Repository navigation

feat(lowering): optionally run multi-query attention against the unrepeated key/value head - #4921

Merged

lanluo-nvidia merged 2 commits into

pytorch:mainfrom

shoumikhin:fold-repeated-kv-matmul

Oct 28, 2026

Merged

feat(lowering): optionally run multi-query attention against the unrepeated key/value head#4921
lanluo-nvidia merged 2 commits into
pytorch:mainfrom
shoumikhin:fold-repeated-kv-matmul

Commits on Oct 26, 2026
- fix(lowering): set meta["val"] on the nodes replace_fused_rms_norm creates
  shoumikhin
  committed
  cce4f08
- feat(lowering): optionally run multi-query attention against the unrepeated key/value head
  shoumikhin
  committed
  a2273a6