- test_fmha_v3_per_row_min.py: minimal per-row test (no C6/C9, no barriers)
Still hangs — likely CuTe DSL issue with logical_divide + explicit loops
- Replaced .load().reduce() on sliced tensors with explicit loops
- Very long compilation times suggest CuTe DSL is struggling
Key conclusion: per-row fix requires correction warp group.
The 6-warp code cant bridge 4 QK rows to 1 PV row per thread.
Need 128 correction threads (1 per output row) reading TMEM vector.
biondizzle
187d9e231c