TorchNeuron: running HF models on AWS Trainium is easier than ever

TorchNeuron is a new native backend for PyTorch that integrates AWS Trainium as a native device. This plugin enables support for eager mode, torch.compile, and native distributed APIs, simplifying the process of running models on this specialized hardware. As a result, Hugging Face libraries like Transformers, Diffusers, and TRL can now run directly on Trainium without the need for the specialized optimum-neuron library. The integration streamlines workflows for both inference and training, making it easier to leverage Trainium for cost-effective AI model development. Code examples are provided for text generation and supervised fine-tuning using standard library scripts.