biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Disable ALL O copies to verify baseline

Browse Source
This commit is contained in:
biondizzle
2026-05-22 19:14:33 +00:00
parent 64d174fdf6
commit e0a604c508
1 changed files with 1 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
tests/fmha_v3_real_softmax.py
View File

@@ -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)