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

Add O normalization with sub-tile TMEM read-modify-write

Browse Source
This commit is contained in:
biondizzle
2026-05-22 18:58:35 +00:00
parent b936c6220d
commit 6b61d5274c
1 changed files with 41 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
tests/fmha_v3_real_softmax.py
View File

@@ -226,7 +226,23 @@ class FmhaV3RealSoftmax:
tScP = cute.make_tensor(tScS.iterator, tScP_layout)
tTMEM_STOREcP = thr_store.partition_S(tScP)
# TODO: O normalize setup (add when O rescale is ready)
# 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
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)
tOcO_i = cute.make_tensor(tOcO.iterator, tOcO_i_layout)
tmem_load_o_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(corr_tile_size)), self.acc_dtype)
tmem_store_o_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(corr_tile_size)), self.acc_dtype)
tiled_tmem_load_o = tcgen05.make_tmem_copy(tmem_load_o_atom, tOtO_i)
tiled_tmem_store_o = tcgen05.make_tmem_copy(tmem_store_o_atom, tOtO_i)
thr_load_o = tiled_tmem_load_o.get_slice(sfw_idx)
thr_store_o = tiled_tmem_store_o.get_slice(sfw_idx)
tTMEM_LOAD_OtO = thr_load_o.partition_S(tOtO_i)
tTMEM_LOAD_OcO = thr_load_o.partition_D(tOcO_i)
tTMEM_STORE_OtO = thr_store_o.partition_D(tOtO_i)
row_max = -Float32.inf
row_sum = Float32(0.0)
@@ -281,10 +297,30 @@ class FmhaV3RealSoftmax:
softmax_done_bar.arrive()
# Final O normalization: O = O / row_sum
# TODO: enable after basic softmax works
# if row_sum != Float32(0.0):
# inv_row_sum = Float32(1.0) / row_sum
# ...
if row_sum != Float32(0.0):
inv_row_sum = Float32(1.0) / row_sum
# Load O from TMEM, multiply by 1/row_sum, write back
n_corr = 128 // corr_tile_size
for ci in range(n_corr):
tOtO_ci = cute.make_tensor(tOtO_i.iterator, tOtO_i.layout)
tOcO_ci = cute.make_tensor(tOcO_i.iterator, tOcO_i.layout)
# Sub-tile index offset
tOtO_sub = tOtO_ci[None, ci, None]
tOcO_sub = tOcO_ci[None, ci, None]
# Use the base partitioned tensors with offset
# Actually, just load the full O sub-tile
pass
# Simple approach: load/store via the partitioned tensors
tTMEM_LOAD_OrO = cute.make_rmem_tensor(tTMEM_LOAD_OcO.shape, self.acc_dtype)
cute.copy(tiled_tmem_load_o, tTMEM_LOAD_OtO, tTMEM_LOAD_OrO)
cute.arch.fence_view_async_tmem_load()
# Iterate with proper indexing
n_o_frg = cute.size(tTMEM_LOAD_OrO, mode=[0])
for fi in range(n_o_frg):
for fj in range(cute.size(tTMEM_LOAD_OrO, mode=[1])):
tTMEM_LOAD_OrO[fi, fj, None] = tTMEM_LOAD_OrO[fi, fj, None] * inv_row_sum
cute.copy(tiled_tmem_store_o, tTMEM_LOAD_OrO, tTMEM_STORE_OtO)
cute.arch.fence_view_async_tmem_store()
# Epilogue: TMEM -> SMEM -> GMEM via TMA store
tCtO_base = cute.make_tensor(tmem_ptr + self.tmem_o0_offset, tCtO_fake.layout)