diff --git a/vllm/nvfp4_cutedsl.py b/vllm/nvfp4_cutedsl.py
index 3f90d2e0..aea180a5 100644
--- a/vllm/nvfp4_cutedsl.py
+++ b/vllm/nvfp4_cutedsl.py
@@ -126,8 +126,8 @@ class CuTeDSLMoERunner:
         # Build expert_offsets: cumulative count of tokens per expert
         # Count how many slots each expert gets
         expert_id_range = torch.arange(num_experts, device=device)
-        tokens_per_expert = (sorted_ids.unsqueeze(1) == expert_id_range.unsqueeze(0)).sum(dim=0)
-        expert_offsets = torch.zeros(num_experts + 1, dtype=torch.int64, device=device)
+        tokens_per_expert = (sorted_ids.unsqueeze(1) == expert_id_range.unsqueeze(0)).sum(dim=0).int()
+        expert_offsets = torch.zeros(num_experts + 1, dtype=torch.int32, device=device)
         expert_offsets[1:] = tokens_per_expert.cumsum(0)
         if expert_offsets[-1] == 0:
             return torch.zeros(num_tokens, hidden_size, dtype=torch.bfloat16, device=device)