Skip to content

Alizaidy/bf16 gemm tuning 081426 - #4749

Open
azaidy wants to merge 10 commits into
mainfrom
alizaidy/bf16-gemm-tuning-081426
Open

Alizaidy/bf16 gemm tuning 081426#4749
azaidy wants to merge 10 commits into
mainfrom
alizaidy/bf16-gemm-tuning-081426

Conversation

@azaidy

@azaidy azaidy commented Aug 14, 2026

Copy link
Copy Markdown
Contributor

GEMM A16W16 tuned kernel time (µs)

N K Shapes Min Max Geomean
64 7168 12 4.61x 12.68x 7.79x
384 7168 14 5.16x 14.22x 8.84x
512 7168 14 5.09x 14.25x 8.69x
1024 7168 15 4.18x 14.11x 8.50x
2048 7168 15 4.29x 16.25x 8.43x
32320 7168 15 2.82x 15.58x 5.90x

@azaidy
azaidy requested review from a team, ahmed-bsod, k50112113 and valarLip August 14, 2026 05:00
@github-actions

Copy link
Copy Markdown
Contributor

🏷️ CI Guide

Runs automatically on every PR:

  • ✅ Pre-checks (submodule verification, code formatting)
  • ✅ Aiter op tests (gfx942 + gfx950)
  • ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label Tests
ci:gfx1250-ffm-triton Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 4749 --add-label <label>

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants