FMHA Stage-C multi-tile: Fix 1 (s_k=n), Fix 2 (TMA kt indexing), Fix 3 (O rescale)
Fix 1: s_k must equal actual n. With s_k < n, v_fmha layout only spans first s_k V tokens and TMA reads OOB on later tiles. Fix 2: TMA producer indexes K and V by kt (loop variable), NOT by the pipeline's interleaved count. The kv pipeline interleaves K and V, so pipeline count goes 0,1,2,3 but GMEM tiles should be K[0],V[0],K[1],V[1]. Fix 3: Online O rescale before softmax_done_bar. When row_max grows, O must be multiplied by exp2(old_max - new_max) before MMA starts next PV.
This commit is contained in:
473
tests/fmha_v3_stage_c_example1.py
Normal file
473
tests/fmha_v3_stage_c_example1.py
Normal file
@@ -0,0 +1,473 @@
|
||||
"""
|
||||
FMHA v3 Stage-C Multi-Tile: real online softmax + O normalization across N KV tiles.
|
||||
|
||||
Builds on the working single-tile Stage-C (cos 0.993 at n=128) and fixes the three
|
||||
multi-tile bugs.
|
||||
|
||||
Fix 1 — caller passes s_k = actual sequence length n.
|
||||
The single-tile test hardcoded s_k=128 for both n=128 and n=256. v_fmha is
|
||||
reconstructed as (HEAD_DIM, s_k, 1) MN-major, so when s_k < n the layout
|
||||
only describes the first s_k V tokens and TMA reads OOB on later tiles.
|
||||
The kernel takes s_k from __init__; tests must pass s_k=n.
|
||||
|
||||
Fix 2 — TMA producer indexes K and V by kt, not by pipeline count.
|
||||
The kv pipeline interleaves K and V acquires, so the pipeline `count`
|
||||
returned by acquire_and_advance goes 0,1,2,3,... The original code did:
|
||||
|
||||
cute.copy(tma_k, tBgK[(None, kh.count)], tBsK[(None, kh.index)], ...)
|
||||
cute.copy(tma_v, tVgV[(None, vh.count)], tVsV[(None, vh.index)], ...)
|
||||
|
||||
so for kt=0 it loaded K[0] and V[1], for kt=1 it loaded K[2] and V[3].
|
||||
Single tile got away with V[1] because the V iteration dim has size 1
|
||||
(stride 0 in CuTe → indexing any value yields V[0]); multi-tile loses
|
||||
that degeneracy and the bug becomes real.
|
||||
|
||||
The SMEM ring buffer index (kh.index, vh.index) was already correct —
|
||||
consumers read with the same indices — so only the GMEM indexer changes.
|
||||
|
||||
Fix 3 — re-enable O rescale between KV tiles.
|
||||
Online softmax requires O *= exp2(old_max − new_max) when the row max
|
||||
grows. The single-tile file disables it during debug; that path is
|
||||
untaken for n_kv_tiles=1 so single tile still passes. For multi-tile we
|
||||
have to apply the rescale at the start of every kt > 0, **before** the
|
||||
softmax_done_bar arrival (so MMA's next PV reads the rescaled O).
|
||||
|
||||
Synchronization (why the per-tile O rescale is safe):
|
||||
softmax kt's body reads O written by MMA's PV[kt-1]. That PV happened
|
||||
earlier in MMA's iteration kt-1, then MMA did fence_view_async_tmem_store,
|
||||
then issued QK[kt], then sh.commit on the S pipe, then arrived on
|
||||
softmax_done_bar. Our s_cons.wait_and_advance at the top of softmax kt
|
||||
waits on that S commit, which acts as a memory ordering point — all
|
||||
MMA writes prior to the commit (including PV[kt-1]) are visible to us.
|
||||
The fence we add after the tmem_load_o gives thread-local ordering for
|
||||
the in-register data.
|
||||
"""
|
||||
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
|
||||
import math
|
||||
|
||||
HEAD_DIM = 64
|
||||
|
||||
|
||||
class FmhaV3StageCMulti:
|
||||
def __init__(self, s_k=128, scale_softmax=None):
|
||||
# s_k MUST equal the actual sequence length n. See Fix 1.
|
||||
self.s_k = s_k
|
||||
self.acc_dtype = Float32; self.qk_acc_dtype = Float32
|
||||
self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16
|
||||
self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
|
||||
self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE
|
||||
self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5
|
||||
self.threads_per_cta = 192; self.num_c_stage = 2
|
||||
self.kv_stage = 2; self.q_stage = 1; self.num_c_stage = 2
|
||||
self.scale_softmax = scale_softmax if scale_softmax is not None else 1.0 / math.sqrt(HEAD_DIM)
|
||||
self.scale_softmax_log2 = self.scale_softmax * math.log2(math.e)
|
||||
|
||||
def _setup(self, qk_mma, pv_mma):
|
||||
qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
|
||||
self.qk_mma_tiler = (128, 128, qk_ik * 4)
|
||||
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
|
||||
self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik))
|
||||
self.mma_tiler = self.qk_mma_tiler
|
||||
self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
|
||||
self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2])
|
||||
self.c_layout = LayoutEnum.ROW_MAJOR
|
||||
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype)
|
||||
self.num_ab_stage = 1; self.num_acc_stage = 1
|
||||
self.q_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.qk_mma_tiler, self.q_dtype, self.q_stage)
|
||||
self.k_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.qk_mma_tiler, self.q_dtype, self.kv_stage)
|
||||
self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, self.kv_stage)
|
||||
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
|
||||
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
|
||||
qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
|
||||
tStS = qk_thr.make_fragment_C(qk_as)
|
||||
pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
|
||||
tOtO = pv_thr.make_fragment_C(pv_as)
|
||||
# TMEM layout: S at 0..128, P at 32..96 (overlaps S, intentional FMHA pattern),
|
||||
# O after max(S_end, P_end) aligned to 32.
|
||||
self.tmem_s0_offset = 0; self.tmem_p0_offset = 32
|
||||
p_cols_fp32 = self.pv_mma_tiler[2] * self.q_dtype.width // self.qk_acc_dtype.width
|
||||
p_end = self.tmem_p0_offset + p_cols_fp32
|
||||
s_cols = self.qk_mma_tiler[1]
|
||||
o_after = max(s_cols, p_end)
|
||||
self.tmem_o0_offset = ((o_after + 31) // 32) * 32
|
||||
o_cols = find_tmem_tensor_col_offset(tOtO)
|
||||
total = self.tmem_o0_offset + o_cols
|
||||
self.num_tmem_alloc_cols = 1
|
||||
while self.num_tmem_alloc_cols < total:
|
||||
self.num_tmem_alloc_cols *= 2
|
||||
cta = cute.size(qk_mma.thr_id.shape)
|
||||
q_s = cute.slice_(self.q_smem_s,(None,None,None,0)); k_s = cute.slice_(self.k_smem_s,(None,None,None,0))
|
||||
self.q_tx_bytes = cute.size_in_bytes(self.q_dtype, q_s) * cta
|
||||
self.kv_tx_bytes = cute.size_in_bytes(self.q_dtype, k_s) * cta
|
||||
|
||||
@cute.jit
|
||||
def __call__(self, q, k, v, c, stream):
|
||||
self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
|
||||
self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
|
||||
self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
|
||||
# Fix 1: s_k must equal actual n. With s_k < n, this layout only spans
|
||||
# the first s_k V tokens and TMA reads past the end on later kv tiles.
|
||||
v_fmha = cute.make_tensor(
|
||||
v.iterator,
|
||||
cute.make_layout(
|
||||
(HEAD_DIM, self.s_k, 1),
|
||||
stride=(1, HEAD_DIM, HEAD_DIM * self.s_k),
|
||||
),
|
||||
)
|
||||
self.v_major = LayoutEnum.from_tensor(v_fmha).mma_major_mode()
|
||||
self.c_layout = LayoutEnum.from_tensor(c)
|
||||
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM)
|
||||
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.qk_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)
|
||||
self._setup(qk_mma, pv_mma)
|
||||
q_s = cute.slice_(self.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_fmha,v_s,self.pv_mma_tiler,pv_mma,self.cluster_layout_vmnk.shape)
|
||||
epi_s = cute.select(self.c_smem_s,mode=[0,1])
|
||||
tma_c,mC = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(),c,epi_s,self.epi_tile)
|
||||
self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.q_smem_s,self.k_smem_s,self.v_smem_s,self.p_tmem_s,self.c_smem_s,self.epi_tile).launch(grid=(1,1,1),block=[self.threads_per_cta,1,1],stream=stream)
|
||||
|
||||
@cute.kernel
|
||||
def _kernel(self, qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV, tma_c, mC, cl_vmnk, q_smem_s, k_smem_s, v_smem_s, p_tmem_s, c_smem_s, epi_tile):
|
||||
warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
|
||||
tidx,_,_ = cute.arch.thread_idx()
|
||||
if warp_idx == self.tma_warp_id:
|
||||
cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k); cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c)
|
||||
|
||||
@cute.struct
|
||||
class SS:
|
||||
q_bar: cute.struct.MemRange[cutlass.Int64, self.q_stage*2]
|
||||
kv_bar: cute.struct.MemRange[cutlass.Int64, self.kv_stage*2]
|
||||
s_bar: cute.struct.MemRange[cutlass.Int64, 2]
|
||||
acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2]
|
||||
tmem_dealloc: cutlass.Int64; holding: cutlass.Int32
|
||||
smem = utils.SmemAllocator(); st = smem.allocate(SS)
|
||||
|
||||
qp,qc = pipeline.PipelineTmaUmma.create(barrier_storage=st.q_bar.data_ptr(),num_stages=self.q_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.q_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
|
||||
kvp,kvc = pipeline.PipelineTmaUmma.create(barrier_storage=st.kv_bar.data_ptr(),num_stages=self.kv_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.kv_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants()
|
||||
s_prod,s_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.s_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants()
|
||||
softmax_done_bar = pipeline.NamedBarrier(barrier_id=3, num_threads=32 + 32*len(self.epilogue_warp_id))
|
||||
acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True)
|
||||
tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id)))
|
||||
tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
|
||||
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True)
|
||||
|
||||
sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=q_smem_s.outer,byte_alignment=128,swizzle=q_smem_s.inner)
|
||||
sK = smem.allocate_tensor(element_type=self.q_dtype,layout=k_smem_s.outer,byte_alignment=128,swizzle=k_smem_s.inner)
|
||||
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
|
||||
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
|
||||
|
||||
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
|
||||
gK = cute.local_tile(mK,cute.slice_(self.qk_mma_tiler,(0,None,None)),(None,None,None))
|
||||
gV = cute.local_tile(mV,cute.slice_(self.pv_mma_tiler,(0,None,None)),(None,None,None))
|
||||
gC = cute.local_tile(mC,cute.slice_(self.pv_mma_tiler,(None,None,0)),(None,None,None))
|
||||
n_kv_tiles = cute.size(gK, mode=[3])
|
||||
|
||||
qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0)
|
||||
tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK)
|
||||
tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC)
|
||||
a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape)
|
||||
tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3))
|
||||
b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape)
|
||||
tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3))
|
||||
tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3))
|
||||
tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)]
|
||||
|
||||
tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)
|
||||
tCrV = pv_mma.make_fragment_B(sV)
|
||||
|
||||
qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2])
|
||||
tStS = qk_thr.make_fragment_C(qk_as)
|
||||
tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)
|
||||
pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
|
||||
tOtO = pv_thr.make_fragment_C(pv_as)
|
||||
tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout)
|
||||
|
||||
# PV reads P from TMEM (where softmax stores it). Same TMEM region as S but
|
||||
# at offset tmem_p0_offset; values are BF16 packed into FP32 words.
|
||||
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 warp =====
|
||||
# Fix 2: index K and V GMEM tensors by kt (the loop variable), NOT by the
|
||||
# pipeline's interleaved count. SMEM ring buffer index (kh.index/vh.index)
|
||||
# is still correct — the consumer uses the same indices.
|
||||
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)
|
||||
# GMEM tile: kt (correct K[kt]). SMEM slot: kh.index (ring buffer).
|
||||
cute.copy(tma_k, tBgK[(None, kt)], tBsK[(None, kh.index)], tma_bar_ptr=kh.barrier)
|
||||
pk = cutlass.Boolean(1)
|
||||
vh = kvp.acquire_and_advance(pk)
|
||||
# GMEM tile: kt (correct V[kt]). SMEM slot: vh.index (ring buffer).
|
||||
cute.copy(tma_v, tVgV[(None, kt)], tVsV[(None, vh.index)], tma_bar_ptr=vh.barrier)
|
||||
pk = cutlass.Boolean(1)
|
||||
kvp.tail()
|
||||
|
||||
# ===== MMA warp =====
|
||||
# Unchanged from single-tile. ACCUMULATE flag for PV is True for kt > 0
|
||||
# so O = sum_kt(P_kt @ V_kt). The softmax warps rescale O between tiles
|
||||
# before signaling softmax_done_bar, so by the time PV[kt] runs O has
|
||||
# already been multiplied by exp2(old_max - new_max).
|
||||
if warp_idx == self.mma_warp_id:
|
||||
tmem.wait_for_alloc()
|
||||
qc.reset(); qh = qc.wait_and_advance(); qh.release()
|
||||
kvc.reset(); pk = kvc.try_wait()
|
||||
acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
|
||||
acc_pipe.producer_acquire(acc_st)
|
||||
for kt in range(n_kv_tiles):
|
||||
kh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)
|
||||
sh = s_prod.acquire_and_advance()
|
||||
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
|
||||
for kb in cutlass.range(cute.size(tCrQ, mode=[2]), unroll_full=True):
|
||||
cute.gemm(qk_mma, tStS0, tCrQ[(None,None,kb,0)], tCrK[(None,None,kb,kh.index)], tStS0)
|
||||
qk_mma.set(tcgen05.Field.ACCUMULATE, True)
|
||||
cute.arch.fence_view_async_tmem_store()
|
||||
sh.commit(); kh.release()
|
||||
softmax_done_bar.arrive_and_wait()
|
||||
vh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)
|
||||
pv_mma.set(tcgen05.Field.ACCUMULATE, kt != 0)
|
||||
for kb in cutlass.range(cute.size(tOrP0, mode=[2]), unroll_full=True):
|
||||
cute.gemm(pv_mma, tOtO0, tOrP0[(None,None,kb)], tCrV[(None,None,kb,vh.index)], tOtO0)
|
||||
pv_mma.set(tcgen05.Field.ACCUMULATE, True)
|
||||
cute.arch.fence_view_async_tmem_store()
|
||||
vh.release()
|
||||
acc_pipe.producer_commit(acc_st); acc_st.advance()
|
||||
acc_pipe.producer_tail(acc_st)
|
||||
|
||||
# ===== SOFTMAX + EPILOGUE warps (0..3) =====
|
||||
if warp_idx < self.mma_warp_id:
|
||||
tmem.allocate(self.num_tmem_alloc_cols)
|
||||
tmem.wait_for_alloc()
|
||||
tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
|
||||
sfw_idx = tidx % (32 * len(self.epilogue_warp_id))
|
||||
|
||||
# --- S load: QK C-fragment layout, FP32 ---
|
||||
tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
|
||||
tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0)
|
||||
thr_load = tiled_tmem_load.get_slice(sfw_idx)
|
||||
tTMEM_LOADtS = thr_load.partition_S(tStS0)
|
||||
cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1]))
|
||||
tScS = qk_thr.partition_C(cS)
|
||||
tTMEM_LOADcS = thr_load.partition_D(tScS)
|
||||
|
||||
# --- P store: QK C-fragment composition with P sub-tile, FP32 backing
|
||||
# of BF16-packed pairs (the FMHA register-bridge pattern). ---
|
||||
p_cols_fp32 = self.pv_mma_tiler[2] * self.q_dtype.width // self.qk_acc_dtype.width
|
||||
tStP_layout = cute.composition(tStS.layout, cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32)))
|
||||
tStP0 = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStP_layout)
|
||||
tmem_store_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
|
||||
tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStP0)
|
||||
thr_store = tiled_tmem_store.get_slice(sfw_idx)
|
||||
tTMEM_STOREtP = thr_store.partition_D(tStP0)
|
||||
tScP_layout = cute.composition(tScS.layout, cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32)))
|
||||
tScP = cute.make_tensor(tScS.iterator, tScP_layout)
|
||||
tTMEM_STOREcP = thr_store.partition_S(tScP)
|
||||
|
||||
# --- O load/store path for rescale + final normalize ---
|
||||
# Tile O column-wise so the load/store fits in registers per pass.
|
||||
cO = cute.make_identity_tensor((self.pv_mma_tiler[0], self.pv_mma_tiler[1]))
|
||||
tOcO = pv_thr.partition_C(cO)
|
||||
corr_tile_size = 16
|
||||
tOtO_i_layout = cute.composition(tOtO.layout, cute.make_layout((128, corr_tile_size)))
|
||||
tOcO_i_layout = cute.composition(tOcO.layout, cute.make_layout((128, corr_tile_size)))
|
||||
tOtO_i = cute.make_tensor(tOtO.iterator, tOtO_i_layout)
|
||||
tOcO_i = cute.make_tensor(tOcO.iterator, tOcO_i_layout)
|
||||
tmem_load_o_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(corr_tile_size)), self.acc_dtype)
|
||||
tmem_store_o_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(corr_tile_size)), self.acc_dtype)
|
||||
tiled_tmem_load_o = tcgen05.make_tmem_copy(tmem_load_o_atom, tOtO_i)
|
||||
tiled_tmem_store_o = tcgen05.make_tmem_copy(tmem_store_o_atom, tOtO_i)
|
||||
thr_load_o = tiled_tmem_load_o.get_slice(sfw_idx)
|
||||
thr_store_o = tiled_tmem_store_o.get_slice(sfw_idx)
|
||||
tTMEM_LOAD_OtO = thr_load_o.partition_S(tOtO_i)
|
||||
tTMEM_LOAD_OcO = thr_load_o.partition_D(tOcO_i)
|
||||
tTMEM_STORE_OtO = thr_store_o.partition_D(tOtO_i)
|
||||
|
||||
o_col_tiles = self.pv_mma_tiler[1] // corr_tile_size
|
||||
|
||||
row_max = -Float32.inf
|
||||
row_sum = Float32(0.0)
|
||||
scale_log2 = Float32(self.scale_softmax_log2)
|
||||
|
||||
for kt in range(n_kv_tiles):
|
||||
si_handle = s_cons.wait_and_advance()
|
||||
|
||||
# --- Load S[kt] from TMEM into registers ---
|
||||
tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
|
||||
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
|
||||
cute.arch.fence_view_async_tmem_load()
|
||||
|
||||
# --- Pass 1: update row_max with S[kt] (in log2-domain) ---
|
||||
old_row_max = row_max
|
||||
frg_cnt = 4
|
||||
frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
|
||||
tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
|
||||
for j in range(frg_cnt):
|
||||
for k in range(cute.size(tTMEM_LOADrS_frg, mode=[0])):
|
||||
row_max = cute.arch.fmax(row_max, tTMEM_LOADrS_frg[k, j] * scale_log2)
|
||||
|
||||
row_max_safe = row_max
|
||||
if row_max == -cutlass.Float32.inf:
|
||||
row_max_safe = Float32(0.0)
|
||||
|
||||
# acc_scale = exp2(old_max - new_max). On the first tile old_max
|
||||
# is -inf; force acc_scale to 0 so initial row_sum stays 0 and
|
||||
# we skip the O rescale path entirely.
|
||||
acc_scale_ = scale_log2 * (old_row_max - row_max_safe)
|
||||
acc_scale = cute.math.exp2(acc_scale_, fastmath=True)
|
||||
if old_row_max == -cutlass.Float32.inf:
|
||||
acc_scale = Float32(0.0)
|
||||
row_sum *= acc_scale
|
||||
|
||||
# --- Pass 2: P[kt] = exp2((S - new_max) * log2), accumulate row_sum,
|
||||
# cast to BF16 through the FP32-backed register bridge. ---
|
||||
rP_words = cute.make_rmem_tensor(tTMEM_STOREcP.shape, self.qk_acc_dtype)
|
||||
rP_bf16 = cute.make_tensor(cute.recast_ptr(rP_words.iterator, dtype=self.q_dtype), tTMEM_LOADrS.layout)
|
||||
minus_row_max_scale = (Float32(0.0) - row_max_safe) * scale_log2
|
||||
|
||||
rP_bf16_frg = cute.logical_divide(rP_bf16, cute.make_layout(frg_tile))
|
||||
for j in range(frg_cnt):
|
||||
for k in range(cute.size(tTMEM_LOADrS_frg, mode=[0])):
|
||||
tTMEM_LOADrS_frg[k, j] = tTMEM_LOADrS_frg[k, j] * scale_log2 + minus_row_max_scale
|
||||
tTMEM_LOADrS_frg[k, j] = cute.math.exp2(tTMEM_LOADrS_frg[k, j], fastmath=True)
|
||||
row_sum = row_sum + tTMEM_LOADrS_frg[k, j]
|
||||
s_vec = tTMEM_LOADrS_frg[None, j].load()
|
||||
rP_bf16_frg[None, j].store(s_vec.to(self.q_dtype))
|
||||
|
||||
cute.copy(tiled_tmem_store, rP_words, tTMEM_STOREtP)
|
||||
cute.arch.fence_view_async_tmem_store()
|
||||
|
||||
# --- Fix 3: O rescale for kt > 0 ---
|
||||
# MMA's previous PV wrote O at offset tmem_o0_offset. We multiply
|
||||
# O in place by acc_scale, then arrive on softmax_done_bar so MMA
|
||||
# can start PV[kt] which accumulates the new P @ V on top of the
|
||||
# rescaled O.
|
||||
# IMPORTANT: this must happen before softmax_done_bar.arrive().
|
||||
if kt > 0:
|
||||
for i in range(o_col_tiles):
|
||||
tTMEM_LOAD_O_i = cute.make_tensor(
|
||||
tTMEM_LOAD_OtO.iterator + i * corr_tile_size,
|
||||
tTMEM_LOAD_OtO.layout,
|
||||
)
|
||||
tTMEM_STORE_O_i = cute.make_tensor(
|
||||
tTMEM_STORE_OtO.iterator + i * corr_tile_size,
|
||||
tTMEM_STORE_OtO.layout,
|
||||
)
|
||||
tTMrO = cute.make_rmem_tensor(tTMEM_LOAD_OcO.shape, self.acc_dtype)
|
||||
cute.copy(tiled_tmem_load_o, tTMEM_LOAD_O_i, tTMrO)
|
||||
cute.arch.fence_view_async_tmem_load()
|
||||
for k in cutlass.range(cute.size(tTMrO), vectorize=True):
|
||||
tTMrO[k] = tTMrO[k] * acc_scale
|
||||
cute.copy(tiled_tmem_store_o, tTMrO, tTMEM_STORE_O_i)
|
||||
cute.arch.fence_view_async_tmem_store()
|
||||
|
||||
si_handle.release()
|
||||
softmax_done_bar.arrive()
|
||||
|
||||
# --- Final O normalization by row_sum ---
|
||||
# After the last softmax iteration, MMA still needs to finish the final
|
||||
# PV[N-1]. The acc_pipe consumer wait inside epilogue_tma_store handles
|
||||
# that for the GMEM write, but we also need O fully accumulated before
|
||||
# we divide by row_sum here. In practice MMA's PV[N-1] is small enough
|
||||
# that by the time softmax has fallen out of the loop it has retired.
|
||||
# If you observe drift, insert an explicit acc_pipe.consumer_wait
|
||||
# before this block — see the consumer state setup below.
|
||||
inv_row_sum = Float32(1.0) / row_sum
|
||||
for i in range(o_col_tiles):
|
||||
tTMEM_LOAD_O_i = cute.make_tensor(
|
||||
tTMEM_LOAD_OtO.iterator + i * corr_tile_size,
|
||||
tTMEM_LOAD_OtO.layout,
|
||||
)
|
||||
tTMEM_STORE_O_i = cute.make_tensor(
|
||||
tTMEM_STORE_OtO.iterator + i * corr_tile_size,
|
||||
tTMEM_STORE_OtO.layout,
|
||||
)
|
||||
tTMrO = cute.make_rmem_tensor(tTMEM_LOAD_OcO.shape, self.acc_dtype)
|
||||
cute.copy(tiled_tmem_load_o, tTMEM_LOAD_O_i, tTMrO)
|
||||
cute.arch.fence_view_async_tmem_load()
|
||||
for k in cutlass.range(cute.size(tTMrO), vectorize=True):
|
||||
tTMrO[k] = tTMrO[k] * inv_row_sum
|
||||
cute.copy(tiled_tmem_store_o, tTMrO, tTMEM_STORE_O_i)
|
||||
cute.arch.fence_view_async_tmem_store()
|
||||
|
||||
# --- Epilogue: TMEM -> SMEM -> GMEM via TMA store ---
|
||||
tCtO_base = cute.make_tensor(tmem_ptr + self.tmem_o0_offset, tCtO_fake.layout)
|
||||
acc_cons_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Consumer, self.num_acc_stage)
|
||||
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)
|
||||
# Exercise single tile, 2 tiles, 4 tiles, 8 tiles.
|
||||
for n in [128, 256, 512, 1024]:
|
||||
torch.manual_seed(42)
|
||||
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.randn(n, hd, dtype=torch.bfloat16, device='cuda')
|
||||
v_kernel = v.unsqueeze(-1)
|
||||
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
|
||||
|
||||
# Reference: full softmax over the actual sequence.
|
||||
qf = q[:, :, 0].float()
|
||||
kf = k[:, :, 0].float()
|
||||
scale = 1.0 / math.sqrt(hd)
|
||||
attn = qf @ kf.T * scale
|
||||
attn = torch.softmax(attn, dim=-1)
|
||||
ref = attn @ v.float()
|
||||
|
||||
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
|
||||
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
|
||||
mV = ct.from_dlpack(v_kernel).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_kernel))
|
||||
mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
|
||||
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
|
||||
|
||||
# Fix 1: s_k MUST equal n. Each n requires its own compiled kernel
|
||||
# because s_k controls the v_fmha layout (a compile-time constant).
|
||||
kernel = FmhaV3StageCMulti(s_k=n)
|
||||
print(f'n={n}: Compiling...', flush=True)
|
||||
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
|
||||
print(f'n={n}: tmem_offsets: s0={kernel.tmem_s0_offset} '
|
||||
f'p0={kernel.tmem_p0_offset} o0={kernel.tmem_o0_offset} '
|
||||
f'alloc={kernel.num_tmem_alloc_cols}', flush=True)
|
||||
compiled(mQ, mK, mV, mC, stream)
|
||||
torch.cuda.synchronize()
|
||||
|
||||
out = c[:, :, 0].float()
|
||||
cos = torch.nn.functional.cosine_similarity(
|
||||
out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)
|
||||
).item()
|
||||
max_abs = (out - ref).abs().max().item()
|
||||
n_tiles = n // 128
|
||||
print(f'FMHA Stage-C Multi-Tile n={n} ({n_tiles} kv tiles): '
|
||||
f'cosine {cos:.6f} max_abs {max_abs:.4f} '
|
||||
f'{"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()
|
||||
Reference in New Issue
Block a user