Enable the deterministic mode option in the backward kernel #1766
+7
−5
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
|
Hi @tridao @sgrigory , may I know if anyone can help with reviewing this PR? We found that flash-attn v3 built from commits this year cannot have deterministic results even when the deterministic mode is ON. Then we deep dived into the issue a bit, and found that the flag was dropped in the BWD template. This PR aims to fix the issue. But please let me know if there are issues other than slowing down the cases with head_dim=256. |
GD06
commented
Aug 13, 2025
| @@ -607,7 +609,7 @@ struct CollectiveMainloopBwdSm90 { | |||
| seqlen_info, n_block, bidb, params.window_size_left, | |||
| params.window_size_right, 0 /*sink_token_length*/); | |||
| // It's possible to have m_block_max <= m_block_min. Exit early | |||
| if constexpr (Is_causal || Is_local || Varlen) { | |||
| if constexpr ((Is_causal || Is_local || Varlen) && !Deterministic) { | |||
|
To check that the output is deterministic, maybe add this test to |
|
there is another PR #1678 which has the same fix but with more unit test coverage. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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 deterministic flag is dropped in the BWD launch template. This PR adds back the deterministic flag, and turn off "Slice_dQKV_Mma" when the deterministic mode is ON for head_dim=256.
NOTE: This could lead to register pressures, and slow down kernel performance for head_dim=256 when running with the deterministic mode.