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 13, 04:30 UTC
[Bugfix] resolve Qwen3-VL GPTQModel quantized model loading failure (#28663)
Signed-off-by: GuanH <guansdrailib@gmail.com> Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn> Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
GuanH committed
Nov 15, 2025 at 02:44 UTC
cec275efcef62a9fb6ea5c3445572dddf9736206
2 files changed
+6
-3
vllm/model_executor/models/qwen3_vl.py
+3
-1
@@ -1138,7 +1138,9 @@ class Qwen3LLMForCausalLM(Qwen3ForCausalLM):
1138
self.config = config
1139
1140
self.quant_config = quant_config
1141
- self.model = Qwen3LLMModel(vllm_config=vllm_config, prefix=prefix)
1141
+ self.model = Qwen3LLMModel(
1142
+ vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
1143
+ )
1144
1145
if get_pp_group().is_last_rank:
1146
if config.tie_word_embeddings:
vllm/model_executor/models/utils.py
+3
-2
@@ -117,9 +117,10 @@ class AutoWeightsLoader:
117
environment variable `VLLM_LOGGING_LEVEL=DEBUG`.
118
"""
119
120
- # Models trained using early version ColossalAI
121
- # may include these tensors in checkpoint. Skip them.
120
+ # Models trained using early version ColossalAI or quantized by
121
+ # GPTQModel may include these tensors in checkpoint. Skip them.
122
ROTARY_EMBEDS_UNUSED_WEIGHTS = [
123
+ "rotary_pos_emb.inv_freq",
124
"rotary_emb.inv_freq",
125
"rotary_emb.cos_cached",
126
"rotary_emb.sin_cached",