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

fix: O rescale uses 2D register tensor pattern (matching CUTLASS correction_rescale)

Browse Source
This commit is contained in:
biondizzle
2026-05-23 01:25:53 +00:00
parent c1f531c0e6
commit e7b2cb5a32
1 changed files with 48 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files

117
tests/unit/test_fmha_v3_stage_c.py
View File

@@ -130,19 +130,10 @@ class FmhaV3StageCMulti:
tma_v,mV = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,pv_mma.thr_id),v_fmha,v_s,self.pv_mma_tiler,pv_mma,self.cluster_layout_vmnk.shape)
epi_s = cute.select(self.c_smem_s,mode=[0,1])
tma_c,mC = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(),c,epi_s,self.epi_tile)
# Pre-compute epilog atoms for correction_epilog pattern
epi_corr_tile_size = 32 * 8 // self.o_dtype.width # 16 for BF16
epi_subtile = (self.epi_tile[0], epi_corr_tile_size)
tmem_load_epi_atom = utils.sm100.get_tmem_load_op(
self.pv_mma_tiler, self.c_layout, self.o_dtype, self.acc_dtype,
epi_subtile, use_2cta_instrs=False,
)
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.q_smem_s,self.k_smem_s,self.v_smem_s,self.p_tmem_s,self.c_smem_s,self.epi_tile,tmem_load_epi_atom).launch(grid=(1,1,1),block=[self.threads_per_cta,1,1],stream=stream)
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.q_smem_s,self.k_smem_s,self.v_smem_s,self.p_tmem_s,self.c_smem_s,self.epi_tile).launch(grid=(1,1,1),block=[self.threads_per_cta,1,1],stream=stream)
@cute.kernel
def _kernel(self, qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV, tma_c, mC, cl_vmnk, q_smem_s, k_smem_s, v_smem_s, p_tmem_s, c_smem_s, epi_tile, tmem_load_epi_atom):
def _kernel(self, qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV, tma_c, mC, cl_vmnk, q_smem_s, k_smem_s, v_smem_s, p_tmem_s, c_smem_s, epi_tile):
warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
tidx,_,_ = cute.arch.thread_idx()
if warp_idx == self.tma_warp_id:
@@ -384,8 +375,18 @@ class FmhaV3StageCMulti:
cute.arch.fence_view_async_tmem_store()
# === Per-tile O rescale: O *= acc_scale for kt > 0 ===
# Uses 2D register tensor pattern (same as CUTLASS correction_rescale
# and our final normalize) to preserve data through TMEM round-trip.
if kt > 0:
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,
@@ -394,12 +395,10 @@ class FmhaV3StageCMulti:
tTMEM_STOREtO.iterator + i * corr_tile_size,
tTMEM_STOREtO.layout,
)
tTMrO = cute.make_rmem_tensor(tTMEM_LOADcO.shape, self.acc_dtype)
cute.copy(tiled_tmem_load_o, tTMEM_LOADtO_i, tTMrO)
cute.arch.fence_view_async_tmem_load()
for k in cutlass.range(cute.size(tTMrO), vectorize=True):
tTMrO[k] = tTMrO[k] * acc_scale
cute.copy(tiled_tmem_store_o, tTMrO, tTMEM_STOREtO_i)
cute.copy(tiled_tmem_load_o, tTMEM_LOADtO_i, tTMrO_i)
for k in cutlass.range(cute.size(tTMrO_i), vectorize=True):
tTMrO_i[k] = tTMrO_i[k] * acc_scale
cute.copy(tiled_tmem_store_o, tTMrO_i, tTMEM_STOREtO_i)
cute.arch.fence_view_async_tmem_store()
si_handle.release()
@@ -408,67 +407,47 @@ class FmhaV3StageCMulti:
# Wait for MMA's PV[N-1] to commit before reading O.
final_o_bar.arrive_and_wait()
# === Final O normalization + epilogue: CUTLASS correction_epilog pattern ===
# ONE-WAY trip: TMEM → reg (normalize + FP32→BF16) → SMEM → TMA → GMEM
# === Final O normalization: O *= 1/row_sum ===
inv_row_sum = Float32(1.0) / row_sum
# Build paired SMEM store atom from the pre-computed TMEM load atom
epi_corr_tile_size = 32 * 8 // self.o_dtype.width # 16 for BF16
tOtO_epi = cute.logical_divide(tOtO0, cute.make_layout((128, epi_corr_tile_size)))
tiled_tmem_load_epi = tcgen05.make_tmem_copy(
tmem_load_epi_atom, tOtO_epi[(None, None), 0]
)
smem_store_epi_atom = utils.sm100.get_smem_store_op(
self.c_layout, self.o_dtype, self.acc_dtype, tiled_tmem_load_epi,
)
tiled_smem_store_epi = cute.make_tiled_copy_D(
smem_store_epi_atom, tiled_tmem_load_epi
tTMrO = cute.make_rmem_tensor(
(tTMEM_LOADcO.shape, 128 // corr_tile_size), self.acc_dtype
)
# Partition SMEM for the epilog output
tOsO = pv_thr.partition_C(sC)
cO_epi = cute.make_identity_tensor((self.pv_mma_tiler[0], self.pv_mma_tiler[1]))
tOcO_epi = pv_thr.partition_C(cO_epi)
tOsO_epi = cute.logical_divide(tOsO, cute.make_layout((128, epi_corr_tile_size)))
tOcO_epi = cute.logical_divide(tOcO_epi, cute.make_layout((128, epi_corr_tile_size)))
thr_tmem_load_epi = tiled_tmem_load_epi.get_slice(sfw_idx)
tTMEM_LOADtO_epi = thr_tmem_load_epi.partition_S(tOtO_epi[(None, None), None])
tTMEM_LOADsO_epi = thr_tmem_load_epi.partition_D(tOsO_epi[(None, None), None])
tTMEM_LOADcO_epi = thr_tmem_load_epi.partition_D(tOcO_epi[(None, None), None])
n_epi_corr_tiles = self.pv_mma_tiler[1] // epi_corr_tile_size
for i in range(n_epi_corr_tiles):
tTMEM_LOADtO_epi_i = tTMEM_LOADtO_epi[None, 0, 0, i]
tTMEM_LOADsO_epi_i = tTMEM_LOADsO_epi[None, 0, 0, i]
tTMrO = cute.make_rmem_tensor(
tTMEM_LOADcO_epi[None, 0, 0, i].shape, 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_epi, tTMEM_LOADtO_epi_i, tTMrO)
for j in range(cute.size(tTMrO)):
tTMrO[j] = tTMrO[j] * inv_row_sum
tSMrO = cute.make_rmem_tensor(tTMrO.shape, self.o_dtype)
o_vec = tTMrO.load()
tSMrO.store(o_vec.to(self.o_dtype))
cute.copy(tiled_smem_store_epi, tSMrO, tTMEM_LOADsO_epi_i)
cute.arch.fence_proxy("async.shared", space="cta")
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)
# TMA store: SMEM → GMEM
softmax_all_bar = pipeline.NamedBarrier(
barrier_id=5, num_threads=32 * len(self.epilogue_warp_id)
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.
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
)
softmax_all_bar.arrive_and_wait()
tCgC_epi = cute.flat_divide(tCgC, self.epi_tile)
tCsC, tCgC_tma = cpasync.tma_partition(
tma_c, 0, cute.make_layout(1),
cute.group_modes(sC, 0, 2),
cute.group_modes(tCgC_epi, 0, 2),
c_grp = pipeline.CooperativeGroup(pipeline.Agent.Thread, 32 * len(self.epilogue_warp_id))
c_pipe = pipeline.PipelineTmaStore.create(num_stages=self.num_c_stage, producer_group=c_grp)
acc_cons_st = utils.gemm.sm100.epilogue_tma_store(
self, tidx, warp_idx, tma_c, tCtO_base, sC, tCgC, epi_tile,
0, const_expr(lambda x: x), (0, 0, 0),
acc_cons_st, acc_pipe, c_pipe,
)
cute.copy(tma_c, tCsC[(None, 0)], tCgC_tma[(None, 0, 0, 0, 0, 0, 0)])
cute.arch.cp_async_bulk_commit_group()
cute.arch.cp_async_bulk_wait_group(0, read=True)
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)