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

diag: skip final normalize, test raw PV output via epilogue_tma_store

Browse Source
This commit is contained in:
biondizzle
2026-05-23 01:27:03 +00:00
parent 13ed779aee
commit 4248589c61
1 changed files with 2 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
tests/unit/test_fmha_v3_stage_c.py
View File

@@ -407,35 +407,8 @@ class FmhaV3StageCMulti:
# Wait for MMA's PV[N-1] to commit before reading O.
final_o_bar.arrive_and_wait()
# === Final O normalization: O *= 1/row_sum ===
inv_row_sum = Float32(1.0) / row_sum
tTMrO = cute.make_rmem_tensor(
(tTMEM_LOADcO.shape, 128 // corr_tile_size), self.acc_dtype
)
for i in range(n_corr_tiles):
tTMrO_i_ = tTMrO[None, i]
tTMrO_i_layout = cute.composition(
tTMrO_i_.layout, cute.make_layout(tTMrO.shape[0])
)
tTMrO_i = cute.make_tensor(tTMrO_i_.iterator, tTMrO_i_layout)
tTMEM_LOADtO_i = cute.make_tensor(
tTMEM_LOADtO.iterator + i * corr_tile_size, tTMEM_LOADtO.layout
)
tTMEM_STOREtO_i = cute.make_tensor(
tTMEM_STOREtO.iterator + i * corr_tile_size, tTMEM_STOREtO.layout
)
cute.copy(tiled_tmem_load_o, tTMEM_LOADtO_i, tTMrO_i)
for j in cutlass.range(cute.size(tTMrO_i), vectorize=True):
tTMrO_i[j] = tTMrO_i[j] * inv_row_sum
cute.copy(tiled_tmem_store_o, tTMrO_i, tTMEM_STOREtO_i)
cute.arch.fence_view_async_tmem_store()
# Standard epilogue: TMEM → SMEM → GMEM via TMA store.
# O in TMEM is now scaled by 1/row_sum.
# DIAG: skip final normalize, just use epilogue_tma_store directly
# to test raw PV output
tCtO_base = cute.make_tensor(tmem_ptr + self.tmem_o0_offset, tCtO_fake.layout)
acc_cons_st = pipeline.make_pipeline_state(
pipeline.PipelineUserType.Consumer, self.num_acc_stage