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

FMHA v1: pv_mma_tiler=(128,64,128) works with V=I, fails with real V (SMEM layout bug)

Browse Source
This commit is contained in:
biondizzle
2026-05-21 10:47:46 +00:00
parent 750f1f09c9
commit d72f854efb
3 changed files with 850 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

354
tests/test_fmha_pipeline.py Normal file
View File

@@ -0,0 +1,354 @@
"""
Stage B — FMHA-style KV-tile interleaved attention kernel.
Following CUTLASS FMHA reference architecture:
- Q: (seq_q, head_dim) — loaded once
- K, V: tiled over sequence dimension, V overwrites K in SMEM (FMHA trick)
- For each KV-tile:
1. TMA load K[tile] into sK SMEM
2. QK MMA: sQ @ sK^T → S in TMEM
3. Softmax: S → P in TMEM (with online softmax rescaling of O in TMEM)
4. V overwrites sK SMEM (after QK, K no longer needed)
5. PV MMA: P @ sV → O in TMEM (accumulate)
- Epilogue: divide O by row_sum, store to GMEM
This properly handles non-(128,128) PV because V SMEM always has the correct
data for the current KV-tile — it's loaded right before PV, not stale from
the beginning.
Warp layout:
Warp 0-3: Softmax (4 warps)
Warp 4: MMA
Warp 5: TMA load
"""
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 FmhaPipelineKernel:
def __init__(self, qk_mma_tiler, pv_mma_tiler):
self.acc_dtype = Float32
self.qk_acc_dtype = Float32
self.q_dtype = BFloat16
self.o_dtype = BFloat16
self.c_dtype = BFloat16
self.qk_mma_tiler = qk_mma_tiler
self.pv_mma_tiler = pv_mma_tiler
self.use_2cta_instrs = False
self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1)
self.cta_group = tcgen05.CtaGroup.ONE
self.softmax_warp_ids = (0, 1, 2, 3)
self.mma_warp_id = 4
self.tma_warp_id = 5
self.threads_per_cta = 192
self.kv_stage = 2 # double-buffered KV
self.q_stage = 1
def _setup(self, qk_mma, pv_mma):
qk_inst_k = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (*self.qk_mma_tiler[:2], qk_inst_k * 4)
pv_inst_k = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (*self.pv_mma_tiler[:2], pv_inst_k * 4)
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.epi_tile = self.pv_mma_tiler[:2]
self.cta_tile_shape_mnk = (
self.qk_mma_tiler[0] // cute.size(qk_mma.thr_id.shape),
self.pv_mma_tiler[1],
self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.num_ab_stage = 1
self.num_acc_stage = 1
self.num_c_stage = 2
self.q_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.qk_mma_tiler, self.q_dtype, self.q_stage)
self.k_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.qk_mma_tiler, self.q_dtype, self.kv_stage)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_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, self.kv_stage)
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.tmem_s0_offset = 0
self.tmem_p0_offset = 32
self.tmem_o0_offset = o_cols
self.tilePlikeFP32 = self.qk_mma_tiler[1] // Float32.width * self.o_dtype.width
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.q_smem_s, (None, None, None, 0))
b_smem = cute.slice_(self.k_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(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.qk_mma_tiler[:2],
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, self.pv_mma_tiler[:2],
tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_smem = cute.slice_(self.q_smem_s, (None, None, None, 0))
k_smem = cute.slice_(self.k_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.qk_mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_k, tma_tk = cute.nvgpu.make_tma_atom_B(
utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn, qk_mma.thr_id),
k, k_smem, self.qk_mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
tma_v, tma_tv = cute.nvgpu.make_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.q_smem_s, self.k_smem_s, self.v_smem_s, self.p_tmem_s, self.c_smem_s, self.epi_tile
).launch(grid=(1, 1, 1), block=[self.threads_per_cta, 1, 1], stream=stream)
@cute.kernel
def _kernel(self, qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV,
tma_c, mC, cl_vmnk, q_smem_s, k_smem_s, v_smem_s, p_tmem_s, c_smem_s, epi_tile):
warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
tidx, _, _ = cute.arch.thread_idx()
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:
q_bar: cute.struct.MemRange[cutlass.Int64, self.q_stage * 2]
kv_bar: cute.struct.MemRange[cutlass.Int64, self.kv_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)
q_prod, q_cons = pipeline.PipelineTmaUmma.create(
barrier_storage=st.q_bar.data_ptr(), num_stages=self.q_stage,
producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),
consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 1),
tx_count=self.num_tma_load_bytes, cta_layout_vmnk=cl_vmnk, defer_sync=True
).make_participants()
kv_prod, kv_cons = pipeline.PipelineTmaUmma.create(
barrier_storage=st.kv_bar.data_ptr(), num_stages=self.kv_stage,
producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),
consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 1),
tx_count=self.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.softmax_warp_ids)),
).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.softmax_warp_ids) * (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.softmax_warp_ids)))
tmem = utils.TmemAllocator(
st.holding.ptr, barrier_for_retrieve=tmem_bar,
allocator_warp_id=self.softmax_warp_ids[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=q_smem_s.outer, byte_alignment=128, swizzle=q_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype, layout=k_smem_s.outer, byte_alignment=128, swizzle=k_smem_s.inner)
# V overwrites K SMEM (FMHA trick)
sV_ptr = cute.recast_ptr(sK.iterator, v_smem_s.inner)
sV = cute.make_tensor(sV_ptr, v_smem_s.outer)
sC = smem.allocate_tensor(element_type=self.o_dtype, layout=c_smem_s.outer, byte_alignment=128, swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ, cute.slice_(self.qk_mma_tiler, (None, 0, None)), (None, None, None))
gK = cute.local_tile(mK, cute.slice_(self.qk_mma_tiler, (0, None, None)), (None, None, None))
gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0, None, None)), (None, None, None))
gC = cute.local_tile(mC, cute.slice_(self.pv_mma_tiler, (None, None, 0)), (None, None, None))
n_kv_tiles = cute.size(gK, mode=[3])
qk_thr = qk_mma.get_slice(0)
pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ)
tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV)
tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk, (0, 0, None, 0)).shape)
tAsQ, tAgQ = cpasync.tma_partition(tma_q, 0, a_lay, cute.group_modes(sQ, 0, 3), cute.group_modes(tCgQ, 0, 3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk, (0, None, 0, 0)).shape)
tBsK, tBgK = cpasync.tma_partition(tma_k, 0, b_lay, cute.group_modes(sK, 0, 3), cute.group_modes(tCgK, 0, 3))
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV, 0, 3), cute.group_modes(tCgV, 0, 3))
tAgQ = tAgQ[(None, 0, None, 0)]
tBgK = tBgK[(None, 0, None, 0)]
tVgV = tVgV[(None, 0, None, 0)]
tCrQ = qk_mma.make_fragment_A(sQ)
tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_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:
# Load Q once
q_prod.reset()
qh = q_prod.acquire_and_advance()
cute.copy(tma_q, tAgQ[(None, qh.count)], tAsQ[(None, qh.index)], tma_bar_ptr=qh.barrier)
q_prod.tail()
# Load KV tiles: for each tile, load K then V
# K and V share SMEM, so V overwrites K after QK consumes it
kv_prod.reset()
peek = kv_prod.try_acquire()
for kt in cutlass.range(n_kv_tiles, unroll=1):
# Load K[tile]
kvh = kv_prod.acquire_and_advance(peek)
cute.copy(tma_k, tBgK[(None, kvh.count)], tBsK[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
# Load V[tile] into the SAME SMEM (overwrites K after QK)
# Wait — we need QK to finish before V overwrites K.
# FMHA uses a SEPARATE pipeline entry for V. The MMA warp
# consumes K first (QK), then V (PV). The pipeline ordering
# ensures V doesn't overwrite K before QK is done.
cute.copy(tma_v, tVgV[(None, kvh.count)], tVsV[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
peek = cutlass.Boolean(1)
if kvh.count + 1 < 2 * n_kv_tiles:
peek = kv_prod.try_acquire()
kv_prod.tail()
# ═══ MMA WARP ═══
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
q_cons.reset()
qh = q_cons.wait_and_advance()
qh.release()
kv_cons.reset()
peek = kv_cons.try_wait()
acc_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
acc_pipe.producer_acquire(acc_prod_st)
for kt in range(n_kv_tiles):
# Wait for K[tile]
kvh = kv_cons.wait_and_advance(peek)
peek = cutlass.Boolean(1)
# ─── QK: Q @ K[tile]^T → S ───
s0_handle = mma_si_prod.acquire_and_advance()
qk_mma.set(tcgen05.Field.ACCUMULATE, kt != 0)
nblk = cute.size(tCrQ, mode=[2])
for kb in cutlass.range(nblk, unroll_full=True):
cute.gemm(qk_mma, tStS0,
tCrQ[(None, None, kb, 0)],
tCrK[(None, None, kb, kvh.index)],
tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
cute.arch.fence_view_async_tmem_store()
s0_handle.commit()
# ─── Wait for softmax: S → P done ───
s0_handle = mma_si_prod.acquire_and_advance()
# ─── Wait for V[tile] ───
vvh = kv_cons.wait_and_advance(peek)
peek = cutlass.Boolean(1)
# ─── PV: P @ V[tile] → O ───
pv_mma.set(tcgen05.Field.ACCUMULATE, kt != 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[(None, None, kb, vvh.index)],
tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
kvh.release()
vvh.release()
acc_pipe.producer_commit(acc_prod_st)
acc_prod_st.advance()
acc_pipe.producer_tail(acc_prod_st)
# ═══ SOFTMAX 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.softmax_warp_ids))
tmem_load_atom = cute.make_copy_atom(
tcgen05.copy.Ld32x32bOp(tcgen

251
tests/test_fmha_v1.py Normal file
View File

@@ -0,0 +1,251 @@
"""
FMHA Pipeline v1: Modified v30 with pv_mma_tiler=(128,64) for head_dim=64.
Step 1: Test if the V SMEM layout works for head_dim=64 PV.
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
HEAD_DIM = 64
class PvHeadDimKernel:
def __init__(self):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
# PV tiler: (M, N, K) = (QK_M, head_dim, QK_N)
# K of PV = sequence per tile = QK's N = 128
pv_nphases = 128 // pv_ik # 128/16 = 8 k-phases
self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * pv_nphases) # (128, 64, 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,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
self.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)
# Diagnostic: print V SMEM size
v_s = cute.slice_(self.v_smem_s, (None,None,None,0))
v_sz = cute.size_in_bytes(self.q_dtype, v_s)
print(f"[DIAG] pv_mma_tiler={self.pv_mma_tiler} V SMEM per stage: {v_sz} bytes ({v_sz//2} BF16)")
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tilePlikeFP32 = self.qk_mma_tiler[1] // Float32.width * self.o_dtype.width
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
a = cute.slice_(self.a_smem_s,(None,None,None,0)); b = cute.slice_(self.b_smem_s,(None,None,None,0))
vs = cute.slice_(self.v_smem_s,(None,None,None,0))
self.num_tma_load_bytes = (cute.size_in_bytes(self.q_dtype,a)+cute.size_in_bytes(self.q_dtype,b)+cute.size_in_bytes(self.q_dtype,vs))*cute.size(qk_mma.thr_id.shape)
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
self.v_major = LayoutEnum.from_tensor(v).mma_major_mode()
self.c_layout = LayoutEnum.from_tensor(c)
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.qk_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.a_smem_s,(None,None,None,0)); k_s = cute.slice_(self.b_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_v,mV = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,pv_mma.thr_id),v,v_s,self.pv_mma_tiler,pv_mma,self.cluster_layout_vmnk.shape)
epi_s = cute.select(self.c_smem_s,mode=[0,1])
tma_c,mC = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(),c,epi_s,self.epi_tile)
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.a_smem_s,self.b_smem_s,self.v_smem_s,self.p_tmem_s,self.c_smem_s,self.epi_tile).launch(grid=(1,1,1),block=[self.threads_per_cta,1,1],stream=stream)
@cute.kernel
def _kernel(self, qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV, tma_c, mC, cl_vmnk, a_smem_s, b_smem_s, v_smem_s, p_tmem_s, c_smem_s, epi_tile):
warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
tidx,_,_ = cute.arch.thread_idx()
if warp_idx == self.tma_warp_id:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k)
cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
ab_bar: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage*2]
mma_si_bar: cute.struct.MemRange[cutlass.Int64, 2]
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
smem = utils.SmemAllocator(); st = smem.allocate(SS)
ab_p,ab_c = pipeline.PipelineTmaUmma.create(barrier_storage=st.ab_bar.data_ptr(),num_stages=self.num_ab_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.num_tma_load_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
mma_si_prod,mma_si_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_si_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)*(2 if cute.size(qk_mma.thr_id.shape)==2 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=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=a_smem_s.outer,byte_alignment=128,swizzle=a_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=b_smem_s.outer,byte_alignment=128,swizzle=b_smem_s.inner)
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
gK = cute.local_tile(mK,cute.slice_(self.qk_mma_tiler,(0,None,None)),(None,None,None))
gC = cute.local_tile(mC,cute.slice_(self.pv_mma_tiler,(None,None,0)),(None,None,None))
k_cnt = cute.size(gQ,mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK); 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)]
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_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator+self.tmem_s0_offset,tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator+self.tmem_o0_offset,tOtO.layout)
tP = cute.make_tensor(tStS.iterator,p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(tOrP.iterator+self.qk_acc_dtype.width//self.q_dtype.width*self.tmem_p0_offset,tOrP.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as,self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as,self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
if warp_idx == self.tma_warp_id:
ab_p.reset(); peek = ab_p.try_acquire()
for kt in cutlass.range(k_cnt,unroll=1):
h = ab_p.acquire_and_advance(peek)
cute.copy(tma_q,tAgQ[(None,h.count)],tAsQ[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_k,tBgK[(None,h.count)],tBsK[(None,h.index)],tma_bar_ptr=h.barrier)
cute.copy(tma_v,tVgV[(None,h.count)],tVsV[(None,h.index)],tma_bar_ptr=h.barrier)
peek = cutlass.Boolean(1)
if h.count+1<k_cnt: peek = ab_p.try_acquire()
ab_p.tail()
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
ab_c.reset(); peek = ab_c.try_wait()
s0_handle = mma_si_prod.acquire_and_advance()
acc_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)
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()
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)
# Full softmax attention with real V
for n in [128]:
m, hd = 128, HEAD_DIM
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
# V in MN-major for PV's B-operand: (K=n, N=hd)
v = torch.ones(n, hd, 1, dtype=torch.bfloat16, device='cuda')
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float(); vf = v[:,:,0].float()
ref = (qf @ kf.T).bfloat16().float() @ vf # identity softmax: S.bf16() @ V
ref = ref.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 = PvHeadDimKernel()
print(f'n={n}: Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print(f'n={n}: 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 head_dim={HEAD_DIM} n={n} softmax: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if cos < 0.99:
print(f' out[0,:4] = {out[0,:4].tolist()}')
print(f' ref[0,:4] = {ref[0,:4].tolist()}')
if __name__ == '__main__':
test()

245
tests/test_fmha_v2.py Normal file
View File

@@ -0,0 +1,245 @@
"""
FMHA Pipeline v2: KV-tile interleaved with V overwriting K in SMEM.
K and V use separate pipeline entries from the same kv pipeline.
MMA: wait for K → QK → wait for softmax → wait for V → PV per KV-tile.
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct
HEAD_DIM = 64
class FmhaKernel:
def __init__(self):
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
self.threads_per_cta = 192; self.num_c_stage = 2
self.kv_stage = 2; self.q_stage = 1
def _setup(self, qk_mma, pv_mma):
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
self.qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik))
self.mma_tiler = self.qk_mma_tiler
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2])
self.c_layout = LayoutEnum.ROW_MAJOR
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
self.num_ab_stage = 1; self.num_acc_stage = 1
self.q_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.qk_mma_tiler, self.q_dtype, self.q_stage)
self.k_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.qk_mma_tiler, self.q_dtype, self.kv_stage)
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, self.kv_stage)
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
self.tilePlikeFP32 = self.qk_mma_tiler[1] // Float32.width * self.o_dtype.width
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
self.tmem_o0_offset = find_tmem_tensor_col_offset(tOtO)
tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
tCO = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS, tCO], arch="sm_100")
a = cute.slice_(self.q_smem_s,(None,None,None,0)); b = cute.slice_(self.k_smem_s,(None,None,None,0))
self.num_tma_load_bytes = (cute.size_in_bytes(self.q_dtype,a)+cute.size_in_bytes(self.q_dtype,b))*cute.size(qk_mma.thr_id.shape)
@cute.jit
def __call__(self, q, k, v, c, stream):
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
self.v_major = LayoutEnum.from_tensor(v).mma_major_mode()
self.c_layout = LayoutEnum.from_tensor(c)
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.qk_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)
self._setup(qk_mma, pv_mma)
q_s = cute.slice_(self.q_smem_s,(None,None,None,0))
k_s = cute.slice_(self.k_smem_s,(None,None,None,0))
v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape)
tma_v,mV = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,pv_mma.thr_id),v,v_s,self.pv_mma_tiler,pv_mma,self.cluster_layout_vmnk.shape)
epi_s = cute.select(self.c_smem_s,mode=[0,1])
tma_c,mC = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(),c,epi_s,self.epi_tile)
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.q_smem_s,self.k_smem_s,self.v_smem_s,self.p_tmem_s,self.c_smem_s,self.epi_tile).launch(grid=(1,1,1),block=[self.threads_per_cta,1,1],stream=stream)
@cute.kernel
def _kernel(self, qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV, tma_c, mC, cl_vmnk, q_smem_s, k_smem_s, v_smem_s, p_tmem_s, c_smem_s, epi_tile):
warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
tidx,_,_ = cute.arch.thread_idx()
if warp_idx == self.tma_warp_id:
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k)
cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
@cute.struct
class SS:
q_bar: cute.struct.MemRange[cutlass.Int64, self.q_stage*2]
kv_bar: cute.struct.MemRange[cutlass.Int64, self.kv_stage*2]
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)
qp,qc = pipeline.PipelineTmaUmma.create(barrier_storage=st.q_bar.data_ptr(),num_stages=self.q_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.num_tma_load_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
kvp,kvc = pipeline.PipelineTmaUmma.create(barrier_storage=st.kv_bar.data_ptr(),num_stages=self.kv_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.num_tma_load_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
mma_si_prod,mma_si_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_si_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=q_smem_s.outer,byte_alignment=128,swizzle=q_smem_s.inner)
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=k_smem_s.outer,byte_alignment=128,swizzle=k_smem_s.inner)
# V overwrites K SMEM
sV = cute.make_tensor(cute.recast_ptr(sK.iterator, v_smem_s.inner), v_smem_s.outer)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
gK = cute.local_tile(mK,cute.slice_(self.qk_mma_tiler,(0,None,None)),(None,None,None))
gV = cute.local_tile(mV,cute.slice_(self.pv_mma_tiler,(0,None,None)),(None,None,None))
gC = cute.local_tile(mC,cute.slice_(self.pv_mma_tiler,(None,None,0)),(None,None,None))
n_kv_tiles = cute.size(gK, mode=[3])
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
tCrV = pv_mma.make_fragment_B(sV)
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
tStS = qk_thr.make_fragment_C(qk_as)
tStS0 = cute.make_tensor(tStS.iterator+self.tmem_s0_offset,tStS.layout)
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
tOtO = pv_thr.make_fragment_C(pv_as)
tOtO0 = cute.make_tensor(tOtO.iterator+self.tmem_o0_offset,tOtO.layout)
tP = cute.make_tensor(tStS.iterator,p_tmem_s.outer)
tOrP_base = pv_thr.make_fragment_A(tP)
tOrP = tOrP_base[(None,None,None,0)]
tOrP0 = cute.make_tensor(tOrP.iterator+self.qk_acc_dtype.width//self.q_dtype.width*self.tmem_p0_offset,tOrP.layout)
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as,self.num_acc_stage))
tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as,self.num_acc_stage))
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ═══ TMA LOAD ═══
if warp_idx == self.tma_warp_id:
qp.reset(); qh = qp.acquire_and_advance()
cute.copy(tma_q,tAgQ[(None,qh.count)],tAsQ[(None,qh.index)],tma_bar_ptr=qh.barrier)
qp.tail()
kvp.reset(); pk = kvp.try_acquire()
for kt in cutlass.range(n_kv_tiles,unroll=1):
kh = kvp.acquire_and_advance(pk)
cute.copy(tma_k,tBgK[(None,kh.count)],tBsK[(None,kh.index)],tma_bar_ptr=kh.barrier)
vh = kvp.acquire_and_advance(cutlass.Boolean(1))
cute.copy(tma_v,tVgV[(None,vh.count)],tVsV[(None,vh.index)],tma_bar_ptr=vh.barrier)
pk = cutlass.Boolean(1)
kvp.tail()
# ═══ MMA ═══
if warp_idx == self.mma_warp_id:
tmem.wait_for_alloc()
qc.reset(); qh = qc.wait_and_advance(); qh.release()
kvc.reset(); pk = kvc.try_wait()
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer,self.num_acc_stage)
acc_pipe.producer_acquire(acc_st)
for kt in range(n_kv_tiles):
kh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)
sh = mma_si_prod.acquire_and_advance()
qk_mma.set(tcgen05.Field.ACCUMULATE, kt != 0)
for kb in cutlass.range(cute.size(tCrQ,mode=[2]),unroll_full=True):
cute.gemm(qk_mma,tStS0,tCrQ[(None,None,kb,0)],tCrK[(None,None,kb,kh.index)],tStS0)
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
cute.arch.fence_view_async_tmem_store(); sh.commit()
sh = mma_si_prod.acquire_and_advance() # wait softmax
vh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1) # wait V
pv_mma.set(tcgen05.Field.ACCUMULATE, kt != 0)
for kb in cutlass.range(cute.size(tOrP0,mode=[2]),unroll_full=True):
cute.gemm(pv_mma,tOtO0,tOrP0[(None,None,kb)],tCrV[(None,None,kb,vh.index)],tOtO0)
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
kh.release(); vh.release()
acc_pipe.producer_commit(acc_st); acc_st.advance(); acc_pipe.producer_tail(acc_st)
# ═══ EPILOGUE ═══
if warp_idx < self.mma_warp_id:
tmem.allocate(self.num_tmem_alloc_cols)
tmem.wait_for_alloc()
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
sfw_idx = tidx % (32 * len(self.epilogue_warp_id))
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)
for kt in range(n_kv_tiles):
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()
tCtO_base = cute.make_tensor(tmem_ptr + self.tmem_o0_offset, tCtO_fake.layout)
acc_cons_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Consumer, self.num_acc_stage)
c_grp = pipeline.CooperativeGroup(pipeline.Agent.Thread, 32 * len(self.epilogue_warp_id))
c_pipe = pipeline.PipelineTmaStore.create(num_stages=self.num_c_stage, producer_group=c_grp)
acc_cons_st = utils.gemm.sm100.epilogue_tma_store(self, tidx, warp_idx, tma_c, tCtO_base, sC, tCgC, epi_tile, 0, const_expr(lambda x: x), (0,0,0), acc_cons_st, acc_pipe, c_pipe)
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()
tmem.free(tmem_ptr)
def test():
torch.manual_seed(42)
for n in [128]:
m, hd = 128, HEAD_DIM
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.ones(n, hd, 1, dtype=torch.bfloat16, device='cuda')
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
qf = q[:,:,0].float(); kf = k[:,:,0].float()
ref = (qf @ kf.T).bfloat16().float() @ v[:,:,0].float()
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
mV = ct.from_dlpack(v).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v))
mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
kernel = FmhaKernel()
print(f'n={n}: Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
print(f'n={n}: 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'FMHA v2 n={n} V=ones: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if cos < 0.99:
print(f' out[0,:4]={out[0,:4].tolist()} ref[0,:4]={ref[0,:4].tolist()}')
if __name__ == '__main__':
test()