[Multimodal] Use fast processor for Qwen2/2.5-VL (vllm-project#19789)

Signed-off-by: minpeter <kali2005611@gmail.com>

Author: WoosukKwon

vllm/model_executor/models/qwen2_5_omni_thinker.py

@@ -150,7 +150,7 @@ def get_hf_processor(
                 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"):

vllm/model_executor/models/qwen2_5_vl.py

@@ -798,7 +798,7 @@ def get_hf_processor(
                 min_pixels=min_pixels,
                 max_pixels=max_pixels,
                 size=size,
-                use_fast=kwargs.get("use_fast")),
+                use_fast=kwargs.get("use_fast", True)),
             **kwargs,
         )
 

vllm/model_executor/models/qwen2_vl.py

@@ -763,7 +763,7 @@ def get_hf_processor(
                 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 @@ def get_image_processor(
         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,