[Qwen-MOE] Fix memory duplication issues in case layers weights are re-assigned during weight loading (#18255)

python/sglang/srt/models/qwen3_moe.py

@@ -1043,10 +1043,9 @@ class Qwen3MoeForCausalLM(nn.Module):
             num_experts=self.config.num_experts,
         )
 
-        # Cache params_dict to avoid repeated expensive traversal of model parameters
+        # Pre-define `params_dict` to avoid repeated expensive traversal of model parameters.
-        if not hasattr(self, "_cached_params_dict"):
+        params_dict = dict(self.named_parameters())
-            self._cached_params_dict = dict(self.named_parameters())
+
-        params_dict = self._cached_params_dict
         for name, loaded_weight in weights:
             layer_id = get_layer_id(name)
             if (

python/sglang/srt/models/qwen3_omni_moe.py

@@ -524,10 +524,8 @@ class Qwen3OmniMoeForConditionalGeneration(PreTrainedModel):
 
         num_experts = self.config.num_experts
 
-        # Cache params_dict to avoid repeated expensive traversal of model parameters
+        # Pre-define `params_dict` to avoid repeated expensive traversal of model parameters.
-        if not hasattr(self, "_cached_params_dict"):
+        params_dict = dict(self.named_parameters())
-            self._cached_params_dict = dict(self.named_parameters())
-        params_dict = self._cached_params_dict
 
         for name, loaded_weight in weights:
             name = name.replace(r"model.language_model.", r"model.")

python/sglang/srt/models/qwen3_vl_moe.py

@@ -226,10 +226,9 @@ class Qwen3VLMoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
 
         num_experts = self.config.num_experts
 
-        # Cache params_dict to avoid repeated expensive traversal of model parameters
+        # Pre-define `params_dict` to avoid repeated expensive traversal of model parameters.
-        if not hasattr(self, "_cached_params_dict"):
+        params_dict = dict(self.named_parameters())
-            self._cached_params_dict = dict(self.named_parameters())
+
-        params_dict = self._cached_params_dict
         for name, loaded_weight in weights:
             name = name.replace(r"model.language_model.", r"model.")