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

fix: interleave_l1_weights granularity_bf16 (not granularity) in SE

Browse Source
This commit is contained in:
biondizzle
2026-06-02 08:29:03 +00:00
parent 553275d810
commit 1726cb64a9
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
dsv4/layers/shared_expert.py
View File

@@ -119,7 +119,7 @@ class Nvfp4SharedExpert:
# The fused kernel's SwiGLU epilogue expects granularity-8 interleaved gate/up. # The fused kernel's SwiGLU epilogue expects granularity-8 interleaved gate/up.
# The unfused path (if _fused_swiglu=False) deinterleaves the GEMM output before splitting. # The unfused path (if _fused_swiglu=False) deinterleaves the GEMM output before splitting.
if self._fused_swiglu: if self._fused_swiglu:
l1_stacked = interleave_l1_weights(l1_stacked, granularity=8) l1_stacked = interleave_l1_weights(l1_stacked, granularity_bf16=8)
# Stack weights and convert to K-major # Stack weights and convert to K-major
self._l1_mat_b = make_b_k_major(l1_stacked) # (1, K_packed, N_packed) self._l1_mat_b = make_b_k_major(l1_stacked) # (1, K_packed, N_packed)
self._l2_mat_b = make_b_k_major(l2_stacked) self._l2_mat_b = make_b_k_major(l2_stacked)