Hugging Face has announced the stable release of Transformers v5, a significant update that introduces several optimizations and new features.

Improved Performance

The new version promises significant performance increases, especially for Mixture-of-Experts (MoE) models, with estimated speedups between 6x and 11x. This should translate into reduced inference times and greater resource efficiency.

Simplified Tokenizers

The API for tokenizers has been simplified, eliminating the distinction between "slow" and "fast" tokenizers. The new approach should make it easier to integrate and use tokenizers, with an explicit backend and improved performance.

Dynamic Weight Loading

Dynamic weight loading has been optimized, making it faster and allowing the use of MoE with quantization, tensor parallelism (tp), and PEFT (Parameter-Efficient Fine-Tuning).

A migration guide is available to facilitate the transition to the new version. Hugging Face encourages users to report any issues encountered while using Transformers v5.