upstream merge sync 2024-03-11 (#108)

SUMMARY: * upstream merge (sync) up to `657061fdced8a33a60c1b09f5da2525de9da8f03` * some minor changes related to `ruff` and `yapf` NOTES: we are now consistently getting out memory of errors when running `tests/models/test_marlin.py`. i've disabled the test and created an ASANA ticket to track the issue. TEST PLAN: runs on remote push --------- Signed-off-by: Tao He <[email protected]> Signed-off-by: Yuan Tang <[email protected]> Co-authored-by: Ronen Schaffer <[email protected]> Co-authored-by: Mustafa Eyceoz <[email protected]> Co-authored-by: Roy <[email protected]> Co-authored-by: Woosuk Kwon <[email protected]> Co-authored-by: Massimiliano Pronesti <[email protected]> Co-authored-by: 44670 <[email protected]> Co-authored-by: zhaoyang-star <[email protected]> Co-authored-by: Harry Mellor <[email protected]> Co-authored-by: Jared Moore <[email protected]> Co-authored-by: Philipp Moritz <[email protected]> Co-authored-by: Cade Daniel <[email protected]> Co-authored-by: 张大成 <[email protected]> Co-authored-by: zhangdacheng <[email protected]> Co-authored-by: Jingru <[email protected]> Co-authored-by: Dylan Hawk <[email protected]> Co-authored-by: Tao He <[email protected]> Co-authored-by: Ganesh Jagadeesan <[email protected]> Co-authored-by: Allen.Dou <[email protected]> Co-authored-by: Liangfu Chen <[email protected]> Co-authored-by: CHU Tianxiang <[email protected]> Co-authored-by: Jae-Won Chung <[email protected]> Co-authored-by: Seonghyeon <[email protected]> Co-authored-by: Billy Cao <[email protected]> Co-authored-by: Nick Hill <[email protected]> Co-authored-by: felixzhu555 <[email protected]> Co-authored-by: br3no <[email protected]> Co-authored-by: simon-mo <[email protected]> Co-authored-by: Sherry <[email protected]> Co-authored-by: Yuan Tang <[email protected]> Co-authored-by: Huarong <[email protected]> Co-authored-by: huohuarong <[email protected]> Co-authored-by: Robert Shaw <114415538+robertgshaw2-neuralmagic@users.noreply.github.com> Co-authored-by: Robert Shaw <[email protected]> Co-authored-by: alexm <[email protected]> Co-authored-by: zixiao <[email protected]> Co-authored-by: cloudhan <[email protected]> Co-authored-by: Sage Moore <[email protected]> Co-authored-by: ElizaWszola <[email protected]> Co-authored-by: Michael Goin <[email protected]> Co-authored-by: Jason Cox <[email protected]> Co-authored-by: Zhuohan Li <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: TianYu GUO <[email protected]> Co-authored-by: Jialun Lyu <[email protected]> Co-authored-by: ttbachyinsda <[email protected]> Co-authored-by: guofangze <[email protected]> Co-authored-by: Antoni Baum <[email protected]> Co-authored-by: Avnish Narayan <[email protected]> Co-authored-by: Chen Wang <[email protected]> Co-authored-by: Hongxia Yang <[email protected]> Co-authored-by: lcskrishna <[email protected]> Co-authored-by: SangBin Cho <[email protected]> Co-authored-by: Chujie Zheng <[email protected]> Co-authored-by: TechxGenus <[email protected]> Co-authored-by: Michael Goin <[email protected]> Co-authored-by: jacobthebanana <[email protected]> Co-authored-by: whyiug <[email protected]> Co-authored-by: Terry <[email protected]> Co-authored-by: Douglas Lehr <[email protected]> Co-authored-by: andy-neuma <[email protected]>
neuralmagic · Mar 12, 2024 · 7bbd2cc · 7bbd2cc
1 parent aebf20b
commit 7bbd2cc
Show file tree

Hide file tree

Showing 126 changed files with 5,123 additions and 813 deletions.
diff --git a/.buildkite/test-pipeline.yaml b/.buildkite/test-pipeline.yaml
@@ -13,6 +13,9 @@ steps:
 
 - label: Basic Correctness Test
   command: pytest -v -s --forked basic_correctness
+
+- label: Core Test
+  command: pytest -v -s core
 
 - label: Distributed Comm Ops Test
   command: pytest -v -s --forked test_comm_ops.py
@@ -25,7 +28,7 @@ steps:
   num_gpus: 2 # only support 1 or 2 for now.
 
 - label: Engine Test
-  command: pytest -v -s engine
+  command: pytest -v -s engine test_sequence.py
 
 - label: Entrypoints Test
   command: pytest -v -s entrypoints
@@ -49,6 +52,9 @@ steps:
 - label: Worker Test
   command: pytest -v -s worker
 
+- label: Speculative decoding tests
+  command: pytest -v -s spec_decode
+
 - label: LoRA Test
   command: pytest -v -s lora --forked
 

diff --git a/.github/workflows/remote-push.yml b/.github/workflows/remote-push.yml
@@ -21,7 +21,7 @@ jobs:
         uses: ./.github/workflows/build-test.yml
         with:
             label: aws-avx2-192G-4-a10g-96G
-            timeout: 180
+            timeout: 240
             gitref: '${{ github.ref }}'
             Gi_per_thread: 4
             python: ${{ matrix.python }}

diff --git a/Dockerfile.rocm b/Dockerfile.rocm
@@ -23,6 +23,9 @@ RUN echo "FA_BRANCH is $FA_BRANCH"
 # In that case, we need to use the python reference attention implementation in vllm
 ARG BUILD_FA="1"
 
+# whether to build cupy on rocm
+ARG BUILD_CUPY="1"
+
 # Install some basic utilities
 RUN apt-get update && apt-get install python3 python3-pip -y
 
@@ -70,16 +73,33 @@ RUN if [ "$BUILD_FA" = "1" ]; then \
     && cd ..; \
     fi
 
-COPY ./ /app/vllm
-
-RUN python3 -m pip install --upgrade pip
-RUN python3 -m pip install xformers==0.0.23 --no-deps
-
 # Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
 # Manually removed it so that later steps of numpy upgrade can continue
 RUN if [ "$BASE_IMAGE" = "rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1" ]; then \
     rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/; fi
 
+# build cupy
+RUN if [ "$BUILD_CUPY" = "1" ]; then \
+    mkdir -p libs \
+    && cd libs \
+    && git clone -b hipgraph_enablement --recursive https://github.com/ROCm/cupy.git \
+    && cd cupy \
+    && pip install mpi4py-mpich \
+    && pip install scipy==1.9.3 \
+    && pip install cython==0.29.* \
+    && env CC=$MPI_HOME/bin/mpicc python -m pip install mpi4py \
+    && export CUPY_INSTALL_USE_HIP=1 \
+    && export ROCM_HOME=/opt/rocm \
+    && export HCC_AMDGPU_TARGET="gfx90a,gfx942,gfx1100" \
+    && pip install . \
+    && cd ..; \
+    fi
+
+COPY ./ /app/vllm
+
+RUN python3 -m pip install --upgrade pip
+RUN python3 -m pip install xformers==0.0.23 --no-deps
+
 RUN cd /app \
     && cd vllm \
     && pip install -U -r requirements-rocm.txt \

diff --git a/README.md b/README.md
@@ -27,7 +27,7 @@ pip install -e .
 
 ## Quickstart
 
-Neural Magic maintains a variety of sparse models on our Hugging Face organization profiles, [neuralmagic](https://huggingface.co/neuralmagic) and [nm-testing](https://huggingface.co/nm-testing). 
+Neural Magic maintains a variety of sparse models on our Hugging Face organization profiles, [neuralmagic](https://huggingface.co/neuralmagic) and [nm-testing](https://huggingface.co/nm-testing).
 
 A collection of ready-to-use SparseGPT and GPTQ models in inference optimized marlin format are [available on Hugging Face](https://huggingface.co/collections/neuralmagic/compressed-llms-for-nm-vllm-65e73e3d51d3200e34b77431)
 
@@ -63,7 +63,7 @@ For a quick demonstration, here's how to run a small [50% sparse llama2-110M](ht
 from vllm import LLM, SamplingParams
 
 model = LLM(
-    "neuralmagic/llama2.c-stories110M-pruned50", 
+    "neuralmagic/llama2.c-stories110M-pruned50",
     sparsity="sparse_w16a16",   # If left off, model will be loaded as dense
 )
 

diff --git a/benchmarks/backend_request_func.py b/benchmarks/backend_request_func.py
@@ -277,10 +277,80 @@ async def async_request_openai_completions(
     return output
 
 
+async def async_request_openai_chat_completions(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith(
+        "v1/chat/completions"
+    ), "OpenAI Chat API URL must end with 'v1/chat/completions'."
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert not request_func_input.use_beam_search
+        payload = {
+            "model": request_func_input.model,
+            "messages": [
+                {
+                    "role": "user",
+                    "content": request_func_input.prompt,
+                },
+            ],
+            "temperature": 0.0,
+            "max_tokens": request_func_input.output_len,
+            "stream": True,
+        }
+        headers = {
+            "Content-Type": "application/json",
+            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"
+        }
+
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        generated_text = ""
+        ttft = 0
+        st = time.perf_counter()
+        try:
+            async with session.post(url=api_url, json=payload,
+                                    headers=headers) as response:
+                if response.status == 200:
+                    async for chunk in response.content:
+                        if ttft == 0:
+                            ttft = time.perf_counter() - st
+                            output.ttft = ttft
+
+                        chunk = chunk.strip()
+                        if not chunk:
+                            continue
+
+                        chunk = chunk.decode("utf-8").lstrip("data: ")
+                        if chunk == "[DONE]":
+                            latency = time.perf_counter() - st
+                        else:
+                            body = json.loads(chunk)
+                            if "content" in body["choices"][0]["delta"]:
+                                generated_text += body["choices"][0]["delta"][
+                                    "content"]
+
+                    output.generated_text = generated_text
+                    output.success = True
+                    output.latency = latency
+                else:
+                    output.success = False
+        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
+            output.success = False
+
+    if pbar:
+        pbar.update(1)
+    return output
+
+
 ASYNC_REQUEST_FUNCS = {
     "tgi": async_request_tgi,
     "vllm": async_request_vllm,
     "deepspeed-mii": async_request_deepspeed_mii,
     "openai": async_request_openai_completions,
+    "openai-chat": async_request_openai_chat_completions,
     "tensorrt-llm": async_request_trt_llm,
 }
diff --git a/benchmarks/benchmark_serving.py b/benchmarks/benchmark_serving.py
@@ -12,7 +12,7 @@
 On the client side, run:
     python benchmarks/benchmark_serving.py \
         --backend <backend> \
-        --tokenizer <your_model> --dataset <target_dataset> \
+        --model <your_model> --dataset <target_dataset> \
         --request-rate <request_rate>
 """
 import argparse
@@ -171,10 +171,10 @@ async def benchmark(
     else:
         raise ValueError(f"Unknown backend: {backend}")
 
-    pbar = None if disable_tqdm else tqdm(total=len(input_requests))
-
     print(f"Traffic request rate: {request_rate}")
 
+    pbar = None if disable_tqdm else tqdm(total=len(input_requests))
+
     benchmark_start_time = time.perf_counter()
     tasks = []
     async for request in get_request(input_requests, request_rate):

diff --git a/csrc/attention/attention_kernels.cu b/csrc/attention/attention_kernels.cu
@@ -15,9 +15,6 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#ifdef USE_ROCM
-#include <hip/hip_runtime.h>
-#endif
 
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
@@ -31,11 +28,6 @@
 
 #include <algorithm>
 
-#ifndef USE_ROCM
-#define WARP_SIZE 32
-#else
-#define WARP_SIZE warpSize
-#endif
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))

diff --git a/csrc/cuda_compat.h b/csrc/cuda_compat.h
@@ -1,5 +1,15 @@
 #pragma once
 
+#ifdef USE_ROCM
+#include <hip/hip_runtime.h>
+#endif
+
+#ifndef USE_ROCM
+  #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
+#endif
+
 #ifndef USE_ROCM
   #define VLLM_LDG(arg) __ldg(arg)
 #else

diff --git a/csrc/punica/bgmv/bgmv_config.h b/csrc/punica/bgmv/bgmv_config.h
@@ -14,13 +14,15 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 128) \
     f(in_T, out_T, W_T, narrow, 256) \
     f(in_T, out_T, W_T, narrow, 512) \
+    f(in_T, out_T, W_T, narrow, 768) \
     f(in_T, out_T, W_T, narrow, 1024) \
     f(in_T, out_T, W_T, narrow, 1280) \
     f(in_T, out_T, W_T, narrow, 1728) \
     f(in_T, out_T, W_T, narrow, 1792) \
     f(in_T, out_T, W_T, narrow, 2048) \
     f(in_T, out_T, W_T, narrow, 2560) \
     f(in_T, out_T, W_T, narrow, 2752) \
+    f(in_T, out_T, W_T, narrow, 2816) \
     f(in_T, out_T, W_T, narrow, 3072) \
     f(in_T, out_T, W_T, narrow, 3456) \
     f(in_T, out_T, W_T, narrow, 3584) \
@@ -36,6 +38,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 10240) \
     f(in_T, out_T, W_T, narrow, 11008) \
     f(in_T, out_T, W_T, narrow, 12288) \
+    f(in_T, out_T, W_T, narrow, 13696) \
     f(in_T, out_T, W_T, narrow, 13824) \
     f(in_T, out_T, W_T, narrow, 14336) \
     f(in_T, out_T, W_T, narrow, 16384) \

diff --git a/csrc/reduction_utils.cuh b/csrc/reduction_utils.cuh
@@ -24,15 +24,15 @@ namespace vllm {
 template<typename T>
 __inline__ __device__ T warpReduceSum(T val) {
 #pragma unroll
-  for (int mask = 16; mask > 0; mask >>= 1)
+  for (int mask = WARP_SIZE/2; mask > 0; mask >>= 1)
     val += VLLM_SHFL_XOR_SYNC(val, mask);
   return val;
 }
 
 /* Calculate the sum of all elements in a block */
 template<typename T>
 __inline__ __device__ T blockReduceSum(T val) {
-  static __shared__ T shared[32];
+  static __shared__ T shared[WARP_SIZE];
   int lane = threadIdx.x & 0x1f;
   int wid = threadIdx.x >> 5;
 
@@ -45,7 +45,7 @@ __inline__ __device__ T blockReduceSum(T val) {
 
   // Modify from blockDim.x << 5 to blockDim.x / 32. to prevent
   // blockDim.x is not divided by 32
-  val = (threadIdx.x < (blockDim.x / 32.f)) ? shared[lane] : (T)(0.0f);
+  val = (threadIdx.x < (blockDim.x / (WARP_SIZE * 1.0f))) ? shared[lane] : (T)(0.0f);
   val = warpReduceSum<T>(val);
   return val;
 }

diff --git a/docs/source/conf.py b/docs/source/conf.py
@@ -72,8 +72,15 @@
 
 # Mock out external dependencies here.
 autodoc_mock_imports = [
-    "torch", "transformers", "psutil", "prometheus_client", "sentencepiece",
-    "vllm.cuda_utils", "vllm._C"
+    "torch",
+    "transformers",
+    "psutil",
+    "prometheus_client",
+    "sentencepiece",
+    "vllm.cuda_utils",
+    "vllm._C",
+    "numpy",
+    "tqdm",
 ]
 
 for mock_target in autodoc_mock_imports:

diff --git a/docs/source/models/lora.rst b/docs/source/models/lora.rst
@@ -90,9 +90,9 @@ Requests can specify the LoRA adapter as if it were any other model via the ``mo
 processed according to the server-wide LoRA configuration (i.e. in parallel with base model requests, and potentially other
 LoRA adapter requests if they were provided and ``max_loras`` is set high enough).
 
-The following is an example request 
+The following is an example request
 
-.. code-block::bash 
+.. code-block::bash
     curl http://localhost:8000/v1/completions \
         -H "Content-Type: application/json" \
         -d '{