biondizzle/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

debug: add shape logging to SF packing

Browse Source
This commit is contained in:
biondizzle
2026-05-11 08:54:14 +00:00
parent dcebe033e2
commit cd7a612175
1 changed files with 4 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
deep_gemm/mega/__init__.py
View File

@@ -196,6 +196,10 @@ def transform_nvfp4_weights_for_mega_moe(
l1_sf_packed = pack_ue4m3_to_int32(l1_sf_32)
l2_sf_packed = pack_ue4m3_to_int32(l2_sf_32)
print(f"[NVFP4-MoE] l1_sf_32: shape={l1_sf_32.shape}, l1_sf_packed: shape={l1_sf_packed.shape}")
print(f"[NVFP4-MoE] l2_sf_32: shape={l2_sf_32.shape}, l2_sf_packed: shape={l2_sf_packed.shape}")
print(f"[NVFP4-MoE] l1_n={l1_n} l1_k={l1_k} l2_n={l2_n} l2_k={l2_k}")
# Transpose to MN-major layout (stride(-2)=1) and make contiguous
# transform_sf_into_required_layout expects MN-major input for TMA stride checks
l1_sf_mn = l1_sf_packed.transpose(-2, -1).contiguous().transpose(-2, -1)