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

Re-enable O rescale + normalize with corr_tile_size=32

Browse Source
This commit is contained in:
biondizzle
2026-05-22 19:11:17 +00:00
parent 0589a14790
commit 59575202a4
1 changed files with 33 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
tests/fmha_v3_real_softmax.py
View File

@@ -229,7 +229,7 @@ class FmhaV3RealSoftmax:
# O normalize setup: sub-tile O for TMEM read-modify-write
cO = cute.make_identity_tensor((self.pv_mma_tiler[0], self.pv_mma_tiler[1]))
tOcO = pv_thr.partition_C(cO)
corr_tile_size = 16
corr_tile_size = 32
tOtO_i_layout = cute.composition(tOtO.layout, cute.make_layout((128, corr_tile_size)))
tOcO_i_layout = cute.composition(tOcO.layout, cute.make_layout((128, corr_tile_size)))
tOtO_i = cute.make_tensor(tOtO0.iterator, tOtO_i_layout)
@@ -276,8 +276,22 @@ class FmhaV3RealSoftmax:
acc_scale = Float32(0.0)
row_sum *= acc_scale
# O rescale: DISABLED — skip for now
# O rescale in TMEM: multiply existing O by acc_scale = exp2(old_max - new_max)
# Only for kt > 0 (first tile: no existing O to rescale)
if kt > 0:
n_corr = HEAD_DIM // corr_tile_size
for ci in range(n_corr):
tTMrO_rs = 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_rs)
for j in cutlass.range(cute.size(tTMrO_rs), vectorize=True):
tTMrO_rs[j] = tTMrO_rs[j] * acc_scale
cute.copy(tiled_tmem_store_o, tTMrO_rs, tTMEM_STORE_OtO_ci)
# Pass 2: P = exp2(S * scale_log2 - row_max), accumulate row_sum
rP_words = cute.make_rmem_tensor(tTMEM_STOREcP.shape, self.qk_acc_dtype)
rP_bf16 = cute.make_tensor(cute.recast_ptr(rP_words.iterator, dtype=self.q_dtype), tTMEM_LOADrS.layout)
@@ -296,7 +310,22 @@ class FmhaV3RealSoftmax:
si_handle.release()
softmax_done_bar.arrive()
# Final O normalization: DISABLED
# Final O normalization: O = O / row_sum
if row_sum != Float32(0.0):
inv_row_sum = Float32(1.0) / row_sum
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)
for j in cutlass.range(cute.size(tTMrO_fn), vectorize=True):
tTMrO_fn[j] = tTMrO_fn[j] * inv_row_sum
cute.copy(tiled_tmem_store_o, tTMrO_fn, tTMEM_STORE_OtO_ci)
# Epilogue: TMEM -> SMEM -> GMEM via TMA store
tCtO_base = cute.make_tensor(tmem_ptr + self.tmem_o0_offset, tCtO_fake.layout)