UPSTREAM PR #16827: Massively Improved ROCm/HIP rocWMMA Performance (pp and tg)#12

Closed

DajanaV wants to merge 2 commits intomainfrom

upstream-PR16827-branch_lhl-rocm-wmma-tune

Commits on Oct 28, 2025

HIP/WMMA: retune WMMA FlashAttention on RDNA3\n\n- Increase block residency on HIP via __launch_bounds__ (min 2 blocks/SM)\n- Adaptive KQ stride on HIP: 128 for D<=128 to reduce LDS footprint\n- Up…
lhl
committed
HIP: use WMMA for prefill only; fix decode regression by enabling TILE and adding a safe fallback\n\n- Do not select WMMA for decode on HIP; fall through to VEC/TILE\n- Remove WMMA TILE pruning on …
lhl
committed