Write Once, Run Everywhere with Pytorch Transformers - Pedro Cuenca, Hugging Face

This talk covers the integration of the Hugging Face transformers library, built on pure PyTorch, into various machine learning architectures. The speaker, Pedro Cuenca from Hugging Face, highlights how model definitions serve as reference implementations for training libraries and deployment engines like vLLM and SGLang. The session details the journey towards simplifying the downstream integration of transformers models, enabling seamless interoperability without code modifications. The discussion also touches on future interoperability features with libraries such as MLX and llama.cpp.