support qwen3-vl handle requests with embeddings (#30037)

Signed-off-by: taoyun <[email protected]>
Signed-off-by: Cyrus Leung <[email protected]>
Co-authored-by: Cyrus Leung <[email protected]>

Author: taoyun951753

docs/features/multimodal_inputs.md

@@ -443,6 +443,8 @@ For Qwen2-VL and MiniCPM-V, we accept additional parameters alongside the embedd
         print(generated_text)
     ```
 
+For Qwen3-VL, the `image_embeds` should contain both the base image embedding and deepstack features.
+
 #### Audio Embeddings
 
 You can pass pre-computed audio embeddings similar to image embeddings:

vllm/model_executor/models/qwen3_vl.py

@@ -103,7 +103,7 @@
     Qwen2_5_VLVideoInputs,
     Qwen2_5_VLVideoPixelInputs,
 )
-from .qwen2_vl import Qwen2VLProcessingInfo
+from .qwen2_vl import Qwen2VLMultiModalDataParser, Qwen2VLProcessingInfo
 from .qwen3 import Qwen3ForCausalLM, Qwen3Model
 from .utils import (
     AutoWeightsLoader,
@@ -884,7 +884,10 @@ def _get_dummy_videos(
 
 class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo]):
     def _get_data_parser(self) -> MultiModalDataParser:
-        return MultiModalDataParser(video_needs_metadata=True)
+        return Qwen2VLMultiModalDataParser(
+            self.info.get_hf_config().vision_config.spatial_merge_size,
+            video_needs_metadata=True,
+        )
 
     def _call_hf_processor(
         self,