Transformers: Standardizing Model Definitions Across the PyTorch Ecosystem - Lysandre Debut & Arthur Zucker, Hugging Face Transformers is relied upon to set the standard of model definitions across the ecosystem. This is relied upon by many tools across the field: - Inference Engines, with vLLM, SGLang, TGI - Training frameworks, with Unsloth, Axolotl, TRL, PyTorch-Lightning, llama-factory, … - and many, many others (400k+ dependents on GitHub) In this session, we’ll go over several aspects: - How are models released? - How are model definitions created from that, and what we’re looking for with transformers acting as the standard-setter - How are these being leveraged across the PyTorch, Python, and general ML ecosystem? - How do we optimize these so that they’re the most friendly towards third-parties? Transformers currently sits at the center of the model-release cycle, with collaborations with the overwhelming majority of model providers. In this session, we aim to shed light on this part of the cycle, so as to give visibility to model releasers, providers, users, and anyone in-between.