JIT compilation support for TVM #880
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
MasterJH5574
force-pushed
the
02-19-tvm-jit
branch
from
e02fb8a to
7636b6b
Compare
This PR introduces the FlashInfer JIT compilation for TVM, with corresponding TVM bindings. Compared with Torch-based JIT which returns the compiled module, the JIT for TVM returns the generated uri and source files directly, which will be compiled and loaded as a TVM runtime module on TVM side. Some notes: * SM90 prefill is not fully enabled due to the layout mismatch of `indptr`. This will be addressed in the near future. * Unit tests are not yet included. We are still working on getting a plan to test TVM bindings in FlashInfer. * The previous TVM bindings in `src/tvm_wrapper.cu` is removed, and AOT compilation for TVM is no longer supported since this PR.
MasterJH5574
force-pushed
the
02-19-tvm-jit
branch
from
7636b6b to
2c27b99
Compare
yzh119
approved these changes
Feb 19, 2025
yzh119
pushed a commit
that referenced
this pull request
Feb 20, 2025
`flashinfer.jit.attention` package is introduced in #880, but it's not added into the packages list when building wheel. Python will complain about the package not found during import. Repro steps: ``` FLASHINFER_ENABLE_AOT=1 python -m build --no-isolation --wheel pip install --force-reinstall dist/flashinfer_python-0.2.1.post2-cp38-abi3-linux_x86_64.whl ``` Log: ``` Python 3.11.11 (main, Dec 06 2024, 17:06:18) [GCC] on linux Type "help", "copyright", "credits" or "license" for more information. >>> import flashinfer Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/build/ktransformers/venv/lib64/python3.11/site-packages/flashinfer/__init__.py", line 18, in <module> from .activation import gelu_and_mul as gelu_and_mul File "/build/ktransformers/venv/lib64/python3.11/site-packages/flashinfer/activation.py", line 21, in <module> from .jit import gen_act_and_mul_module, has_prebuilt_ops, load_cuda_ops # noqa: F401 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/build/ktransformers/venv/lib64/python3.11/site-packages/flashinfer/jit/__init__.py", line 20, in <module> from .attention import gen_batch_decode_mla_module as gen_batch_decode_mla_module ModuleNotFoundError: No module named 'flashinfer.jit.attention' ```
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR introduces the FlashInfer JIT compilation for TVM, with corresponding TVM bindings. Compared with Torch-based JIT which returns the compiled module, the JIT for TVM returns the generated uri and source files directly, which will be compiled and loaded as a TVM runtime module on TVM side.
Some notes:
indptr. This will be addressed in the near future.src/tvm_wrapper.cuis removed, and AOT compilation for TVM is no longer supported since this PR.