Add SDPA decomposition for Metal backend with unsupported head_dim#221
Open
seyeong-han wants to merge 1 commit intohuggingface:mainfrom
Open
Add SDPA decomposition for Metal backend with unsupported head_dim#221seyeong-han wants to merge 1 commit intohuggingface:mainfrom
seyeong-han wants to merge 1 commit intohuggingface:mainfrom
Conversation
The Metal SDPA kernel only supports head_dim in {64, 96, 128}. Models
like Gemma3 (head_dim=256) crash at runtime. This decomposes SDPA into
matmul + softmax when the model's head_dim is unsupported, following
the pattern from voxtral_realtime's Metal export.
Changes:
- metal.py: Add _sdpa_decomposition and _linear_bias_decomposition,
applied via run_decompositions() before lowering. Conditional on
head_dim not in {64, 96, 128}. Force use_custom_sdpa=False for Metal.
- integrations.py: Guard get_custom_sdpa_for_ring_kv_cache() and
RemoveRedundantTransposes imports behind use_custom_sdpa check to
avoid triggering torchao import chain when custom SDPA is not used.
This PR was authored with the assistance of Claude.
Made-with: Cursor
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 Metal SDPA kernel only supports head_dim in {64, 96, 128}. Models like
Gemma3(head_dim=256) crash at runtime. This decomposes SDPA into matmul + softmax when the model's head_dim is unsupported, following the pattern from voxtral_realtime's Metal export.Changes: