diff --git a/sgl-kernel/python/sgl_kernel/moe.py b/sgl-kernel/python/sgl_kernel/moe.py
index 24066424f..1067a1760 100644
--- a/sgl-kernel/python/sgl_kernel/moe.py
+++ b/sgl-kernel/python/sgl_kernel/moe.py
@@ -41,6 +41,6 @@ def moe_fused_gate(input_tensor, bias, num_expert_group, topk_group, topk):
     # the #experts is decided by the input tensor shape and we currently only support power of 2 #experts
     # and #experts should be divisible by num_expert_group. #expert/num_expert_group <= 32 is limitted for now.
     # for non-supported case, we suggestion to use the biased_grouped_topk func in sglang.srt.layers.moe.topk
-    return torch.ops.sgl_kernel.moe_fused_gate(
+    return torch.ops.sgl_kernel.moe_fused_gate.default(
         input_tensor, bias, num_expert_group, topk_group, topk
     )