Stage B Bug 4b debugging: P/A alias proven working, V layout issue for (128,64) PV

Key findings:
- P/A alias WORKS: PV reads non-zero P from TMEM at offset 32 (proven by no-softmax test)
- V mode bug: V=(128,64) only loads 64 K-values, PV needs 128. Output = sum(S[:,:64]) = 0.67 cosine
- FMHA-style V reconstruction (hd,n,1) stride (1,hd) gives NaN for (128,64) PV
- K-major V (64,128) contiguous gives NaN for (128,64) PV
- Square (128,128) PV works with ALL V approaches (cosine 0.999999)
- Non-square PV consistently broken regardless of V layout

Test files:
- test_128_128_fmha_v.py: (128,128) with FMHA V - PASS
- test_pv64_fmha_v.py: (128,64) with FMHA V - NaN
- test_pv64_kmajor_v.py: (128,64) with K-major V - NaN
- test_pv64_with_softmax.py: (128,64) with original V - 0.67
- test_pv64_no_softmax.py: proves P/A alias works
- test_fmha_v3.py: full pipeline with QK C-fragment composition store
This commit is contained in:
2026-05-21 15:20:14 +00:00
parent 81d5d8d04c
commit 4564758466
10 changed files with 2286 additions and 37 deletions

View File

@@ -0,0 +1,236 @@
"""Test (128,128) PV with FMHA-style V reconstruction. If this works, the FMHA V layout is fine."""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
class Test128FmhaV:
def __init__(self):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
self.num_ab_stage = 1; self.num_acc_stage = 1
self.hd = 128 # square
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (128, self.hd, pv_ik * (128 // pv_ik))
self.mma_tiler = self.qk_mma_tiler
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), self.hd, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.a_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.b_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tilePlikeFP32 = self.qk_mma_tiler[1] // Float32.width * self.o_dtype.width
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
a_s = cute.slice_(self.a_smem_s,(None,None,None,0)); b_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
self.num_tma_load_bytes = (cute.size_in_bytes(self.q_dtype,a_s)+cute.size_in_bytes(self.q_dtype,b_s)+cute.size_in_bytes(self.q_dtype,v_s))*cute.size(qk_mma.thr_id.shape)
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
# FMHA-style V: (hd, 128, 1) stride (1, hd, hd*128)
v_fmha = cute.make_tensor(
v.iterator,
cute.make_layout(
(self.hd, 128, 1),
stride=(1, self.hd, self.hd * 128),
),
)
self.v_major = LayoutEnum.from_tensor(v_fmha).mma_major_mode()
self.c_layout = LayoutEnum.from_tensor(c)
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.qk_acc_dtype, self.cta_group, (128,self.hd), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.a_smem_s,(None,None,None,0)); k_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
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)
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.a_smem_s,self.b_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, a_smem_s, b_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:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k)
cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
ab_bar: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage*2]
mma_si_bar: cute.struct.MemRange[cutlass.Int64, 2]
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
smem = utils.SmemAllocator(); st = smem.allocate(SS)
ab_p,ab_c = pipeline.PipelineTmaUmma.create(barrier_storage=st.ab_bar.data_ptr(),num_stages=self.num_ab_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.num_tma_load_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
mma_si_prod,mma_si_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_si_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
softmax_done_bar = pipeline.NamedBarrier(barrier_id=3, num_threads=32 + 32*len(self.epilogue_warp_id))
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=a_smem_s.outer,byte_alignment=128,swizzle=a_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=b_smem_s.outer,byte_alignment=128,swizzle=b_smem_s.inner)
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
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))
k_cnt = cute.size(gQ, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator+self.tmem_s0_offset,tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator+self.tmem_o0_offset,tOtO.layout)
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(tOrP.iterator+self.qk_acc_dtype.width//self.q_dtype.width*self.tmem_p0_offset,tOrP.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as,self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as,self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
if warp_idx == self.tma_warp_id:
ab_p.reset(); peek = ab_p.try_acquire()
for kt in cutlass.range(k_cnt,unroll=1):
h = ab_p.acquire_and_advance(peek)
cute.copy(tma_q,tAgQ[(None,h.count)],tAsQ[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_k,tBgK[(None,h.count)],tBsK[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_v,tVgV[(None,h.count)],tVsV[(None,h.index)],tma_bar_ptr=h.barrier)
peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_p.try_acquire()
ab_p.tail()
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
ab_c.reset(); peek = ab_c.try_wait()
s0_handle = mma_si_prod.acquire_and_advance()
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer,self.num_acc_stage)
acc_pipe.producer_acquire(acc_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
for kb in cutlass.range(cute.size(tCrQ,mode=[2]),unroll_full=True):
cute.gemm(qk_mma,tStS0,tCrQ[(None,None,kb,h.index)],tCrK[(None,None,kb,h.index)],tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
h.release(); peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_c.try_wait()
cute.arch.fence_view_async_tmem_store(); s0_handle.commit()
softmax_done_bar.arrive_and_wait()
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None,None,None,0)]
for kb in cutlass.range(cute.size(tOrP0,mode=[2]),unroll_full=True):
cute.gemm(pv_mma,tOtO0,tOrP0[(None,None,kb)],tCrV_s[(None,None,kb)],tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
acc_pipe.producer_commit(acc_st); acc_st.advance(); acc_pipe.producer_tail(acc_st)
if warp_idx < self.mma_warp_id:
tmem.allocate(self.num_tmem_alloc_cols)
tmem.wait_for_alloc()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
sfw_idx = tidx % (32 * len(self.epilogue_warp_id))
tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0)
thr_load = tiled_tmem_load.get_slice(sfw_idx)
tTMEM_LOADtS = thr_load.partition_S(tStS0)
cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1]))
tScS = qk_thr.partition_C(cS)
tTMEM_LOADcS = thr_load.partition_D(tScS)
tStS_P_layout = cute.composition(tStS.layout, cute.make_layout((128, self.tilePlikeFP32)))
tStS_P = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStS_P_layout)
tmem_store_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStS_P)
thr_store = tiled_tmem_store.get_slice(sfw_idx)
tTMEM_STOREtS_x4 = thr_store.partition_D(tStS_P)
tScS_P_layout = cute.composition(tScS.layout, cute.make_layout((128, self.tilePlikeFP32)))
tScS_P = cute.make_tensor(tScS.iterator, tScS_P_layout)
tTMEM_STOREcS = thr_store.partition_S(tScS_P)
si_handle = mma_si_cons.wait_and_advance()
tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
tTMEM_STORErS_x4 = cute.make_rmem_tensor(tTMEM_STOREcS.shape, self.qk_acc_dtype)
tTMEM_STORErS_x4_e = cute.make_tensor(cute.recast_ptr(tTMEM_STORErS_x4.iterator, dtype=self.q_dtype), tTMEM_LOADrS.layout)
frg_cnt = 4; frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
tTMEM_STORErS_x4_e_frg = cute.logical_divide(tTMEM_STORErS_x4_e, cute.make_layout(frg_tile))
for j in range(frg_cnt):
s_vec = tTMEM_LOADrS_frg[None, j].load()
tTMEM_STORErS_x4_e_frg[None, j].store(s_vec.to(self.q_dtype))
cute.copy(tiled_tmem_store, tTMEM_STORErS_x4, tTMEM_STOREtS_x4)
cute.arch.fence_view_async_tmem_store()
si_handle.release()
softmax_done_bar.arrive()
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)
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)
def test():
torch.manual_seed(42)
m, n, hd = 128, 128, 128
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.ones(n, hd, dtype=torch.bfloat16, device='cuda')
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
ref = (qf @ kf.T).bfloat16().float() @ v.float()
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
mV = ct.from_dlpack(v).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v))
mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
kernel = Test128FmhaV()
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
compiled(mQ, mK, mV, mC, stream)
torch.cuda.synchronize()
out = c[:,:,0].float()
cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item()
print(f'128x128 FMHA-V: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if __name__ == '__main__':
test()

View File

@@ -1,7 +1,8 @@
"""
FMHA v3: QK -> softmax -> PV with KV-tile interleaving.
Bug 4b fix: tP uses QK C-fragment layout so PV reads from same TMEM columns softmax writes.
V=ones test: O[i,j] = sum_j(S[i,:].bf16())
Bug 4b fix (FMHA pattern): P store uses QK C-fragment layout composition,
NOT PV A-fragment layout. Register bridge: FP32 backing (store partition shape)
recast to BF16 view (QK-load layout).
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
@@ -43,7 +44,6 @@ class FmhaV3:
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tilePlikeFP32 = self.qk_mma_tiler[1] // Float32.width * self.o_dtype.width
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
@@ -128,10 +128,7 @@ class FmhaV3:
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout)
# BUG 4b FIX: Use QK C-fragment layout for tP so PV A-fragment reads
# from the same TMEM columns that softmax writes P to.
# Softmax writes P using tStS_P (QK C-fragment composition).
# BUG 4b: Use p_tmem_s layout for tP so PV A-fragment reads from correct TMEM columns
# --- PV read view (for MMA only, NOT for softmax store) ---
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
@@ -191,55 +188,83 @@ class FmhaV3:
tmem.wait_for_alloc()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
sfw_idx = tidx % (32 * len(self.epilogue_warp_id))
# S load
# --- S load (QK C-fragment layout) ---
tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0)
thr_load = tiled_tmem_load.get_slice(sfw_idx)
tTMEM_LOADtS = thr_load.partition_S(tStS0)
# S coordinate tensor (QK C-fragment)
cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1]))
tScS = qk_thr.partition_C(cS)
tTMEM_LOADcS = thr_load.partition_D(tScS)
# BUG 4b FIX: P store uses PV A-fragment's TMEM layout from tOrP0.
# tOrP0 is 3D — flatten to 2D for make_tmem_store compatibility.
# We create a 2D tensor from tOrP0's pointer with a flat layout.
tilePlikeFP32_pv = self.pv_mma_tiler[1] // Float32.width * self.o_dtype.width
# Flatten tOrP0 from 3D to 2D: (M, K_frag*K_blocks) → (M, K_total)
tOrP0_shape = tOrP0.shape
k_total = cute.size(tOrP0_shape, mode=[1]) * cute.size(tOrP0_shape, mode=[2])
m_total = cute.size(tOrP0_shape, mode=[0])
# Recast to FP32: BF16 pairs → FP32, so K_FP32 = K_total / 2
k_fp32 = k_total // 2
tOrP0_2d_fp32 = cute.make_tensor(
cute.recast_ptr(tOrP0.iterator, dtype=self.qk_acc_dtype),
cute.composition(
cute.make_layout((m_total, k_fp32)),
tOrP0.layout
),
# --- P store (QK C-fragment layout composition, FMHA pattern) ---
# P logical columns = PV K = QK N = pv_mma_tiler[2]
# Packed FP32 columns: BF16 pairs packed into FP32 words
p_cols_fp32 = self.pv_mma_tiler[2] * self.q_dtype.width // self.qk_acc_dtype.width
# BF16: 128 * 16 / 32 = 64
# P TMEM destination: QK C-fragment layout composed with P sub-tile
tStP_layout = cute.composition(
tStS.layout,
cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32)),
)
tmem_store_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tOrP0_2d_fp32)
tStP0 = cute.make_tensor(
tStS.iterator + self.tmem_p0_offset,
tStP_layout,
)
# P TMEM store atom and tiled copy
tmem_store_atom = cute.make_copy_atom(
tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)),
self.qk_acc_dtype,
)
tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStP0)
thr_store = tiled_tmem_store.get_slice(sfw_idx)
tTMEM_STOREtP = thr_store.partition_D(tOrP0_2d_fp32)
# 2D coordinate tensor — use thr_store's partition (NOT qk_thr)
# so source and dest have matching element distributions
cP = cute.make_identity_tensor((m_total, k_fp32))
tTMEM_STOREcP = thr_store.partition_S(cP)
tTMEM_STOREtP = thr_store.partition_D(tStP0)
# P coordinate tensor: QK C-fragment coordinate composed with P sub-tile
tScP_layout = cute.composition(
tScS.layout,
cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32)),
)
tScP = cute.make_tensor(tScS.iterator, tScP_layout)
tTMEM_STOREcP = thr_store.partition_S(tScP)
for kt in range(n_kv_tiles):
si_handle = s_cons.wait_and_advance()
# Load S from TMEM (FP32, QK C-fragment layout)
tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
tTMEM_STORErP = cute.make_rmem_tensor(tTMEM_STOREcP.shape, self.qk_acc_dtype)
tTMEM_STORErP_e = cute.make_tensor(cute.recast_ptr(tTMEM_STORErP.iterator, dtype=self.q_dtype), tTMEM_LOADrS.layout)
frg_cnt = 4; frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
# Register bridge (FMHA pattern):
# rP_words: FP32 backing store with store-partition shape
# rP_bf16: BF16 view over same registers using QK-load layout
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,
)
# Fragmented load→convert→store:
# Load S as FP32, convert to BF16, store through rP_bf16 view
frg_cnt = 4
frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
tTMEM_STORErP_e_frg = cute.logical_divide(tTMEM_STORErP_e, cute.make_layout(frg_tile))
rP_bf16_frg = cute.logical_divide(rP_bf16, cute.make_layout(frg_tile))
for j in range(frg_cnt):
s_vec = tTMEM_LOADrS_frg[None, j].load()
tTMEM_STORErP_e_frg[None, j].store(s_vec.to(self.q_dtype))
cute.copy(tiled_tmem_store, tTMEM_STORErP, tTMEM_STOREtP)
rP_bf16_frg[None, j].store(s_vec.to(self.q_dtype))
# Copy packed FP32 backing registers to TMEM
cute.copy(tiled_tmem_store, rP_words, tTMEM_STOREtP)
cute.arch.fence_view_async_tmem_store()
si_handle.release()
softmax_done_bar.arrive()
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)
c_grp = pipeline.CooperativeGroup(pipeline.Agent.Thread, 32 * len(self.epilogue_warp_id))

284
tests/test_fmha_v3_debug.py Normal file
View File

@@ -0,0 +1,284 @@
"""
FMHA v3 debug: Test with HEAD_DIM=128 first (PV 128,128) to verify plumbing,
then HEAD_DIM=64 to debug Bug 4b.
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
class FmhaV3Debug:
def __init__(self, head_dim=64):
self.head_dim = head_dim
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
self.kv_stage = 2; self.q_stage = 1; self.num_c_stage = 2
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (128, self.head_dim, pv_ik * (128 // pv_ik))
self.mma_tiler = self.qk_mma_tiler
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), self.head_dim, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
self.q_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.qk_mma_tiler, self.q_dtype, self.q_stage)
self.k_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.qk_mma_tiler, self.q_dtype, self.kv_stage)
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, self.kv_stage)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
cta = cute.size(qk_mma.thr_id.shape)
q_s = cute.slice_(self.q_smem_s,(None,None,None,0)); k_s = cute.slice_(self.k_smem_s,(None,None,None,0))
self.q_tx_bytes = cute.size_in_bytes(self.q_dtype, q_s) * cta
self.kv_tx_bytes = cute.size_in_bytes(self.q_dtype, k_s) * cta
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
a_major = LayoutEnum.from_tensor(q).mma_major_mode()
b_major = LayoutEnum.from_tensor(k).mma_major_mode()
v_major = LayoutEnum.from_tensor(v).mma_major_mode()
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, a_major, b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, v_major, self.qk_acc_dtype, self.cta_group, (128,self.head_dim), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.q_smem_s,(None,None,None,0)); k_s = cute.slice_(self.k_smem_s,(None,None,None,0)); v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
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,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)
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())
tidx,_,_ = cute.arch.thread_idx()
if warp_idx == self.tma_warp_id:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k); cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
q_bar: cute.struct.MemRange[cutlass.Int64, self.q_stage*2]
kv_bar: cute.struct.MemRange[cutlass.Int64, self.kv_stage*2]
s_bar: cute.struct.MemRange[cutlass.Int64, 2]
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
smem = utils.SmemAllocator(); st = smem.allocate(SS)
qp,qc = pipeline.PipelineTmaUmma.create(barrier_storage=st.q_bar.data_ptr(),num_stages=self.q_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.q_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
kvp,kvc = pipeline.PipelineTmaUmma.create(barrier_storage=st.kv_bar.data_ptr(),num_stages=self.kv_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.kv_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
s_prod,s_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.s_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
softmax_done_bar = pipeline.NamedBarrier(barrier_id=3, num_threads=32 + 32*len(self.epilogue_warp_id))
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=q_smem_s.outer,byte_alignment=128,swizzle=q_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=k_smem_s.outer,byte_alignment=128,swizzle=k_smem_s.inner)
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
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))
n_kv_tiles = cute.size(gK, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout)
# PV read view (MMA only)
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(
tOrP.iterator + self.qk_acc_dtype.width // self.q_dtype.width * self.tmem_p0_offset,
tOrP.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# TMA LOAD
if warp_idx == self.tma_warp_id:
qp.reset(); qh = qp.acquire_and_advance()
cute.copy(tma_q,tAgQ[(None,qh.count)],tAsQ[(None,qh.index)],tma_bar_ptr=qh.barrier)
qp.tail()
kvp.reset(); pk = kvp.try_acquire()
for kt in cutlass.range(n_kv_tiles,unroll=1):
kh = kvp.acquire_and_advance(pk)
cute.copy(tma_k,tBgK[(None,kh.count)],tBsK[(None,kh.index)],tma_bar_ptr=kh.barrier)
pk = cutlass.Boolean(1)
vh = kvp.acquire_and_advance(pk)
cute.copy(tma_v,tVgV[(None,vh.count)],tVsV[(None,vh.index)],tma_bar_ptr=vh.barrier)
pk = cutlass.Boolean(1)
kvp.tail()
# MMA
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
qc.reset(); qh = qc.wait_and_advance(); qh.release()
kvc.reset(); pk = kvc.try_wait()
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_st)
for kt in range(n_kv_tiles):
kh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)
sh = s_prod.acquire_and_advance()
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kb in cutlass.range(cute.size(tCrQ,mode=[2]), unroll_full=True):
cute.gemm(qk_mma, tStS0, tCrQ[(None,None,kb,0)], tCrK[(None,None,kb,kh.index)], tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
cute.arch.fence_view_async_tmem_store()
sh.commit(); kh.release()
softmax_done_bar.arrive_and_wait()
vh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)
pv_mma.set(tcgen05.Field.ACCUMULATE, kt != 0)
for kb in cutlass.range(cute.size(tOrP0,mode=[2]), unroll_full=True):
cute.gemm(pv_mma, tOtO0, tOrP0[(None,None,kb)], tCrV[(None,None,kb,vh.index)], tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
cute.arch.fence_view_async_tmem_store()
vh.release()
acc_pipe.producer_commit(acc_st); acc_st.advance()
acc_pipe.producer_tail(acc_st)
# EPILOGUE
if warp_idx < self.mma_warp_id:
tmem.allocate(self.num_tmem_alloc_cols)
tmem.wait_for_alloc()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
sfw_idx = tidx % (32 * len(self.epilogue_warp_id))
# S load
tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0)
thr_load = tiled_tmem_load.get_slice(sfw_idx)
tTMEM_LOADtS = thr_load.partition_S(tStS0)
cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1]))
tScS = qk_thr.partition_C(cS)
tTMEM_LOADcS = thr_load.partition_D(tScS)
# P store — FMHA pattern: QK C-fragment layout composition
p_cols_fp32 = self.pv_mma_tiler[2] * self.q_dtype.width // self.qk_acc_dtype.width
tStP_layout = cute.composition(
tStS.layout,
cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32)),
)
tStP0 = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStP_layout)
tmem_store_atom = cute.make_copy_atom(
tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)),
self.qk_acc_dtype,
)
tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStP0)
thr_store = tiled_tmem_store.get_slice(sfw_idx)
tTMEM_STOREtP = thr_store.partition_D(tStP0)
tScP_layout = cute.composition(
tScS.layout,
cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32)),
)
tScP = cute.make_tensor(tScS.iterator, tScP_layout)
tTMEM_STOREcP = thr_store.partition_S(tScP)
for kt in range(n_kv_tiles):
si_handle = s_cons.wait_and_advance()
tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
# Register bridge
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,
)
frg_cnt = 4
frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
rP_bf16_frg = cute.logical_divide(rP_bf16, cute.make_layout(frg_tile))
for j in range(frg_cnt):
s_vec = tTMEM_LOADrS_frg[None, j].load()
rP_bf16_frg[None, j].store(s_vec.to(self.q_dtype))
cute.copy(tiled_tmem_store, rP_words, tTMEM_STOREtP)
cute.arch.fence_view_async_tmem_store()
si_handle.release()
softmax_done_bar.arrive()
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)
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)
def test():
torch.manual_seed(42)
for hd in [128, 64]:
m, n = 128, 128
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.ones(n, hd, dtype=torch.bfloat16, device='cuda')
v = v.as_strided((n, hd), (1, n)).unsqueeze(-1)
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
ref = (qf @ kf.T).bfloat16().float() @ v[:,:,0].float()
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
mV = ct.from_dlpack(v).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v))
mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
kernel = FmhaV3Debug(head_dim=hd)
print(f'hd={hd}: Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print(f'hd={hd}: pv_mma_tiler={kernel.pv_mma_tiler} p0={kernel.tmem_p0_offset} o0={kernel.tmem_o0_offset} alloc={kernel.num_tmem_alloc_cols}', flush=True)
compiled(mQ, mK, mV, mC, stream)
torch.cuda.synchronize()
out = c[:,:,0].float()
cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item()
print(f'hd={hd} V=ones: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if cos < 0.99:
print(f' out[0,:4]={out[0,:4].tolist()} ref[0,:4]={ref[0,:4].tolist()}')
if __name__ == '__main__':
test()

258
tests/test_pv64_fmha_v.py Normal file
View File

@@ -0,0 +1,258 @@
"""
Test (128,64) PV with FMHA-style V reconstruction.
V storage is (n, hd) row-major. Inside CuTe, we reconstruct as (hd, n, 1) MN-major.
This makes PV B operand see N=hd=64, K=n=128 with all 128 K-values.
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
HEAD_DIM = 64
class Pv64FmhaV:
def __init__(self):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
self.num_ab_stage = 1; self.num_acc_stage = 1
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik))
self.mma_tiler = self.qk_mma_tiler
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.a_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.b_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tilePlikeFP32 = self.qk_mma_tiler[1] // Float32.width * self.o_dtype.width
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
a_s = cute.slice_(self.a_smem_s,(None,None,None,0)); b_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
self.num_tma_load_bytes = (cute.size_in_bytes(self.q_dtype,a_s)+cute.size_in_bytes(self.q_dtype,b_s)+cute.size_in_bytes(self.q_dtype,v_s))*cute.size(qk_mma.thr_id.shape)
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
# FMHA-style V reconstruction: logical shape (HEAD_DIM, s_k, 1) MN-major
# Physical V storage is (n, hd) row-major. Reinterpret as (hd, n, 1) with stride (1, hd, hd*n).
s_k = 128 # sequence length (QK N dimension)
d = HEAD_DIM
v_fmha = cute.make_tensor(
v.iterator,
cute.make_layout(
(d, s_k, 1),
stride=(1, d, d * s_k),
),
)
self.v_major = LayoutEnum.from_tensor(v_fmha).mma_major_mode()
self.c_layout = LayoutEnum.from_tensor(c)
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.qk_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.a_smem_s,(None,None,None,0)); k_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
# TMA for V uses v_fmha (reconstructed layout), not original v
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)
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.a_smem_s,self.b_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, a_smem_s, b_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:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k)
cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
ab_bar: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage*2]
mma_si_bar: cute.struct.MemRange[cutlass.Int64, 2]
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
smem = utils.SmemAllocator(); st = smem.allocate(SS)
ab_p,ab_c = pipeline.PipelineTmaUmma.create(barrier_storage=st.ab_bar.data_ptr(),num_stages=self.num_ab_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.num_tma_load_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
mma_si_prod,mma_si_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_si_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
softmax_done_bar = pipeline.NamedBarrier(barrier_id=3, num_threads=32 + 32*len(self.epilogue_warp_id))
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=a_smem_s.outer,byte_alignment=128,swizzle=a_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=b_smem_s.outer,byte_alignment=128,swizzle=b_smem_s.inner)
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
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))
k_cnt = cute.size(gQ, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator+self.tmem_s0_offset,tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator+self.tmem_o0_offset,tOtO.layout)
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(tOrP.iterator+self.qk_acc_dtype.width//self.q_dtype.width*self.tmem_p0_offset,tOrP.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as,self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as,self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# TMA LOAD
if warp_idx == self.tma_warp_id:
ab_p.reset(); peek = ab_p.try_acquire()
for kt in cutlass.range(k_cnt,unroll=1):
h = ab_p.acquire_and_advance(peek)
cute.copy(tma_q,tAgQ[(None,h.count)],tAsQ[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_k,tBgK[(None,h.count)],tBsK[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_v,tVgV[(None,h.count)],tVsV[(None,h.index)],tma_bar_ptr=h.barrier)
peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_p.try_acquire()
ab_p.tail()
# MMA
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
ab_c.reset(); peek = ab_c.try_wait()
s0_handle = mma_si_prod.acquire_and_advance()
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer,self.num_acc_stage)
acc_pipe.producer_acquire(acc_st)
# QK
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
for kb in cutlass.range(cute.size(tCrQ,mode=[2]),unroll_full=True):
cute.gemm(qk_mma,tStS0,tCrQ[(None,None,kb,h.index)],tCrK[(None,None,kb,h.index)],tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
h.release(); peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_c.try_wait()
cute.arch.fence_view_async_tmem_store(); s0_handle.commit()
softmax_done_bar.arrive_and_wait()
# PV
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None,None,None,0)]
for kb in cutlass.range(cute.size(tOrP0,mode=[2]),unroll_full=True):
cute.gemm(pv_mma,tOtO0,tOrP0[(None,None,kb)],tCrV_s[(None,None,kb)],tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
acc_pipe.producer_commit(acc_st); acc_st.advance(); acc_pipe.producer_tail(acc_st)
# EPILOGUE (softmax + epilogue)
if warp_idx < self.mma_warp_id:
tmem.allocate(self.num_tmem_alloc_cols)
tmem.wait_for_alloc()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
sfw_idx = tidx % (32 * len(self.epilogue_warp_id))
# S load
tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0)
thr_load = tiled_tmem_load.get_slice(sfw_idx)
tTMEM_LOADtS = thr_load.partition_S(tStS0)
cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1]))
tScS = qk_thr.partition_C(cS)
tTMEM_LOADcS = thr_load.partition_D(tScS)
# P store (QK C-fragment composition)
tStS_P_layout = cute.composition(tStS.layout, cute.make_layout((128, self.tilePlikeFP32)))
tStS_P = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStS_P_layout)
tmem_store_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStS_P)
thr_store = tiled_tmem_store.get_slice(sfw_idx)
tTMEM_STOREtS_x4 = thr_store.partition_D(tStS_P)
tScS_P_layout = cute.composition(tScS.layout, cute.make_layout((128, self.tilePlikeFP32)))
tScS_P = cute.make_tensor(tScS.iterator, tScS_P_layout)
tTMEM_STOREcS = thr_store.partition_S(tScS_P)
# Softmax
si_handle = mma_si_cons.wait_and_advance()
tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
tTMEM_STORErS_x4 = cute.make_rmem_tensor(tTMEM_STOREcS.shape, self.qk_acc_dtype)
tTMEM_STORErS_x4_e = cute.make_tensor(cute.recast_ptr(tTMEM_STORErS_x4.iterator, dtype=self.q_dtype), tTMEM_LOADrS.layout)
frg_cnt = 4; frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
tTMEM_STORErS_x4_e_frg = cute.logical_divide(tTMEM_STORErS_x4_e, cute.make_layout(frg_tile))
for j in range(frg_cnt):
s_vec = tTMEM_LOADrS_frg[None, j].load()
tTMEM_STORErS_x4_e_frg[None, j].store(s_vec.to(self.q_dtype))
cute.copy(tiled_tmem_store, tTMEM_STORErS_x4, tTMEM_STOREtS_x4)
cute.arch.fence_view_async_tmem_store()
si_handle.release()
softmax_done_bar.arrive()
# Epilogue
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)
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)
def test():
torch.manual_seed(42)
m, n, hd = 128, 128, HEAD_DIM
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
# V as (n, hd) row-major — FMHA-style: reconstruct inside CuTe as (hd, n, 1) MN-major
v = torch.ones(n, hd, dtype=torch.bfloat16, device='cuda')
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
ref = (qf @ kf.T).bfloat16().float() @ v.float()
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
mV = ct.from_dlpack(v).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v))
mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
kernel = Pv64FmhaV()
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print(f'tilePlikeFP32={kernel.tilePlikeFP32} v_major={kernel.v_major}', flush=True)
compiled(mQ, mK, mV, mC, stream)
torch.cuda.synchronize()
out = c[:,:,0].float()
cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item()
print(f'PV64 FMHA-V: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if cos < 0.99:
print(f' out[0,:4]={out[0,:4].tolist()} ref[0,:4]={ref[0,:4].tolist()}')
if __name__ == '__main__':
test()

238
tests/test_pv64_kmajor_v.py Normal file
View File

@@ -0,0 +1,238 @@
"""
Test (128,64) PV with V=(64,128) K-major (row-major contiguous).
PV MMA uses OperandMajorMode.K for V (auto-detected).
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
HEAD_DIM = 64
class Pv64KMajorV:
def __init__(self):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
self.num_ab_stage = 1; self.num_acc_stage = 1
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik))
self.mma_tiler = self.qk_mma_tiler
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.a_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.b_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tilePlikeFP32 = self.qk_mma_tiler[1] // Float32.width * self.o_dtype.width
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
a_s = cute.slice_(self.a_smem_s,(None,None,None,0)); b_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
self.num_tma_load_bytes = (cute.size_in_bytes(self.q_dtype,a_s)+cute.size_in_bytes(self.q_dtype,b_s)+cute.size_in_bytes(self.q_dtype,v_s))*cute.size(qk_mma.thr_id.shape)
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
# V is (64, 128) K-major — auto-detect
self.v_major = LayoutEnum.from_tensor(v).mma_major_mode()
self.c_layout = LayoutEnum.from_tensor(c)
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.qk_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.a_smem_s,(None,None,None,0)); k_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
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,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)
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.a_smem_s,self.b_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, a_smem_s, b_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:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k)
cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
ab_bar: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage*2]
mma_si_bar: cute.struct.MemRange[cutlass.Int64, 2]
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
smem = utils.SmemAllocator(); st = smem.allocate(SS)
ab_p,ab_c = pipeline.PipelineTmaUmma.create(barrier_storage=st.ab_bar.data_ptr(),num_stages=self.num_ab_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.num_tma_load_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
mma_si_prod,mma_si_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_si_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
softmax_done_bar = pipeline.NamedBarrier(barrier_id=3, num_threads=32 + 32*len(self.epilogue_warp_id))
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=a_smem_s.outer,byte_alignment=128,swizzle=a_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=b_smem_s.outer,byte_alignment=128,swizzle=b_smem_s.inner)
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
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))
k_cnt = cute.size(gQ, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator+self.tmem_s0_offset,tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator+self.tmem_o0_offset,tOtO.layout)
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(tOrP.iterator+self.qk_acc_dtype.width//self.q_dtype.width*self.tmem_p0_offset,tOrP.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as,self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as,self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
if warp_idx == self.tma_warp_id:
ab_p.reset(); peek = ab_p.try_acquire()
for kt in cutlass.range(k_cnt,unroll=1):
h = ab_p.acquire_and_advance(peek)
cute.copy(tma_q,tAgQ[(None,h.count)],tAsQ[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_k,tBgK[(None,h.count)],tBsK[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_v,tVgV[(None,h.count)],tVsV[(None,h.index)],tma_bar_ptr=h.barrier)
peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_p.try_acquire()
ab_p.tail()
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
ab_c.reset(); peek = ab_c.try_wait()
s0_handle = mma_si_prod.acquire_and_advance()
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer,self.num_acc_stage)
acc_pipe.producer_acquire(acc_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
for kb in cutlass.range(cute.size(tCrQ,mode=[2]),unroll_full=True):
cute.gemm(qk_mma,tStS0,tCrQ[(None,None,kb,h.index)],tCrK[(None,None,kb,h.index)],tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
h.release(); peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_c.try_wait()
cute.arch.fence_view_async_tmem_store(); s0_handle.commit()
softmax_done_bar.arrive_and_wait()
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None,None,None,0)]
for kb in cutlass.range(cute.size(tOrP0,mode=[2]),unroll_full=True):
cute.gemm(pv_mma,tOtO0,tOrP0[(None,None,kb)],tCrV_s[(None,None,kb)],tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
acc_pipe.producer_commit(acc_st); acc_st.advance(); acc_pipe.producer_tail(acc_st)
if warp_idx < self.mma_warp_id:
tmem.allocate(self.num_tmem_alloc_cols)
tmem.wait_for_alloc()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
sfw_idx = tidx % (32 * len(self.epilogue_warp_id))
tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0)
thr_load = tiled_tmem_load.get_slice(sfw_idx)
tTMEM_LOADtS = thr_load.partition_S(tStS0)
cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1]))
tScS = qk_thr.partition_C(cS)
tTMEM_LOADcS = thr_load.partition_D(tScS)
tStS_P_layout = cute.composition(tStS.layout, cute.make_layout((128, self.tilePlikeFP32)))
tStS_P = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStS_P_layout)
tmem_store_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStS_P)
thr_store = tiled_tmem_store.get_slice(sfw_idx)
tTMEM_STOREtS_x4 = thr_store.partition_D(tStS_P)
tScS_P_layout = cute.composition(tScS.layout, cute.make_layout((128, self.tilePlikeFP32)))
tScS_P = cute.make_tensor(tScS.iterator, tScS_P_layout)
tTMEM_STOREcS = thr_store.partition_S(tScS_P)
si_handle = mma_si_cons.wait_and_advance()
tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
tTMEM_STORErS_x4 = cute.make_rmem_tensor(tTMEM_STOREcS.shape, self.qk_acc_dtype)
tTMEM_STORErS_x4_e = cute.make_tensor(cute.recast_ptr(tTMEM_STORErS_x4.iterator, dtype=self.q_dtype), tTMEM_LOADrS.layout)
frg_cnt = 4; frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
tTMEM_STORErS_x4_e_frg = cute.logical_divide(tTMEM_STORErS_x4_e, cute.make_layout(frg_tile))
for j in range(frg_cnt):
s_vec = tTMEM_LOADrS_frg[None, j].load()
tTMEM_STORErS_x4_e_frg[None, j].store(s_vec.to(self.q_dtype))
cute.copy(tiled_tmem_store, tTMEM_STORErS_x4, tTMEM_STOREtS_x4)
cute.arch.fence_view_async_tmem_store()
si_handle.release()
softmax_done_bar.arrive()
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)
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)
def test():
torch.manual_seed(42)
m, n, hd = 128, 128, HEAD_DIM
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
v_base = torch.ones(n, hd, dtype=torch.bfloat16, device='cuda')
# V as (hd, n, 1) row-major = K-major
v = v_base.T.contiguous().unsqueeze(-1)
print(f'V: shape={v.shape} stride={v.stride()}', flush=True)
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
ref = (qf @ kf.T).bfloat16().float() @ v_base.float()
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
mV = ct.from_dlpack(v).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v))
mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
kernel = Pv64KMajorV()
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
compiled(mQ, mK, mV, mC, stream)
torch.cuda.synchronize()
out = c[:,:,0].float()
cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item()
print(f'PV64 K-major V: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if cos < 0.99:
print(f' out[0,:4]={out[0,:4].tolist()} ref[0,:4]={ref[0,:4].tolist()}')
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,222 @@
"""
Test (128,64) PV WITHOUT softmax.
QK writes S to TMEM. Then PV uses S directly as P (no BF16 conversion).
If the C-fragment store path works, PV should read S and produce output.
If PV reads zeros, the P/A alias is broken for (128,64).
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
HEAD_DIM = 64
class Pv64NoSoftmax:
def __init__(self):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
self.num_ab_stage = 1; self.num_acc_stage = 1
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik))
self.mma_tiler = self.qk_mma_tiler
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.a_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.b_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tmem_s0_offset = 0; self.tmem_p0_offset =32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
a_s = cute.slice_(self.a_smem_s,(None,None,None,0)); b_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
self.num_tma_load_bytes = (cute.size_in_bytes(self.q_dtype,a_s)+cute.size_in_bytes(self.q_dtype,b_s)+cute.size_in_bytes(self.q_dtype,v_s))*cute.size(qk_mma.thr_id.shape)
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
self.v_major = LayoutEnum.from_tensor(v).mma_major_mode()
self.c_layout = LayoutEnum.from_tensor(c)
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.qk_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.a_smem_s,(None,None,None,0)); k_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
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,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)
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.a_smem_s,self.b_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, a_smem_s, b_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:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k)
cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
ab_bar: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage*2]
mma_si_bar: cute.struct.MemRange[cutlass.Int64, 2]
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
smem = utils.SmemAllocator(); st = smem.allocate(SS)
ab_p,ab_c = pipeline.PipelineTmaUmma.create(barrier_storage=st.ab_bar.data_ptr(),num_stages=self.num_ab_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.num_tma_load_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
mma_si_prod,mma_si_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_si_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=a_smem_s.outer,byte_alignment=128,swizzle=a_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=b_smem_s.outer,byte_alignment=128,swizzle=b_smem_s.inner)
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
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))
k_cnt = cute.size(gQ, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator+self.tmem_s0_offset,tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator+self.tmem_o0_offset,tOtO.layout)
# PV reads from S offset (no separate P, no softmax)
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(
tOrP.iterator + self.qk_acc_dtype.width // self.q_dtype.width * self.tmem_p0_offset,
tOrP.layout) # reads from S offset = 0
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# TMA LOAD
if warp_idx == self.tma_warp_id:
ab_p.reset(); peek = ab_p.try_acquire()
for kt in cutlass.range(k_cnt,unroll=1):
h = ab_p.acquire_and_advance(peek)
cute.copy(tma_q,tAgQ[(None,h.count)],tAsQ[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_k,tBgK[(None,h.count)],tBsK[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_v,tVgV[(None,h.count)],tVsV[(None,h.index)],tma_bar_ptr=h.barrier)
peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_p.try_acquire()
ab_p.tail()
# MMA — QK then PV, no softmax, PV reads S directly
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
ab_c.reset(); peek = ab_c.try_wait()
s0_handle = mma_si_prod.acquire_and_advance()
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer,self.num_acc_stage)
acc_pipe.producer_acquire(acc_st)
# QK
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
for kb in cutlass.range(cute.size(tCrQ,mode=[2]),unroll_full=True):
cute.gemm(qk_mma,tStS0,tCrQ[(None,None,kb,h.index)],tCrK[(None,None,kb,h.index)],tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
h.release(); peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_c.try_wait()
cute.arch.fence_view_async_tmem_store(); s0_handle.commit()
# No softmax — just signal epilogue to proceed
s0_handle = mma_si_prod.acquire_and_advance()
# PV — reads S directly from tmem_s0_offset as P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None,None,None,0)]
for kb in cutlass.range(cute.size(tOrP0,mode=[2]),unroll_full=True):
cute.gemm(pv_mma,tOtO0,tOrP0[(None,None,kb)],tCrV_s[(None,None,kb)],tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
acc_pipe.producer_commit(acc_st); acc_st.advance(); acc_pipe.producer_tail(acc_st)
# EPILOGUE — no softmax, just drain the si pipeline and do epilogue
if warp_idx < self.mma_warp_id:
tmem.allocate(self.num_tmem_alloc_cols)
tmem.wait_for_alloc()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
# Drain the si pipeline (no actual softmax work)
si_handle = mma_si_cons.wait_and_advance()
si_handle.release()
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)
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)
def test():
torch.manual_seed(42)
m, n, hd = 128, 128, HEAD_DIM
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.ones(n, hd, dtype=torch.bfloat16, device='cuda')
v = v.as_strided((n, hd), (1, n)).unsqueeze(-1)
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# No softmax: O = S_fp32_as_bf16 @ V (treating FP32 S as BF16 pairs for P)
# This is a correctness test for the P/A alias, not a useful computation
ref = (qf @ kf.T).bfloat16().float() @ v[:,:,0].float()
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
mV = ct.from_dlpack(v).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v))
mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
kernel = Pv64NoSoftmax()
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print(f'tmem_offsets: s0={kernel.tmem_s0_offset} p0={kernel.tmem_p0_offset} o0={kernel.tmem_o0_offset}', flush=True)
compiled(mQ, mK, mV, mC, stream)
torch.cuda.synchronize()
out = c[:,:,0].float()
# PV reads FP32 S as BF16 P — this is NOT numerically meaningful
# but if the alias works, we should get NON-ZERO output
nonzero = (out != 0).sum().item()
print(f'PV64 no-softmax: nonzero={nonzero}/{out.numel()} out[0,:4]={out[0,:4].tolist()}')
if nonzero > 0:
print('P/A alias works for (128,64) — PV reads non-zero data from TMEM')
else:
print('P/A alias BROKEN for (128,64) — PV reads all zeros from TMEM')
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,207 @@
"""
Test (128,64) PV with FMHA-style V reconstruction, no softmax.
V storage is (n, hd) row-major. Reconstructed inside CuTe as (hd, n, 1) MN-major.
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
HEAD_DIM = 64
class Pv64NoSoftmaxFmhaV:
def __init__(self):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
self.num_ab_stage = 1; self.num_acc_stage = 1
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik))
self.mma_tiler = self.qk_mma_tiler
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.a_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.b_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tmem_s0_offset = 0; self.tmem_p0_offset = 0
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
a_s = cute.slice_(self.a_smem_s,(None,None,None,0)); b_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
self.num_tma_load_bytes = (cute.size_in_bytes(self.q_dtype,a_s)+cute.size_in_bytes(self.q_dtype,b_s)+cute.size_in_bytes(self.q_dtype,v_s))*cute.size(qk_mma.thr_id.shape)
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
# FMHA-style V: logical (HEAD_DIM, s_k, 1) with stride (1, HEAD_DIM, HEAD_DIM*s_k)
v_fmha = cute.make_tensor(
v.iterator,
cute.make_layout(
(HEAD_DIM, 128, 1),
stride=(1, HEAD_DIM, HEAD_DIM * 128),
),
)
self.v_major = LayoutEnum.from_tensor(v_fmha).mma_major_mode()
self.c_layout = LayoutEnum.from_tensor(c)
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.qk_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.a_smem_s,(None,None,None,0)); k_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
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)
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.a_smem_s,self.b_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, a_smem_s, b_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:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k)
cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
ab_bar: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage*2]
mma_si_bar: cute.struct.MemRange[cutlass.Int64, 2]
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
smem = utils.SmemAllocator(); st = smem.allocate(SS)
ab_p,ab_c = pipeline.PipelineTmaUmma.create(barrier_storage=st.ab_bar.data_ptr(),num_stages=self.num_ab_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.num_tma_load_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
mma_si_prod,mma_si_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_si_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=a_smem_s.outer,byte_alignment=128,swizzle=a_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=b_smem_s.outer,byte_alignment=128,swizzle=b_smem_s.inner)
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
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))
k_cnt = cute.size(gQ, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator+self.tmem_s0_offset,tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator+self.tmem_o0_offset,tOtO.layout)
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(tOrP.iterator+self.qk_acc_dtype.width//self.q_dtype.width*self.tmem_s0_offset,tOrP.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
if warp_idx == self.tma_warp_id:
ab_p.reset(); peek = ab_p.try_acquire()
for kt in cutlass.range(k_cnt,unroll=1):
h = ab_p.acquire_and_advance(peek)
cute.copy(tma_q,tAgQ[(None,h.count)],tAsQ[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_k,tBgK[(None,h.count)],tBsK[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_v,tVgV[(None,h.count)],tVsV[(None,h.index)],tma_bar_ptr=h.barrier)
peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_p.try_acquire()
ab_p.tail()
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
ab_c.reset(); peek = ab_c.try_wait()
s0_handle = mma_si_prod.acquire_and_advance()
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer,self.num_acc_stage)
acc_pipe.producer_acquire(acc_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
for kb in cutlass.range(cute.size(tCrQ,mode=[2]),unroll_full=True):
cute.gemm(qk_mma,tStS0,tCrQ[(None,None,kb,h.index)],tCrK[(None,None,kb,h.index)],tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
h.release(); peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_c.try_wait()
cute.arch.fence_view_async_tmem_store(); s0_handle.commit()
s0_handle = mma_si_prod.acquire_and_advance()
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None,None,None,0)]
for kb in cutlass.range(cute.size(tOrP0,mode=[2]),unroll_full=True):
cute.gemm(pv_mma,tOtO0,tOrP0[(None,None,kb)],tCrV_s[(None,None,kb)],tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
acc_pipe.producer_commit(acc_st); acc_st.advance(); acc_pipe.producer_tail(acc_st)
if warp_idx < self.mma_warp_id:
tmem.allocate(self.num_tmem_alloc_cols)
tmem.wait_for_alloc()
si_handle = mma_si_cons.wait_and_advance()
si_handle.release()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
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)
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)
def test():
torch.manual_seed(42)
m, n, hd = 128, 128, HEAD_DIM
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.ones(n, hd, dtype=torch.bfloat16, device='cuda')
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
mV = ct.from_dlpack(v).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v))
mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
kernel = Pv64NoSoftmaxFmhaV()
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
compiled(mQ, mK, mV, mC, stream)
torch.cuda.synchronize()
out = c[:,:,0].float()
nonzero = (out != 0).sum().item()
has_nan = torch.isnan(out).sum().item()
print(f'No-softmax FMHA-V: nonzero={nonzero} nan={has_nan} out[0,:4]={out[0,:4].tolist()}')
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,248 @@
"""
Test (128,64) PV WITH identity softmax, single AB pipeline.
This is test_pv64.py but we add a print to see if softmax actually writes.
The key: if P/A alias works (proven above), then 0.67 must be from softmax
writing to wrong columns or V being wrong.
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
HEAD_DIM = 64
class Pv64WithSoftmax:
def __init__(self):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
self.num_ab_stage = 1; self.num_acc_stage = 1
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik))
self.mma_tiler = self.qk_mma_tiler
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.a_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.b_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.mma_tiler, self.q_dtype, 1)
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tilePlikeFP32 = self.qk_mma_tiler[1] // Float32.width * self.o_dtype.width
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
a_s = cute.slice_(self.a_smem_s,(None,None,None,0)); b_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
self.num_tma_load_bytes = (cute.size_in_bytes(self.q_dtype,a_s)+cute.size_in_bytes(self.q_dtype,b_s)+cute.size_in_bytes(self.q_dtype,v_s))*cute.size(qk_mma.thr_id.shape)
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
self.v_major = LayoutEnum.from_tensor(v).mma_major_mode()
self.c_layout = LayoutEnum.from_tensor(c)
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.qk_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.a_smem_s,(None,None,None,0)); k_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
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,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)
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.a_smem_s,self.b_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, a_smem_s, b_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:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k)
cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
ab_bar: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage*2]
mma_si_bar: cute.struct.MemRange[cutlass.Int64, 2]
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
smem = utils.SmemAllocator(); st = smem.allocate(SS)
ab_p,ab_c = pipeline.PipelineTmaUmma.create(barrier_storage=st.ab_bar.data_ptr(),num_stages=self.num_ab_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.num_tma_load_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
mma_si_prod,mma_si_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_si_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
softmax_done_bar = pipeline.NamedBarrier(barrier_id=3, num_threads=32 + 32*len(self.epilogue_warp_id))
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=a_smem_s.outer,byte_alignment=128,swizzle=a_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=b_smem_s.outer,byte_alignment=128,swizzle=b_smem_s.inner)
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
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))
k_cnt = cute.size(gQ, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator+self.tmem_s0_offset,tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator+self.tmem_o0_offset,tOtO.layout)
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(tOrP.iterator+self.qk_acc_dtype.width//self.q_dtype.width*self.tmem_p0_offset,tOrP.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as,self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as,self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# TMA LOAD
if warp_idx == self.tma_warp_id:
ab_p.reset(); peek = ab_p.try_acquire()
for kt in cutlass.range(k_cnt,unroll=1):
h = ab_p.acquire_and_advance(peek)
cute.copy(tma_q,tAgQ[(None,h.count)],tAsQ[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_k,tBgK[(None,h.count)],tBsK[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_v,tVgV[(None,h.count)],tVsV[(None,h.index)],tma_bar_ptr=h.barrier)
peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_p.try_acquire()
ab_p.tail()
# MMA
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
ab_c.reset(); peek = ab_c.try_wait()
s0_handle = mma_si_prod.acquire_and_advance()
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer,self.num_acc_stage)
acc_pipe.producer_acquire(acc_st)
# QK
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
for kb in cutlass.range(cute.size(tCrQ,mode=[2]),unroll_full=True):
cute.gemm(qk_mma,tStS0,tCrQ[(None,None,kb,h.index)],tCrK[(None,None,kb,h.index)],tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
h.release(); peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_c.try_wait()
cute.arch.fence_view_async_tmem_store(); s0_handle.commit()
softmax_done_bar.arrive_and_wait()
# PV
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None,None,None,0)]
if tidx == 0: print(f"PV64: nblk={int(cute.size(tOrP0,mode=[2]))} tOrP0_s={int(cute.size(tOrP0))}")
for kb in cutlass.range(cute.size(tOrP0,mode=[2]),unroll_full=True):
cute.gemm(pv_mma,tOtO0,tOrP0[(None,None,kb)],tCrV_s[(None,None,kb)],tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
acc_pipe.producer_commit(acc_st); acc_st.advance(); acc_pipe.producer_tail(acc_st)
# EPILOGUE (softmax + epilogue)
if warp_idx < self.mma_warp_id:
tmem.allocate(self.num_tmem_alloc_cols)
tmem.wait_for_alloc()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
sfw_idx = tidx % (32 * len(self.epilogue_warp_id))
# S load
tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0)
thr_load = tiled_tmem_load.get_slice(sfw_idx)
tTMEM_LOADtS = thr_load.partition_S(tStS0)
cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1]))
tScS = qk_thr.partition_C(cS)
tTMEM_LOADcS = thr_load.partition_D(tScS)
# P store (QK C-fragment composition)
tStS_P_layout = cute.composition(tStS.layout, cute.make_layout((128, self.tilePlikeFP32)))
tStS_P = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStS_P_layout)
tmem_store_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStS_P)
thr_store = tiled_tmem_store.get_slice(sfw_idx)
tTMEM_STOREtS_x4 = thr_store.partition_D(tStS_P)
tScS_P_layout = cute.composition(tScS.layout, cute.make_layout((128, self.tilePlikeFP32)))
tScS_P = cute.make_tensor(tScS.iterator, tScS_P_layout)
tTMEM_STOREcS = thr_store.partition_S(tScS_P)
# Softmax
si_handle = mma_si_cons.wait_and_advance()
tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
tTMEM_STORErS_x4 = cute.make_rmem_tensor(tTMEM_STOREcS.shape, self.qk_acc_dtype)
tTMEM_STORErS_x4_e = cute.make_tensor(cute.recast_ptr(tTMEM_STORErS_x4.iterator, dtype=self.q_dtype), tTMEM_LOADrS.layout)
frg_cnt = 4; frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
tTMEM_STORErS_x4_e_frg = cute.logical_divide(tTMEM_STORErS_x4_e, cute.make_layout(frg_tile))
for j in range(frg_cnt):
s_vec = tTMEM_LOADrS_frg[None, j].load()
tTMEM_STORErS_x4_e_frg[None, j].store(s_vec.to(self.q_dtype))
cute.copy(tiled_tmem_store, tTMEM_STORErS_x4, tTMEM_STOREtS_x4)
cute.arch.fence_view_async_tmem_store()
si_handle.release()
softmax_done_bar.arrive()
# Epilogue
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)
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)
def test():
torch.manual_seed(42)
m, n, hd = 128, 128, HEAD_DIM
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.ones(n, hd, dtype=torch.bfloat16, device='cuda')
v = v.as_strided((n, hd), (1, n)).unsqueeze(-1)
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
ref = (qf @ kf.T).bfloat16().float() @ v[:,:,0].float()
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
mV = ct.from_dlpack(v).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v))
mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
kernel = Pv64WithSoftmax()
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print(f'tilePlikeFP32={kernel.tilePlikeFP32} s0={kernel.tmem_s0_offset} p0={kernel.tmem_p0_offset} o0={kernel.tmem_o0_offset}', flush=True)
compiled(mQ, mK, mV, mC, stream)
torch.cuda.synchronize()
out = c[:,:,0].float()
cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item()
print(f'PV64 with softmax: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if cos < 0.99:
print(f' out[0,:4]={out[0,:4].tolist()} ref[0,:4]={ref[0,:4].tolist()}')
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,247 @@
"""
Debug: Verify softmax store actually writes non-zero P to TMEM.
Test: QK only + identity softmax + read back S region (should be partially overwritten by P).
If S region is modified after softmax, the store is working.
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
HEAD_DIM = 64
class SoftmaxStoreDebug:
def __init__(self):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
self.kv_stage = 2; self.q_stage = 1; self.num_c_stage = 2
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik))
self.mma_tiler = self.qk_mma_tiler
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
self.q_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.qk_mma_tiler, self.q_dtype, self.q_stage)
self.k_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.qk_mma_tiler, self.q_dtype, self.kv_stage)
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, self.kv_stage)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
cta = cute.size(qk_mma.thr_id.shape)
q_s = cute.slice_(self.q_smem_s,(None,None,None,0)); k_s = cute.slice_(self.k_smem_s,(None,None,None,0))
self.q_tx_bytes = cute.size_in_bytes(self.q_dtype, q_s) * cta
self.kv_tx_bytes = cute.size_in_bytes(self.q_dtype, k_s) * cta
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type
a_major = LayoutEnum.from_tensor(q).mma_major_mode()
b_major = LayoutEnum.from_tensor(k).mma_major_mode()
v_major = LayoutEnum.from_tensor(v).mma_major_mode()
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, a_major, b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, v_major, self.qk_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.q_smem_s,(None,None,None,0)); k_s = cute.slice_(self.k_smem_s,(None,None,None,0)); v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
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,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)
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())
tidx,_,_ = cute.arch.thread_idx()
if warp_idx == self.tma_warp_id:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k); cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
q_bar: cute.struct.MemRange[cutlass.Int64, self.q_stage*2]
kv_bar: cute.struct.MemRange[cutlass.Int64, self.kv_stage*2]
s_bar: cute.struct.MemRange[cutlass.Int64, 2]
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
smem = utils.SmemAllocator(); st = smem.allocate(SS)
qp,qc = pipeline.PipelineTmaUmma.create(barrier_storage=st.q_bar.data_ptr(),num_stages=self.q_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.q_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
kvp,kvc = pipeline.PipelineTmaUmma.create(barrier_storage=st.kv_bar.data_ptr(),num_stages=self.kv_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.kv_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
s_prod,s_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.s_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
softmax_done_bar = pipeline.NamedBarrier(barrier_id=3, num_threads=32 + 32*len(self.epilogue_warp_id))
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=q_smem_s.outer,byte_alignment=128,swizzle=q_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=k_smem_s.outer,byte_alignment=128,swizzle=k_smem_s.inner)
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
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))
n_kv_tiles = cute.size(gK, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout)
# PV read view (MMA only)
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(
tOrP.iterator + self.qk_acc_dtype.width // self.q_dtype.width * self.tmem_p0_offset,
tOrP.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# TMA LOAD
if warp_idx == self.tma_warp_id:
qp.reset(); qh = qp.acquire_and_advance()
cute.copy(tma_q,tAgQ[(None,qh.count)],tAsQ[(None,qh.index)],tma_bar_ptr=qh.barrier)
qp.tail()
kvp.reset(); pk = kvp.try_acquire()
for kt in cutlass.range(n_kv_tiles,unroll=1):
kh = kvp.acquire_and_advance(pk)
cute.copy(tma_k,tBgK[(None,kh.count)],tBsK[(None,kh.index)],tma_bar_ptr=kh.barrier)
pk = cutlass.Boolean(1)
vh = kvp.acquire_and_advance(pk)
cute.copy(tma_v,tVgV[(None,vh.count)],tVsV[(None,vh.index)],tma_bar_ptr=vh.barrier)
pk = cutlass.Boolean(1)
kvp.tail()
# MMA — same as v3
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
qc.reset(); qh = qc.wait_and_advance(); qh.release()
kvc.reset(); pk = kvc.try_wait()
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_st)
for kt in range(n_kv_tiles):
kh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)
sh = s_prod.acquire_and_advance()
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kb in cutlass.range(cute.size(tCrQ,mode=[2]), unroll_full=True):
cute.gemm(qk_mma, tStS0, tCrQ[(None,None,kb,0)], tCrK[(None,None,kb,kh.index)], tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
cute.arch.fence_view_async_tmem_store()
sh.commit(); kh.release()
softmax_done_bar.arrive_and_wait()
vh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)
pv_mma.set(tcgen05.Field.ACCUMULATE, kt != 0)
for kb in cutlass.range(cute.size(tOrP0,mode=[2]), unroll_full=True):
cute.gemm(pv_mma, tOtO0, tOrP0[(None,None,kb)], tCrV[(None,None,kb,vh.index)], tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
cute.arch.fence_async_view_tmem_store()
vh.release()
acc_pipe.producer_commit(acc_st); acc_st.advance()
acc_pipe.producer_tail(acc_st)
# EPILOGUE — debug: dump S+P region to GMEM instead of doing PV
if warp_idx < self.mma_warp_id:
tmem.allocate(self.num_tmem_alloc_cols)
tmem.wait_for_alloc()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
sfw_idx = tidx % (32 * len(self.epilogue_warp_id))
# S load + P store (same as v3)
tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0)
thr_load = tiled_tmem_load.get_slice(sfw_idx)
tTMEM_LOADtS = thr_load.partition_S(tStS0)
cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1]))
tScS = qk_thr.partition_C(cS)
tTMEM_LOADcS = thr_load.partition_D(tScS)
p_cols_fp32 = self.pv_mma_tiler[2] * self.q_dtype.width // self.qk_acc_dtype.width
tStP_layout = cute.composition(tStS.layout, cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32)))
tStP0 = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStP_layout)
tmem_store_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStP0)
thr_store = tiled_tmem_store.get_slice(sfw_idx)
tTMEM_STOREtP = thr_store.partition_D(tStP0)
tScP_layout = cute.composition(tScS.layout, cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32)))
tScP = cute.make_tensor(tScS.iterator, tScP_layout)
tTMEM_STOREcP = thr_store.partition_S(tScP)
for kt in range(n_kv_tiles):
si_handle = s_cons.wait_and_advance()
tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
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)
frg_cnt = 4; frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
rP_bf16_frg = cute.logical_divide(rP_bf16, cute.make_layout(frg_tile))
for j in range(frg_cnt):
s_vec = tTMEM_LOADrS_frg[None, j].load()
rP_bf16_frg[None, j].store(s_vec.to(self.q_dtype))
cute.copy(tiled_tmem_store, rP_words, tTMEM_STOREtP)
cute.arch.fence_view_async_tmem_store()
si_handle.release()
softmax_done_bar.arrive()
# Now read back the P region directly
# Read from tStP0 (the same region we wrote to)
tmem_read_P_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_read_P = tcgen05.make_tmem_copy(tmem_read_P_atom, tStP0)
thr_read_P = tiled_tmem_read_P.get_slice(sfw_idx)
tTMEM_READtP = thr_read_P.partition_S(tStP0)
rP_read = cute.make_rmem_tensor(thr_read_P.partition_D(tScP).shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_read_P, tTMEM_READtP, rP_read)
# Print the first few values
if sfw_idx == 0:
print(f"DEBUG P readback: size={cute.size(rP_read)} val0={float(rP_read.load()[0])}")
# Epilogue as before
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)
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)

284
tests/test_v_mode_fix.py Normal file
View File

@@ -0,0 +1,284 @@
"""
Test: V shape (HEAD_DIM, n, 1) instead of (n, HEAD_DIM, 1).
PV B operand wants (N=HEAD_DIM, K=seq_len). V must be MN-major (d, s_k).
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
HEAD_DIM = 64
class FmhaV3:
def __init__(self):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
self.kv_stage = 2; self.q_stage = 1; self.num_c_stage = 2
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik))
self.mma_tiler = self.qk_mma_tiler
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
self.q_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.qk_mma_tiler, self.q_dtype, self.q_stage)
self.k_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.qk_mma_tiler, self.q_dtype, self.kv_stage)
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, self.kv_stage)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
cta = cute.size(qk_mma.thr_id.shape)
q_s = cute.slice_(self.q_smem_s,(None,None,None,0)); k_s = cute.slice_(self.k_smem_s,(None,None,None,0))
self.q_tx_bytes = cute.size_in_bytes(self.q_dtype, q_s) * cta
self.kv_tx_bytes = cute.size_in_bytes(self.q_dtype, k_s) * cta
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
self.v_major = LayoutEnum.from_tensor(v).mma_major_mode()
self.c_layout = LayoutEnum.from_tensor(c)
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.qk_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.q_smem_s,(None,None,None,0)); k_s = cute.slice_(self.k_smem_s,(None,None,None,0)); v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
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,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)
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())
tidx,_,_ = cute.arch.thread_idx()
if warp_idx == self.tma_warp_id:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k); cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
q_bar: cute.struct.MemRange[cutlass.Int64, self.q_stage*2]
kv_bar: cute.struct.MemRange[cutlass.Int64, self.kv_stage*2]
s_bar: cute.struct.MemRange[cutlass.Int64, 2]
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
smem = utils.SmemAllocator(); st = smem.allocate(SS)
qp,qc = pipeline.PipelineTmaUmma.create(barrier_storage=st.q_bar.data_ptr(),num_stages=self.q_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.q_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
kvp,kvc = pipeline.PipelineTmaUmma.create(barrier_storage=st.kv_bar.data_ptr(),num_stages=self.kv_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.kv_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
s_prod,s_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.s_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
softmax_done_bar = pipeline.NamedBarrier(barrier_id=3, num_threads=32 + 32*len(self.epilogue_warp_id))
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=q_smem_s.outer,byte_alignment=128,swizzle=q_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=k_smem_s.outer,byte_alignment=128,swizzle=k_smem_s.inner)
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
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))
n_kv_tiles = cute.size(gK, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout)
# PV read view (MMA only)
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(
tOrP.iterator + self.qk_acc_dtype.width // self.q_dtype.width * self.tmem_p0_offset,
tOrP.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# TMA LOAD
if warp_idx == self.tma_warp_id:
qp.reset(); qh = qp.acquire_and_advance()
cute.copy(tma_q,tAgQ[(None,qh.count)],tAsQ[(None,qh.index)],tma_bar_ptr=qh.barrier)
qp.tail()
kvp.reset(); pk = kvp.try_acquire()
for kt in cutlass.range(n_kv_tiles,unroll=1):
kh = kvp.acquire_and_advance(pk)
cute.copy(tma_k,tBgK[(None,kh.count)],tBsK[(None,kh.index)],tma_bar_ptr=kh.barrier)
pk = cutlass.Boolean(1)
vh = kvp.acquire_and_advance(pk)
cute.copy(tma_v,tVgV[(None,vh.count)],tVsV[(None,vh.index)],tma_bar_ptr=vh.barrier)
pk = cutlass.Boolean(1)
kvp.tail()
# MMA
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
qc.reset(); qh = qc.wait_and_advance(); qh.release()
kvc.reset(); pk = kvc.try_wait()
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_st)
for kt in range(n_kv_tiles):
kh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)
sh = s_prod.acquire_and_advance()
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kb in cutlass.range(cute.size(tCrQ,mode=[2]), unroll_full=True):
cute.gemm(qk_mma, tStS0, tCrQ[(None,None,kb,0)], tCrK[(None,None,kb,kh.index)], tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
cute.arch.fence_view_async_tmem_store()
sh.commit(); kh.release()
softmax_done_bar.arrive_and_wait()
vh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)
pv_mma.set(tcgen05.Field.ACCUMULATE, kt != 0)
for kb in cutlass.range(cute.size(tOrP0,mode=[2]), unroll_full=True):
cute.gemm(pv_mma, tOtO0, tOrP0[(None,None,kb)], tCrV[(None,None,kb,vh.index)], tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
cute.arch.fence_view_async_tmem_store()
vh.release()
acc_pipe.producer_commit(acc_st); acc_st.advance()
acc_pipe.producer_tail(acc_st)
# EPILOGUE
if warp_idx < self.mma_warp_id:
tmem.allocate(self.num_tmem_alloc_cols)
tmem.wait_for_alloc()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
sfw_idx = tidx % (32 * len(self.epilogue_warp_id))
# S load
tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0)
thr_load = tiled_tmem_load.get_slice(sfw_idx)
tTMEM_LOADtS = thr_load.partition_S(tStS0)
cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1]))
tScS = qk_thr.partition_C(cS)
tTMEM_LOADcS = thr_load.partition_D(tScS)
# P store — QK C-fragment composition (FMHA pattern)
p_cols_fp32 = self.pv_mma_tiler[2] * self.q_dtype.width // self.qk_acc_dtype.width
tStP_layout = cute.composition(
tStS.layout,
cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32)),
)
tStP0 = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStP_layout)
tmem_store_atom = cute.make_copy_atom(
tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)),
self.qk_acc_dtype,
)
tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStP0)
thr_store = tiled_tmem_store.get_slice(sfw_idx)
tTMEM_STOREtP = thr_store.partition_D(tStP0)
tScP_layout = cute.composition(
tScS.layout,
cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32)),
)
tScP = cute.make_tensor(tScS.iterator, tScP_layout)
tTMEM_STOREcP = thr_store.partition_S(tScP)
for kt in range(n_kv_tiles):
si_handle = s_cons.wait_and_advance()
tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
# Register bridge
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,
)
frg_cnt = 4
frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
rP_bf16_frg = cute.logical_divide(rP_bf16, cute.make_layout(frg_tile))
for j in range(frg_cnt):
s_vec = tTMEM_LOADrS_frg[None, j].load()
rP_bf16_frg[None, j].store(s_vec.to(self.q_dtype))
cute.copy(tiled_tmem_store, rP_words, tTMEM_STOREtP)
cute.arch.fence_view_async_tmem_store()
si_handle.release()
softmax_done_bar.arrive()
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)
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)
def test():
torch.manual_seed(42)
m, n, hd = 128, 128, HEAD_DIM
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
# FIX: V is (HEAD_DIM, n, 1) — PV B operand wants (d, s_k) = (HEAD_DIM, seq_len)
v_base = torch.ones(n, hd, dtype=torch.bfloat16, device='cuda')
v_for_kernel = v_base.T.unsqueeze(-1) # shape (hd, n, 1), stride (1, hd)
print(f"V shape: {v_for_kernel.shape} stride: {v_for_kernel.stride()}", flush=True)
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
ref = (qf @ kf.T).bfloat16().float() @ v_base.float()
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
mV = ct.from_dlpack(v_for_kernel).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_for_kernel))
mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
kernel = FmhaV3()
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print(f'pv_mma_tiler={kernel.pv_mma_tiler} p0={kernel.tmem_p0_offset} o0={kernel.tmem_o0_offset} alloc={kernel.num_tmem_alloc_cols}', flush=True)
compiled(mQ, mK, mV, mC, stream)
torch.cuda.synchronize()
out = c[:,:,0].float()
cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item()
print(f'V=(hd,n) cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if cos < 0.99:
print(f' out[0,:4]={out[0,:4].tolist()} ref[0,:4]={ref[0,:4].tolist()}')
if __name__ == '__main__':
test()