From e0a604c5085d02471f15529b66c5203f0855675b Mon Sep 17 00:00:00 2001 From: biondizzle Date: Fri, 22 May 2026 19:14:33 +0000 Subject: [PATCH] Disable ALL O copies to verify baseline --- tests/fmha_v3_real_softmax.py | 15 +-------------- 1 file changed, 1 insertion(+), 14 deletions(-) diff --git a/tests/fmha_v3_real_softmax.py b/tests/fmha_v3_real_softmax.py index d8b235b5..aa3096ac 100644 --- a/tests/fmha_v3_real_softmax.py +++ b/tests/fmha_v3_real_softmax.py @@ -299,20 +299,7 @@ class FmhaV3RealSoftmax: si_handle.release() softmax_done_bar.arrive() - # Final O normalization: load, NO-OP write (debug copy) - # If this matches unnormalized output, the TMEM copy is correct. - n_corr = HEAD_DIM // corr_tile_size - for ci in range(n_corr): - tTMrO_fn = cute.make_rmem_tensor(tTMEM_LOAD_OcO.shape, self.acc_dtype) - tTMEM_LOAD_OtO_ci = cute.make_tensor( - tTMEM_LOAD_OtO.iterator + ci * corr_tile_size, tTMEM_LOAD_OtO.layout - ) - tTMEM_STORE_OtO_ci = cute.make_tensor( - tTMEM_STORE_OtO.iterator + ci * corr_tile_size, tTMEM_STORE_OtO.layout - ) - cute.copy(tiled_tmem_load_o, tTMEM_LOAD_OtO_ci, tTMrO_fn) - # NO-OP: write back without modification - cute.copy(tiled_tmem_store_o, tTMrO_fn, tTMEM_STORE_OtO_ci) + # Final O normalization: DISABLED for NO-OP test # Epilogue: TMEM -> SMEM -> GMEM via TMA store tCtO_base = cute.make_tensor(tmem_ptr + self.tmem_o0_offset, tCtO_fake.layout)