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 15, 04:45 UTC
[Multimodal] Use fast processor for Qwen2/2.5-VL (#19789)
Woosuk Kwon committed
Jun 18, 2025 at 15:49 UTC
d49adea1f9f10eec0250a9ff34624b51fb7e3ee9
3 files changed
+4
-3
vllm/model_executor/models/qwen2_5_omni_thinker.py
+1
-1
@@ -150,7 +150,7 @@ class Qwen2_5OmniThinkerProcessingInfo(Qwen2AudioProcessingInfo,
150
min_pixels=min_pixels,
151
max_pixels=max_pixels,
152
size=size,
153
- use_fast=kwargs.get("use_fast")),
153
+ use_fast=kwargs.get("use_fast", True)),
154
**kwargs,
155
)
156
if not hasattr(processor, "audio_token"):
vllm/model_executor/models/qwen2_5_vl.py
+1
-1
@@ -798,7 +798,7 @@ class Qwen2_5_VLProcessingInfo(Qwen2VLProcessingInfo):
798
min_pixels=min_pixels,
799
max_pixels=max_pixels,
800
size=size,
801
- use_fast=kwargs.get("use_fast")),
801
+ use_fast=kwargs.get("use_fast", True)),
802
**kwargs,
803
)
804
vllm/model_executor/models/qwen2_vl.py
+2
-1
@@ -763,7 +763,7 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo):
763
min_pixels=min_pixels,
764
max_pixels=max_pixels,
765
size=size,
766
- use_fast=kwargs.get("use_fast")),
766
+ use_fast=kwargs.get("use_fast", True)),
767
**kwargs,
768
)
769
@@ -808,6 +808,7 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo):
808
size: Optional[dict[str, int]] = None,
809
**kwargs: object,
810
) -> Qwen2VLImageProcessor:
811
+ kwargs["use_fast"] = kwargs.get("use_fast", True)
812
return cached_image_processor_from_config(
813
self.ctx.model_config,
814
**self._get_image_processor_kwargs(min_pixels=min_pixels,