Repository navigation

perf(cuda): release MMA stages earlier and flatten native weight tiles - #137

Merged

jinzhen-lin merged 6 commits into

mainfrom

feat/flat-native-weights

Oct 13, 2026

Merged

perf(cuda): release MMA stages earlier and flatten native weight tiles#137
jinzhen-lin merged 6 commits into
mainfrom
feat/flat-native-weights

Commits on Oct 12, 2026
- perf(cuda): release consumed MMA stages before computation
  
  jinzhen-lin
  and
  codex
  committed
  9a7b388
- perf(cuda): use flat tiles for native MMA weights
  
  jinzhen-lin
  and
  codex
  committed
  6e2c9e3
- perf(cuda): relax Stream-K stage alignment for flat weights
  
  jinzhen-lin
  and
  codex
  committed
  691ec04
- perf(cuda): transpose flat weight groups and separate raw layouts
  
  jinzhen-lin
  and
  codex
  committed
  3804574
Commits on Oct 13, 2026
- merge: integrate main into flat native weights
  
  jinzhen-lin
  and
  codex
  committed
  f78e17d
- fix(sm90): mask cluster rank bits in WGMMA descriptors
  
  jinzhen-lin
  and
  codex
  committed
  37c386e