1]) + latency_s = (self.output_token_ts_s_lst[i] - + self.output_token_ts_s_lst[i - 1]) + latency_s_lst.append(latency_s) + return latency_s_lst + + @property + def num_output_tokens(self) -> int: + return len(self.output_token_ts_s_lst) + + @property + def is_finished(self) -> bool: + return self.finished_ts_s is not None + + def update_from(self, update: "RequestStatsUpdate"): + RequestStatsUpdate.check_valid_update(update, self.last_update_type, + self.last_updated_ts_s) + ts = update.monotonic_ts_s + self.last_updated_ts_s = ts + self.last_update_type = update.type + if update.type == RequestStatsUpdate.Type.ARRIVED: + self.arrival_ts_s = ts + elif update.type == RequestStatsUpdate.Type.INPUT_PROCESSED: + self.input_processor_end_ts_s = ts + self.sampling_params = update.sampling_params + self.num_prompt_tokens = update.num_prompt_tokens + elif update.type == RequestStatsUpdate.Type.QUEUED: + self.queued_ts_s = ts + elif update.type == RequestStatsUpdate.Type.PREFILLING: + self.prefill_start_ts_s_lst.append(ts) + self.num_cached_tokens = update.num_cached_tokens or 0 + self.num_computed_tokens = update.num_computed_tokens or 0 + elif update.type == RequestStatsUpdate.Type.PREEMPTED: + self._reset_for_preemption(ts) + elif update.type == RequestStatsUpdate.Type.DECODING: + self.decoding_ts_s_lst.append(ts) + elif update.type == RequestStatsUpdate.Type.DETOKENIZED: + self._record_detokenized_output( + ts, + update.num_new_tokens or 0, + ) + elif update.type == RequestStatsUpdate.Type.FINISHED: + self.finished_ts_s = ts + self.finish_reason = update.finish_reason + else: + raise ValueError(f"Unknown update type: {update.type}") + + def _record_detokenized_output( + self, + ts_s: float, + num_new_tokens: int, + ): + # Update if first output token is generated. + if len(self.output_token_ts_s_lst) == 0: + self.first_token_ts_s = ts_s + assert ( + self.prefill_ts_s is not None + ), "Request must be running before generating output tokens." + + # Some X new tokens were generated at the ts. + self.output_token_ts_s_lst.extend([ts_s] * num_new_tokens) + + def _reset_for_preemption(self, ts_s: float): + self.preempted_ts_s_lst.append(ts_s) + # Reset the computed tokens since it might restart the prefill. + self.num_computed_tokens = 0 + # Cached token count might also change when resumed. + self.num_cached_tokens = 0 + # These stats don't change since they happen before request running. + # - arrival_ts_s + # - input_processor_end_ts_s + # - sampling_params + # - num_prompt_tokens + # - first_token_ts_s + # + # These stats are accumulated over preemptions: + # - output_token_ts_s_lst + # - prefill_start_ts_s_lst (after preemption, it will prefill the + # original prefill tokens and any output tokens generated before + # preemption.) + + +@dataclass +class KVCacheStats: + # KV Cache Usage in % + gpu_cache_usage_sys: float = 0.0 + gpu_prefix_cache_hit_rate: float = 0.0 + + +@dataclass +class SchedulerStats: + """Stats associated with the scheduler.""" + + # Number of requests currently running. + num_running_reqs: int = 0 + # Number of requests currently waiting. + num_waiting_reqs: int = 0 + + kv_cache_stats: KVCacheStats = dataclass_field( + default_factory=KVCacheStats) + + +@dataclass +class EngineCoreProcessStats: + """Stats associated with the engine core process.""" + + # Number of requests currently in the input queue. None if the engine core + # is not running in multiprocess mode. + input_queue_size: Optional[int] = None + # Number of outputs currently in the output queue. None if the engine core + # is not running in multiprocess mode. + output_queue_size: Optional[int] = None + + +class EngineCoreStatsSnapshot( + msgspec.Struct, # type: ignore + array_like=True, + omit_defaults=True, + gc=False): + """ + A snapshot of the EngineCore's current stats over a period of time. + """ + + # Snapshot of the scheduler stats. + scheduler_stats: SchedulerStats = msgspec_field( + default_factory=SchedulerStats) + + # Per request stats updates. + requests_stats_updates: List[RequestStatsUpdate] = msgspec_field( + default_factory=list) + + # Engine core's queue stats. + engine_core_process_stats: EngineCoreProcessStats = msgspec_field( + default_factory=EngineCoreProcessStats) + + # TODO(rickyx): Add other components' stats, + # e.g. model runner/worker and etc. diff --git a/vllm/v1/utils.py b/vllm/v1/utils.py index b0a7affbebb7e..8dfcf2dd78606 100644 --- a/vllm/v1/utils.py +++ b/vllm/v1/utils.py @@ -1,13 +1,20 @@ import multiprocessing import os import weakref +from collections import defaultdict from collections.abc import Sequence -from typing import (Any, Callable, Dict, Generic, List, Optional, TypeVar, - Union, overload) +from typing import (TYPE_CHECKING, Any, Callable, Dict, Generic, List, + Optional, TypeVar, Union, overload) + +import torch from vllm.logger import init_logger +from vllm.model_executor.models.utils import extract_layer_index from vllm.utils import get_mp_context, kill_process_tree +if TYPE_CHECKING: + from vllm.attention.layer import Attention + logger = init_logger(__name__) T = TypeVar("T") @@ -134,3 +141,48 @@ def shutdown(proc: multiprocessing.Process, input_path: str, output_path: str): socket_file = ipc_socket.replace("ipc://", "") if os and os.path.exists(socket_file): os.remove(socket_file) + + +def bind_kv_cache( + kv_caches: Dict[str, torch.Tensor], + forward_context: Dict[str, "Attention"], + runner_kv_caches: List[torch.Tensor], +) -> None: + """ + Bind the allocated KV cache to both ModelRunner and forward context so + that the KV cache can be used in the forward pass. + + This function: + 1) Fills the ModelRunner's kv cache list (`runner_kv_caches`) with + kv_caches. + 2) Associates each attention layer in the `forward_context` with its + corresponding KV cache in kv_caches. + + Args: + kv_caches: The allocated kv_caches with layer names as keys. + forward_context: The global forward context containing all Attention + layers with layer names as keys. + runner_kv_caches: The kv_cache declared by ModelRunner. + """ + # Bind kv_caches to ModelRunner + assert len(runner_kv_caches) == 0 + + # Convert kv_caches dict to a list of tensors in the order of layer_index. + index2name = defaultdict(list) + for layer_name in kv_caches: + index2name[extract_layer_index(layer_name)].append(layer_name) + + for layer_index in sorted(index2name.keys()): + layer_names = index2name[layer_index] + if len(layer_names) > 1: + # One typical case is encoder-decoder model, e.g., bart. + # The cross attention and self attention in the same decoder layer + # has different layer_name but the same layer_index. + raise NotImplementedError + layer_name = layer_names[0] + runner_kv_caches.append(kv_caches[layer_name]) + + # Bind kv_caches to forward context + for layer_name, kv_cache in kv_caches.items(): + # NOTE: Use list because of v0 PP virtual engine. + forward_context[layer_name].kv_cache = [kv_cache] diff --git a/vllm/v1/worker/gpu_input_batch.py b/vllm/v1/worker/gpu_input_batch.py index 40494e64b22f0..28d8e39053874 100644 --- a/vllm/v1/worker/gpu_input_batch.py +++ b/vllm/v1/worker/gpu_input_batch.py @@ -30,6 +30,9 @@ class CachedRequestState: num_computed_tokens: int output_token_ids: List[int] + mrope_positions: Optional[torch.Tensor] = None + mrope_position_delta: Optional[int] = None + @property def num_tokens(self) -> int: return len(self.prompt_token_ids) + len(self.output_token_ids) diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py index fb87dc5a8222a..9d7e30079dfbb 100644 --- a/vllm/v1/worker/gpu_model_runner.py +++ b/vllm/v1/worker/gpu_model_runner.py @@ -1,28 +1,35 @@ import gc import time -from typing import TYPE_CHECKING, Dict, List, Tuple, cast +from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, cast import numpy as np import torch import torch.distributed import torch.nn as nn +from vllm.attention.backends.abstract import AttentionType +from vllm.attention.layer import Attention from vllm.config import CompilationLevel, VllmConfig from vllm.distributed.parallel_state import graph_capture from vllm.forward_context import set_forward_context from vllm.inputs import INPUT_REGISTRY from vllm.logger import init_logger +from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding from vllm.model_executor.model_loader import get_model from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs +from vllm.multimodal.utils import group_mm_inputs_by_modality from vllm.sampling_params import SamplingType from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler, - LayerBlockType, bind_kv_cache, cdiv, - is_pin_memory_available) + LayerBlockType, cdiv, is_pin_memory_available) from vllm.v1.attention.backends.flash_attn import (FlashAttentionBackend, FlashAttentionMetadata) +from vllm.v1.core.encoder_cache_manager import compute_encoder_budget from vllm.v1.engine.mm_input_mapper import MMInputMapperClient +from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig, + KVCacheSpec) from vllm.v1.outputs import ModelRunnerOutput from vllm.v1.sample.metadata import SamplingMetadata +from vllm.v1.utils import bind_kv_cache from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch if TYPE_CHECKING: @@ -88,8 +95,12 @@ def __init__( self.mm_input_mapper_profiling = MMInputMapperClient(self.model_config) self.mm_input_mapper_profiling.use_cache = False - self.max_num_encoder_input_tokens = self.scheduler_config.max_num_encoder_input_tokens # noqa: E501 - self.encoder_cache_size = self.scheduler_config.encoder_cache_size + encoder_compute_budget, encoder_cache_size = compute_encoder_budget( + model_config=model_config, + scheduler_config=scheduler_config, + ) + self.max_num_encoder_input_tokens = encoder_compute_budget + self.encoder_cache_size = encoder_cache_size # Lazy initialization # self.model: nn.Module # Set after load_model @@ -117,7 +128,8 @@ def __init__( # self.cudagraph_batch_sizes sorts in ascending order. # The batch sizes in the config are in descending order. self.cudagraph_batch_sizes = list( - reversed(self.vllm_config.compilation_config.capture_sizes)) + reversed( + self.vllm_config.compilation_config.cudagraph_capture_sizes)) # Cache the device properties. self.device_properties = torch.cuda.get_device_properties(self.device) @@ -130,6 +142,28 @@ def __init__( self.positions = torch.zeros(self.max_num_tokens, dtype=torch.int64, device=self.device) + + # Only relevant for models using M-RoPE (e.g, Qwen2-VL) + if self.model_config.uses_mrope: + # NOTE: `mrope_positions` is implemented with one additional dummy + # position on purpose to make it non-contiguous so that it can work + # with torch compile. + # See detailed explanation in https://github.com/vllm-project/vllm/pull/12128#discussion_r1926431923 + + # NOTE: When M-RoPE is enabled, position ids are 3D regardless of + # the modality of inputs. For text-only inputs, each dimension has + # identical position IDs, making M-RoPE functionally equivalent to + # 1D-RoPE. + # See page 5 of https://arxiv.org/abs/2409.12191 + self.mrope_positions = torch.zeros((3, self.max_num_tokens + 1), + dtype=torch.int64, + device=self.device) + self.mrope_positions_cpu = torch.zeros( + (3, self.max_num_tokens + 1), + dtype=torch.int64, + device="cpu", + pin_memory=self.pin_memory) + self.inputs_embeds = torch.zeros( (self.max_num_tokens, self.hidden_size), dtype=self.dtype, @@ -137,7 +171,8 @@ def __init__( # OPTIMIZATION: Cache the tensors rather than creating them every step. self.arange_np = np.arange(max(self.max_num_reqs + 1, - self.max_model_len), + self.max_model_len, + self.max_num_tokens), dtype=np.int32) # NOTE(woosuk): These tensors are "stateless", i.e., they are literally # a faster version of creating a new tensor every time. Thus, we should @@ -162,11 +197,11 @@ def __init__( device="cpu", pin_memory=self.pin_memory) self.query_start_loc_np = self.query_start_loc_cpu.numpy() - self.seq_start_loc_cpu = torch.zeros(self.max_num_reqs + 1, - dtype=torch.int32, - device="cpu", - pin_memory=self.pin_memory) - self.seq_start_loc_np = self.seq_start_loc_cpu.numpy() + self.seq_lens_cpu = torch.zeros(self.max_num_reqs, + dtype=torch.int32, + device="cpu", + pin_memory=self.pin_memory) + self.seq_lens_np = self.seq_lens_cpu.numpy() def _update_states(self, scheduler_output: "SchedulerOutput") -> None: # Remove stopped requests from the cached states. @@ -237,6 +272,35 @@ def _update_states(self, scheduler_output: "SchedulerOutput") -> None: num_computed_tokens=new_req_data.num_computed_tokens, output_token_ids=[], ) + + # Only relevant for models using M-RoPE (e.g, Qwen2-VL) + if self.model_config.uses_mrope: + image_grid_thw = [] + video_grid_thw = [] + for mm_input in self.requests[req_id].mm_inputs: + if mm_input.get("image_grid_thw") is not None: + image_grid_thw.extend( + mm_input["image_grid_thw"].tolist()) + if mm_input.get("video_grid_thw") is not None: + video_grid_thw.extend( + mm_input["video_grid_thw"].tolist()) + + hf_config = self.model_config.hf_config + + self.requests[req_id].mrope_positions, \ + self.requests[req_id].mrope_position_delta = \ + MRotaryEmbedding.get_input_positions_tensor( + self.requests[req_id].prompt_token_ids, + image_grid_thw=image_grid_thw, + video_grid_thw=video_grid_thw, + image_token_id=hf_config.image_token_id, + video_token_id=hf_config.video_token_id, + vision_start_token_id=hf_config.vision_start_token_id, + vision_end_token_id=hf_config.vision_end_token_id, + spatial_merge_size=hf_config.vision_config. + spatial_merge_size, + ) + req_ids_to_add.append(req_id) # Update the cached states of the resumed requests. @@ -295,8 +359,15 @@ def _prepare_inputs(self, scheduler_output: "SchedulerOutput"): # Get batched arange. # E.g., [2, 5, 3] -> [0, 1, 0, 1, 2, 3, 4, 0, 1, 2] - arange = np.concatenate( - [self.arange_np[:n] for n in num_scheduled_tokens]) + # Equivalent to but faster than: + # np.concatenate([np.arange(n) for n in num_scheduled_tokens]) + # Step 1. [2, 5, 3] -> [2, 7, 10] + cu_num_tokens = np.cumsum(num_scheduled_tokens) + # Step 2. [2, 7, 10] -> [0, 0, 2, 2, 2, 2, 2, 7, 7, 7] + cumsums_offsets = np.repeat(cu_num_tokens - num_scheduled_tokens, + num_scheduled_tokens) + # Step 3. [0, 1, 0, 1, 2, 3, 4, 0, 1, 2] + arange = self.arange_np[:total_num_scheduled_tokens] - cumsums_offsets # Get positions. positions_np = self.positions_np[:total_num_scheduled_tokens] @@ -304,6 +375,11 @@ def _prepare_inputs(self, scheduler_output: "SchedulerOutput"): arange, out=positions_np) + # Calculate M-RoPE positions. + # Only relevant for models using M-RoPE (e.g, Qwen2-VL) + if self.model_config.uses_mrope: + self._calc_mrope_positions(scheduler_output) + # Get token indices. # E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2] # -> [0, 1, M, M + 1, M + 2, M + 3, M + 4, 2 * M, 2 * M + 1, 2 * M + 2] @@ -338,24 +414,30 @@ def _prepare_inputs(self, scheduler_output: "SchedulerOutput"): # Prepare the attention metadata. self.query_start_loc_np[0] = 0 - np.cumsum(num_scheduled_tokens, - out=self.query_start_loc_np[1:num_reqs + 1]) + self.query_start_loc_np[1:num_reqs + 1] = cu_num_tokens - seq_lens = (self.input_batch.num_computed_tokens_cpu[:num_reqs] + - num_scheduled_tokens) - max_seq_len = seq_lens.max() - self.seq_start_loc_np[0] = 0 - np.cumsum(seq_lens, out=self.seq_start_loc_np[1:num_reqs + 1]) + self.seq_lens_np[:num_reqs] = ( + self.input_batch.num_computed_tokens_cpu[:num_reqs] + + num_scheduled_tokens) + max_seq_len = self.seq_lens_np[:num_reqs].max() # Copy the tensors to the GPU. self.input_ids[:total_num_scheduled_tokens].copy_( self.input_ids_cpu[:total_num_scheduled_tokens], non_blocking=True) - self.positions[:total_num_scheduled_tokens].copy_( - self.positions_cpu[:total_num_scheduled_tokens], non_blocking=True) + if self.model_config.uses_mrope: + # Only relevant for models using M-RoPE (e.g, Qwen2-VL) + self.mrope_positions[:, :total_num_scheduled_tokens].copy_( + self.mrope_positions_cpu[:, :total_num_scheduled_tokens], + non_blocking=True) + else: + # Common case (1D positions) + self.positions[:total_num_scheduled_tokens].copy_( + self.positions_cpu[:total_num_scheduled_tokens], + non_blocking=True) query_start_loc = self.query_start_loc_cpu[:num_reqs + 1].to( self.device, non_blocking=True) - seq_start_loc = self.seq_start_loc_cpu[:num_reqs + 1].to( - self.device, non_blocking=True) + seq_lens = self.seq_lens_cpu[:num_reqs].to(self.device, + non_blocking=True) slot_mapping = self.slot_mapping_cpu[:total_num_scheduled_tokens].to( self.device, non_blocking=True).long() @@ -427,33 +509,30 @@ def _prepare_inputs(self, scheduler_output: "SchedulerOutput"): [0, total_num_scheduled_tokens], dtype=torch.int32, device=self.device) - cu_prefix_kv_lens = torch.tensor([0, common_prefix_len], - dtype=torch.int32, - device=self.device) - cu_suffix_kv_lens = ( - self.seq_start_loc_np[:num_reqs + 1] - - self.arange_np[:num_reqs + 1] * common_prefix_len) - cu_suffix_kv_lens = torch.from_numpy(cu_suffix_kv_lens).to( - self.device) + prefix_kv_lens = torch.tensor([common_prefix_len], + dtype=torch.int32, + device=self.device) + suffix_kv_lens = (self.seq_lens_np[:num_reqs] - common_prefix_len) + suffix_kv_lens = torch.from_numpy(suffix_kv_lens).to(self.device) else: cu_prefix_query_lens = None - cu_prefix_kv_lens = None - cu_suffix_kv_lens = None + prefix_kv_lens = None + suffix_kv_lens = None attn_metadata = FlashAttentionMetadata( num_actual_tokens=total_num_scheduled_tokens, max_query_len=max_num_scheduled_tokens, query_start_loc=query_start_loc, max_seq_len=max_seq_len, - seq_start_loc=seq_start_loc, + seq_lens=seq_lens, block_table=( self.input_batch.block_table.get_device_tensor()[:num_reqs]), slot_mapping=slot_mapping, use_cascade=use_cascade, common_prefix_len=common_prefix_len, cu_prefix_query_lens=cu_prefix_query_lens, - cu_prefix_kv_lens=cu_prefix_kv_lens, - cu_suffix_kv_lens=cu_suffix_kv_lens, + prefix_kv_lens=prefix_kv_lens, + suffix_kv_lens=suffix_kv_lens, ) # NOTE(woosuk): Due to chunked prefills, there can be at most 1 partial # request in the batch. While we should not sample any token from this @@ -463,6 +542,61 @@ def _prepare_inputs(self, scheduler_output: "SchedulerOutput"): logits_indices = query_start_loc[1:] - 1 return attn_metadata, logits_indices + def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"): + mrope_pos_ptr = 0 + num_reqs = self.input_batch.num_reqs + for index, req_id in enumerate(self.input_batch.req_ids[:num_reqs]): + assert req_id is not None + + req = self.requests[req_id] + assert req.mrope_positions is not None + + num_computed_tokens = \ + self.input_batch.num_computed_tokens_cpu[index] + num_scheduled_tokens = \ + scheduler_output.num_scheduled_tokens[req_id] + num_prompt_tokens = len(req.prompt_token_ids) + + if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens: + prompt_part_len = max(0, + num_prompt_tokens - num_computed_tokens) + completion_part_len = max( + 0, num_scheduled_tokens - prompt_part_len) + else: + prompt_part_len = num_scheduled_tokens + completion_part_len = 0 + + assert num_scheduled_tokens == prompt_part_len + completion_part_len + + if prompt_part_len > 0: + # prompt's mrope_positions are pre-computed + dst_start = mrope_pos_ptr + dst_end = mrope_pos_ptr + prompt_part_len + src_start = num_computed_tokens + src_end = num_computed_tokens + prompt_part_len + + self.mrope_positions_cpu[:, dst_start:dst_end] = \ + req.mrope_positions[:,src_start:src_end] + + mrope_pos_ptr += prompt_part_len + + if completion_part_len > 0: + # compute completion's mrope_positions on-the-fly + dst_start = mrope_pos_ptr + dst_end = mrope_pos_ptr + completion_part_len + + self.mrope_positions_cpu[:, dst_start:dst_end] = \ + MRotaryEmbedding.get_next_input_positions_tensor( + req.mrope_position_delta, + context_len=num_computed_tokens + + prompt_part_len, + seq_len=num_computed_tokens + + prompt_part_len + + completion_part_len, + ) + + mrope_pos_ptr += completion_part_len + def _prepare_sampling( self, scheduler_output: "SchedulerOutput", @@ -496,19 +630,34 @@ def _execute_encoder(self, scheduler_output: "SchedulerOutput"): for input_id in encoder_input_ids: mm_inputs.append(req_state.mm_inputs[input_id]) req_input_ids.append((req_id, input_id)) - batched_mm_inputs = MultiModalKwargs.batch(mm_inputs) - batched_mm_inputs = MultiModalKwargs.as_kwargs(batched_mm_inputs, - device=self.device) - - # Run the encoder. - # `encoder_outputs` is either of the following: - # 1. A tensor of shape [num_images, feature_size, hidden_size] - # in case when feature_size is fixed across all images. - # 2. A list (length: num_images) of tensors, each of shape - # [feature_size, hidden_size] in case when the feature size is - # dynamic depending on input images. - encoder_outputs = self.model.get_multimodal_embeddings( - **batched_mm_inputs) + + # Batch mm inputs as much as we can: if a request in the batch has + # multiple modalities or a different modality than the previous one, + # we process it separately to preserve item order. + # FIXME(ywang96): This is a hacky way to deal with multiple modalities + # in the same batch while still being able to benefit from batching + # multimodal inputs. The proper solution should be reordering the + # encoder outputs. + grouped_mm_inputs_list = group_mm_inputs_by_modality(mm_inputs) + + encoder_outputs = [] + for grouped_mm_inputs in grouped_mm_inputs_list: + batched_mm_inputs = MultiModalKwargs.batch(grouped_mm_inputs) + batched_mm_inputs = MultiModalKwargs.as_kwargs(batched_mm_inputs, + device=self.device) + + # Run the encoder. + # `curr_group_outputs` is either of the following: + # 1. A tensor of shape (num_items, feature_size, hidden_size) + # in case feature_size is fixed across all multimodal items. + # 2. A list or tuple (length: num_items) of tensors, each of shape + # (feature_size, hidden_size) in case the feature size is dynamic + # depending on the input multimodal items. + curr_group_outputs = self.model.get_multimodal_embeddings( + **batched_mm_inputs) + + for output in curr_group_outputs: + encoder_outputs.append(output) # Cache the encoder outputs. for (req_id, input_id), output in zip(req_input_ids, encoder_outputs): @@ -556,6 +705,9 @@ def _gather_encoder_outputs( encoder_outputs.append(encoder_output[start_idx:end_idx]) return encoder_outputs + def get_model(self) -> nn.Module: + return self.model + @torch.inference_mode() def execute_model( self, @@ -609,9 +761,12 @@ def execute_model( # Run the decoder. # Use persistent buffers for CUDA graphs. with set_forward_context(attn_metadata, self.vllm_config): + positions = self.mrope_positions[:, :num_input_tokens] \ + if self.model_config.uses_mrope \ + else self.positions[:num_input_tokens] hidden_states = self.model( input_ids=input_ids, - positions=self.positions[:num_input_tokens], + positions=positions, kv_caches=self.kv_caches, attn_metadata=None, inputs_embeds=inputs_embeds, @@ -627,10 +782,10 @@ def execute_model( sampling_metadata=sampling_metadata, ) - sampled_token_ids = sampler_output.sampled_token_ids # TODO(woosuk): The following loop can be slow since it iterates over # the requests one by one. Optimize. num_reqs = self.input_batch.num_reqs + request_seq_lens: List[Tuple[int, CachedRequestState, int]] = [] for i, req_id in enumerate(self.input_batch.req_ids[:num_reqs]): assert req_id is not None req_state = self.requests[req_id] @@ -639,10 +794,10 @@ def execute_model( assert seq_len <= req_state.num_tokens if seq_len == req_state.num_tokens: # Append the sampled token to the output token ids. - token_id = sampled_token_ids[i] - self.input_batch.token_ids_cpu[i, seq_len] = token_id self.input_batch.num_tokens[i] += 1 - req_state.output_token_ids.append(token_id) + # OPTIMIZATION: Priming the state updates for later updates. + req_state.output_token_ids.append(0) + request_seq_lens.append((i, req_state, seq_len)) else: # Ignore the sampled token from the partial request. # Rewind the generator state as if the token was not sampled. @@ -651,6 +806,21 @@ def execute_model( # This relies on cuda-specific torch-internal impl details generator.set_offset(generator.get_offset() - 4) + # num_reqs entries should be non-None + assert all( + req_id is not None for req_id in + self.input_batch.req_ids[:num_reqs]), "req_ids contains None" + req_ids = cast(List[str], self.input_batch.req_ids[:num_reqs]) + + # NOTE: GPU -> CPU Sync happens here. + # Move as many CPU operations as possible before this sync point. + sampled_token_ids = sampler_output.sampled_token_ids.tolist() + # Update with the actual token ids + for i, req_state, seq_len in request_seq_lens: + token_id = sampled_token_ids[i] + self.input_batch.token_ids_cpu[i, seq_len] = token_id + req_state.output_token_ids[-1] = token_id + if sampler_output.logprob_token_ids is None: logprob_token_ids = None else: @@ -660,12 +830,6 @@ def execute_model( else: logprobs = sampler_output.logprobs.cpu() - # num_reqs entries should be non-None - assert all( - req_id is not None for req_id in - self.input_batch.req_ids[:num_reqs]), "req_ids contains None" - req_ids = cast(List[str], self.input_batch.req_ids[:num_reqs]) - model_runner_output = ModelRunnerOutput( req_ids=req_ids, req_id_to_index=self.input_batch.req_id_to_index, @@ -687,10 +851,12 @@ def load_model(self) -> None: @torch.inference_mode() def _dummy_run( self, - model: nn.Module, num_tokens: int, - kv_caches: List[torch.Tensor], + kv_caches: Optional[List[torch.Tensor]] = None, ) -> torch.Tensor: + model = self.model + if kv_caches is None: + kv_caches = self.kv_caches if self.is_multimodal_model: input_ids = None inputs_embeds = self.inputs_embeds[:num_tokens] @@ -698,9 +864,12 @@ def _dummy_run( input_ids = self.input_ids[:num_tokens] inputs_embeds = None with set_forward_context(None, self.vllm_config): + positions = self.mrope_positions[:, :num_tokens] \ + if self.model_config.uses_mrope \ + else self.positions[:num_tokens] hidden_states = model( input_ids=input_ids, - positions=self.positions[:num_tokens], + positions=positions, kv_caches=kv_caches, attn_metadata=None, inputs_embeds=inputs_embeds, @@ -721,44 +890,30 @@ def profile_run(self) -> None: ] # Profile with multimodal encoder & encoder cache. - if self.is_multimodal_model: - - # Create dummy batch of multimodal inputs. - dummy_request_data = self.input_registry.dummy_data_for_profiling( - model_config=self.model_config, - seq_len=self.max_num_tokens, - mm_registry=self.mm_registry, - ) - dummy_mm_data = dummy_request_data.multi_modal_data + # TODO: handle encoder-decoder models once we support them. + if (self.is_multimodal_model and self.max_num_encoder_input_tokens > 0 + and self.encoder_cache_size > 0): # NOTE: Currently model is profiled with a single non-text # modality with the max possible input tokens even when # it supports multiple. - max_tokens_by_modality_dict = self.mm_registry.get_max_tokens_per_item_by_modality( # noqa: E501 + max_tokens_by_modality_dict = MULTIMODAL_REGISTRY.get_max_tokens_per_item_by_nonzero_modality( # noqa: E501 self.model_config) - dummy_data_modality, max_tokens_per_mm_item = max( max_tokens_by_modality_dict.items(), key=lambda item: item[1]) # Check how many items of this modality can be supported by - # the encoder cache budget. - encoder_cache_budget = min(self.max_num_encoder_input_tokens, - self.encoder_cache_size) - max_num_mm_items_encoder_budget = encoder_cache_budget // \ - max_tokens_per_mm_item - - # TODO: Allow users to set encoder_cache_budget in case this - # happens. - assert max_num_mm_items_encoder_budget > 0, ( - f"Encoder cache budget={encoder_cache_budget} is too small to " - f"support the maximum possible size of multimodal embeddings" - f"={max_tokens_per_mm_item}.") + # the encoder budget. + encoder_budget = min(self.max_num_encoder_input_tokens, + self.encoder_cache_size) + + max_num_mm_items_encoder_budget = cdiv(encoder_budget, + max_tokens_per_mm_item) # Check how many items of this modality can be supported by # the decoder budget. - max_mm_items_per_req = max( - self.mm_registry.get_mm_limits_per_prompt( - self.model_config).values()) + max_mm_items_per_req = self.mm_registry.get_mm_limits_per_prompt( + self.model_config)[dummy_data_modality] # NOTE: We do not consider max_num_batched_tokens on purpose # because the multimodal embeddings can be generated in advance @@ -769,6 +924,19 @@ def profile_run(self) -> None: max_num_mm_items = min(max_num_mm_items_encoder_budget, max_num_mm_items_decoder_budget) + logger.info( + "Encoder cache will be initialized with a budget of %s tokens," + " and profiled with %s %s items of the maximum feature size.", + encoder_budget, max_num_mm_items, dummy_data_modality) + + # Create dummy batch of multimodal inputs. + dummy_request_data = self.input_registry.dummy_data_for_profiling( + model_config=self.model_config, + seq_len=self.max_num_tokens, + mm_registry=self.mm_registry, + ) + dummy_mm_data = dummy_request_data.multi_modal_data + # Dummy data definition in V0 may contain multiple multimodal items # (e.g, multiple images) for a single request, therefore here we # always replicate first item by max_num_mm_items times since in V1 @@ -814,8 +982,7 @@ def profile_run(self) -> None: self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs)) # Trigger compilation for general shape. - hidden_states = self._dummy_run(self.model, self.max_num_tokens, - dummy_kv_caches) + hidden_states = self._dummy_run(self.max_num_tokens, dummy_kv_caches) logits = self.model.compute_logits(hidden_states, None) logits = logits[:self.max_num_tokens] # TODO(woosuk): Consider the memory usage of the sampler. @@ -841,8 +1008,8 @@ def capture_model(self) -> None: for num_tokens in reversed(self.cudagraph_batch_sizes): for _ in range(self.vllm_config.compilation_config. cudagraph_num_of_warmups): - self._dummy_run(self.model, num_tokens, self.kv_caches) - self._dummy_run(self.model, num_tokens, self.kv_caches) + self._dummy_run(num_tokens) + self._dummy_run(num_tokens) end_time = time.perf_counter() end_free_gpu_memory = torch.cuda.mem_get_info()[0] @@ -852,15 +1019,71 @@ def capture_model(self) -> None: logger.info("Graph capturing finished in %.0f secs, took %.2f GiB", elapsed_time, cuda_graph_size / (1 << 30)) - def initialize_kv_cache(self, num_blocks: int) -> None: - assert len(self.kv_caches) == 0 - kv_cache_shape = FlashAttentionBackend.get_kv_cache_shape( - num_blocks, self.block_size, self.num_kv_heads, self.head_size) - for _ in range(self.num_attn_layers): - self.kv_caches.append( - torch.zeros(kv_cache_shape, - dtype=self.kv_cache_dtype, - device=self.device)) + def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None: + """ + Initialize KV cache based on `kv_cache_config`. + Args: + kv_cache_config: Configuration for the KV cache, including the KV + cache size of each layer + """ + if len(kv_cache_config.groups) > 1: + raise NotImplementedError( + "Hybrid models with more than one KV cache type are not " + "supported yet.") + + kv_caches: Dict[str, torch.Tensor] = {} + + for layer_name, layer_spec in kv_cache_config.kv_cache_spec.items(): + tensor_config = kv_cache_config.tensors[layer_name] + assert tensor_config.size % layer_spec.page_size_bytes == 0 + num_blocks = tensor_config.size // layer_spec.page_size_bytes + if isinstance(layer_spec, FullAttentionSpec): + kv_cache_shape = FlashAttentionBackend.get_kv_cache_shape( + num_blocks, layer_spec.block_size, layer_spec.num_kv_heads, + layer_spec.head_size) + dtype = layer_spec.dtype + kv_caches[layer_name] = torch.zeros(kv_cache_shape, + dtype=dtype, + device=self.device) + else: + raise NotImplementedError + bind_kv_cache( + kv_caches, self.vllm_config.compilation_config.static_forward_context, - [self.kv_caches]) + self.kv_caches) + + def get_kv_cache_spec(self) -> KVCacheSpec: + """ + Generates the KVCacheSpec by parsing the kv cache format from each + Attention module in the static forward context. + Returns: + KVCacheSpec: A dictionary mapping layer names to their KV cache + format. Layers that do not need KV cache are not included. + """ + + forward_ctx = self.vllm_config.compilation_config.static_forward_context + block_size = self.vllm_config.cache_config.block_size + kv_cache_spec: KVCacheSpec = {} + for layer_name, attn_module in forward_ctx.items(): + # TODO: Support other attention modules, e.g., sliding window, + # cross-attention, MLA. + assert isinstance(attn_module, Attention) + if attn_module.attn_type == AttentionType.DECODER: + kv_cache_spec[layer_name] = FullAttentionSpec( + block_size=block_size, + num_kv_heads=attn_module.num_kv_heads, + head_size=attn_module.head_size, + dtype=attn_module.dtype, + ) + elif attn_module.attn_type in (AttentionType.ENCODER, + AttentionType.ENCODER_ONLY): + # encoder-only attention does not need KV cache. + continue + elif attn_module.attn_type == AttentionType.ENCODER_DECODER: + raise NotImplementedError + else: + raise ValueError( + f"Unknown attention type: {attn_module.attn_type}") + + return kv_cache_spec diff --git a/vllm/v1/worker/gpu_worker.py b/vllm/v1/worker/gpu_worker.py index e83bce4283555..a8cf0aec3f17b 100644 --- a/vllm/v1/worker/gpu_worker.py +++ b/vllm/v1/worker/gpu_worker.py @@ -1,21 +1,24 @@ """A GPU worker class.""" import gc import os -from typing import TYPE_CHECKING, Optional, Tuple +from typing import TYPE_CHECKING, Optional import torch import torch.distributed +import torch.nn as nn import vllm.envs as envs -from vllm.config import CacheConfig, ModelConfig, ParallelConfig, VllmConfig +from vllm.config import ParallelConfig, VllmConfig +from vllm.device_allocator.cumem import CuMemAllocator from vllm.distributed import (ensure_model_parallel_initialized, init_distributed_environment, set_custom_all_reduce) from vllm.logger import init_logger from vllm.model_executor import set_random_seed from vllm.platforms import current_platform -from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, LayerBlockType, get_dtype_size +from vllm.utils import GiB_bytes from vllm.v1.core.scheduler import SchedulerOutput +from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec from vllm.v1.outputs import ModelRunnerOutput from vllm.v1.worker.gpu_model_runner import GPUModelRunner @@ -33,6 +36,7 @@ def __init__( local_rank: int, rank: int, distributed_init_method: str, + is_driver_worker: bool = False, ): # TODO: use WorkerBase.__init__(self, vllm_config=vllm_config) @@ -75,7 +79,24 @@ def __init__( else: self.profiler = None - def initialize(self): + def sleep(self, level: int = 1) -> None: + free_bytes_before_sleep = torch.cuda.mem_get_info()[0] + allocator = CuMemAllocator.get_instance() + allocator.sleep(offload_tags=("weights", ) if level == 1 else tuple()) + free_bytes_after_sleep, total = torch.cuda.mem_get_info() + freed_bytes = free_bytes_after_sleep - free_bytes_before_sleep + used_bytes = total - free_bytes_after_sleep + assert freed_bytes >= 0, "Memory usage increased after sleeping." + logger.info( + "Sleep mode freed %.2f GiB memory, " + "%.2f GiB memory is still in use.", freed_bytes / GiB_bytes, + used_bytes / GiB_bytes) + + def wake_up(self) -> None: + allocator = CuMemAllocator.get_instance() + allocator.wake_up() + + def init_device(self): if self.device_config.device.type == "cuda": # torch.distributed.all_reduce does not free the input tensor until # the synchronization point. This causes the memory usage to grow @@ -108,23 +129,31 @@ def initialize(self): self.model_runner = GPUModelRunner(self.vllm_config, self.device) def load_model(self) -> None: - self.model_runner.load_model() + if self.vllm_config.model_config.enable_sleep_mode: + allocator = CuMemAllocator.get_instance() + assert allocator.get_current_usage() == 0, ( + "Sleep mode can only be " + "used for one instance per process.") + context = allocator.use_memory_pool(tag="weights") + else: + from contextlib import nullcontext + context = nullcontext() + with context: + self.model_runner.load_model() @torch.inference_mode() - def determine_num_available_blocks(self) -> Tuple[int, int]: - """Profiles the peak memory usage of the model to determine how many - KV blocks may be allocated without OOMs. + def determine_available_memory(self) -> int: + """Profiles the peak memory usage of the model to determine how much + memory can be used for KV cache without OOMs. The engine will first conduct a profiling of the existing memory usage. - Then, it calculate the maximum possible number of GPU and CPU blocks - that can be allocated with the remaining free memory. + Then, it calculate the free memory that can be used for KV cache in + bytes. .. tip:: You may limit the usage of GPU memory by adjusting the `gpu_memory_utilization` parameter. """ - # Profile the memory usage of the model and get the maximum number of - # cache blocks that can be allocated with the remaining free memory. torch.cuda.empty_cache() torch.cuda.reset_peak_memory_stats() @@ -160,46 +189,49 @@ def determine_num_available_blocks(self) -> Tuple[int, int]: total_gpu_memory * self.cache_config.gpu_memory_utilization - peak_memory) - # Calculate the number of blocks that can be allocated with the - # profiled peak memory. - cache_block_size = _get_cache_block_size(self.cache_config, - self.model_config, - self.parallel_config) - num_gpu_blocks = int(available_kv_cache_memory // cache_block_size) - num_gpu_blocks = max(num_gpu_blocks, 0) - return num_gpu_blocks, 0 - - def initialize_cache(self, num_gpu_blocks: int) -> None: - """Allocate GPU and CPU KV cache with the specified number of blocks.""" - if num_gpu_blocks <= 0: - raise ValueError("No available memory for the cache blocks. " - "Try increasing `gpu_memory_utilization` when " - "initializing the engine.") - - max_seq_len = self.cache_config.block_size * num_gpu_blocks - max_model_len = self.model_config.max_model_len - if max_model_len > max_seq_len: - raise ValueError( - f"The model's max seq len ({max_model_len}) " - "is larger than the maximum number of tokens that can be " - f"stored in KV cache ({max_seq_len}). Try increasing " - "`gpu_memory_utilization` or decreasing `max_model_len` when " - "initializing the engine.") + return int(available_kv_cache_memory) - self.model_runner.initialize_kv_cache(num_gpu_blocks) + def get_kv_cache_spec(self) -> KVCacheSpec: + return self.model_runner.get_kv_cache_spec() + + def initialize_cache(self, kv_cache_config: KVCacheConfig) -> None: + """Allocate GPU KV cache with the specified kv_cache_config.""" + if self.vllm_config.model_config.enable_sleep_mode: + allocator = CuMemAllocator.get_instance() + context = allocator.use_memory_pool(tag="kv_cache") + else: + from contextlib import nullcontext + context = nullcontext() + with context: + self.model_runner.initialize_kv_cache(kv_cache_config) def compile_or_warm_up_model(self) -> None: + # warm up sizes that are not in cudagraph capture sizes, + # but users still want to compile for better performance, + # e.g. for the max-num-batched token size in chunked prefill. + warmup_sizes = self.vllm_config.compilation_config.compile_sizes.copy() + if not self.model_config.enforce_eager: + warmup_sizes = [ + x for x in warmup_sizes if x not in + self.vllm_config.compilation_config.cudagraph_capture_sizes + ] + for size in sorted(warmup_sizes, reverse=True): + logger.info("Compile and warming up model for size %d", size) + self.model_runner._dummy_run(size) if not self.model_config.enforce_eager: self.model_runner.capture_model() # Reset the seed to ensure that the random state is not affected by # the model initialization and profiling. set_random_seed(self.model_config.seed) + def get_model(self) -> nn.Module: + return self.model_runner.get_model() + @torch.inference_mode() def execute_model( self, scheduler_output: "SchedulerOutput", - ) -> ModelRunnerOutput: + ) -> Optional[ModelRunnerOutput]: output = self.model_runner.execute_model(scheduler_output) return output if self.rank == 0 else None @@ -250,24 +282,3 @@ def _check_if_gpu_supports_dtype(torch_dtype: torch.dtype): f"of at least 8.0. Your {gpu_name} GPU {compute_str}. " "You can use float16 instead by explicitly setting the" "`dtype` flag in CLI, for example: --dtype=half.") - - -def _get_cache_block_size( - cache_config: CacheConfig, - model_config: ModelConfig, - parallel_config: ParallelConfig, -) -> int: - head_size = model_config.get_head_size() - num_heads = model_config.get_num_kv_heads(parallel_config) - num_attention_layers = model_config.get_num_layers_by_block_type( - parallel_config, LayerBlockType.attention) - - key_cache_block = cache_config.block_size * num_heads * head_size - value_cache_block = key_cache_block - total = num_attention_layers * (key_cache_block + value_cache_block) - if cache_config.cache_dtype == "auto": - dtype = model_config.dtype - else: - dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype] - dtype_size = get_dtype_size(dtype) - return dtype_size * total diff --git a/vllm/worker/cpu_model_runner.py b/vllm/worker/cpu_model_runner.py index 303d9a15e9c3c..4b429b67b36f8 100644 --- a/vllm/worker/cpu_model_runner.py +++ b/vllm/worker/cpu_model_runner.py @@ -144,9 +144,7 @@ def __init__(self, runner: "CPUModelRunner", finished_requests_ids: Optional[List[str]] = None) -> None: super().__init__() - self.seq_group_metadata_list: List[SequenceGroupMetadata] = [] self.runner = runner - self.chunked_prefill = (runner.scheduler_config.chunked_prefill_enabled or runner.cache_config.enable_prefix_caching) self.model_input_cls = self.runner._model_input_cls @@ -156,10 +154,17 @@ def __init__(self, self.device = self.runner.device self.multi_modal_input_mapper = self.runner.multi_modal_input_mapper self.enable_lora = self.runner.lora_config is not None + if self.runner.attn_backend is not None: + # spec decode (e.g. Medusa) does not have atten backend + attn_backend = self.runner.attn_backend + self.att_metadata_builder = attn_backend.get_builder_cls()(self) + + def prepare(self, + finished_requests_ids: Optional[List[str]] = None) -> None: + self.seq_group_metadata_list: List[SequenceGroupMetadata] = [] self.input_data = ModelInputForCPUBuilder.ModelInputData( self.runner.model_config.uses_mrope) - self.att_metadata_builder = self.runner.attn_backend.get_builder_cls()( - self) + self.att_metadata_builder.prepare() def add_seq_group(self, seq_group_metadata: SequenceGroupMetadata): self.seq_group_metadata_list.append(seq_group_metadata) @@ -431,6 +436,7 @@ class CPUModelRunnerBase(ModelRunnerBase[TModelInputForCPU]): """ _model_input_cls: Type[TModelInputForCPU] _builder_cls: Type[ModelInputForCPUBuilder] + builder: ModelInputForCPUBuilder def __init__( self, @@ -477,6 +483,10 @@ def __init__( # Set after load_model. self.lora_manager: Optional[LRUCacheWorkerLoRAManager] = None + if hasattr(self, "_builder_cls"): + # multi-step model runner does not have `_builder_cls` + self.builder = self._builder_cls(weakref.proxy(self)) + def load_model(self) -> None: self.model = get_model(vllm_config=self.vllm_config) @@ -509,6 +519,9 @@ def load_model(self) -> None: ) self.model = self.lora_manager.create_lora_manager(self.model) + def get_model(self) -> nn.Module: + return self.model + def _prepare_model_input_tensors( self, seq_group_metadata_list: List[SequenceGroupMetadata], @@ -519,10 +532,10 @@ def _prepare_model_input_tensors( metadata for possible additional steps, e.g., sampling. """ - builder = self._builder_cls(weakref.proxy(self), finished_requests_ids) - builder.set_seq_group_list(seq_group_metadata_list) + self.builder.prepare(finished_requests_ids) + self.builder.set_seq_group_list(seq_group_metadata_list) - return builder.build() # type: ignore + return self.builder.build() # type: ignore # sampler property will be used by spec_decode_worker @property diff --git a/vllm/worker/hpu_model_runner.py b/vllm/worker/hpu_model_runner.py index 090c460597083..9c4dc7e8571af 100644 --- a/vllm/worker/hpu_model_runner.py +++ b/vllm/worker/hpu_model_runner.py @@ -21,6 +21,7 @@ import habana_frameworks.torch as htorch import habana_frameworks.torch.internal.bridge_config as bc import torch +import torch.nn as nn from vllm_hpu_extension.ops import LoraMask as LoraMask from vllm_hpu_extension.profiler import (HabanaHighLevelProfiler, HabanaMemoryProfiler, format_bytes) @@ -289,12 +290,14 @@ def forward_hook(module, args, output): class HpuModelAdapter: - def __init__(self, model, block_size, dtype, enforce_eager): + def __init__(self, model, vllm_config): self.model = model self.prefill_use_fusedsdpa = os.getenv('VLLM_PROMPT_USE_FUSEDSDPA', '0').lower() in ['1', 'true'] - self.block_size = block_size - self.dtype = dtype + self.vllm_config = vllm_config + self.block_size = vllm_config.cache_config.block_size + self.dtype = vllm_config.model_config.dtype + enforce_eager = vllm_config.model_config.enforce_eager if not htorch.utils.internal.is_lazy() and not enforce_eager: self.model = torch.compile(self.model, backend='hpu_backend', @@ -353,14 +356,20 @@ def forward(self, *args, **kwargs): selected_token_indices = kwargs.pop('selected_token_indices') if 'warmup_mode' in kwargs: kwargs.pop('warmup_mode') + virtual_engine = 0 + if 'virtual_engine' in kwargs: + virtual_engine = kwargs.pop('virtual_engine') input_ids = kwargs['input_ids'] kwargs['attn_metadata'] = self._update_metadata( kwargs['attn_metadata'], input_ids.size(0), input_ids.size(1), input_ids.device, self.dtype) LoraMask.setLoraMask(kwargs.pop('lora_mask')) - hidden_states = self.model(*args, **kwargs) - hidden_states = hidden_states.view(-1, hidden_states.shape[-1]) - hidden_states = hidden_states.index_select(0, selected_token_indices) + with set_forward_context(kwargs['attn_metadata'], self.vllm_config, + virtual_engine): + hidden_states = self.model(*args, **kwargs) + hidden_states = hidden_states.view(-1, hidden_states.shape[-1]) + hidden_states = hidden_states.index_select(0, + selected_token_indices) return hidden_states def compute_logits(self, *args, **kwargs): @@ -660,10 +669,7 @@ def load_model(self) -> None: with HabanaMemoryProfiler() as m_wrap: self.model = _maybe_wrap_in_hpu_graph( - self.model, - self.block_size, - dtype=self.model_config.dtype, - enforce_eager=self.enforce_eager) + self.model, vllm_config=self.vllm_config) msg = f"Wrapping in HPU Graph took {m_wrap.get_summary_string()}" logger.info(msg) @@ -671,6 +677,9 @@ def load_model(self) -> None: msg = f"Loading model weights took in total {m.get_summary_string()}" logger.info(msg) + def get_model(self) -> nn.Module: + return self.model + def _use_graphs(self, batch_size, seq_len, is_prompt): if self.enforce_eager: return False @@ -1937,6 +1946,7 @@ def execute_model( "attn_metadata": self.trim_attn_metadata(attn_metadata), "intermediate_tensors": intermediate_tensors, "lora_mask": lora_mask, + "virtual_engine": model_input.virtual_engine, **(model_input.multi_modal_kwargs or {}), } if htorch.utils.internal.is_lazy(): diff --git a/vllm/worker/hpu_worker.py b/vllm/worker/hpu_worker.py index 8b2d8aaed2803..3c570212625c4 100644 --- a/vllm/worker/hpu_worker.py +++ b/vllm/worker/hpu_worker.py @@ -2,6 +2,7 @@ # Copyright (C) 2024 Habana Labs, Ltd. an Intel Company ############################################################################### +import contextlib import gc import os from typing import List, Optional, Set, Tuple, Type @@ -18,6 +19,7 @@ from vllm.logger import init_logger from vllm.lora.request import LoRARequest from vllm.model_executor import set_random_seed +from vllm.model_executor.layers.sampler import SamplerOutput from vllm.prompt_adapter.request import PromptAdapterRequest from vllm.sequence import ExecuteModelRequest from vllm.utils import bind_kv_cache @@ -124,6 +126,70 @@ def init_device(self) -> None: def load_model(self): self.model_runner.load_model() + def execute_model( + self, + execute_model_req: Optional[ExecuteModelRequest] = None, + ) -> Optional[List[SamplerOutput]]: + # VLLM_HPU_LOG_STEP_GRAPH_COMPILATION - will log graph compilations per engine step, only when there was any - highly recommended to use alongside PT_HPU_METRICS_GC_DETAILS! # noqa:E501 + # VLLM_HPU_LOG_STEP_GRAPH_COMPILATION_ALL - will log graph compilations per engine step, always, even if there were none # noqa:E501 + # VLLM_HPU_LOG_STEP_CPU_FALLBACKS - will log cpu fallbacks per engine step, only when there was any # noqa:E501 + # VLLM_HPU_LOG_STEP_CPU_FALLBACKS_ALL - will log cpu fallbacks per engine step, always, even if there were none # noqa:E501 + log_graph_compilation_all = os.environ.get( + 'VLLM_HPU_LOG_STEP_GRAPH_COMPILATION_ALL', '0') != '0' + log_graph_compilation = os.environ.get( + 'VLLM_HPU_LOG_STEP_GRAPH_COMPILATION', + '0') != '0' or log_graph_compilation_all + log_cpu_fallbacks_all = os.environ.get( + 'VLLM_HPU_LOG_STEP_CPU_FALLBACKS_ALL', '0') != '0' + log_cpu_fallbacks = os.environ.get('VLLM_HPU_LOG_STEP_CPU_FALLBACKS', + '0') != '0' or log_cpu_fallbacks_all + if (log_graph_compilation or log_cpu_fallbacks) and \ + execute_model_req is not None: + from habana_frameworks.torch.hpu.metrics import metric_localcontext + seq_group_metadata_list = execute_model_req.seq_group_metadata_list + is_prompt = any([ + seq_group_metadata.is_prompt + for seq_group_metadata in seq_group_metadata_list + ]) + max_context_len = max([ + max([ + len(v.prompt_token_ids) + len(v.output_token_ids) + for v in seq_group_metadata.seq_data.values() + ]) for seq_group_metadata in seq_group_metadata_list + ]) # whoa, that's some spicy stuff right here + max_num_blocks = ( + (max_context_len - 1) // self.cache_config.block_size) + 1 + input_stats = (f'is_prompt: {is_prompt}, ' + f'num_seqs: {len(seq_group_metadata_list)}, ' + f'max_context_len: {max_context_len}, ' + f'max_num_blocks {max_num_blocks}') + gc_ctx = metric_localcontext( + "graph_compilation" + ) if log_graph_compilation else contextlib.nullcontext() + cpu_fallback_ctx = metric_localcontext( + "cpu_fallback" + ) if log_cpu_fallbacks else contextlib.nullcontext() + with gc_ctx as gc_local_metric, \ + cpu_fallback_ctx as cpu_fallback_local_metric: + output = LocalOrDistributedWorkerBase.execute_model( + self, execute_model_req) + if (log_graph_compilation and gc_local_metric.stats()[0][1] > 0 + ) or log_graph_compilation_all: + msg = ("VLLM_HPU_STEP_GRAPH_COMPILATION: " + f"{gc_local_metric.stats()}, {input_stats}") + logger.warning(msg) + if (log_cpu_fallbacks and cpu_fallback_local_metric.stats()[0][1] > + 0) or log_cpu_fallbacks_all: + msg = ("VLLM_HPU_STEP_CPU_FALLBACK: " + f"{cpu_fallback_local_metric.stats()}, {input_stats}") + logger.warning(msg) + + return output + + output = LocalOrDistributedWorkerBase.execute_model( + self, execute_model_req) + return output + @torch.inference_mode() def determine_num_available_blocks(self) -> Tuple[int, int]: """Profiles the peak memory usage of the model to determine how many diff --git a/vllm/worker/model_runner.py b/vllm/worker/model_runner.py index faf2703e03649..9c6ec7818385d 100644 --- a/vllm/worker/model_runner.py +++ b/vllm/worker/model_runner.py @@ -4,6 +4,7 @@ import itertools import time import weakref +from contextlib import contextmanager from dataclasses import dataclass from typing import (TYPE_CHECKING, Any, Callable, Dict, List, Optional, Set, Tuple, Type, TypeVar, Union) @@ -47,8 +48,8 @@ from vllm.sequence import IntermediateTensors, SequenceGroupMetadata from vllm.utils import (DeviceMemoryProfiler, GiB_bytes, PyObjectCache, async_tensor_h2d, flatten_2d_lists, - is_pin_memory_available, rpd_mark, supports_dynamo, - weak_ref_tensor) + is_pin_memory_available, rpd_mark, rpd_user_marker, + supports_dynamo, weak_ref_tensor) from vllm.worker.model_runner_base import ( ModelRunnerBase, ModelRunnerInputBase, ModelRunnerInputBuilderBase, _add_attn_metadata_broadcastable_dict, @@ -454,17 +455,12 @@ def __init__(self, self.enable_prompt_adapter = (self.runner.prompt_adapter_config is not None) self.multi_modal_input_mapper = self.runner.multi_modal_input_mapper - self.finished_requests_ids = finished_requests_ids - self.decode_only = True - - # Intermediate data (data in CPU before going to GPU) for - # the current sequence group. - self.inter_data_list: List[ - ModelInputForGPUBuilder.InterDataForSeqGroup] = [] # Attention metadata inputs. - self.attn_metadata_builder = self.attn_backend.make_metadata_builder( - weakref.proxy(self)) + if self.attn_backend is not None: + # spec decode (e.g. Medusa) does not have atten backend + self.attn_metadata_builder = self.attn_backend.get_builder_cls()( + weakref.proxy(self)) # Engine/Model configurations. self.chunked_prefill_enabled = ( @@ -476,6 +472,21 @@ def __init__(self, self.block_aligned_sliding_window = \ self.sliding_window_blocks * self.block_size + def prepare(self, + finished_requests_ids: Optional[List[str]] = None) -> None: + self.finished_requests_ids = finished_requests_ids + + # if the current batch is decode-only. + # will be set to False if there is any non-decode request. + self.decode_only = True + + # Intermediate data (data in CPU before going to GPU) for + # the current sequence group. + self.inter_data_list: List[ + ModelInputForGPUBuilder.InterDataForSeqGroup] = [] + + self.attn_metadata_builder.prepare() + def _compute_lens(self, inter_data: InterDataForSeqGroup, seq_idx: int, seq_group_metadata: SequenceGroupMetadata): """Compute context length, sequence length and tokens @@ -990,6 +1001,7 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]): """ _model_input_cls: Type[TModelInputForGPU] _builder_cls: Type[ModelInputForGPUBuilder] + builder: ModelInputForGPUBuilder def __init__( self, @@ -1026,6 +1038,8 @@ def __init__( self.has_inner_state = model_config.has_inner_state + self.in_profile_run = False + # When using CUDA graph, the input block tables must be padded to # max_seq_len_to_capture. However, creating the block table in # Python can be expensive. To optimize this, we cache the block table @@ -1088,6 +1102,10 @@ def __init__( SamplingMetadataCache() \ if self.parallel_config.pipeline_parallel_size == 1 else None + if hasattr(self, "_builder_cls"): + # multi-step model runner does not have `_builder_cls` + self.builder = self._builder_cls(weakref.proxy(self)) + def load_model(self) -> None: logger.info("Starting to load model %s...", self.model_config.model) with DeviceMemoryProfiler() as m: @@ -1143,6 +1161,9 @@ def load_model(self) -> None: fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE, backend=backend) + def get_model(self) -> nn.Module: + return self.model + def save_sharded_state( self, path: str, @@ -1190,122 +1211,141 @@ def _prepare_model_input_tensors( If cuda graph is required, this API automatically pads inputs. """ - builder = self._builder_cls(weakref.proxy(self), finished_requests_ids) + self.builder.prepare(finished_requests_ids) for seq_group_metadata in seq_group_metadata_list: - builder.add_seq_group(seq_group_metadata) + self.builder.add_seq_group(seq_group_metadata) + + self.builder.reset_cached_inter_data() - builder.reset_cached_inter_data() + return self.builder.build() # type: ignore - return builder.build() # type: ignore + @contextmanager + def set_in_profile_run(self): + self.in_profile_run = True + try: + yield + finally: + self.in_profile_run = False @torch.inference_mode() def profile_run(self) -> None: - # Enable top-k sampling to reflect the accurate memory usage. - sampling_params = SamplingParams(top_p=0.99, top_k=self.vocab_size - 1) - max_num_batched_tokens = self.scheduler_config.max_num_batched_tokens + max_num_batched_tokens = \ + self.scheduler_config.max_num_batched_tokens max_num_seqs = self.scheduler_config.max_num_seqs - # This represents the maximum number of different requests - # that will have unique loras, an therefore the max amount of memory - # consumption create dummy lora request copies from the lora request - # passed in, which contains a lora from the lora warmup path. - dummy_lora_requests: List[LoRARequest] = [] - dummy_lora_requests_per_seq: List[LoRARequest] = [] - if self.lora_config: - assert self.lora_manager is not None - with self.lora_manager.dummy_lora_cache(): - for idx in range(self.lora_config.max_loras): - lora_id = idx + 1 - dummy_lora_request = LoRARequest( - lora_name=f"warmup_{lora_id}", - lora_int_id=lora_id, - lora_path="/not/a/real/path", - ) - self.lora_manager.add_dummy_lora(dummy_lora_request, - rank=LORA_WARMUP_RANK) - dummy_lora_requests.append(dummy_lora_request) - dummy_lora_requests_per_seq = [ - dummy_lora_requests[idx % len(dummy_lora_requests)] - for idx in range(max_num_seqs) - ] - - # Profile memory usage with max_num_sequences sequences and the total - # number of tokens equal to max_num_batched_tokens. - seqs: List[SequenceGroupMetadata] = [] - # Additional GPU memory may be needed for multi-modal encoding, which - # needs to be accounted for when calculating the GPU blocks for - # vLLM blocker manager. - # To exercise the worst scenario for GPU memory consumption, - # the number of seqs (batch_size) is chosen to maximize the number - # of images processed. - - max_mm_tokens = self.mm_registry.get_max_multimodal_tokens( - self.model_config) - if max_mm_tokens > 0: - max_num_seqs_orig = max_num_seqs - max_num_seqs = min(max_num_seqs, - max_num_batched_tokens // max_mm_tokens) - if max_num_seqs < 1: - expr = (f"min({max_num_seqs_orig}, " - f"{max_num_batched_tokens} // {max_mm_tokens})") - logger.warning( - "Computed max_num_seqs (%s) to be less than 1. " - "Setting it to the minimum value of 1.", expr) - max_num_seqs = 1 - - batch_size = 0 - for group_id in range(max_num_seqs): - seq_len = (max_num_batched_tokens // max_num_seqs + - (group_id < max_num_batched_tokens % max_num_seqs)) - batch_size += seq_len - - dummy_data = self.input_registry \ - .dummy_data_for_profiling(self.model_config, - seq_len, - self.mm_registry) - - seq = SequenceGroupMetadata( - request_id=str(group_id), - is_prompt=True, - seq_data={group_id: dummy_data.seq_data}, - sampling_params=sampling_params, - block_tables=None, - lora_request=dummy_lora_requests_per_seq[group_id] - if dummy_lora_requests_per_seq else None, - multi_modal_data=dummy_data.multi_modal_data, - multi_modal_placeholders=dummy_data.multi_modal_placeholders, - ) - seqs.append(seq) - - # Run the model with the dummy inputs. - num_layers = self.model_config.get_num_layers(self.parallel_config) - # use an empty tensor instead of `None`` to force Dynamo to pass - # it by reference, rather by specializing on the value ``None``. - # the `dtype` argument does not matter, and we use `float32` as - # a placeholder (it has wide hardware support). - # it is important to create tensors inside the loop, rather than - # multiplying the list, to avoid Dynamo from treating them as - # tensor aliasing. - kv_caches = [ - torch.tensor([], dtype=torch.float32, device=self.device) - for _ in range(num_layers) - ] - finished_requests_ids = [seq.request_id for seq in seqs] - model_input = self.prepare_model_input( - seqs, finished_requests_ids=finished_requests_ids) - intermediate_tensors = None - if not get_pp_group().is_first_rank: - intermediate_tensors = self.model.make_empty_intermediate_tensors( - batch_size=batch_size, - dtype=self.model_config.dtype, - device=self.device) - - # Disable KV Scale Calculation for dummy data during profile run - if model_input.attn_metadata is not None: - model_input.attn_metadata.enable_kv_scales_calculation = False - - self.execute_model(model_input, kv_caches, intermediate_tensors) - torch.cuda.synchronize() - return + self._dummy_run(max_num_batched_tokens, max_num_seqs) + + def _dummy_run(self, + max_num_batched_tokens: int, + max_num_seqs: int = 1) -> None: + with self.set_in_profile_run(): + # Enable top-k sampling to reflect the accurate memory usage. + sampling_params = \ + SamplingParams(top_p=0.99, top_k=self.vocab_size - 1) + + # This represents the maximum number of different requests + # that will have unique loras, an therefore the max amount of memory + # consumption create dummy lora request copies from the lora request + # passed in, which contains a lora from the lora warmup path. + dummy_lora_requests: List[LoRARequest] = [] + dummy_lora_requests_per_seq: List[LoRARequest] = [] + if self.lora_config: + assert self.lora_manager is not None + with self.lora_manager.dummy_lora_cache(): + for idx in range(self.lora_config.max_loras): + lora_id = idx + 1 + dummy_lora_request = LoRARequest( + lora_name=f"warmup_{lora_id}", + lora_int_id=lora_id, + lora_path="/not/a/real/path", + ) + self.lora_manager.add_dummy_lora(dummy_lora_request, + rank=LORA_WARMUP_RANK) + dummy_lora_requests.append(dummy_lora_request) + dummy_lora_requests_per_seq = [ + dummy_lora_requests[idx % len(dummy_lora_requests)] + for idx in range(max_num_seqs) + ] + + # Profile memory usage with max_num_sequences sequences and the + # total number of tokens equal to max_num_batched_tokens. + seqs: List[SequenceGroupMetadata] = [] + # Additional GPU memory may be needed for multi-modal encoding, + # which needs to be accounted for when calculating the GPU blocks + # for vLLM blocker manager. + # To exercise the worst scenario for GPU memory consumption, + # the number of seqs (batch_size) is chosen to maximize the number + # of images processed. + + max_mm_tokens = self.mm_registry.get_max_multimodal_tokens( + self.model_config) + if max_mm_tokens > 0: + max_num_seqs_orig = max_num_seqs + max_num_seqs = min(max_num_seqs, + max_num_batched_tokens // max_mm_tokens) + if max_num_seqs < 1: + expr = (f"min({max_num_seqs_orig}, " + f"{max_num_batched_tokens} // {max_mm_tokens})") + logger.warning( + "Computed max_num_seqs (%s) to be less than 1. " + "Setting it to the minimum value of 1.", expr) + max_num_seqs = 1 + + batch_size = 0 + for group_id in range(max_num_seqs): + seq_len = (max_num_batched_tokens // max_num_seqs + + (group_id < max_num_batched_tokens % max_num_seqs)) + batch_size += seq_len + + dummy_data = self.input_registry \ + .dummy_data_for_profiling(self.model_config, + seq_len, + self.mm_registry) + + seq = SequenceGroupMetadata( + request_id=str(group_id), + is_prompt=True, + seq_data={group_id: dummy_data.seq_data}, + sampling_params=sampling_params, + block_tables=None, + lora_request=dummy_lora_requests_per_seq[group_id] + if dummy_lora_requests_per_seq else None, + multi_modal_data=dummy_data.multi_modal_data, + multi_modal_placeholders=dummy_data. + multi_modal_placeholders, + ) + seqs.append(seq) + + # Run the model with the dummy inputs. + num_layers = self.model_config.get_num_layers(self.parallel_config) + # use an empty tensor instead of `None`` to force Dynamo to pass + # it by reference, rather by specializing on the value ``None``. + # the `dtype` argument does not matter, and we use `float32` as + # a placeholder (it has wide hardware support). + # it is important to create tensors inside the loop, rather than + # multiplying the list, to avoid Dynamo from treating them as + # tensor aliasing. + kv_caches = [ + torch.tensor([], dtype=torch.float32, device=self.device) + for _ in range(num_layers) + ] + finished_requests_ids = [seq.request_id for seq in seqs] + model_input = self.prepare_model_input( + seqs, finished_requests_ids=finished_requests_ids) + intermediate_tensors = None + if not get_pp_group().is_first_rank: + intermediate_tensors = \ + self.model.make_empty_intermediate_tensors( + batch_size=batch_size, + dtype=self.model_config.dtype, + device=self.device) + + # Disable KV Scale Calculation for dummy data during profile run + if model_input.attn_metadata is not None: + model_input.attn_metadata.enable_kv_scales_calculation = False + + self.execute_model(model_input, kv_caches, intermediate_tensors) + torch.cuda.synchronize() + return def remove_all_loras(self): if not self.lora_manager: @@ -1434,13 +1474,14 @@ def capture_model(self, kv_caches: List[List[torch.Tensor]]) -> None: for virtual_engine in range( self.parallel_config.pipeline_parallel_size): # Only rank 0 should print progress bar during capture - capture_sizes = ( - tqdm( - self.vllm_config.compilation_config.capture_sizes, - desc="Capturing CUDA graph shapes", - ) if get_tensor_model_parallel_rank() == 0 else - self.vllm_config.compilation_config.capture_sizes) - for batch_size in capture_sizes: + cudagraph_capture_sizes = (tqdm( + self.vllm_config.compilation_config. + cudagraph_capture_sizes, + desc="Capturing CUDA graph shapes", + ) if get_tensor_model_parallel_rank() == 0 else + self.vllm_config.compilation_config. + cudagraph_capture_sizes) + for batch_size in cudagraph_capture_sizes: attn_metadata = ( self.attn_state.graph_capture_get_metadata_for_batch( batch_size, @@ -1630,6 +1671,12 @@ def execute_model( assert model_input.attn_metadata is not None prefill_meta = model_input.attn_metadata.prefill_metadata decode_meta = model_input.attn_metadata.decode_metadata + if prefill_meta: + marker_instance = rpd_user_marker(name="Prefill") + else: + marker_instance = rpd_user_marker(name="Decode") + + marker_instance.start() # TODO(andoorve): We can remove this once all # virtual engines share the same kv cache. virtual_engine = model_input.virtual_engine @@ -1765,6 +1812,7 @@ def execute_model( output.hidden_states = hidden_states + marker_instance.end() return [output] def need_recv_kv(self, model_input, kv_caches) -> bool: diff --git a/vllm/worker/model_runner_base.py b/vllm/worker/model_runner_base.py index c7abad7e0258d..aef4bdcdd4bf9 100644 --- a/vllm/worker/model_runner_base.py +++ b/vllm/worker/model_runner_base.py @@ -7,6 +7,7 @@ Optional, Type, TypeVar) import torch +import torch.nn as nn from torch import is_tensor from vllm.config import VllmConfig @@ -199,6 +200,11 @@ class ModelRunnerInputBuilderBase(ABC, Generic[T]): """A builder to create ModelRunnerInputBase objects. """ + @abstractmethod + def prepare(self, + finished_requests_ids: Optional[List[str]] = None) -> None: + raise NotImplementedError + @abstractmethod def add_seq_group(self, seq_group_metadata): """TBA""" @@ -264,6 +270,10 @@ def prepare_model_input( """ raise NotImplementedError + @abstractmethod + def get_model(self) -> nn.Module: + raise NotImplementedError + def execute_model( self, model_input: T, @@ -297,9 +307,9 @@ class ModelRunnerWrapperBase: def __init__( self, - moderl_runner: ModelRunnerBase, + model_runner: ModelRunnerBase, ) -> None: - self.model_runner: ModelRunnerBase = moderl_runner + self.model_runner: ModelRunnerBase = model_runner def __getattr__(self, attr): return getattr(self.model_runner, attr) diff --git a/vllm/worker/multi_step_model_runner.py b/vllm/worker/multi_step_model_runner.py index acce923498d7e..38b2223f260dc 100644 --- a/vllm/worker/multi_step_model_runner.py +++ b/vllm/worker/multi_step_model_runner.py @@ -32,7 +32,7 @@ MULTI_STEP_ATTENTION_BACKENDS = [ "FLASH_ATTN", "ROCM_FLASH", "FLASHINFER", "NO_ATTENTION" ] -MULTI_STEP_CHUNKED_PREFILL_ATTENTION_BACKENDS = ["FLASH_ATTN"] +MULTI_STEP_CHUNKED_PREFILL_ATTENTION_BACKENDS = ["FLASH_ATTN", "FLASHINFER"] def _get_supported_attention_backends(chunked_prefill_enabled: bool) \ -> List[str]: @@ -904,4 +904,4 @@ def _pythonize_sampler_output( seq_outputs, (group_prompt_logprobs if any_logprobs_are_requested else None))) - assert len(output.outputs) > 0 + assert len(output.outputs) > 0 \ No newline at end of file diff --git a/vllm/worker/neuron_model_runner.py b/vllm/worker/neuron_model_runner.py index a35f5467e1a1f..596c26eac28bd 100644 --- a/vllm/worker/neuron_model_runner.py +++ b/vllm/worker/neuron_model_runner.py @@ -113,6 +113,9 @@ def load_model(self) -> None: raise NotImplementedError( "Supports only Transformer-NeuronX based models.") + def get_model(self) -> nn.Module: + return self.model + def _prepare_prompt( self, seq_group_metadata_list: List[SequenceGroupMetadata], diff --git a/vllm/worker/neuron_worker.py b/vllm/worker/neuron_worker.py index 3f6269684ac93..e02c72faace70 100644 --- a/vllm/worker/neuron_worker.py +++ b/vllm/worker/neuron_worker.py @@ -8,6 +8,7 @@ from vllm.distributed import (ensure_model_parallel_initialized, init_distributed_environment) from vllm.model_executor import set_random_seed +from vllm.model_executor.layers.sampler import SamplerOutput from vllm.sequence import ExecuteModelRequest from vllm.worker.neuron_model_runner import NeuronModelRunner from vllm.worker.worker_base import (LocalOrDistributedWorkerBase, @@ -25,6 +26,7 @@ def __init__( local_rank: int, rank: int, distributed_init_method: str, + is_driver_worker: bool = True, ) -> None: WorkerBase.__init__(self, vllm_config=vllm_config) self.local_rank = local_rank @@ -37,7 +39,22 @@ def __init__( self.model_runner: NeuronModelRunner = NeuronModelRunner( vllm_config=vllm_config) - self.is_driver_worker = True + self.is_driver_worker = is_driver_worker + + def execute_model( + self, + execute_model_req: Optional[ExecuteModelRequest] = None, + ) -> Optional[List[SamplerOutput]]: + assert execute_model_req is not None + assert (not execute_model_req.blocks_to_swap_in + and not execute_model_req.blocks_to_swap_out + and not execute_model_req.blocks_to_copy), ( + "Cache operations are not supported for Neuron backend.") + assert execute_model_req.num_lookahead_slots == 0, ( + "lookahead not supported for Neuron backend.") + output = LocalOrDistributedWorkerBase.execute_model( + self, execute_model_req) + return output def init_device(self) -> None: self.init_distributed_environment() @@ -103,13 +120,14 @@ def get_cache_block_size_bytes(self) -> int: def init_distributed_environment(self): """Neuron uses transformers-neuronx for tensor parallelism. - - vLLM still needs the environment inited when TP/PP > 1 + It has only one process to control multiple devices. + vLLM still needs the environment initialized when TP/PP > 1, + so we initialize a distributed environment with one process. """ init_distributed_environment( world_size=1, - rank=self.rank, - local_rank=self.local_rank, + rank=0, + local_rank=0, distributed_init_method=self.distributed_init_method, backend="gloo", ) diff --git a/vllm/worker/openvino_model_runner.py b/vllm/worker/openvino_model_runner.py index 9a69e7dce2bc5..42fe2cf668ad8 100644 --- a/vllm/worker/openvino_model_runner.py +++ b/vllm/worker/openvino_model_runner.py @@ -84,6 +84,9 @@ def load_model(self) -> None: kv_cache_dtype=self.kv_cache_dtype, ov_core=self.ov_core) + def get_model(self) -> nn.Module: + return self.model + def _prepare_model_input( self, seq_group_metadata_list: List[SequenceGroupMetadata], diff --git a/vllm/worker/openvino_worker.py b/vllm/worker/openvino_worker.py index 3482073566215..f5b46cde3969c 100644 --- a/vllm/worker/openvino_worker.py +++ b/vllm/worker/openvino_worker.py @@ -4,6 +4,7 @@ import openvino as ov import torch import torch.distributed +import torch.nn as nn import vllm.envs as envs from vllm.attention import get_attn_backend @@ -211,16 +212,14 @@ class OpenVINOWorker(LoraNotSupportedWorkerBase): def __init__( self, - ov_core: ov.Core, vllm_config: VllmConfig, local_rank: int, rank: int, distributed_init_method: str, - kv_cache_dtype: Optional[ov.Type] = ov.Type.undefined, is_driver_worker: bool = False, ) -> None: - self.ov_core = ov_core WorkerBase.__init__(self, vllm_config) + self.ov_core = ov.Core() self.parallel_config.rank = rank self.local_rank = local_rank self.rank = rank @@ -237,7 +236,7 @@ def __init__( self.model_runner = OpenVINOModelRunner( self.ov_core, vllm_config=self.vllm_config, - kv_cache_dtype=kv_cache_dtype, + kv_cache_dtype=self.vllm_config.cache_config.cache_dtype, is_driver_worker=is_driver_worker, ) # Uninitialized cache engine. Will be initialized by @@ -364,6 +363,9 @@ def cache_copy( ) -> None: self.cache_engine.copy(blocks_to_copy) # type: ignore + def get_model(self) -> nn.Module: + return self.model_runner.get_model() + @torch.inference_mode() def execute_model( self, diff --git a/vllm/worker/tpu_model_runner.py b/vllm/worker/tpu_model_runner.py index 436bf86f98220..a3f648f4cc645 100644 --- a/vllm/worker/tpu_model_runner.py +++ b/vllm/worker/tpu_model_runner.py @@ -158,6 +158,9 @@ def load_model(self) -> None: fullgraph=True, dynamic=False) + def get_model(self) -> nn.Module: + return self.model.model + def _dummy_run( self, batch_size: int, diff --git a/vllm/worker/worker.py b/vllm/worker/worker.py index 037cded60fda5..c115dc1d2a4e1 100644 --- a/vllm/worker/worker.py +++ b/vllm/worker/worker.py @@ -9,6 +9,7 @@ import vllm.envs as envs from vllm.config import VllmConfig +from vllm.device_allocator.cumem import CuMemAllocator from vllm.distributed import (ensure_kv_transfer_initialized, ensure_model_parallel_initialized, init_distributed_environment, @@ -22,7 +23,8 @@ from vllm.prompt_adapter.request import PromptAdapterRequest from vllm.sequence import (ExecuteModelRequest, IntermediateTensors, SequenceGroupMetadata, SequenceGroupMetadataDelta) -from vllm.utils import GiB_bytes, bind_kv_cache, memory_profiling +from vllm.utils import (GiB_bytes, MemorySnapshot, bind_kv_cache, + memory_profiling) from vllm.worker.cache_engine import CacheEngine from vllm.worker.enc_dec_model_runner import EncoderDecoderModelRunner from vllm.worker.model_runner import GPUModelRunnerBase, ModelRunner @@ -56,9 +58,6 @@ def __init__( self.rank = rank self.distributed_init_method = distributed_init_method self.is_driver_worker = is_driver_worker - if is_driver_worker: - assert rank % self.parallel_config.tensor_parallel_size == 0, \ - "Driver worker should be rank 0 of tensor parallel group." if self.model_config.trust_remote_code: # note: lazy import to avoid importing torch before initializing from vllm.utils import init_cached_hf_modules @@ -150,6 +149,23 @@ def stop_profile(self): else: self.profiler.stop() + def sleep(self, level: int = 1) -> None: + free_bytes_before_sleep = torch.cuda.mem_get_info()[0] + allocator = CuMemAllocator.get_instance() + allocator.sleep(offload_tags=("weights", ) if level == 1 else tuple()) + free_bytes_after_sleep, total = torch.cuda.mem_get_info() + freed_bytes = free_bytes_after_sleep - free_bytes_before_sleep + used_bytes = total - free_bytes_after_sleep + assert freed_bytes >= 0, "Memory usage increased after sleeping." + logger.info( + "Sleep mode freed %.2f GiB memory, " + "%.2f GiB memory is still in use.", freed_bytes / GiB_bytes, + used_bytes / GiB_bytes) + + def wake_up(self) -> None: + allocator = CuMemAllocator.get_instance() + allocator.wake_up() + def init_device(self) -> None: if self.device_config.device.type == "cuda": # torch.distributed.all_reduce does not free the input tensor until @@ -168,7 +184,8 @@ def init_device(self) -> None: _check_if_gpu_supports_dtype(self.model_config.dtype) gc.collect() torch.cuda.empty_cache() - self.init_gpu_memory = torch.cuda.mem_get_info()[0] + torch.cuda.reset_peak_memory_stats() + self.baseline_snapshot = MemorySnapshot() else: raise RuntimeError( f"Not support device type: {self.device_config.device}") @@ -180,7 +197,17 @@ def init_device(self) -> None: set_random_seed(self.model_config.seed) def load_model(self): - self.model_runner.load_model() + if self.vllm_config.model_config.enable_sleep_mode: + allocator = CuMemAllocator.get_instance() + assert allocator.get_current_usage() == 0, ( + "Sleep mode can only be " + "used for one instance per process.") + context = allocator.use_memory_pool(tag="weights") + else: + from contextlib import nullcontext + context = nullcontext() + with context: + self.model_runner.load_model() def save_sharded_state( self, @@ -223,10 +250,9 @@ def determine_num_available_blocks(self) -> Tuple[int, int]: # Execute a forward pass with dummy inputs to profile the memory usage # of the model. - with memory_profiling(baseline_memory_in_bytes=total_gpu_memory - - self.init_gpu_memory, - weights_memory_in_bytes=self.model_runner. - model_memory_usage) as result: + with memory_profiling( + self.baseline_snapshot, + weights_memory=self.model_runner.model_memory_usage) as result: self.model_runner.profile_run() self._assert_memory_footprint_increased_during_profiling() @@ -234,7 +260,7 @@ def determine_num_available_blocks(self) -> Tuple[int, int]: memory_for_current_instance = total_gpu_memory * \ self.cache_config.gpu_memory_utilization available_kv_cache_memory = (memory_for_current_instance - - result.non_kv_cache_memory_in_bytes) + result.non_kv_cache_memory) # Calculate the number of blocks that can be allocated with the # profiled peak memory. @@ -257,11 +283,11 @@ def determine_num_available_blocks(self) -> Tuple[int, int]: f"({self.cache_config.gpu_memory_utilization:.2f})" f" = {(memory_for_current_instance / GiB_bytes):.2f}GiB\n" "model weights take " - f"{(result.weights_memory_in_bytes / GiB_bytes):.2f}GiB;" + f"{(result.weights_memory / GiB_bytes):.2f}GiB;" " non_torch_memory takes " - f"{(result.non_torch_increase_in_bytes / GiB_bytes):.2f}GiB;" + f"{(result.non_torch_increase / GiB_bytes):.2f}GiB;" " PyTorch activation peak memory takes " - f"{(result.torch_peak_increase_in_bytes / GiB_bytes):.2f}GiB;" + f"{(result.torch_peak_increase / GiB_bytes):.2f}GiB;" " the rest of the memory reserved for KV Cache is " f"{(available_kv_cache_memory / GiB_bytes):.2f}GiB.") @@ -277,11 +303,13 @@ def determine_num_available_blocks(self) -> Tuple[int, int]: def _assert_memory_footprint_increased_during_profiling(self): # NOTE(woosuk): Here we assume that the other processes using the same # GPU did not change their memory usage during the profiling. - free_gpu_memory, _ = torch.cuda.mem_get_info() - assert self.init_gpu_memory - free_gpu_memory > 0, ( + free_gpu_memory, total = torch.cuda.mem_get_info() + cuda_memory = total - free_gpu_memory + assert self.baseline_snapshot.cuda_memory < cuda_memory, ( "Error in memory profiling. " - f"Initial free memory {self.init_gpu_memory}, current free memory" - f" {free_gpu_memory}. This happens when the GPU memory was " + f"Initial used memory {self.baseline_snapshot.cuda_memory}, " + f"currently used memory {cuda_memory}. " + f"This happens when the GPU memory was " "not properly cleaned up before initializing the vLLM instance.") def initialize_cache(self, num_gpu_blocks: int, @@ -298,7 +326,14 @@ def initialize_cache(self, num_gpu_blocks: int, self.cache_config.num_gpu_blocks = num_gpu_blocks self.cache_config.num_cpu_blocks = num_cpu_blocks - self._init_cache_engine() + if self.vllm_config.model_config.enable_sleep_mode: + allocator = CuMemAllocator.get_instance() + context = allocator.use_memory_pool(tag="kv_cache") + else: + from contextlib import nullcontext + context = nullcontext() + with context: + self._init_cache_engine() self._warm_up_model() def _init_cache_engine(self): @@ -316,6 +351,18 @@ def _init_cache_engine(self): self.gpu_cache) def _warm_up_model(self) -> None: + # warm up sizes that are not in cudagraph capture sizes, + # but users still want to compile for better performance, + # e.g. for the max-num-batched token size in chunked prefill. + warmup_sizes = self.vllm_config.compilation_config.compile_sizes.copy() + if not self.model_config.enforce_eager: + warmup_sizes = [ + x for x in warmup_sizes if x not in + self.vllm_config.compilation_config.cudagraph_capture_sizes + ] + for size in sorted(warmup_sizes, reverse=True): + logger.info("Compile and warming up model for size %d", size) + self.model_runner._dummy_run(size) if not self.model_config.enforce_eager: self.model_runner.capture_model(self.gpu_cache) # Reset the seed to ensure that the random state is not affected by diff --git a/vllm/worker/worker_base.py b/vllm/worker/worker_base.py index a835718e1db19..c3edce584142e 100644 --- a/vllm/worker/worker_base.py +++ b/vllm/worker/worker_base.py @@ -4,16 +4,20 @@ from abc import ABC, abstractmethod from typing import Any, Dict, List, Optional, Set, Tuple, Type, Union +import cloudpickle import torch +import torch.nn as nn -from vllm.config import ObservabilityConfig, VllmConfig +from vllm.config import (ObservabilityConfig, VllmConfig, + set_current_vllm_config) from vllm.distributed import broadcast_tensor_dict, get_pp_group, get_tp_group from vllm.logger import init_logger from vllm.lora.request import LoRARequest from vllm.model_executor.layers.sampler import SamplerOutput from vllm.sequence import ExecuteModelRequest, IntermediateTensors from vllm.utils import (enable_trace_function_call_for_thread, - resolve_obj_by_qualname, update_environment_variables) + resolve_obj_by_qualname, run_method, + update_environment_variables) from vllm.worker.model_runner_base import (BroadcastableModelInput, ModelRunnerBase, ModelRunnerInputBase) @@ -88,6 +92,10 @@ def start_worker_execution_loop(self) -> None: if output is None: return None + @abstractmethod + def get_model(self) -> nn.Module: + raise NotImplementedError + @abstractmethod def execute_model( self, @@ -119,6 +127,61 @@ def list_loras(self) -> Set[int]: raise NotImplementedError +class DelegateWorkerBase(WorkerBase): + """ + A class that delegates all methods to another WorkerBase instance. This is + useful for creating a WorkerBase that wraps another WorkerBase instance, + e.g. speculative decoding. + """ + worker: WorkerBase + + def __init__( + self, + *args, + **kwargs, + ) -> None: + vllm_config: VllmConfig = kwargs.get("vllm_config") + cls = resolve_obj_by_qualname(vllm_config.parallel_config.worker_cls) + self.worker = cls(*args, **kwargs) + + def init_device(self) -> None: + self.worker.init_device() + + def determine_num_available_blocks(self) -> Tuple[int, int]: + return self.worker.determine_num_available_blocks() + + def initialize_cache(self, num_gpu_blocks: int, + num_cpu_blocks: int) -> None: + self.worker.initialize_cache(num_gpu_blocks, num_cpu_blocks) + + def get_model(self) -> nn.Module: + return self.worker.get_model() + + def execute_model( + self, + execute_model_req: Optional[ExecuteModelRequest] = None + ) -> Optional[List[SamplerOutput]]: + return self.worker.execute_model(execute_model_req) + + def get_cache_block_size_bytes(self) -> int: + return self.worker.get_cache_block_size_bytes() + + def add_lora(self, lora_request: LoRARequest) -> bool: + return self.worker.add_lora(lora_request) + + def remove_lora(self, lora_id: int) -> bool: + return self.worker.remove_lora(lora_id) + + def pin_lora(self, lora_id: int) -> bool: + return self.worker.pin_lora(lora_id) + + def list_loras(self) -> Set[int]: + return self.worker.list_loras() + + def __getattr__(self, attr): + return getattr(self.worker, attr) + + class LoraNotSupportedWorkerBase(WorkerBase): """Partial implementation of WorkerBase that raises exceptions when LoRA methods are invoked. @@ -310,6 +373,9 @@ def prepare_input( else: return self._get_worker_input_from_broadcast() + def get_model(self) -> nn.Module: + return self.model_runner.get_model() + def execute_model( self, execute_model_req: Optional[ExecuteModelRequest] = None, @@ -410,7 +476,8 @@ def _execute_model_spmd( class WorkerWrapperBase: """ - The whole point of this class is to lazily initialize the worker. + This class represents one process in an executor/engine. It is responsible + for lazily initializing the worker and handling the worker's lifecycle. We first instantiate the WorkerWrapper, which remembers the worker module and class name. Then, when we call `update_environment_variables`, and the real initialization happens in `init_worker`. @@ -419,17 +486,44 @@ class WorkerWrapperBase: def __init__( self, vllm_config: VllmConfig, + rpc_rank: int = 0, ) -> None: - self.vllm_config = vllm_config - trust_remote_code = vllm_config.model_config.trust_remote_code + """ + Initialize the worker wrapper with the given vllm_config and rpc_rank. + Note: rpc_rank is the rank of the worker in the executor. In most cases, + it is also the rank of the worker in the distributed group. However, + when multiple executors work together, they can be different. + e.g. in the case of SPMD-style offline inference with TP=2, + users can launch 2 engines/executors, each with only 1 worker. + All workers have rpc_rank=0, but they have different ranks in the TP + group. + """ + self.rpc_rank = rpc_rank self.worker: Optional[WorkerBase] = None - if trust_remote_code: - # note: lazy import to avoid importing torch before initializing - from vllm.utils import init_cached_hf_modules - init_cached_hf_modules() + # do not store this `vllm_config`, `init_worker` will set the final + # one. TODO: investigate if we can remove this field in + # `WorkerWrapperBase`, `init_cached_hf_modules` should be + # unnecessary now. + if vllm_config.model_config is not None: + # it can be None in tests + trust_remote_code = vllm_config.model_config.trust_remote_code + if trust_remote_code: + # note: lazy import to avoid importing torch before initializing + from vllm.utils import init_cached_hf_modules + init_cached_hf_modules() + + def adjust_rank(self, rank_mapping: Dict[int, int]) -> None: + """ + Adjust the rpc_rank based on the given mapping. + It is only used during the initialization of the executor, + to adjust the rpc_rank of workers after we create all workers. + """ + if self.rpc_rank in rank_mapping: + self.rpc_rank = rank_mapping[self.rpc_rank] - @staticmethod - def update_environment_variables(envs: Dict[str, str]) -> None: + def update_environment_variables(self, envs_list: List[Dict[str, + str]]) -> None: + envs = envs_list[self.rpc_rank] key = 'CUDA_VISIBLE_DEVICES' if key in envs and key in os.environ: # overwriting CUDA_VISIBLE_DEVICES is desired behavior @@ -437,35 +531,43 @@ def update_environment_variables(envs: Dict[str, str]) -> None: del os.environ[key] update_environment_variables(envs) - def init_worker(self, *args, **kwargs): + def init_worker(self, all_kwargs: List[Dict[str, Any]]) -> None: """ Here we inject some common logic before initializing the worker. Arguments are passed to the worker class constructor. """ + kwargs = all_kwargs[self.rpc_rank] + self.vllm_config = kwargs.get("vllm_config", None) + assert self.vllm_config is not None, ( + "vllm_config is required to initialize the worker") enable_trace_function_call_for_thread(self.vllm_config) - # see https://github.com/NVIDIA/nccl/issues/1234 - os.environ['NCCL_CUMEM_ENABLE'] = '0' - from vllm.plugins import load_general_plugins load_general_plugins() - worker_class = resolve_obj_by_qualname( - self.vllm_config.parallel_config.worker_cls) - self.worker = worker_class(*args, **kwargs) - assert self.worker is not None - - def execute_method(self, method: str, *args, **kwargs): + if isinstance(self.vllm_config.parallel_config.worker_cls, str): + worker_class = resolve_obj_by_qualname( + self.vllm_config.parallel_config.worker_cls) + else: + assert isinstance(self.vllm_config.parallel_config.worker_cls, + bytes) + worker_class = cloudpickle.loads( + self.vllm_config.parallel_config.worker_cls) + with set_current_vllm_config(self.vllm_config): + # To make vLLM config available during worker initialization + self.worker = worker_class(**kwargs) + assert self.worker is not None + + def execute_method(self, method: Union[str, bytes], *args, **kwargs): try: target = self if self.worker is None else self.worker - executor = getattr(target, method) - return executor(*args, **kwargs) + return run_method(target, method, args, kwargs) except Exception as e: # if the driver worker also execute methods, # exceptions in the rest worker may cause deadlock in rpc like ray # see https://github.com/vllm-project/vllm/issues/3455 # print the error and inform the user to solve the error - msg = (f"Error executing method {method}. " + msg = (f"Error executing method {method!r}. " "This might cause deadlock in distributed execution.") logger.exception(msg) raise e @@ -491,4 +593,4 @@ def extract_previous_hidden_states( output["previous_hidden_states"] = data.previous_hidden_states\ .hidden_states - return output + return output \ No newline at end of file diff --git a/vllm/worker/xpu_model_runner.py b/vllm/worker/xpu_model_runner.py index 5f8acfb31765b..b7b7b7227b22c 100644 --- a/vllm/worker/xpu_model_runner.py +++ b/vllm/worker/xpu_model_runner.py @@ -113,7 +113,6 @@ def __init__(self, runner: "XPUModelRunner", finished_requests_ids: Optional[List[str]] = None) -> None: super().__init__() - self.seq_group_metadata_list: List[SequenceGroupMetadata] = [] self.runner = runner self.model_input_cls = self.runner._model_input_cls self.attn_backend = self.runner.attn_backend @@ -121,6 +120,10 @@ def __init__(self, self.block_size = self.runner.block_size self.device = self.runner.device + def prepare(self, + finished_requests_ids: Optional[List[str]] = None) -> None: + self.seq_group_metadata_list: List[SequenceGroupMetadata] = [] + def add_seq_group(self, seq_group_metadata: SequenceGroupMetadata): self.seq_group_metadata_list.append(seq_group_metadata) @@ -410,6 +413,8 @@ def __init__( SamplingMetadataCache() \ if self.parallel_config.pipeline_parallel_size == 1 else None + self.builder = self._builder_cls(weakref.proxy(self)) + def load_model(self) -> None: with DeviceMemoryProfiler() as m: self.model = get_model(vllm_config=self.vllm_config) @@ -418,6 +423,9 @@ def load_model(self) -> None: logger.info("Loading model weights took %.4f GB", self.model_memory_usage / float(2**30)) + def get_model(self) -> nn.Module: + return self.model + @property def vocab_size(self) -> int: return self.model_config.get_vocab_size() @@ -516,7 +524,8 @@ def _prepare_model_input_tensors( metadata for possible additional steps, e.g., sampling. """ - builder = self._builder_cls(weakref.proxy(self), finished_requests_ids) + builder = self.builder + builder.prepare(finished_requests_ids) for seq_group_metadata in seq_group_metadata_list: builder.add_seq_group(seq_group_metadata)