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, 18:45 UTC
[Bugfix] Qwen 3 VL Embedding loading (#30303)
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io> Signed-off-by: wang.yuqi <noooop@126.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
wang.yuqi committed
Dec 9, 2025 at 16:04 UTC
9c32df6101b24040c824dee5b0e5543a94dc6f45
1 file changed
+8
-3
vllm/model_executor/models/adapters.py
+8
-3
@@ -175,9 +175,14 @@ def _create_pooling_model_cls(orig_cls: _T) -> _T:
175
self.vllm_config = vllm_config
176
177
# These are not used in pooling models
178
- for attr in ("lm_head", "logits_processor"):
179
- if hasattr(self, attr):
180
- delattr(self, attr)
178
+ objects_to_clean = [self]
179
+ if language_model := getattr(self, "language_model", None):
180
+ objects_to_clean.append(language_model)
181
+
182
+ for obj in objects_to_clean:
183
+ for attr in ("lm_head", "logits_processor"):
184
+ if hasattr(obj, attr):
185
+ delattr(obj, attr)
186
187
# If the model already defines a pooler instance, don't overwrite it
188
if not getattr(self, "pooler", None):