GRPO fail for inference

[John1231983]

I trained GRPO and ran inference on the trained checkpoint, however I got the error

  File "qwen2_vl/image_processing_qwen2_vl_fast.py", lin
e 163, in _preprocess_image_like_inputs
    batch_feature = self._preprocess(images, **kwargs)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "qwen2_vl/image_processing_qwen2_vl_fast.py", lin
e 242, in _preprocess
    patches = patches.view(
              ^^^^^^^^^^^^^
RuntimeError: shape '[1, 1, 2, 3, 35, 2, 16, 35, 2, 16]' is invalid for input of size 7907424

I used both Qwen3VL and Qwen2.5VL and got the same issue. How can I fix the issue?

[2U1]

It seems the setting for the inference and training is different. Check for the patch size and other things.

[John1231983]

I cannot find patch setting. In my script, I changed

--max_completion_length 1024
--max_prompt_length 2048
--image_min_pixels $((16 * 32 * 32))
--image_max_pixels $((2000 * 32 * 32)) \

[John1231983]

Same issue #206

[John1231983]

@2U1 it fixed by

 processor = AutoProcessor.from_pretrained(
        processor_path,
        trust_remote_code=True, 
        min_pixels=args.min_pixels, 
        max_pixels=args.max_pixels,
    )
    processor.image_processor.do_resize = True

I’m not sure why it’s necessary to set processor.image_processor.do_resize = True in the inference code for GRPO mode, while in the SFT phase I didn’t include this line and the code still worked. Do you know why it works?

[John1231983]

@2U1 why did you set processor.image_processor.do_resize = False in GRPO? It’s different from the SFT setup.

[2U1]

Because if you see here, sft uses the resize in the dataset it self.

Qwen-VL-Series-Finetune/src/dataset/sft_dataset.py

Lines 191 to 200 in 93fef86

	inputs = processor(
	text=[user_input],
	images=images,
	videos=video_datas_for_turn,
	padding=False,
	do_resize=False,
	return_tensors='pt',
	**video_kwargs,
	video_metadata=video_metadatas_for_turn,
	)

However, in GRPO this process is held in the trainer so, I've set the option outside the trainer so it can act the same way as SFT.