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

BREAKTHROUGH: unnormalized P@V cosine 0.999998 for n=128!

Browse Source 
The softmax math (exp2, P store, PV) is correct for single-tile.
The bug is ONLY in C6/C9 normalization: applying inv_row_sum
using PV partition instead of QK partition.

n=128 (single tile): cosine 0.999998 PASS
n=256/384 (multi-tile): C6 O-rescale using wrong partition = FAIL

Fix: normalize O using QK row coordinates, not PV row coordinates.
Can use TMEM vector to bridge QK partition to PV partition.
This commit is contained in:
biondizzle
2026-05-21 18:55:00 +00:00
parent a983a8fb41
commit d7aa4da686
1 changed files with 5 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
tests/unit/test_fmha_v3_softmax.py
View File

@@ -373,37 +373,9 @@ class FmhaV3Softmax:
row_sum = row_sum + tile_sum
# --- C9: Final normalization via O TMEM rescale ---
# --- C9: SKIPPED for debug (no normalization) ---
pv_done_bar.arrive_and_wait()
# Store final row_sum to TMEM vector (per-row, using QK partition)
tTMEM_STORE_VECrS_final = cute.make_rmem_tensor(tTMEM_STORE_VECcS.shape, self.qk_acc_dtype)
tTMEM_STORE_VECrS_final[0] = row_sum
cute.copy(tiled_tmem_store_vec, tTMEM_STORE_VECrS_final, tTMEM_STORE_VECtS)
cute.arch.fence_view_async_tmem_store()
# Read vector back: per-row row_sum using QK partition coordinates
tTMEM_LOAD_VECrS = cute.make_rmem_tensor(tTMEM_LOAD_VECcS.shape, self.qk_acc_dtype)
cute.copy(tiled_tmem_load_vec, tTMEM_LOAD_VECtS, tTMEM_LOAD_VECrS)
cute.arch.fence_view_async_tmem_load()
inv_row_sum = cutlass.Float32(1.0) / tTMEM_LOAD_VECrS[0]
# Normalize O in TMEM
tTMrO_final = cute.make_rmem_tensor((tTMEM_LOADcO.shape, o_col_tiles), self.qk_acc_dtype)
for i in range(o_col_tiles):
tTMrO_i_ = tTMrO_final[None, i]
tTMrO_i_layout = cute.composition(tTMrO_i_.layout, cute.make_layout(tTMrO_final.shape[0]))
tTMrO_i = cute.make_tensor(tTMrO_i_.iterator, tTMrO_i_layout)
tTMEM_LOADtO_i = cute.make_tensor(
tTMEM_LOADtO.iterator + i * corr_tile_size, tTMEM_LOADtO.layout)
tTMEM_STOREtO_i = cute.make_tensor(
tTMEM_STOREtO.iterator + i * corr_tile_size, tTMEM_STOREtO.layout)
cute.copy(o_tiled_tmem_load, tTMEM_LOADtO_i, tTMrO_i)
for j in cutlass.range(cute.size(tTMrO_i), vectorize=True):
tTMrO_i[j] = tTMrO_i[j] * inv_row_sum
cute.copy(o_tiled_tmem_store, tTMrO_i, tTMEM_STOREtO_i)
cute.arch.fence_view_async_tmem_store()
# Now O in TMEM is normalized. Use standard epilogue_tma_store with identity.
# O is unnormalized in TMEM. Use standard epilogue_tma_store with identity.
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))
@@ -429,7 +401,8 @@ def test():
c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device="cuda")
qf = q[:,:,0].float(); kf = k[:,:,0].float()
attn = qf @ kf.T / math.sqrt(hd)
ref = torch.softmax(attn, dim=-1) @ v.float()
P = torch.exp(attn - attn.max(dim=-1, keepdim=True)[0])
ref = P @ v.float() # unnormalized P@V
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))
@@ -445,7 +418,7 @@ def test():
out = c[:,:,0].float()
cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item()
max_err = (out - ref).abs().max().item()
print(f"FMHA softmax n={n}: cosine {cos:.6f} max_err {max_err:.6f} {'PASS' if cos >= 0.999 else 'FAIL'}", flush=True)
print(f"FMHA softmax (no C9 norm) n={n}: cosine {cos:.6f} max_err {max_err:.6f} {'PASS' if cos >= 0.999 else 'FAIL'}", flush=True)
if __name__ == "__main__":
test()