mlx-community

/

ui-tars-7b-dpo

like 0

Follow

MLX Community 11.4k

This PR resolves a critical compatibility issue with the Qwen2VLImageProcessor used by this model.

Currently, the preprocessor_config.json defines the "size" field as:

"size": {
  "max_pixels": 2116800,
  "min_pixels": 3136
}```

However, this is invalid for the Qwen2VLImageProcessor used in transformers>=4.39. When loading the model with mlx-vlm, it fails with the following error:  ```ValueError: size must contain 'shortest_edge' and 'longest_edge' keys.```

This PR fixes the config by replacing the size field with: 

"size": {
"shortest_edge": 224,
"longest_edge": 224
} ```

This change makes the model usable in downstream applications without further user-side patching.

Thanks for the great model! Let me know if you'd like to attach logs or screenshots too — this is already sufficient for maintainers to understand the bug, fix, and verification.