Simplify and standardize flex attention interface #90
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
Renames parameter names for consistency and adds tensor transpositions to match expected input format. Removes attention weights/LSE computation and return value to streamline the interface, keeping only the attention output. Disables compilation for debugging purposes and updates parameter name from scaling to scale.
Renames function parameters to follow more conventional naming patterns: - `causal` becomes `is_causal` for boolean clarity - `q`, `k`, `v` become `query`, `key`, `value` for readability - `mask`, `bias` become `attn_mask`, `attn_bias` for specificity Updates function signatures, internal usage, and wrapper function to maintain consistency throughout the codebase.
LoserCheems
requested review from
Evanwu1125,
SNHuan,
Copilot and
wubingheng111
and removed request for
Copilot
Contributor
There was a problem hiding this comment.
Pull Request Overview
This PR streamlines the flex attention interface by standardizing parameter names and removing unnecessary return values. The changes focus on improving clarity and consistency across the attention implementation.
Key changes:
- Rename parameters from
causaltois_causaland standardize attention parameter names (q/k/vtoquery/key/value,mask/biastoattn_mask/attn_bias) - Remove unnecessary return values from flex attention function (attention weights/LSE)
- Add tensor transpositions to match expected input formats and disable compilation for debugging
Reviewed Changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
| flash_dmattn/flash_dmattn_triton.py | Standardizes parameter names throughout the triton implementation and creates wrapper function |
| flash_dmattn/flash_dmattn_flex.py | Adds tensor transpositions, removes attention weights return value, and disables compilation |
Comment on lines
+20
to
+21
| attn_mask = attn_mask[:, :, :, : key.shape[-2]] | ||
| attn_bias = attn_bias[:, :, :, : key.shape[-2]] |
There was a problem hiding this comment.
The tensor slicing uses key.shape[-2] but after transposition on line 18, the key tensor shape has changed. This should use the sequence length dimension from the transposed tensor, which would be key.shape[2] instead of key.shape[-2].
Suggested change
| attn_mask = attn_mask[:, :, :, : key.shape[-2]] | |
| attn_bias = attn_bias[:, :, :, : key.shape[-2]] | |
| attn_mask = attn_mask[:, :, :, : key.shape[2]] | |
| attn_bias = attn_bias[:, :, :, : key.shape[2]] |
| key = key.transpose(1, 2).contiguous() # [B, H, KV_LEN, D] | ||
| value = value.transpose(1, 2).contiguous() # [B, H, KV_LEN, D] | ||
| attn_mask = attn_mask[:, :, :, : key.shape[-2]] | ||
| attn_bias = attn_bias[:, :, :, : key.shape[-2]] |
There was a problem hiding this comment.
Similar to the attn_mask slicing, this uses key.shape[-2] but should use key.shape[2] after the tensor transposition performed on line 18.
Suggested change
| attn_bias = attn_bias[:, :, :, : key.shape[-2]] | |
| attn_bias = attn_bias[:, :, :, : key.shape[2]] |
Activates the compile flag to improve performance through kernel optimization during flex attention computation.
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.
Streamline the flex attention interface by removing unnecessary return values and standardizing parameter names for clarity and consistency. Adjustments include renaming parameters and ensuring tensor transpositions match expected input formats. Disabling compilation aids in debugging.