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 9, 18:15 UTC
[Bugfix] Fix getting vision features in Transformer Multimodal backend (#32933)
Signed-off-by: raushan <raushan@huggingface.co>
Raushan Turganbay committed
Jan 23, 2026 at 14:34 UTC
d95d65076292df5cfb4d06ed0bdba7496825c499
1 file changed
+9
vllm/model_executor/models/transformers/multimodal.py
+9
@@ -376,6 +376,15 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
376
pixel_values, **kwargs
377
)
378
379
+ # Transformers `v5`, `self.get_image_features` returns a tuple
380
+ # containing the features and optionally attentions/hidden_states
381
+ # After v5 is settled, we can enable qwen3-vl with several outputs
382
+ # from `self.get_image_features`
383
+ if isinstance(vision_embeddings, tuple):
384
+ vision_embeddings = vision_embeddings[0]
385
+ elif isinstance(vision_embeddings, dict):
386
+ vision_embeddings = vision_embeddings.pooler_output
387
+
388
if isinstance(vision_embeddings, torch.Tensor):
389
if vision_embeddings.ndim == 2:
390
vision_embeddings = vision_embeddings.unsqueeze(0)