[Bugfix] Fix GLM-4 MoE router logits dtype for data parallel chunking (#31055)

Signed-off-by: ReinforcedKnowledge <reinforced.knowledge@gmail.com>

vllm/model_executor/models/glm4_moe.py

@@ -197,6 +197,7 @@ class Glm4MoE(nn.Module):
             e_score_correction_bias=self.gate.e_score_correction_bias,
             enable_eplb=self.enable_eplb,
             num_redundant_experts=self.n_redundant_experts,
+            router_logits_dtype=torch.float32,
         )
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: