Transformers Integration Support for timm Models

TimmWrapper has been released, enabling the vast model collection of timm (PyTorch Image Models) to be leveraged within the Hugging Face transformers ecosystem. This allows the latest computer vision models that are not natively supported by transformers to be immediately integrated into existing workflows.

The key features are as follows:

• Pipeline API Usage: Inference tasks such as image classification can be easily performed with timm models via pipeline.
• Auto Classes Compatibility: Fully compatible with transformers' automatic class APIs such as AutoModel.
• Efficient Optimization: Quantization is possible with just a few lines of code, and inference acceleration via torch.compile is supported.
• Training and Fine-tuning: Fine-tuning can be performed using the Trainer API, and adapter techniques such as LoRA can also be applied.
• Bidirectional Workflow: Models fine-tuned in transformers can be brought back into the timm environment for use.