From 6dcb07f676ae473703b4ab69ef15d93f51d24ce8 Mon Sep 17 00:00:00 2001 From: Tao Yun <30410832+taoyun951753@users.noreply.github.com> Date: Fri, 5 Dec 2025 01:34:06 +0800 Subject: [PATCH] support qwen3-vl handle requests with embeddings (#30037) Signed-off-by: taoyun <1069423820@qq.com> Signed-off-by: Cyrus Leung Co-authored-by: Cyrus Leung --- docs/features/multimodal_inputs.md | 2 ++ vllm/model_executor/models/qwen3_vl.py | 7 +++++-- 2 files changed, 7 insertions(+), 2 deletions(-) diff --git a/docs/features/multimodal_inputs.md b/docs/features/multimodal_inputs.md index 2b25dc7666..0adb32a7ac 100644 --- a/docs/features/multimodal_inputs.md +++ b/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: diff --git a/vllm/model_executor/models/qwen3_vl.py b/vllm/model_executor/models/qwen3_vl.py index 1d3929b936..58721303df 100644 --- a/vllm/model_executor/models/qwen3_vl.py +++ b/vllm/model_executor/models/qwen3_vl.py @@ -103,7 +103,7 @@ from .qwen2_5_vl import ( 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 @@ class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]): 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,