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 9, 12:30 UTC
Don't use `min_pixels`/`max_pixels` from Qwen2VL's processor (#33208)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Jan 28, 2026 at 05:02 UTC
35fb0b8613742f39e68b3ac8a0060a55fde3abb2
1 file changed
+3
-5
vllm/model_executor/models/qwen2_vl.py
+3
-5
@@ -841,8 +841,8 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo):
841
height=image_height,
842
width=image_width,
843
factor=patch_size * merge_size,
844
- min_pixels=image_processor.min_pixels,
845
- max_pixels=image_processor.max_pixels,
844
+ min_pixels=image_processor.size["shortest_edge"],
845
+ max_pixels=image_processor.size["longest_edge"],
846
)
847
preprocessed_size = ImageSize(width=resized_width, height=resized_height)
848
else:
@@ -914,9 +914,7 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo):
914
merge_size = vision_config.spatial_merge_size
915
if max_pixels is None:
916
image_processor = self.get_image_processor()
917
- max_pixels = (
918
- image_processor.max_pixels or image_processor.size["longest_edge"]
919
- )
917
+ max_pixels = image_processor.size["longest_edge"]
918
unit = patch_size * merge_size
919
max_seq_len = max_pixels // (unit * unit)
920