[Misc] Skip loading extra bias for Qwen2-MOE GPTQ models (#8329)

vllm/model_executor/models/qwen2_moe.py

@@ -469,7 +469,8 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                     continue
                 name = name.replace(weight_name, param_name)
                 # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
+                if ((name.endswith(".bias") or name.endswith("_bias"))
+                        and name not in params_dict):
                     continue
                 # Skip layers on other devices.
                 if is_pp_missing_parameter(name, self):
@@ -490,6 +491,10 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                     # Skip layers on other devices.
                     if is_pp_missing_parameter(name, self):
                         continue
+                    # Skip loading extra bias for GPTQ models.
+                    if ((name.endswith(".bias") or name.endswith("_bias"))
+                            and name not in params_dict):
+                        continue
                     param = params_dict[name]
                     weight_loader = param.weight_loader
                     weight_loader(param,
@@ -500,7 +505,8 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                     break
                 else:
                     # Skip loading extra bias for GPTQ models.
-                    if name.endswith(".bias") and name not in params_dict:
+                    if ((name.endswith(".bias") or name.endswith("_bias"))
+                            and name not in params_dict):
                         continue
                     # Skip layers on other devices.
                     if is_pp_missing_parameter(name, self):