torch custom_op support: norm #552
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
yzh119
approved these changes
Oct 24, 2024
yzh119
pushed a commit
that referenced
this pull request
Oct 25, 2024
Follow up of #552. This PR adds torch library annotation to all FlashInfer kernels so that torch.compile can recognize the kernels. Most changes are tedious. I manually ran subsets of pytest test cases when I made these changes, but since there are too many of them and also some of them didn't pass even before I made the change, I cannot guarantee it's all working. To run tests with torch.compile, pass `FLASHINFER_TEST_TORCH_COMPILE=1` env. ```bash # With torch.compile FLASHINFER_TEST_TORCH_COMPILE=1 pytest -svx tests/test_norm.py # Without torch.compile pytest -svx tests/test_norm.py ``` Notable changes: * For the prefill and decode pybind, it used to return `Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]` depending on `return_lse`. This causes trouble for `torch.compile`. I changed the pybind interface to accept a `maybe_lse: Optional[torch.Tensor]` and only return one tensor. The allocation of the lse tensor is moved to Python side. The Python API does not change. * `chain_speculative_sampling` pybind: Move the allocation of `accepted` and `emitted` from C++ to Python. This is because `torch.compile` doesn't like returning input tensor as output tensor. The Python API does not change. Piggyback changes: * `BatchPrefillWithRaggedKVCacheWrapper.plan`: Bugfix qo_indptr not on CPU * `merge_state`: Fix typo in docs * Change `run_return_lse(...)` to `run(..., return_lse=True)` because torch.compile does not recognize `functools.partial`. * In tests, change `flashinfer.xxx()` to `flashinfer.<module>.xxx()` so that the monkeypatch works. Unsupported for torch.compile: * `flashinfer.quantization.segment_packbits`: Because it's data dependent. Untouched: * `sparse.py`: Tests didn't pass beforehand, so I skiped this. Also, it doesn't seem like need custom_op annotations, as it does not have CUDA kernels. Failed test cases: * batch_decode non contiguous kv: `test_batch_decode_with_paged_kv_cache[False-kv_dtype0-q_dtype0-True-0.0-NONE-NHD-128-4-4-1-54-12]`
Merged
yzh119
pushed a commit
that referenced
this pull request
Oct 31, 2024
Here's the reason why docs fail to build after #552: As specified in `conf.py`, Sphinx mocks `torch`. The mock makes the following predicate behave badly: `TorchVersion(torch_version) < TorchVersion("2.4")`. The fix is to explicitly pass in an env var indicating docs building. Also changing the way that `prefill.py` imports compiled `_kernels` so that it's consistent with other files.
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.
Add torch custom_op (aka, torch library, torch.compile) support for
norm.py. It should be a no-op for PyTorch < 2.4.Testing is done by
torch.compile-- as we expect the custom_op marks can isolate out our kernels during torch.compile. To avoid changes to tests, I introduced some magic that replaces the kernels with atorch.compile-ed version. For example, to run with/without torch.compile:If this PR looks good, I'll add it to more kernels.