Repository navigation
Conversation
qsa_sparse_supported() warned that the Triton kernel is only tested on CUDA, then returned True anyway. On Ascend, Triton still imports, so Qwen4 training with packing or context parallelism compiles the CUDA tile sizes and aborts with a UB/Cc overflow. The kernel is now left uninstalled on non-CUDA devices. CP == 1 without packing still uses the bool-mask path. Packing and CP raise until QSA_SPARSE_KERNEL=0, which is the full-attention fallback. That fallback is not the sparse result, so it stays opt-in.
Collaborator
|
Thanks for contribution. Have you tested this on NPUs? |
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
qsa_sparse_supported()warned that the Triton kernel is only tested on CUDA, then returned True. On Ascend, Triton still imports, so Qwen4 training with packing or context parallelism compiles the CUDA tile sizes and aborts in the compiler with a UB/Cc overflow.The kernel is no longer installed when CUDA is unavailable. With context parallel size 1 and packing off, QSA still uses the bool-mask path. Packing and context parallelism keep the explicit
QSA_SPARSE_KERNEL=0fallback to full attention. That fallback is not the sparse result, so it stays opt-in. The raised error now says a non-CUDA device is one reason the kernel is absent.Checked with
python3 -m pytest tests/test_qsa_sparse_supported.py(3 passed): non-CUDA returns false, a power-of-two head on CUDA stays enabled, and a non-power-of-two head stays disabled.flake8is clean on the changed files.