Fix FMHA bwd_nl kernel issue for batch_size == 3 #1975
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
The bwd_nl (no-loop backward) kernel produces incorrect results for batch_size == 3 due to a num_chunks configuration issue in the CUDA kernel. For batch_size == 3, the default num_chunks = 2 is used, but this configuration is buggy (produces 0.45 error vs expected 0.001). This fix changes the condition from `batch_size < 4` to `batch_size <= 2` for using the bwd_nl path, falling back to the regular bwd function for batch_size >= 3 which works correctly. Test results before fix: - b=2: bwd_nl works (max diff: 0.001) - b=3: bwd_nl FAILS (max diff: 0.452) - b=4: bwd works (max diff: 0.0005) Test results after fix: - b=2: bwd_nl works (max diff: 0.001) - b=3: bwd works (max diff: 0.0005) - b=4: bwd works (max diff: 0.0005) Signed-off-by: Masaki Kozuki <mkozuki@nvidia.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.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.
The bwd_nl (no-loop backward) kernel produces incorrect results for batch_size == 3 due to a num_chunks configuration issue in the CUDA kernel. For batch_size == 3, the default num_chunks = 2 is used, but this configuration is buggy (produces 0.45 error vs expected 0.001).
This fix changes the condition from
batch_size < 4tobatch_size <= 2for using the bwd_nl path, falling back to the regular bwd function for batch_size >= 3 which works correctly.Test results before fix:
Test results after fix:
Used Claude Opus 4.5