[XPU] Add flash_attn2 support for XPU #41956
Draft
+143
−44
Conversation
This file contains hidden or 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
yao-matrix
reviewed
Oct 30, 2025
|
[For maintainers] Suggested jobs to run (before merge) run-slow: bamba, diffllama, ernie4_5_moe, esm, glm4, gpt2, jamba, jetmoe, kosmos2_5, m2m_100, ministral, mixtral, modernbert, seed_oss, zamba, zamba2 |
|
Once huggingface/kernels-community/pull/59 has been merged and the binary files has been uploaded to kernels-community/flash-attn2, we can begin the review. Cc @yao-matrix . |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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.
What does this PR do?
XPU has now implemented the basic functionality of
flash_attn2in kernels-community/flash-attn2. So this PR adds the function call to transformers.Main contributions:
flash_attn2support for XPU;flash_attn2UTs on XPU.Now we can use
attn_implementation="flash_attention_2"orattn_implementation="kernels-community/flash-attn"on XPU to invoke this feature.Note (CI):
XXX::test_flash_attn_2_inference_equivalenceandXXX::test_flash_attn_2_equivalence, etc. , will sometimes pass and sometimes fail due to some non-determinism in theflash-attn2kernel implementation. I have observed this phenomenon on both CUDA (A100) and XPU.tests/models/kosmos2/test_modeling_kosmos2.py::Kosmos2ModelTest::test_eager_matches_fa2_generate, CUDA will trigger aRuntimeError: cu_seqlens_q must have shape (batch_size + 1), while XPU will directly cause aAborted (core dumped). This appears to be caused by the robustness of the underlying data reading function, but the root cause is the test case issue.