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

Add debug prints for SMEM-P partition layouts

Browse Source
This commit is contained in:
biondizzle
2026-05-23 19:13:07 +00:00
parent b4b11db0fa
commit c9448aca03
1 changed files with 7 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
dsv4/kernels/attention/fmha.py
View File

@@ -271,8 +271,11 @@ class FmhaKernel:
)
tiled_smem_copy = cute.make_tiled_copy_C(smem_copy_atom, qk_mma)
thr_smem_copy = tiled_smem_copy.get_slice(sfw_idx)
sP_2d = cute.group_modes(sP, 0, 3)
tSMEM_CPYsP = thr_smem_copy.partition_D(sP_2d) # destination (SMEM)
# Debug sP shape
print(f"[SMEM-P PROPER] sP shape: {cute.shape(sP)} rank: {len(cute.shape(sP))}")
# Try without group_modes first
tSMEM_CPYsP = thr_smem_copy.partition_D(sP) # destination (SMEM)
print(f"[SMEM-P PROPER] After partition_D: tSMEM_CPYsP layout: {tSMEM_CPYsP.layout}")
row_max = -Float32.inf
row_sum = Float32(0.0)
@@ -352,7 +355,9 @@ class FmhaKernel:
rP_bf16_qk = cute.make_tensor(cute.recast_ptr(rP_bf16.iterator, dtype=self.q_dtype), rP_qk_layout)
# Partition source with QK layout
print(f"[SMEM-P PROPER] Before partition_S: rP_bf16_qk shape: {cute.shape(rP_bf16_qk)} layout: {rP_bf16_qk.layout}")
tSMEM_CPYrP_qk = thr_smem_copy.partition_S(rP_bf16_qk)
print(f"[SMEM-P PROPER] After partition_S: tSMEM_CPYrP_qk layout: {tSMEM_CPYrP_qk.layout}")
# Debug shapes
print(f"[SMEM-P PROPER] rP_bf16 shape: {cute.shape(rP_bf16)}, layout: TMEM")