Stage B N-tiling: (128,16) PV MMA compiles and runs, cosine 0.36 (TMEM alias mismatch bug). FMHA head_dim=64 passes. Debugging TMEM layout alignment.

This commit is contained in:
2026-05-21 08:45:49 +00:00
parent 5e37ea56e4
commit ad3f63033d
18 changed files with 6941 additions and 0 deletions

385
tests/test_128_16_full.py Normal file
View File

@@ -0,0 +1,385 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 Test128x16Tiler:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
(self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2]), False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_((self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2]), (None,0,0)), (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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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)
saved_cta_tile = self.cta_tile_shape_mnk
self.cta_tile_shape_mnk = (int(self.pv_mma_tiler[0]), int(self.pv_mma_tiler[1]), int(self.pv_mma_tiler[2]))
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)
self.cta_tile_shape_mnk = saved_cta_tile
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
def test():
torch.manual_seed(42)
m, n, head_dim = 128, 128, 64
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity[:16,:] MN-major: (16,128) with strides (1,16)
v_data = torch.eye(128, dtype=torch.bfloat16, device='cuda')
v = v_data[:16, :].as_strided((16, 128), (1, 16)).unsqueeze(-1)
c = torch.zeros(m, 16, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().float()[:, :16]
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 = Test128x16Tiler(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV(128,16) full: cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,385 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 MinimalV8Kernel:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
(self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2]), False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.pv_mma_tiler, (None,0,0)), (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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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)
saved_cta_tile = self.cta_tile_shape_mnk
self.cta_tile_shape_mnk = (self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2])
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)
self.cta_tile_shape_mnk = saved_cta_tile
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
def test():
torch.manual_seed(42)
m, n, head_dim = 128, 128, 16
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity[:16,:] MN-major: (16,128) with strides (1,16)
v_data = torch.eye(128, dtype=torch.bfloat16, device='cuda')
v = v_data[:16, :].as_strided((16, 128), (1, 16)).unsqueeze(-1)
c = torch.zeros(m, 16, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# V=I[:16,:] -> O = P[:,:16] = (Q@K^T).bf16()[:,:16]
ref = (qf @ kf.T).bfloat16().float()[:, :16]
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 = MinimalV8Kernel(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV(128,16) minimal: cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

383
tests/test_128_16_nogC.py Normal file
View File

@@ -0,0 +1,383 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 Test128x16Tiler:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
(self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2]), False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.qk_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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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, head_dim = 128, 128, 64
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity (128x128) in MN-major: (128,128) with strides (1,128)
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')
# MN-major: (128,128) with strides (1,128) — row is fast dim
v = v.as_strided((128, 128), (1, 128)).unsqueeze(-1)
c = torch.zeros(m, n, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().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 = Test128x16Tiler(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV diag (V=I): cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

383
tests/test_128_16_nopack.py Normal file
View File

@@ -0,0 +1,383 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 NoPackV8Kernel:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
(self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2]), False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.qk_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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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, head_dim = 128, 128, 64
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity (128x128) in MN-major: (128,128) with strides (1,128)
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')
# MN-major: (128,128) with strides (1,128) — row is fast dim
v = v.as_strided((128, 128), (1, 128)).unsqueeze(-1)
c = torch.zeros(m, n, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().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 = NoPackV8Kernel(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV diag (V=I): cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,383 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 Test128x16NoSoftmax:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], self.qk_mma_tiler[1], self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], 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.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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.qk_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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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, head_dim = 128, 128, 64
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity (128x128) in MN-major: (128,128) with strides (1,128)
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')
# MN-major: (128,128) with strides (1,128) — row is fast dim
v = v.as_strided((128, 128), (1, 128)).unsqueeze(-1)
c = torch.zeros(m, n, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().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 = Test128x16NoSoftmax(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV diag (V=I): cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

383
tests/test_128_16_stepA.py Normal file
View File

@@ -0,0 +1,383 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 Test128x16Tiler:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], 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.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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.qk_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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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, head_dim = 128, 128, 64
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity (128x128) in MN-major: (128,128) with strides (1,128)
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')
# MN-major: (128,128) with strides (1,128) — row is fast dim
v = v.as_strided((128, 128), (1, 128)).unsqueeze(-1)
c = torch.zeros(m, n, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().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 = Test128x16Tiler(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV diag (V=I): cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

383
tests/test_128_16_stepB.py Normal file
View File

@@ -0,0 +1,383 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 Test128x16Tiler:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
(self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2]), False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.qk_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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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, head_dim = 128, 128, 64
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity (128x128) in MN-major: (128,128) with strides (1,128)
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')
# MN-major: (128,128) with strides (1,128) — row is fast dim
v = v.as_strided((128, 128), (1, 128)).unsqueeze(-1)
c = torch.zeros(m, n, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().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 = Test128x16Tiler(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV diag (V=I): cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

383
tests/test_128_16_stepC.py Normal file
View File

@@ -0,0 +1,383 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 Test128x16Tiler:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
(self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2]), False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.pv_mma_tiler, (None,0,0)), (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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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, head_dim = 128, 128, 64
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity (128x128) in MN-major: (128,128) with strides (1,128)
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')
# MN-major: (128,128) with strides (1,128) — row is fast dim
v = v.as_strided((128, 128), (1, 128)).unsqueeze(-1)
c = torch.zeros(m, n, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().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 = Test128x16Tiler(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV diag (V=I): cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

386
tests/test_128_16_stepD.py Normal file
View File

@@ -0,0 +1,386 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 Test128x16Tiler:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
(self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2]), False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.pv_mma_tiler, (None,0,0)), (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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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)
saved_cta_tile = self.cta_tile_shape_mnk
self.cta_tile_shape_mnk = (self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2])
acc_cons_st = utils.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)
self.cta_tile_shape_mnk = saved_cta_tile
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
def test():
torch.manual_seed(42)
m, n, head_dim = 128, 128, 64
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity (128x128) in MN-major: (128,128) with strides (1,128)
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')
# MN-major: (128,128) with strides (1,128) — row is fast dim
v = v.as_strided((128, 128), (1, 128)).unsqueeze(-1)
c = torch.zeros(m, n, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().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 = Test128x16Tiler(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV diag (V=I): cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

385
tests/test_128_16_stepE.py Normal file
View File

@@ -0,0 +1,385 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 Test128x16Tiler:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
(self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2]), False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.pv_mma_tiler, (None,0,0)), (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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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)
saved_cta_tile = self.cta_tile_shape_mnk
self.cta_tile_shape_mnk = (self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2])
acc_cons_st = utils.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)
self.cta_tile_shape_mnk = saved_cta_tile
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
def test():
torch.manual_seed(42)
m, n, head_dim = 128, 128, 16
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity[:16,:] MN-major: (16,128) with strides (1,16)
v_data = torch.eye(128, dtype=torch.bfloat16, device='cuda')
v = v_data[:16, :].as_strided((16, 128), (1, 16)).unsqueeze(-1)
c = torch.zeros(m, 16, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().float()[:, :16]
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 = Test128x16Tiler(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV diag (V=I): cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

383
tests/test_128_16_tiler.py Normal file
View File

@@ -0,0 +1,383 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 Test128x16Tiler:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], 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.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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, self.mma_tiler_mn, tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.qk_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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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, head_dim = 128, 128, 64
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity (128x128) in MN-major: (128,128) with strides (1,128)
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')
# MN-major: (128,128) with strides (1,128) — row is fast dim
v = v.as_strided((128, 128), (1, 128)).unsqueeze(-1)
c = torch.zeros(m, n, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().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 = Test128x16Tiler(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV diag (V=I): cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

385
tests/test_128_16_v8.py Normal file
View File

@@ -0,0 +1,385 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 Test128x16Tiler:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) since V is 128x128
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
(self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2]), False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.pv_mma_tiler, (None,0,0)), (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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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)
saved_cta_tile = self.cta_tile_shape_mnk
self.cta_tile_shape_mnk = (self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2])
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)
self.cta_tile_shape_mnk = saved_cta_tile
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
def test():
torch.manual_seed(42)
m, n, head_dim = 128, 128, 16
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity[:16,:] MN-major: (16,128) with strides (1,16)
v_data = torch.eye(128, dtype=torch.bfloat16, device='cuda')
v = v_data[:16, :].as_strided((16, 128), (1, 16)).unsqueeze(-1)
c = torch.zeros(m, 16, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I and identity softmax: O = (Q@K^T).bf16() @ I = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().float()[:, :16]
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 = Test128x16Tiler(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV(128,16) v8: cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,399 @@
"""
Stage B with N-tiling: QK + identity softmax + PV with (128,16) PV MMA.
Single subtile test: verify the C→A TMEM bridge works for (128,16) PV.
pv_mma_tiler = (128, 16, 128). K=128 matches QK's N=128 → C/A bridge preserved.
V is (16, 128) MN-major — first 16 rows of the head_dim.
Output is (128, 16) — should be the first 16 columns of (Q@K^T).bf16().
If this passes, the TMEM layout mismatch (Bug 4) is fixed by N-tiling.
Then we extend to the full (128, head_dim) output with the N-subtile loop.
"""
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 StageBSingleSubtileKernel:
"""Single (128,16) PV subtile test. Output should be (Q@K^T).bf16()[:, :16]."""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV MMA tiler: (M=128, N=16, K=128)
# N=16 = qk_inst_k (one subtile of head_dim)
# K=128 = qk_mma_tiler[1] (QK's N → PV's K, preserves C/A bridge)
self.pv_mma_tiler = (
self.qk_mma_tiler[0], # M = 128
qk_inst_k, # N = 16
self.qk_mma_tiler[1], # K = 128
)
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,))
# For the epilogue: use PV's cta_tile_shape_mnk and epi_tile
self.cta_tile_shape_mnk = (
self.pv_mma_tiler[0] // cute.size(pv_mma.thr_id.shape),
self.pv_mma_tiler[1], self.pv_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.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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
self.tilePlikeFP32 = self.qk_mma_tiler[1] // Float32.width * self.o_dtype.width # = 64
self.tmem_s0_offset = 0
self.tmem_p0_offset = 32
self.tmem_o0_offset = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem)
+ cute.size_in_bytes(self.q_dtype, b_smem)
+ cute.size_in_bytes(self.q_dtype, v_smem)
) * 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)
# PV MMA tiler: (M=128, N=16, K=128) — N=16 is one subtile, K=128 matches QK N
# This must be computed before make_trivial_tiled_mma
self.pv_mma_tiler_mn = (128, 16) # (M, N) for PV MMA
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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV MMA with (128, 16) output, P from TMEM
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, self.pv_mma_tiler_mn, tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
# Output is (128, 16) — one PV subtile
gC = cute.local_tile(mC, cute.select(self.pv_mma_tiler, mode=[0,1]), (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)
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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
# V: pv_mma_tiler = (128, 16, 128). V is (16, 128) MN-major.
# TMA loads V using pv_mma_tiler's B-dimension slice.
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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 accumulator
qk_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
# PV accumulator: (128, 16) FP32 in TMEM
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout)
# P A-fragment for PV MMA
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
# QK MMA
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V → O (128, 16)
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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))
# Softmax packing: ld FP32 S → BF16 P → st to TMEM
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()
# Output epilogue: Try writing QK scores (S) first to verify pipeline works
# Use QK accumulator which we know is correct
tCtS_base = cute.make_tensor(tmem_ptr + self.tmem_s0_offset, tCtS_fake.layout)
# Write S to GMEM — using the original QK cta tile for the epilogue
self.cta_tile_shape_mnk = (
self.qk_mma_tiler[0] // cute.size(qk_mma.thr_id.shape),
self.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
# Need to re-partition gC for QK shape (128, 128)
gC_qk = cute.local_tile(mC, cute.select(self.qk_mma_tiler, mode=[0,1]), (None,None,None))
tCgC_qk = qk_thr.partition_C(gC_qk)
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, tCtS_base, sC, tCgC_qk,
self.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 = 128, 128
q = torch.randn(m, 128, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, 128, 1, dtype=torch.bfloat16, device='cuda')
# V = full 128x128 identity in MN-major, same as v30
# (128, 128) with strides (1, 128) — row (head_dim) is fast dim = MN-major
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')
v = v.as_strided((128, 128), (1, 128)).unsqueeze(-1)
# Output: (128, 16) — only one subtile of the output
c = torch.zeros(m, 16, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# V=I(128,128) with (128,16) PV MMA and single subtile:
# P@V subtile 0 = P[:, :16] (first 16 cols)
ref = (qf @ kf.T).bfloat16()[:, :16].float()
# Test 1: Output S = Q@K^T (bypass PV) to verify QK + epilogue still work
ref_s = (qf @ kf.T).bfloat16().float()
c_s = torch.zeros(m, 128, 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_s = ct.from_dlpack(c_s).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c_s))
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
kernel = StageBSingleSubtileKernel(mma_tiler_mn=(128, 128))
print('Compiling QK-only test...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC_s, stream)
print('Running QK-only...', flush=True)
compiled(mQ, mK, mV, mC_s, stream)
torch.cuda.synchronize()
out_s = c_s[:,:,0].float()
cos_s = torch.nn.functional.cosine_similarity(out_s.flatten().unsqueeze(0), ref_s.flatten().unsqueeze(0)).item()
print(f'QK scores: cosine {cos_s:.6f} {"PASS" if cos_s >= 0.99 else "FAIL"}')
return cos
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,357 @@
"""
Stage B N-tiling v3: (128,16) PV MMA. Minimal changes from v30 test_pv_diag.py.
Key changes:
1. pv_mma_tiler = (128, 16, 128) instead of (128, 128, 128)
2. epi_tile from pv_mma_tiler[:2] = (128, 16)
3. cta_tile_shape_mnk from PV for the epilogue
4. V = first 16 rows of identity (16, 128) MN-major
5. Output (128, 16)
"""
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 PvDiag16Kernel:
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
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,))
# CTA tile shape from PV for epilogue (FMHA pattern)
self.cta_tile_shape_mnk = (
self.pv_mma_tiler[0] // cute.size(pv_mma.thr_id.shape),
self.pv_mma_tiler[1], self.pv_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
# epi_tile from PV tiler, matching FMHA: self.epi_tile = self.pv_mma_tiler[:2]
self.epi_tile = self.pv_mma_tiler[:2]
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, 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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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.select(self.pv_mma_tiler, mode=[0,1]), (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)
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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output epilogue using PV accumulator
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 = 128, 128
q = torch.randn(m, 128, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, 128, 1, dtype=torch.bfloat16, device='cuda')
# V = first 16 rows of 128x128 identity, MN-major
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')[:16, :]
v = v.as_strided((16, 128), (1, 16)).unsqueeze(-1)
c = torch.zeros(m, 16, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# V=I[:16,:] → P@V = P[:,:16] (first 16 cols of (Q@K^T).bf16())
ref = (qf @ kf.T).bfloat16()[:, :16].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 = PvDiag16Kernel(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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'(128,16) PV MMA: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if cos < 0.5:
print(f' out max={out.max().item():.4f} min={out.min().item():.4f}')
print(f' ref max={ref.max().item():.4f} min={ref.min().item():.4f}')
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,415 @@
"""
Stage B N-tiling v5: (128,16) PV MMA with proper FMHA-style output.
TMEM → register → SMEM → TMA → GMEM. No shortcuts. Blackwell all the way.
V = I[:16,:] → output = (Q@K^T).bf16()[:,:16]
pv_mma_tiler = (128, 16, 128). K=128 matches QK's N → C/A bridge preserved.
"""
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 PvDiag16V5Kernel:
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
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.threads_per_warp = 32
def _setup(self, qk_mma, pv_mma):
qk_inst_k = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
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,))
# FMHA: epi_tile = pv_mma_tiler[:2]
self.epi_tile = self.pv_mma_tiler[:2] # (128, 16)
self.c_layout = LayoutEnum.ROW_MAJOR
self.num_ab_stage = 1; self.num_acc_stage = 1
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, self.num_c_stage)
qk_thr = qk_mma.get_slice(0)
qk_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, 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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
o_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
# TMA store for output — FMHA pattern: cpasync.make_tiled_tma_atom
tma_o, tma_to = cpasync.make_tiled_tma_atom(
cpasync.CopyBulkTensorTileS2GOp(), c, o_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_o, tma_to, self.cluster_layout_vmnk,
self.a_smem_s, self.b_smem_s, self.v_smem_s, self.p_tmem_s, self.c_smem_s
).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_o, mO, cl_vmnk, a_smem_s, b_smem_s, v_smem_s, p_tmem_s, c_smem_s):
warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
tidx, _, _ = cute.arch.thread_idx()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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_o)
@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) * (2 if use_2cta else 1)),
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=use_2cta,
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)
sO = 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))
k_cnt = cute.size(gQ, mode=[3])
# Output TMA partition — FMHA pattern: flat_divide + group_modes
gO_qdl = cute.flat_divide(mO, cute.select(self.pv_mma_tiler, mode=[0, 1]))
gO = gO_qdl[None, None, None, 0, 0] # single tile, coord (0,0)
o_smem_layout = cute.select(c_smem_s, mode=[0, 1])
tOsO, tOgO = cpasync.tma_partition(
tma_o, 0, cute.make_layout(1),
cute.group_modes(sO, 0, 2),
cute.group_modes(gO, 0, 2))
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)
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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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))
# ─── Softmax packing: FP32 S → BF16 P ───
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()
# ─── PV output epilogue: TMEM → register → SMEM → TMA → GMEM ───
# Wait for PV MMA to finish
acc_cons_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Consumer, self.num_acc_stage)
acc_pipe.consumer_wait(acc_cons_st)
# FMHA correction_epilog: TMEM → register → SMEM
corr_tile_size = 16 # 32*8//o_dtype.width for BF16
tOtO_i = cute.logical_divide(tOtO0, cute.make_layout((128, corr_tile_size)))
pv_tiled_mma_shape = (self.pv_mma_tiler[0], self.pv_mma_tiler[1])
cO = cute.make_identity_tensor(pv_tiled_mma_shape)
tOcO = pv_thr.partition_C(cO)
tOcO_i = cute.logical_divide(tOcO, cute.make_layout((128, corr_tile_size)))
tOsO_e = pv_thr.partition_C(sO)
tOsO_i = cute.logical_divide(tOsO_e, cute.make_layout((128, corr_tile_size)))
thread_idx = tidx % (self.threads_per_warp * len(self.epilogue_warp_id))
# TMEM load atom for O (correction epilog pattern)
epi_subtile = (self.epi_tile[0], corr_tile_size)
tmem_load_o_atom = cute.make_copy_atom(
tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(16)), self.qk_acc_dtype)
tiled_tmem_load_o = tcgen05.make_tmem_copy(tmem_load_o_atom, tOtO_i[(None, None), 0])
thr_tmem_load_o = tiled_tmem_load_o.get_slice(thread_idx)
# SMEM store atom (FMHA: get_smem_store_op)
smem_copy_atom = utils.sm100.get_smem_store_op(
self.c_layout, self.o_dtype, self.qk_acc_dtype, tiled_tmem_load_o)
tiled_smem_store = cute.make_tiled_copy_D(smem_copy_atom, tiled_tmem_load_o)
tTMEM_LOADtO = thr_tmem_load_o.partition_S(tOtO_i[(None, None), None])
tTMEM_LOADsO = thr_tmem_load_o.partition_D(tOsO_i[(None, None), None])
tTMEM_LOADoO = thr_tmem_load_o.partition_D(tOcO_i[(None, None), None])
# Loop over correction subtiles
for i in range(self.pv_mma_tiler[1] // corr_tile_size):
tTMEM_LOADtO_i = tTMEM_LOADtO[None, 0, 0, i]
tTMEM_LOADsO_i = tTMEM_LOADsO[None, 0, 0, i]
tTMrO = cute.make_rmem_tensor(tTMEM_LOADoO[None, 0, 0, i].shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_load_o, tTMEM_LOADtO_i, tTMrO)
cute.arch.fence_view_async_tmem_load()
# Cast FP32 → BF16
tSMrO = cute.make_rmem_tensor(tTMrO.shape, self.o_dtype)
o_vec = tTMrO.load()
tSMrO.store(o_vec.to(self.o_dtype))
# Store to SMEM
cute.copy(tiled_smem_store, tSMrO, tTMEM_LOADsO_i)
# Fence SMEM writes, then barrier so all epilogue warps are done writing
cute.arch.fence_proxy("async.shared", space="cta")
epilog_sync_barrier = pipeline.NamedBarrier(
barrier_id=self.epilog_sync_bar_id,
num_threads=32 * len(self.epilogue_warp_id))
epilog_sync_barrier.arrive_and_wait()
# TMA store: SMEM → GMEM (FMHA pattern)
cute.copy(tma_o, tOsO[None, 0], tOgO[None, 0])
cute.arch.cp_async_bulk_commit_group()
cute.arch.cp_async_bulk_wait_group(0, read=True)
acc_pipe.consumer_release(acc_cons_st)
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
def test():
torch.manual_seed(42)
m, n = 128, 128
q = torch.randn(m, 128, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, 128, 1, dtype=torch.bfloat16, device='cuda')
# V = first 16 rows of identity, MN-major
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')[:16, :]
v = v.as_strided((16, 128), (1, 16)).unsqueeze(-1)
c = torch.zeros(m, 16, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# V=I[:16,:] → P@V = P[:,:16] (first 16 cols of (Q@K^T).bf16())
ref = (qf @ kf.T).bfloat16()[:, :16].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 = PvDiag16V5Kernel(mma_tiler_mn=(128, 128))
print('Compiling v5...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running v5...', 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'(128,16) PV MMA v5: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if cos < 0.5:
print(f' out max={out.max().item():.4f} min={out.min().item():.4f}')
print(f' ref max={ref.max().item():.4f} min={ref.min().item():.4f}')
print(f' out[:4,:4]:\n{out[:4,:4]}')
print(f' ref[:4,:4]:\n{ref[:4,:4]}')
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,400 @@
"""
Stage B N-tiling v6: Run BOTH (128,128) and (128,16) PV MMAs in the same kernel.
Use the working (128,128) epilogue for output. Compare first 16 columns.
This tests whether the (128,16) PV MMA's C→A TMEM bridge is correct
without needing a (128,16) output epilogue (which is broken).
The (128,16) result is written to a separate TMEM region, then
manually padded and copied to the (128,128) output buffer.
"""
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 DualPvKernel:
"""Run QK → softmax → PV(128,128) and PV(128,16) side by side.
Output PV(128,128) result using working epilogue_tma_store.
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
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
def _setup(self, qk_mma, pv128_mma, pv16_mma):
qk_inst_k = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV(128,128): same as v30
self.pv128_mma_tiler = (self.qk_mma_tiler[0], self.qk_mma_tiler[1], self.qk_mma_tiler[1])
# PV(128,16): N-tiling with N=16
self.pv16_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
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.qk_mma_tiler[1], 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.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(pv128_mma, self.pv128_mma_tiler, self.q_dtype, 1)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv128_mma, self.pv128_mma_tiler, self.q_dtype, 1)
self.p16_tmem_s = utils.sm100.make_smem_layout_a(pv16_mma, self.pv16_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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv128_thr = pv128_mma.get_slice(0)
pv128_acc_shape = pv128_thr.partition_shape_C(self.pv128_mma_tiler[:2])
tOtO = pv128_thr.make_fragment_C(pv128_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
# PV(128,16) accumulator
pv16_thr = pv16_mma.get_slice(0)
pv16_acc_shape = pv16_thr.partition_shape_C(self.pv16_mma_tiler[:2])
tOtO16 = pv16_thr.make_fragment_C(pv16_acc_shape)
o16_cols = find_tmem_tensor_col_offset(tOtO16)
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 = s_cols # PV(128,128) accumulator
self.tmem_o16_offset = o_cols # PV(128,16) accumulator (separate TMEM)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv128_mma.make_fragment_C(cute.append(pv128_acc_shape, self.num_acc_stage))
tCtO16_fake = pv16_mma.make_fragment_C(cute.append(pv16_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake, tCtO16_fake], arch="sm_100")
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
pv128_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, self.mma_tiler_mn, tcgen05.OperandSource.TMEM)
pv16_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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv128_mma, pv16_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = cute.nvgpu.make_tiled_tma_atom_B(
utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn, pv128_mma.thr_id),
v, v_smem, self.pv128_mma_tiler, pv128_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv128_mma, pv16_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, self.cluster_layout_vmnk,
self.a_smem_s, self.b_smem_s, self.v_smem_s, self.p_tmem_s, self.p16_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, pv128_mma, pv16_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, p16_tmem_s, c_smem_s, epi_tile):
warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
tidx, _, _ = cute.arch.thread_idx()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.qk_mma_tiler, (None,None,0)), (None,None,None))
k_cnt = cute.size(gQ, mode=[3])
qk_thr = qk_mma.get_slice(0)
pv128_thr = pv128_mma.get_slice(0)
pv16_thr = pv16_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv128_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv128_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv128_mma.make_fragment_B(sV)
# QK accumulator
qk_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
# PV(128,128) accumulator
pv128_acc_shape = pv128_thr.partition_shape_C(self.pv128_mma_tiler[:2])
tOtO = pv128_thr.make_fragment_C(pv128_acc_shape)
tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout)
# PV(128,16) accumulator
pv16_acc_shape = pv16_thr.partition_shape_C(self.pv16_mma_tiler[:2])
tOtO16 = pv16_thr.make_fragment_C(pv16_acc_shape)
tOtO16_0 = cute.make_tensor(tOtO16.iterator + self.tmem_o16_offset, tOtO16.layout)
# P A-fragment for PV(128,128) — reads from softmax P in TMEM
tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
tOrP_base = pv128_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)
# P A-fragment for PV(128,16) — reads from the SAME softmax P in TMEM
tP16 = cute.make_tensor(tStS.iterator, p16_tmem_s.outer)
tOrP16_base = pv16_thr.make_fragment_A(tP16)
tOrP16 = tOrP16_base[(None, None, None, 0)]
tOrP16_0 = cute.make_tensor(
tOrP16.iterator + self.qk_acc_dtype.width // self.q_dtype.width * self.tmem_p0_offset,
tOrP16.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv128_mma.make_fragment_C(cute.append(pv128_acc_shape, self.num_acc_stage))
tCtO16_fake = pv16_mma.make_fragment_C(cute.append(pv16_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
# QK MMA
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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(128,128) MMA: P @ V=I → O = P
pv128_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, unroll_full=True):
cute.gemm(pv128_mma, tOtO0, tOrP0[(None,None,kb)], tCrV_s[(None,None,kb)], tOtO0)
pv128_mma.set(tcgen05.Field.ACCUMULATE, True)
# PV(128,16) MMA: P @ V[:16,:] → O16 = P[:,:16]
# Note: PV(128,16) reads P from the same TMEM location as PV(128,128)
# but the A-fragment is for (128,16) PV MMA which has different layout
pv16_mma.set(tcgen05.Field.ACCUMULATE, False)
for kb in cutlass.range(cute.size(tOrP16_0, mode=[2]), unroll_full=True):
cute.gemm(pv16_mma, tOtO16_0, tOrP16_0[(None,None,kb)], tCrV_s[(None,None,kb)], tOtO16_0)
pv16_mma.set(tcgen05.Field.ACCUMULATE, True)
acc_pipe.producer_commit(acc_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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))
# Softmax packing: FP32 S → BF16 P
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()
# Output: PV(128,128) result using working 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 = 128, 128
q = torch.randn(m, 128, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, 128, 1, dtype=torch.bfloat16, device='cuda')
# V = identity (128,128) MN-major — same as v30
v = torch.eye(128, dtype=torch.bfloat16, device='cuda')
v = v.as_strided((128, 128), (1, 128)).unsqueeze(-1)
c = torch.zeros(m, 128, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# V=I → O = P = (Q@K^T).bf16()
ref = (qf @ kf.T).bfloat16().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 = DualPvKernel(mma_tiler_mn=(128, 128))
print('Compiling dual-PV test...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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'PV(128,128) output: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
# If PV(128,128) still works, the QK+softmax pipeline is fine
# The PV(128,16) result is in TMEM at tmem_o16_offset but not yet output
# We need to read it to compare
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,376 @@
"""
Stage B N-tiling v7: PV(128,16) single N-tile test (head_dim=16).
Based on test_pv_diag (128,128 PV) with minimal changes:
- PV MMA tiler: (128, 16, 128) instead of (128, 128, 128)
- V: (16, 128) identity MN-major instead of (128, 128)
- Output: (128, 16) instead of (128, 128)
- epi_tile uses PV's cta_tile_shape_mnk for compute_epilogue_tile_shape
With V=I(16,128): O = P[:, :16] = (Q@K^T).bf16()[:, :16]
"""
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 NTileV7Kernel:
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
# PV with N=16 (single N-tile, head_dim=16)
self.pv_mma_tiler = (self.qk_mma_tiler[0], self.qk_mma_tiler[1], self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 128, 128) — same as QK, but PV MMA is (128, 16) with TMEM source
# Wait, pv_mma_tiler must match the PV MMA shape, not QK's.
# For PV(128,16): pv_mma_tiler = (M_pv, N_pv, K_pv) = (128, 16, 128)
# In fmha: pv_mma_tiler = (qk_mma_tiler[0], qk_mma_tiler[2], qk_mma_tiler[1])
# = (128, inst_k*4, 128) — but we want N=16, not inst_k*4.
# The pv_mma_tiler is used for V SMEM layout and TMA, NOT for the MMA atom.
# The actual MMA atom is (128, 16) set in make_trivial_tiled_mma.
# pv_mma_tiler must match: (128, 16, 128) for the B-operand to tile V as (16, 128).
self.pv_mma_tiler = (self.qk_mma_tiler[0], 16, self.qk_mma_tiler[1])
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
# epi_tile must match PV's output shape (128, 16)
self.pv_cta_tile_shape_mnk = (
self.pv_mma_tiler[0] // cute.size(qk_mma.thr_id.shape),
self.pv_mma_tiler[1], self.pv_mma_tiler[2])
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
self.pv_cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with (128, 16) output, P from TMEM
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.qk_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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I(16,128) → O = P[:, :16]
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output epilogue — use PV's cta_tile for the 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)
# Must use PV's cta_tile_shape_mnk for the epilogue
saved = self.cta_tile_shape_mnk
self.cta_tile_shape_mnk = self.pv_cta_tile_shape_mnk
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)
self.cta_tile_shape_mnk = saved
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
def test():
torch.manual_seed(42)
m, n, head_dim = 128, 128, 16
q = torch.randn(m, 128, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, 128, 1, dtype=torch.bfloat16, device='cuda')
# V = identity (16, 128) MN-major: strides (1, 16)
v_data = torch.zeros(16 * 128, dtype=torch.bfloat16, device='cuda')
for i in range(16):
v_data[i + 16 * i] = 1.0
v = torch.as_strided(v_data, (16, 128), (1, 16)).unsqueeze(-1)
c = torch.zeros(m, 16, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# With V=I(16,128) MN-major: O = P[:, :16] = (Q@K^T).bf16()[:, :16]
ref = (qf @ kf.T).bfloat16().float()
ref_out = ref[:, :16]
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 = NTileV7Kernel(mma_tiler_mn=(128, 128))
print('Compiling PV(128,16) single-tile test...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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_out.flatten().unsqueeze(0)).item()
print('PV(128,16) single-tile: cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if cos < 0.99:
print(' out[0,:8] =', out[0,:8].tolist())
print(' ref[0,:8] =', ref_out[0,:8].tolist())
if __name__ == '__main__':
test()

View File

@@ -0,0 +1,387 @@
"""
Minimal PV-only test: Load P from GMEM to TMEM via QK-style MMA, then PV from TMEM.
Step 1: QK MMA writes FP32 S to TMEM (we know this works)
Step 2: Softmax packing writes BF16 P to TMEM (test this)
Step 3: PV MMA reads BF16 P from TMEM and V from SMEM, produces O
But to isolate the bug, let me test just the PV MMA in isolation.
I'll write known BF16 values to TMEM using the softmax packing path,
then immediately read them back using the PV A-fragment path,
and compare.
Actually, the simplest isolation test:
1. Do QK MMA to get S in TMEM (cosine 0.999999 verified)
2. Do softmax packing: S → P in TMEM (at offset 32)
3. Skip PV entirely — read P from TMEM using the C-fragment composition LOAD path
4. Output P to GMEM and compare against S.to(BF16)
This tests whether the softmax packing writes P correctly to the same TMEM
that the PV would read from.
But we can't easily read P from TMEM using the standard epilogue path
because the epilogue expects FP32 accumulator data.
Alternative: Use the PV MMA with V=I (identity). If P is correct,
then P @ I = P. But V needs to be MN-major and (128, 128), not (128, 64).
The output would be (128, 128) which doesn't match our (128, 64) c tensor.
Let me use V that selects the first 64 columns: V[k, n] = delta(k, n) for k in [0,63].
This gives P @ V = P[:, :64], and the output is (128, 64).
But V is (128, 128) in the MMA K,N dims. V[k, n] for k in [0,127], n in [0,63].
Hmm, this is getting complicated. Let me just do the identity approach with a (128, 128) output.
"""
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 NTileV8Kernel:
"""QK + softmax packing + PV with V=I to isolate PV MMA correctness.
Output should be P = S.to(BF16), i.e. (Q@K^T).bfloat16()
With V=I, O = P @ I = P.
But V is (K=128, N=128) in the MMA. We need a 128x128 identity in MN-major.
Output tensor is (128, 128).
"""
def __init__(self, mma_tiler_mn):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
self.use_2cta_instrs = False # needed by epilogue_tma_store
self.epilog_sync_bar_id = 1 # needed by epilogue_tma_store
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
def _setup(self, qk_mma, pv_mma):
qk_inst_k = int(cute.size(qk_mma.shape_mnk, mode=[2]))
self.qk_mma_tiler = (*self.mma_tiler_mn, int(qk_inst_k * 4))
# PV with V=I: output is (128, 128), same as QK
self.pv_mma_tiler = (self.qk_mma_tiler[0], qk_inst_k, self.qk_mma_tiler[1])
# pv_mma_tiler = (128, 16, 128) FMHA: (qk_M, qk_K_inst, qk_N)
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.qk_mma_tiler[1], self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
pv_cta_tile = (self.pv_mma_tiler[0], self.pv_mma_tiler[1], self.pv_mma_tiler[2])
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
pv_cta_tile, False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
s_cols = find_tmem_tensor_col_offset(tStS)
pv_thr = pv_mma.get_slice(0)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
o_cols = find_tmem_tensor_col_offset(tOtO)
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 = s_cols
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")
# ⛔⛔⛔ CRITICAL: num_tma_load_bytes MUST include ALL TMA-loaded tensors (Q + K + V). Missing V → DEADLOCK. See FOOTGUN #0 in README.
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
self.num_tma_load_bytes = (
cute.size_in_bytes(self.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem) +
cute.size_in_bytes(self.q_dtype, v_smem)
) * 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, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
# PV with 128x128 output (V=I)
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, 16), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))
tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = 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_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)
epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)
self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
tma_c, tma_tc, 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()
use_2cta = cute.size(qk_mma.thr_id.shape) == 2
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) * (2 if use_2cta else 1)),
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=use_2cta,
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))
gC = cute.local_tile(mC, cute.slice_(self.pv_mma_tiler, (None,0,0)), (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); tCgC = qk_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))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
tCgV = pv_thr.partition_B(gV)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
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_acc_shape = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_acc_shape)
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_acc_shape)
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_acc_shape, self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD WARP ═══
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 WARP ═══
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_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
for kt in range(k_cnt):
h = ab_c.wait_and_advance(peek)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, 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: P @ V where V=I → O = P
pv_mma.set(tcgen05.Field.ACCUMULATE, False)
tCrV_s = tCrV[(None, None, None, 0)]
nblk_pv = cute.size(tOrP0, mode=[2])
for kb in cutlass.range(nblk_pv, 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_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ EPILOGUE WARPS ═══
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()
# Output 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)
# Swap cta_tile to PV's for epilogue
saved_cta_tile = self.cta_tile_shape_mnk
self.cta_tile_shape_mnk = (int(self.pv_mma_tiler[0]), int(self.pv_mma_tiler[1]), int(self.pv_mma_tiler[2]))
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)
self.cta_tile_shape_mnk = saved_cta_tile
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
def test():
torch.manual_seed(42)
m, n, head_dim = 128, 128, 64
q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
# V = identity[:16,:] MN-major: (16,128) with strides (1,16)
v_data = torch.eye(128, dtype=torch.bfloat16, device='cuda')
v = v_data[:16, :].as_strided((16, 128), (1, 16)).unsqueeze(-1)
c = torch.zeros(m, 16, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
# V=I[:16,:] -> O = P[:,:16] = (Q@K^T).bf16()[:,:16]
ref = (qf @ kf.T).bfloat16().float()[:, :16]
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 = NTileV8Kernel(mma_tiler_mn=(128, 128))
print('Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print('Running...', 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('PV(128,16) v8: cosine {:.6f} {}'.format(cos, 'PASS' if cos >= 0.99 else 'FAIL'))
if __name__ == '__main__':
test()