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

fix: pass expert_offsets without leading 0 to GEMM (matches pipeline)

Browse Source
This commit is contained in:
biondizzle
2026-05-17 07:59:00 +00:00
parent a1e6f5f891
commit ae6b879d38
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/nvfp4_cutedsl.py
View File

@@ -274,7 +274,7 @@ class CuTeDSLMoERunner:
l1_out = run_nvfp4_grouped_gemm(
mat_a=x_fp4, mat_b=self._l1_mat_b,
scale_a=l1_scale_a, scale_b=self._l1_scale_b,
expert_offsets=expert_offsets[:self.num_experts + 1],
expert_offsets=expert_offsets[1:self.num_experts + 1],
global_scale_a=l1_gsa, global_scale_b=self._l1_gsb,
)
@@ -300,7 +300,7 @@ class CuTeDSLMoERunner:
l2_out = run_nvfp4_grouped_gemm(
mat_a=l2_x_fp4, mat_b=self._l2_mat_b,
scale_a=l2_scale_a, scale_b=self._l2_scale_b,
expert_offsets=expert_offsets[:self.num_experts + 1],
expert_offsets=expert_offsets[1:self.num_experts + 1],
global_scale_a=l2_gsa, global_scale_b=self._l2_gsb,
)