diff --git a/vllm/nvfp4_cutedsl.py b/vllm/nvfp4_cutedsl.py
index 6c05e53e..46625cef 100644
--- a/vllm/nvfp4_cutedsl.py
+++ b/vllm/nvfp4_cutedsl.py
@@ -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,
         )