Fix routing_bias dtype (#25711)

Signed-off-by: Shu Wang. <shuw@nvidia.com>

vllm/model_executor/layers/quantization/modelopt.py

@@ -1454,10 +1454,13 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
             routing_method_type = flashinfer.RoutingMethodType.DeepSeekV3
             if use_llama4_routing:
                 routing_method_type = flashinfer.RoutingMethodType.Llama4
+            routing_bias = e_score_correction_bias
+            if routing_bias is not None:
+                routing_bias = routing_bias.to(torch.bfloat16)
             out = flashinfer.fused_moe.trtllm_fp4_block_scale_moe(
                 routing_logits=router_logits
                 if use_llama4_routing else router_logits.to(torch.float32),
-                routing_bias=e_score_correction_bias,
+                routing_bias=routing_bias,
                 hidden_states=hidden_states_fp4,
                 hidden_states_scale=hidden_states_scale_linear_fp4.view(
                     torch.float8_e4m3fn).flatten(),