[Triton/Gluon] gemm A16w16 tune for disable amdgpu-trackers - #5137
Open
yanxuer-999 wants to merge 3 commits into
Open
[Triton/Gluon] gemm A16w16 tune for disable amdgpu-trackers#5137yanxuer-999 wants to merge 3 commits into
yanxuer-999 wants to merge 3 commits into
Conversation
…42% on llama3 GQA decode (triton 3.8 VGPR 160->202 occupancy drop) Co-authored-by: Cursor <cursoragent@cursor.com> (cherry picked from commit 429d355)
…Q_32 8->4, M_LEQ_256 6->4) to remove tracker-off VGPR spill Cherry-picked from release/0829/aiter 34fae60; file remapped to the new config layout (configs/gfx950/triton/gemm/gemm_a16w16/) since the flat configs/gemm/gfx950-*.json path was reorganized on main. Co-authored-by: Cursor <cursoragent@cursor.com>
Contributor
🏷️ CI GuideRuns automatically on every PR:
Extended tests (opt-in via labels):
PR title tags & labels: |
Dewei-Wang-sh
approved these changes
Aug 31, 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
Restore two triton 3.8 performance regressions on gfx950 by cherry-picking validated tuning fixes: paged-attention GQA decode (llama3, ~-18%) and GEMM-A16W16 at N=1280/K=8192 (tracker-off VGPR spill).
Technical Details
pa_decode.py: set waves_per_eu=2 on the two v1 w_dot decode kernels (non-quant + per-token FP8), so regalloc gets the full 256-VGPR budget instead of triton 3.8's inflated 202-VGPR/occupancy-2 state (s_waitcnt lgkmcnt(0) 72→50, s_nop 36→9); v2 w_dot is intentionally left unchanged because waves_per_eu=2 there spills and regresses.
GEMM-A16W16-N=1280-K=8192.json: new gfx950 config lowering waves_per_eu (M_LEQ_32 8→4, M_LEQ_256 6→4) to remove the tracker-off VGPR spill on this shape.
Test Result
llama3 pa_decode 8B/70B/405B recovered +43/+43/+42% (deepseek-V3/wo_dot unaffected at 0%); GEMM-A16W16 N=1280 problem M sizes recovered +242/+151/+157% (M=32/192/256) with no regression on untuned shapes (<2% spread).