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 12, 01:00 UTC
support qwen3-vl handle requests with embeddings (#30037)
Signed-off-by: taoyun <1069423820@qq.com> Signed-off-by: Cyrus Leung <cyrus.tl.leung@gmail.com> Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Tao Yun committed
Dec 5, 2025 at 01:34 UTC
6dcb07f676ae473703b4ab69ef15d93f51d24ce8
2 files changed
+7
-2
docs/features/multimodal_inputs.md
+2
@@ -443,6 +443,8 @@ For Qwen2-VL and MiniCPM-V, we accept additional parameters alongside the embedd
443
print(generated_text)
444
```
445
446
+For Qwen3-VL, the `image_embeds` should contain both the base image embedding and deepstack features.
447
+
448
#### Audio Embeddings
449
450
You can pass pre-computed audio embeddings similar to image embeddings:
vllm/model_executor/models/qwen3_vl.py
+5
-2
@@ -103,7 +103,7 @@ from .qwen2_5_vl import (
103
Qwen2_5_VLVideoInputs,
104
Qwen2_5_VLVideoPixelInputs,
105
)
106
-from .qwen2_vl import Qwen2VLProcessingInfo
106
+from .qwen2_vl import Qwen2VLMultiModalDataParser, Qwen2VLProcessingInfo
107
from .qwen3 import Qwen3ForCausalLM, Qwen3Model
108
from .utils import (
109
AutoWeightsLoader,
@@ -884,7 +884,10 @@ class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]):
884
885
class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo]):
886
def _get_data_parser(self) -> MultiModalDataParser:
887
- return MultiModalDataParser(video_needs_metadata=True)
887
+ return Qwen2VLMultiModalDataParser(
888
+ self.info.get_hf_config().vision_config.spatial_merge_size,
889
+ video_needs_metadata=True,
890
+ )
891
892
def _call_hf_processor(
893
self,