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

fix: inline epilogue with paired atoms + inv_row_sum normalize, no TMEM round-trip

Browse Source
This commit is contained in:
biondizzle
2026-05-23 02:13:52 +00:00
parent 043b66406a
commit 6ee28d8423
1 changed files with 76 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files

149
tests/unit/test_fmha_v3_stage_c.py
View File

@@ -130,75 +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 paired TMEM load atom for correction_epilog
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)
def _correction_epilog(self, pv_thr, tOtO, scale, sC, tCgC, tma_c, sfw_idx, tidx, warp_idx, epi_tile, tmem_load_epi_atom):
"""CUTLASS correction_epilog: read O from TMEM, normalize, convert, write SMEM→GMEM."""
epi_corr_tile_size = 32 * 8 // self.o_dtype.width # 16 for BF16
cO = cute.make_identity_tensor((self.pv_mma_tiler[0], self.pv_mma_tiler[1]))
tOcO = pv_thr.partition_C(cO)
tOsO = pv_thr.partition_C(sC)
tOtO_i = cute.logical_divide(tOtO, cute.make_layout((128, epi_corr_tile_size)))
tOcO_i = cute.logical_divide(tOcO, cute.make_layout((128, epi_corr_tile_size)))
tOsO_i = cute.logical_divide(tOsO, cute.make_layout((128, epi_corr_tile_size)))
tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_epi_atom, tOtO_i[(None, None), 0])
smem_copy_atom = utils.sm100.get_smem_store_op(
self.c_layout, self.o_dtype, self.acc_dtype, tiled_tmem_load
)
tiled_smem_store = cute.make_tiled_copy_D(smem_copy_atom, tiled_tmem_load)
thr_tmem_load = tiled_tmem_load.get_slice(sfw_idx)
tTMEM_LOADtO = thr_tmem_load.partition_S(tOtO_i[(None, None), None])
tTMEM_LOADsO = thr_tmem_load.partition_D(tOsO_i[(None, None), None])
tTMEM_LOADcO = thr_tmem_load.partition_D(tOcO_i[(None, None), None])
n_corr_tiles = self.pv_mma_tiler[1] // epi_corr_tile_size
for i in range(n_corr_tiles):
tTMrO = cute.make_rmem_tensor(tTMEM_LOADcO[None, 0, 0, i].shape, self.acc_dtype)
cute.copy(tiled_tmem_load, tTMEM_LOADtO[None, 0, 0, i], tTMrO)
for j in range(cute.size(tTMrO)):
tTMrO[j] = tTMrO[j] * scale
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, tSMrO, tTMEM_LOADsO[None, 0, 0, i])
cute.arch.fence_proxy("async.shared", space="cta")
# TMA store SMEM → GMEM (same pattern as epilogue_tma_store)
tCgC_epi = cute.flat_divide(tCgC, 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),
)
epilog_sync_bar = pipeline.NamedBarrier(
barrier_id=self.epilog_sync_bar_id,
num_threads=32 * len(self.epilogue_warp_id),
)
epilog_sync_bar.arrive_and_wait()
# All warps in the epilogue group do the TMA store together
c_buffer = 0
cute.copy(tma_c, tCsC[(None, c_buffer)], 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)
epilog_sync_bar.arrive_and_wait()
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:
@@ -464,12 +399,80 @@ 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 → SMEM with normalize ===
# === Correction epilog: one-way TMEM → reg → SMEM → GMEM with normalize ===
# Uses get_tmem_load_op + get_smem_store_op paired atoms (same as CUTLASS correction_epilog).
# NO TMEM round-trip — hand-constructed Ld32x32bOp/St32x32bOp atoms corrupt data.
inv_row_sum = Float32(1.0) / row_sum
self._correction_epilog(
pv_thr, tOtO0, inv_row_sum, sC, tCgC, tma_c, sfw_idx,
tidx, warp_idx, epi_tile, tmem_load_epi_atom,
# Build the TMEM→reg and reg→SMEM tiled copies using paired atoms
tCtO_base = cute.make_tensor(tmem_ptr + self.tmem_o0_offset, tCtO_fake.layout)
tCtO = utils.gemm.sm100.transform_partitioned_tensor_layout(tCtO_base)
tiled_copy_t2r, tTR_tO, tTR_rO = utils.gemm.sm100.epilogue_tmem_copy_and_partition(
self, tidx, tCtO, tCgC, epi_tile, self.use_2cta_instrs
)
tTR_rC = cute.make_rmem_tensor(tTR_rO.shape, self.c_dtype)
tiled_copy_r2s, tRS_rC, tRS_sC = utils.gemm.sm100.epilogue_smem_copy_and_partition(
self, tiled_copy_t2r, tTR_rC, tidx, sC
)
tCgC_epi = cute.flat_divide(tCgC, epi_tile)
bSG_sC, bSG_gC_partitioned = cpasync.tma_partition(
tma_c, 0, cute.make_layout(1),
cute.group_modes(sC, 0, 2),
cute.group_modes(tCgC_epi, 0, 2),
)
epilog_sync_bar = pipeline.NamedBarrier(
barrier_id=self.epilog_sync_bar_id,
num_threads=32 * len(self.epilogue_warp_id),
)
# Consume the accumulator pipeline
acc_cons_st = pipeline.make_pipeline_state(
pipeline.PipelineUserType.Consumer, self.num_acc_stage
)
c_pipe = pipeline.PipelineTmaStore.create(
num_stages=self.num_c_stage,
producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 32 * len(self.epilogue_warp_id)),
)
acc_pipe.consumer_wait(acc_cons_st)
# Slice to the current tile
tTR_tO_tile = tTR_tO[(None, None, None, None, None, acc_cons_st.index)]
bSG_gC = bSG_gC_partitioned[(None, None, None, Int32(0), Int32(0), Int32(0))]
tTR_tO_tile = cute.group_modes(tTR_tO_tile, 3, cute.rank(tTR_tO_tile))
bSG_gC = cute.group_modes(bSG_gC, 1, cute.rank(bSG_gC))
# Store O to global memory in subtiles, applying 1/row_sum normalize
subtile_cnt = cute.size(tTR_tO_tile.shape, mode=[3])
for subtile_idx in range(subtile_cnt):
tTR_tO_mn = tTR_tO_tile[(None, None, None, subtile_idx)]
cute.copy(tiled_copy_t2r, tTR_tO_mn, tTR_rO)
# Apply normalize: multiply by inv_row_sum, then convert to BF16
acc_vec = tiled_copy_r2s.retile(tTR_rO).load()
# acc_vec is in FP32 — apply scale before conversion
# We can't directly scale the vector, but we can scale the register tensor
for j in cutlass.range(cute.size(tTR_rO), vectorize=True):
tTR_rO[j] = tTR_rO[j] * inv_row_sum
acc_vec = tiled_copy_r2s.retile(tTR_rO).load()
acc_vec = acc_vec.to(self.c_dtype)
tRS_rC.store(acc_vec)
c_buffer = subtile_cnt * 0 + subtile_idx # num_prev_subtiles = 0
c_buffer = c_buffer % self.num_c_stage
cute.copy(tiled_copy_r2s, tRS_rC, tRS_sC[(None, None, None, c_buffer)])
cute.arch.fence_proxy("async.shared", space="cta")
epilog_sync_bar.arrive_and_wait()
if warp_idx == self.epilogue_warp_id[0]:
cute.copy(tma_c, bSG_sC[(None, c_buffer)], bSG_gC[(None, subtile_idx)])
c_pipe.producer_commit()
c_pipe.producer_acquire()
epilog_sync_bar.arrive_and_wait()
epilog_sync_bar.arrive_and_wait()
acc_pipe.consumer_release(acc_cons_st)
acc_cons_st.advance()
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
@@ -489,8 +492,8 @@ def test():
qf = q[:, :, 0].float()
kf = k[:, :, 0].float()
scale = 1.0 / math.sqrt(hd)
attn_raw = qf @ kf.T * scale
attn = torch.softmax(attn_raw, dim=-1)
attn = qf @ kf.T * scale
attn = torch.softmax(attn, dim=-1)
ref = attn @ v.float()
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))