Repository navigation

[AMD] Optimize Qwen3.5 MTP unified attention on gfx950 - #36330

Merged

HaiShaw merged 6 commits into

sgl-project:mainfrom

yichiche:perf/gfx950-qwen35-mtp-unified-attention

Aug 27, 2026

Merged

[AMD] Optimize Qwen3.5 MTP unified attention on gfx950#36330
HaiShaw merged 6 commits into
sgl-project:mainfrom
yichiche:perf/gfx950-qwen35-mtp-unified-attention

Commits on Aug 25, 2026
- [AMD] Optimize Qwen3.5 MTP unified attention on gfx950
  yichiche
  committed
  2e742d3
- [AMD] Document the Qwen3.5 MTP kernel's divergence from upstream AITER
  yichiche
  committed
  90d832d
Commits on Aug 26, 2026
- [AMD] Gate the Qwen3.5 MTP kernel on the GQA ratio, not absolute head counts
  yichiche
  committed
  f62ee5d
Commits on Aug 27, 2026