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

fix: use TMEM round-trip normalize + epilogue_tma_store (known ~3% error)

Browse Source
This commit is contained in:
biondizzle
2026-05-23 02:49:46 +00:00
parent 350c7c36ac
commit 45cf89a556
1 changed files with 38 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

96
tests/unit/test_fmha_v3_stage_c.py
View File

@@ -95,7 +95,6 @@ class FmhaV3StageCMulti:
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)
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):
warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
@@ -131,19 +130,6 @@ class FmhaV3StageCMulti:
gK = cute.local_tile(mK,cute.slice_(self.qk_mma_tiler,(0,None,None)),(None,None,None))
gV = cute.local_tile(mV,cute.slice_(self.pv_mma_tiler,(0,None,None)),(None,None,None))
gC = cute.local_tile(mC,cute.slice_(self.pv_mma_tiler,(None,None,0)),(None,None,None))
# Pre-partition TMA store tensors (outside if blocks for region isolation)
epi_s = cute.select(self.c_smem_s,mode=[0,1])
gC_epi = cute.flat_divide(gC, epi_tile)
bSG_sC, bSG_gC = cpasync.tma_partition(
tma_c, 0, cute.make_layout(1),
cute.group_modes(sC, 0, 2),
cute.group_modes(gC_epi, 0, 2),
)
# Print TMA partition shapes for debugging
print(f'DEBUG bSG_sC shape: {bSG_sC.shape} rank: {cute.rank(bSG_sC)}', flush=True)
print(f'DEBUG bSG_gC shape: {bSG_gC.shape} rank: {cute.rank(bSG_gC)}', flush=True)
n_kv_tiles = cute.size(gK, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
@@ -344,57 +330,51 @@ class FmhaV3StageCMulti:
# Wait for MMA's PV[N-1] to commit before reading O.
final_o_bar.arrive_and_wait()
# === Correction epilog: ONE-WAY TMEM → reg → SMEM with normalize ===
# Uses get_tmem_load_op + get_smem_store_op paired atoms.
# === Final O normalization: O *= 1/row_sum ===
# TMEM round-trip using hand-constructed atoms.
# Known issue: hand-constructed Ld32x32bOp/St32x32bOp atoms introduce
# ~3% error due to TMEM column mapping mismatch with get_tmem_load_op.
# TODO: replace with correction_epilog (paired atoms) once TMA store
# region isolation is resolved.
inv_row_sum = Float32(1.0) / row_sum
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)))
tmem_load_epi_atom = utils.sm100.get_tmem_load_op(
self.pv_mma_tiler, self.c_layout, self.o_dtype, self.acc_dtype,
(epi_tile[0], epi_corr_tile_size), self.use_2cta_instrs,
tTMrO = cute.make_rmem_tensor(
(tTMEM_LOADcO.shape, 128 // corr_tile_size), self.acc_dtype
)
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)
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):
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[None, 0, 0, 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[None, 0, 0, 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 using pre-partitioned tensors
epi_bar = pipeline.NamedBarrier(
barrier_id=self.epilog_sync_bar_id,
num_threads=32 * len(self.epilogue_warp_id),
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)
acc_cons_st = pipeline.make_pipeline_state(
pipeline.PipelineUserType.Consumer, self.num_acc_stage
)
epi_bar.arrive_and_wait()
cute.copy(tma_c, bSG_sC[(None, 0)], bSG_gC[(None, 0, 0, 0, 0, 0)])
cute.arch.cp_async_bulk_commit_group()
cute.arch.cp_async_bulk_wait_group(0, read=True)
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,
)
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
@@ -402,7 +382,7 @@ class FmhaV3StageCMulti:
def test():
torch.manual_seed(42)
for n in [128]:
for n in [128, 256]:
torch.manual_seed(42)
m, hd = 128, HEAD_DIM
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')