biondizzle 5c2d9ad312 FMHA v3: per-row patch from Mike + deadlock fix + V layout fix ...

- test_fmha_v3_per_row.py: Mike's per-row patch with deadlock fix
  (moved C6 O-rescale after softmax_done_bar, fixed pv_done_bar for kt=0)
  Still GPU hangs — needs further debugging
- test_fmha_v3_fixed_v.py: s_k parameter + acc_pipe consumer fix
  Same cosine as original (V TMA handles data shape correctly)
- Baseline: n=128→0.993, n=256→0.725, n=384→0.620

Key insight: QK TMEM load fragment has 4 rows × 32 cols per thread.
Fragment-level row_max/row_sum is wrong for per-row operations.
Per-row tracking (4 separate row_max/row_sum per thread) is needed.

2026-05-22 07:09:52 +00:00

32 KiB

Raw Blame History

View Raw