diff --git a/python/sglang/srt/models/minimax_m2.py b/python/sglang/srt/models/minimax_m2.py
index 011e45d8d..3ac16102f 100644
--- a/python/sglang/srt/models/minimax_m2.py
+++ b/python/sglang/srt/models/minimax_m2.py
@@ -222,7 +222,7 @@ class MiniMaxM2MoE(nn.Module):
         if hidden_states.shape[0] > 0:
             # router_logits: (num_tokens, n_experts)
             router_logits, _ = self.gate(hidden_states.to(torch.float32))
-            topk_weights, topk_idx, _ = self.topk(
+            topk_output = self.topk(
                 hidden_states,
                 router_logits,
                 num_token_non_padded=forward_batch.num_token_non_padded,
@@ -231,14 +231,10 @@ class MiniMaxM2MoE(nn.Module):
                 ),
             )
         else:
-            topk_weights, topk_idx, _ = self.topk.empty_topk_output(
-                hidden_states.shape[0], self.top_k
-            )
+            topk_output = self.topk.empty_topk_output(device=hidden_states.device)
         final_hidden_states = self.experts(
             hidden_states=hidden_states,
-            topk_idx=topk_idx,
-            topk_weights=topk_weights,
-            forward_batch=forward_batch,
+            topk_output=topk_output,
         )
 
         return final_hidden_states