[PD Disaggregation] Support Qwen3-MoE use PD + EP inference. (#4691)

support Qwen-MoE PD/EP

fastdeploy/config.py

@@ -212,7 +212,6 @@ class ModelConfig:
         # set attribute from pretrained_config
         for key, value in pretrained_config.items():
             setattr(self, key, value)
-
         # we need set default value when not exist
         for key, value in PRETRAINED_INIT_CONFIGURATION.items():
             if not hasattr(self, key):
@@ -300,6 +299,9 @@ class ModelConfig:
         if not hasattr(self, "mla_use_absorb"):
             self.mla_use_absorb = False
 
+        if hasattr(self, "num_experts") and getattr(self, "moe_num_experts") is None:
+            self.moe_num_experts = self.num_experts
+
     def read_from_env(self):
         """
         Read configuration information from environment variables and update the object's attributes.