[release/2.7] Optimize Flex-Attention occupancy for head_dim=128#2957
Merged
jataylo merged 1 commit intorelease/2.7from Feb 11, 2026
Merged
Conversation
Adjust n_warps from 8 to 4 for head_dim=128 configurations to improve performance stability across different attention patterns. - Forward speedup: ~1.07x geomean uplift. - Backward speedup: 1.27x to 1.60x geomean uplift. - Validated with a filtered sweep of 320 unique cases.
tomjen12
changed the title
|
Jenkins build for 4b7cecf5fbae5ab9ef60a3b7ae25abfa63074041 commit finished as FAILURE |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
[ROCm] Optimize Flex-Attention occupancy for head_dim=128
Summary
This PR adjusts
n_warpsfrom 8 to 4 forhead_dim=128configurations on ROCm for gfx942.Performance Impact (320 Valid Cases)
The following table shows the Geometric Mean (Geomean) speedups compared to the current
n_warps=8baseline:Benchmark Coverage:
window_sizeandprefix_lm∈ [128, 256, 512, 1024, 2048].