[Fix][Relax][Metal] Constrain wide-head prefill tiling - #20235
Merged
tlopex merged 3 commits intoSep 2, 2026
Conversation
tlopex
requested changes
Sep 1, 2026
Apply the existing low-storage WebGPU prefill configuration to Metal for wide attention heads. This keeps generated threadgroup allocations below Metal device limits, including Gemma 4 global attention heads.
Choose Metal prefill tiling from the target shared-memory limit and the kernel's exact Q/K/V layout. Account for unequal ragged-attention dimensions and MLA's merged KV buffer, and reject configurations that do not satisfy the scheduler's factorization constraints.\n\nShare configuration and factorization helpers with both tree-attention paths while preserving existing WebGPU behavior. Add allocation-level coverage for standard, ragged, MLA, and tree-attention kernels, custom Metal limits, and impossible limits.
Search query tiles around the heuristic when adjusting warp and key tile sizes cannot satisfy Metal shared-memory and scheduler constraints. Recheck factorization for every candidate and keep the reduction tile equal to the real head dimension. Add Metal codegen regressions for non-power-of-two head dimensions, nearby larger query tiles, and key-tile expansion.
akaashrp
force-pushed
the
upstream/metal-wide-head-prefill-tiling
branch
from
September 1, 2026 23:43
a243318 to
c10c00f
Compare
tlopex
approved these changes
Sep 2, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Apply the existing low-storage WebGPU prefill configuration to Metal for wide attention heads. This keeps generated threadgroup allocations below Metal device limits, including Gemma 4 global attention heads.