fix(mm_plugin): regularize images before processor in MiniCPMV4_6Plugin

[ZMXJJ]

What does this PR do?

MiniCPMV4_6Plugin._get_mm_inputs feeds raw image inputs (file paths / bytes / dicts)
directly into image_processor, unlike every other plugin — including the older
MiniCPMVPlugin — which first calls _regularize_images() to load them into PIL
objects and apply image_max_pixels / image_min_pixels resizing.

This has two consequences for MiniCPM-V-4.6:

1. image_max_pixels / image_min_pixels are silently ignored, so images are
   never resized according to the configured pixel bounds.
2. Training/inference crashes on torchvision builds without libjpeg. With
   transformers>=5.7, the (torchvision-backed) image processor decodes path inputs
   through torchvision. On wheels built without libjpeg (e.g. the ROCm torchvision
   builds) this fails during preprocessing with:

   RuntimeError: decode_jpeg: torchvision not compiled with libjpeg support
   

Pre-loading the images to PIL via _regularize_images() makes the processor pass
them through unchanged, which both restores the resize behavior and avoids the
torchvision decode path. The video branch is intentionally left untouched, since
videos are decoded via PyAV (_regularize_videos), not torchvision.

How was this tested?

Reproduced while LoRA fine-tuning MiniCPM-V-4.6 (template: minicpm_v_4_6) on an
AMD ROCm setup (transformers==5.7.0, ROCm torchvision without libjpeg). Before this
change, llamafactory-cli train crashed in dataset preprocessing with the
decode_jpeg error above. After the change, SFT runs to completion and the trained
adapter produces correct image-grounded answers.

Fixes # (no issue)

Before submitting

• Did you read the contributor guideline?
• Did you write any new necessary tests?

[gemini-code-assist]

Code Review

This pull request updates src/llamafactory/data/mm_plugin.py to pre-load images into PIL objects using _regularize_images before passing them to the image processor. This prevents torchvision from decoding image paths directly, which can fail on environments like ROCm where torchvision might be built without libjpeg support. There are no review comments, and I have no feedback to provide.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.