Version v1.15.0 was released in the huggingface/accelerate repository on GitHub. Accelerate is a Python library that lets users run PyTorch training scripts they write themselves in single or distributed CPU, GPU, and TPU environments. When running the same script on multiple devices, the library handles the initialization and launch steps that would otherwise be required.
The README explains that Accelerate abstracts away the boilerplate code needed for multi-GPU and multi-TPU training and for mixed precision (automatic mixed precision) while leaving the rest of the training code intact. The repository description sets the goal of running and training PyTorch models on nearly any device and distributed configuration, and highlights mixed precision including fp8 along with FSDP and DeepSpeed support.
It also plays the role of connecting training infrastructure such as DeepSpeed while wrapping the distributed initialization and launchers. Environment configuration and script execution are handled by an optional command-line tool. The license is Apache-2.0, and the repository has 9922 stars. The repository was created on October 30, 2020.
However, this material does not include release notes for v1.15.0, so what changed in this version cannot be confirmed. The features listed in the repository description and the README are also documented content, not independently reproduced results. The momentum index was 2.1 out of 10, and no figures were given for the novelty, technology, and quality metrics. How much easier distributed training configurations actually become is a separate matter that needs to be checked.