A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 12:30 UTC
[Bugfix] Qwen3-next with --hf-overrides \{\"num_hidden_layers\":8\} (#30433)
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Chen Zhang committed
Dec 13, 2025 at 06:12 UTC
ace34e3783208a31b185968a1e92c79ac8f633cb
1 file changed
+7
vllm/model_executor/models/qwen3_next.py
+7
@@ -1092,6 +1092,8 @@ class Qwen3NextModel(nn.Module):
1092
name.endswith(".bias") or name.endswith("_bias")
1093
) and name not in params_dict:
1094
continue
1095
+ if name not in params_dict:
1096
+ continue
1097
param = params_dict[name]
1098
weight_loader = param.weight_loader
1099
weight_loader(
@@ -1108,6 +1110,11 @@ class Qwen3NextModel(nn.Module):
1110
continue
1111
if is_pp_missing_parameter(name, self):
1112
continue
1113
+ if name not in params_dict:
1114
+ logger.warning_once(
1115
+ f"Parameter {name} not found in params_dict, skip loading"
1116
+ )
1117
+ continue
1118
param = params_dict[name]
1119
weight_loader = getattr(
1120
param, "weight_loader", default_weight_loader