diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/layer.py b/python/sglang/srt/layers/moe/fused_moe_triton/layer.py
index 224ab3c7f1e6..6e0487a1ddc8 100644
--- a/python/sglang/srt/layers/moe/fused_moe_triton/layer.py
+++ b/python/sglang/srt/layers/moe/fused_moe_triton/layer.py
@@ -1117,7 +1117,7 @@ def forward(self, hidden_states: torch.Tensor, topk_output: TopKOutput):
         topk_config = topk_output.topk_config
 
         hs_fp4, hs_scale_linear = self._quantize_hidden_states_fp4(hidden_states)
-
+        router_logits = router_logits.to(torch.float32)
         routing_method_type = self.routing_method_type
         assert (
             routing_method_type is not None