[Qwen2VL] Fix missing min_pixels/max_pixels attributes in fast image processor

[tomaszcichy98]

What does this PR do?

Fixes a bug where Qwen2VLImageProcessorFast doesn't set min_pixels and max_pixels instance attributes, breaking compatibility with code that expects these attributes.

The Problem

The slow processor (Qwen2VLImageProcessor) sets self.min_pixels and self.max_pixels as instance attributes. The fast processor converts these to size["shortest_edge"] and size["longest_edge"] but never sets the original attributes.

This causes vLLM (and potentially other libraries) to fail when loading Qwen2-VL/Qwen2.5-Omni models:

File "transformers/models/qwen2_vl/image_processing_qwen2_vl.py", line 92, in smart_resize
    if h_bar * w_bar > max_pixels:
TypeError: '>' not supported between instances of 'int' and 'NoneType'

Why config values don't help

Even if preprocessor_config.json contains min_pixels and max_pixels values, this doesn't fix the issue because:

1. use_fast: false in the config is ignored - transformers 5.0 always auto-upgrades to the fast processor
2. The fast processor receives these values as kwargs but only stores them in self.size dict, not as instance attributes
3. Code accessing processor.min_pixels gets None instead of the config value

The Fix

Set min_pixels and max_pixels attributes after super().__init__() to maintain backward compatibility:

super().__init__(size=size, **kwargs)
self.min_pixels = self.size.get("shortest_edge")
self.max_pixels = self.size.get("longest_edge")

Verification

from transformers import AutoProcessor

# Before fix
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
print(processor.image_processor.min_pixels)  # None
print(processor.image_processor.max_pixels)  # None

# After fix
print(processor.image_processor.min_pixels)  # 3136
print(processor.image_processor.max_pixels)  # 12845056

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: glm_image, qwen2_vl, video_llama_3

[Rocketknight1]

cc @zucchini-nlp maybe?

[zucchini-nlp]

It was removed as part of v5 release in favor of self.size. I believe the error is raised from within vLLM code?

We can ask vLLM team to update the code on their side, not sure if they already bumped to v5

[zucchini-nlp]

fyi @hmellor ,can be added in vllm-project/vllm#30566 if it's indeed not working

[tomaszcichy98]

It was removed as part of v5 release in favor of self.size. I believe the error is raised from within vLLM code?

We can ask vLLM team to update the code on their side, not sure if they already bumped to v5

Yeah it is coming from vLLM @zucchini-nlp

[hmellor]

This is already fixed in vLLM vllm-project/vllm#33208