[webgpu] Optimize FlashAttention for prefill #25395
Merged
+33
−8
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
This PR enhances unidirectional `FlashAttention` by applying causal masking inside the main loop. This optimization eliminates unnecessary memory loads by avoiding future entries in the KV cache. Testing on Lunar Lake shows up to a 20% performance improvement for `phi-4-mini-accuracy4` (with a prompt of 4096). Similar performance gains were also observed for other models, including `Qwen3-0.6B-accuracy4`. This PR now uses the more readable `unidirectional` attribute instead of `is_gpa`, to control causal masking.
|
Lunar Lake, Phi-4-mini-accuracy4:
|
|
@sushraja-msft @qjia7 pls take a look. |
qjia7
reviewed
Jul 15, 2025
qjia7
reviewed
Jul 15, 2025
|
/azp run Linux QNN CI Pipeline, Win_TRT_Minimal_CUDA_Test_CI, Windows ARM64 QNN CI Pipeline, Windows GPU Doc Gen CI Pipeline, Windows x64 QNN CI Pipeline |
fs-eire
approved these changes
Jul 22, 2025
|
Azure Pipelines successfully started running 5 pipeline(s). |
sanketkaleoss
pushed a commit
to sanketkaleoss/onnxruntime
that referenced
this pull request
Aug 11, 2025
### Description This PR enhances unidirectional `FlashAttention` by applying causal masking inside the main loop. This optimization eliminates unnecessary memory loads by avoiding future entries in the KV cache. Testing on Lunar Lake shows up to a 20% performance improvement for `phi-4-mini-accuracy4` (with a prompt of 4096). Similar performance gains were also observed for other models, including `Qwen3-0.6B-accuracy4`. This PR now uses the more readable `unidirectional` attribute instead of `is_gqa`, to control causal masking. ### Motivation and Context See above.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
Description
This PR enhances unidirectional
FlashAttentionby applying causal masking inside the main loop. This optimization eliminates unnecessary memory loads by avoiding future entries in the KV cache.Testing on Lunar Lake shows up to a 20% performance improvement for
phi-4-mini-accuracy4(with a prompt of 4096). Similar performance gains were also observed for other models, includingQwen3-0.6B-accuracy4.This PR now uses the more readable
unidirectionalattribute instead ofis_gqa, to control causal masking.Motivation and Context
See above.