diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py index 1513254d..008c31f7 100644 --- a/dsv4/kernels/attention/fmha.py +++ b/dsv4/kernels/attention/fmha.py @@ -418,7 +418,7 @@ class FmhaKernel: # sub-tiles, and building BOTH copies from the SAME tensor, ensures the # column mappings agree on round-trip. # ============================================================ - corr_tile_size = 16 # Must be power of 2, divides head_dim + corr_tile_size = 32 # Must be power of 2, divides head_dim. Try 32 instead of 16. tOtO_i_layout = cute.composition( tOtO0.layout, cute.make_layout((128, corr_tile_size)) )