[Triton/Gluon] [MI350] Optimize FP8 MQA logits kernel - #5216

Merged

cagrikymk merged 13 commits into

mainfrom

cagri/opt_mqa_logits

Sep 9, 2026

Commits on Aug 28, 2026

fix(triton): correct use_buffer_store bound in fp8_mqa_logits

cagrikymk
and
claude
committed
8e7e428
perf(triton): retune fp8_mqa_logits gfx950 launch config

cagrikymk
and
claude
committed
1876e43

Commits on Aug 31, 2026

perf(triton): chunked head fold + 32x32 MFMA for fp8_mqa_logits on gfx950

cagrikymk
and
claude
committed
c97caaf

Commits on Sep 1, 2026

more updates
cagrikymk
committed
75ab310

Commits on Sep 2, 2026

perf(triton): drop the BLOCK_M=2 store mask and the barriers on gfx950

cagrikymk
and
claude
committed
b8a1e9a
cleanup
cagrikymk
committed
7c75429
comments
cagrikymk
committed
3189b61
Merge branch 'main' into cagri/opt_mqa_logits

cagrikymk
and
claude
committed
690e298
comment cleanup
cagrikymk
committed
021d570
formatting
cagrikymk
committed
15602d3
triton 3.8 gate
cagrikymk
committed
a1b963f

Commits on Sep 9, 2026

Fix comment formatting in fp8_mqa_logits.py

cagrikymk
and
Copilot
authored
903f989
Merge branch 'main' into cagri/opt_mqa_logits
cagrikymk
authored
6a476db