From d49adea1f9f10eec0250a9ff34624b51fb7e3ee9 Mon Sep 17 00:00:00 2001 From: Woosuk Kwon Date: Wed, 18 Jun 2025 15:49:40 -0700 Subject: [PATCH] [Multimodal] Use fast processor for Qwen2/2.5-VL (#19789) --- vllm/model_executor/models/qwen2_5_omni_thinker.py | 2 +- vllm/model_executor/models/qwen2_5_vl.py | 2 +- vllm/model_executor/models/qwen2_vl.py | 3 ++- 3 files changed, 4 insertions(+), 3 deletions(-) diff --git a/vllm/model_executor/models/qwen2_5_omni_thinker.py b/vllm/model_executor/models/qwen2_5_omni_thinker.py index 9344bf8e03..c0ed473103 100644 --- a/vllm/model_executor/models/qwen2_5_omni_thinker.py +++ b/vllm/model_executor/models/qwen2_5_omni_thinker.py @@ -150,7 +150,7 @@ class Qwen2_5OmniThinkerProcessingInfo(Qwen2AudioProcessingInfo, min_pixels=min_pixels, max_pixels=max_pixels, size=size, - use_fast=kwargs.get("use_fast")), + use_fast=kwargs.get("use_fast", True)), **kwargs, ) if not hasattr(processor, "audio_token"): diff --git a/vllm/model_executor/models/qwen2_5_vl.py b/vllm/model_executor/models/qwen2_5_vl.py index 01e85ae805..4faa0d2c36 100644 --- a/vllm/model_executor/models/qwen2_5_vl.py +++ b/vllm/model_executor/models/qwen2_5_vl.py @@ -798,7 +798,7 @@ class Qwen2_5_VLProcessingInfo(Qwen2VLProcessingInfo): min_pixels=min_pixels, max_pixels=max_pixels, size=size, - use_fast=kwargs.get("use_fast")), + use_fast=kwargs.get("use_fast", True)), **kwargs, ) diff --git a/vllm/model_executor/models/qwen2_vl.py b/vllm/model_executor/models/qwen2_vl.py index d5e297ea66..3b939a43e9 100644 --- a/vllm/model_executor/models/qwen2_vl.py +++ b/vllm/model_executor/models/qwen2_vl.py @@ -763,7 +763,7 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo): min_pixels=min_pixels, max_pixels=max_pixels, size=size, - use_fast=kwargs.get("use_fast")), + use_fast=kwargs.get("use_fast", True)), **kwargs, ) @@ -808,6 +808,7 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo): size: Optional[dict[str, int]] = None, **kwargs: object, ) -> Qwen2VLImageProcessor: + kwargs["use_fast"] = kwargs.get("use_fast", True) return cached_image_processor_from_config( self.ctx.model_config, **self._get_image_processor_kwargs(min_pixels=min_pixels,