self.gate dtype update for GLM-4.5 (vllm-project#22203)

Signed-off-by: zRzRzRzRzRzRzR <2448370773@qq.com>
Signed-off-by: x22x22 <wadeking@qq.com>

Author: zRzRzRzRzRzRzR

tests/models/registry.py

@@ -383,7 +383,7 @@ def check_available_online(
                                         trust_remote_code=True,
                                         hf_overrides={"architectures": ["GLM4VForCausalLM"]}),  # noqa: E501
     "Glm4vForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.1V-9B-Thinking"),  # noqa: E501
-    "Glm4v_moeForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.5V-Air",
+    "Glm4v_moeForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.5V",
                                           is_available_online=False),   # noqa: E501
     "H2OVLChatModel": _HfExamplesInfo("h2oai/h2ovl-mississippi-800m",
                                       trust_remote_code=True,

vllm/model_executor/models/glm4_moe.py

@@ -123,6 +123,7 @@ def __init__(
                                      config.n_routed_experts,
                                      bias=False,
                                      quant_config=None,
+                                     params_dtype=torch.float32,
                                      prefix=f"{prefix}.gate")
 
         self.gate.e_score_correction_bias = nn.Parameter(
@@ -180,7 +181,7 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
 
         if self.n_shared_experts is not None:
             shared_output = self.shared_experts(hidden_states)
-        router_logits, _ = self.gate(hidden_states)
+        router_logits, _ = self.gate(hidden_states.to(dtype=torch.float32))
         final_hidden_states = self.experts(
             hidden_states=hidden_states,
             router_logits=router_logits) * self.routed_scaling_factor