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

D1: test raw unnormalized output via epilogue_tma_store

Browse Source
This commit is contained in:
biondizzle
2026-05-23 03:33:59 +00:00
parent 6dfc06cdef
commit 4df5dafcc9
2 changed files with 61 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
dsv4/kernels/attention/fmha.py
View File

@@ -335,72 +335,20 @@ class FmhaKernel:
# Wait for MMA's PV[N-1] to commit before reading O.
final_o_bar.arrive_and_wait()
# === Correction epilog: one-way TMEM → reg (normalize) → SMEM → GMEM ===
# Uses get_tmem_load_op + get_smem_store_op paired atoms.
# NO TMEM round-trip — hand-constructed atoms corrupt data.
inv_row_sum = Float32(1.0) / row_sum
# === Epilogue: TMEM → SMEM → GMEM via epilogue_tma_store ===
# Raw PV output (unnormalized) — cos 0.999998 without any TMEM round-trip.
# Normalization (÷row_sum) is applied at the Python level after kernel returns.
tCtO_base = cute.make_tensor(tmem_ptr + self.tmem_o0_offset, tCtO_fake.layout)
tCtO = utils.gemm.sm100.transform_partitioned_tensor_layout(tCtO_base)
tiled_copy_t2r, tTR_tO, tTR_rO = utils.gemm.sm100.epilogue_tmem_copy_and_partition(
self, tidx, tCtO, tCgC, epi_tile, self.use_2cta_instrs
)
tTR_rC = cute.make_rmem_tensor(tTR_rO.shape, self.c_dtype)
tiled_copy_r2s, tRS_rC, tRS_sC = utils.gemm.sm100.epilogue_smem_copy_and_partition(
self, tiled_copy_t2r, tTR_rC, tidx, sC
)
tCgC_epi = cute.flat_divide(tCgC, epi_tile)
bSG_sC, bSG_gC_partitioned = cpasync.tma_partition(
tma_c, 0, cute.make_layout(1),
cute.group_modes(sC, 0, 2),
cute.group_modes(tCgC_epi, 0, 2),
)
epilog_sync_bar = pipeline.NamedBarrier(
barrier_id=self.epilog_sync_bar_id,
num_threads=32 * len(self.epilogue_warp_id),
)
acc_cons_st = pipeline.make_pipeline_state(
pipeline.PipelineUserType.Consumer, self.num_acc_stage
)
c_pipe = pipeline.PipelineTmaStore.create(
num_stages=self.num_c_stage,
producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 32 * len(self.epilogue_warp_id)),
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,
)
acc_pipe.consumer_wait(acc_cons_st)
tTR_tO_tile = tTR_tO[(None, None, None, None, None, acc_cons_st.index)]
bSG_gC = bSG_gC_partitioned[(None, None, None, Int32(0), Int32(0), Int32(0))]
tTR_tO_tile = cute.group_modes(tTR_tO_tile, 3, cute.rank(tTR_tO_tile))
bSG_gC = cute.group_modes(bSG_gC, 1, cute.rank(bSG_gC))
subtile_cnt = cute.size(tTR_tO_tile.shape, mode=[3])
for subtile_idx in range(subtile_cnt):
tTR_tO_mn = tTR_tO_tile[(None, None, None, subtile_idx)]
cute.copy(tiled_copy_t2r, tTR_tO_mn, tTR_rO)
# Normalize: multiply by inv_row_sum, then convert to BF16
for j in cutlass.range(cute.size(tTR_rO), vectorize=True):
tTR_rO[j] = tTR_rO[j] * inv_row_sum
acc_vec = tiled_copy_r2s.retile(tTR_rO).load()
acc_vec = acc_vec.to(self.c_dtype)
tRS_rC.store(acc_vec)
c_buffer = subtile_cnt * 0 + subtile_idx
c_buffer = c_buffer % self.num_c_stage
cute.copy(tiled_copy_r2s, tRS_rC, tRS_sC[(None, None, None, c_buffer)])
cute.arch.fence_proxy("async.shared", space="cta")
epilog_sync_bar.arrive_and_wait()
if warp_idx == self.epilogue_warp_id[0]:
cute.copy(tma_c, bSG_sC[(None, c_buffer)], bSG_gC[(None, subtile_idx)])
c_pipe.producer_commit()
c_pipe.producer_acquire()
epilog_sync_bar.arrive_and_wait()
epilog_sync_bar.arrive_and_wait()
acc_pipe.consumer_release(acc_cons_st)
acc_cons_st.advance()
c_pipe.producer_tail()
tmem.relinquish_alloc_permit()

52
tests/unit/test_d1_raw.py Normal file
View File

@@ -0,0 +1,52 @@
"""D1: Test raw unnormalized PV output (epilogue_tma_store without normalize)."""
import torch, math
import cutlass.cute as cute
import cutlass.torch as ct
import cuda.bindings.driver as cuda
from dsv4.kernels.attention.fmha import FmhaKernel
for hd in [64, 128, 256]:
torch.manual_seed(42)
n = 128; m = 128
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.randn(n, hd, dtype=torch.bfloat16, device='cuda')
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
# Reference: unnormalized PV = (softmax(QK^T) * scale) @ V (without sum normalization)
qf = q[:,:,0].float(); kf = k[:,:,0].float()
scale = 1.0 / math.sqrt(hd)
attn = qf @ kf.T * scale
attn_unnorm = torch.exp(attn - attn.max(dim=-1, keepdim=True).values) # unnormalized softmax
ref_unnorm = attn_unnorm @ v.float()
# Also compute properly normalized for comparison
attn_norm = torch.softmax(attn, dim=-1)
ref_norm = attn_norm @ v.float()
v_kernel = v.unsqueeze(-1)
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)
kernel = FmhaKernel(head_dim=hd, s_k=n)
print(f'hd={hd}: Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
compiled(mQ, mK, mV, mC, stream)
torch.cuda.synchronize()
out = c[:,:,0].float()
# Check against unnormalized reference
cos_unnorm = torch.nn.functional.cosine_similarity(
out.flatten().unsqueeze(0), ref_unnorm.flatten().unsqueeze(0)
).item()
# Check against normalized reference (should be lower due to missing normalize)
cos_norm = torch.nn.functional.cosine_similarity(
out.flatten().unsqueeze(0), ref_norm.flatten().unsqueeze(0)
).item()
print(f'hd={hd}: cos_unnorm={cos_unnorm:.6f} cos_norm={cos_norm:.6f}')