[Transformers backend] Ignore MTP weights when num_nextn_predict_layers=0 (#34888)

Signed-off-by: SteadfastAsArt <695488173@qq.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

vllm/model_executor/models/transformers/base.py

@@ -300,14 +300,26 @@ class Base(
             for child_name, child_module in module.named_children():
                 new_module = child_module
                 qual_name = maybe_prefix(prefix, child_name)
-                # Populate Eagle3 attrs
                 if (
                     isinstance(module, nn.ModuleList)
                     and len(module) == self.text_config.num_hidden_layers
                 ):
+                    # Populate Eagle3 attrs
                     self._target_class = type(child_module)
                     layer_name = qual_name.removeprefix("model.")
                     self._layer_names[int(child_name)] = layer_name
+                    # MTP weights should not be loaded into the base model
+                    num_hidden_layers = self.text_config.num_hidden_layers
+                    names = (
+                        "n_predict",  # Override from SpeculativeConfig
+                        "num_nextn_predict_layers",  # Most models
+                        "mtp_num_hidden_layers",  # Qwen 3.5
+                    )
+                    n_predict = getattr_iter(self.text_config, names, 0)
+                    for i in range(num_hidden_layers, num_hidden_layers + n_predict):
+                        mtp_prefix = f"{prefix}.{i}."
+                        if mtp_prefix not in self.ignore_unexpected_prefixes:
+                            self.ignore_unexpected_prefixes.append(mtp_prefix)
                 # Replace modules as needed
                 if isinstance(child_module, nn.Linear):
                     generator = (p for p in tp_plan if re.match(p, qual_name))

vllm/model_executor/models/utils.py

@@ -311,8 +311,9 @@ class AutoWeightsLoader:
 
                     continue
 
+                named_parameters = module.named_parameters(recurse=True)
                 desc_param_keys = {
-                    base_prefix + k for k, _ in module.named_parameters(recurse=True)
+                    maybe_prefix(base_prefix, k) for k, _ in named_parameters
                 }
                 msg = (
                     f"There is no module or parameter named {prefix!r} "