Composer makes it easy to train models faster at the algorithmic level. Use our collection of speedup methods in your own training loop or—for the best experience—with our Composer trainer.
Reproducing science is hard. We draw from the latest research to create a collection of speedup methods that help you stay at the cutting edge. We compose these methods into recipes that make training as fast as possible.
We automatically handle the complexity of adding speed-up methods to your workflow. Experiment with composing these methods by just adding a few lines of code.
Composer comes with standard models and datasets, integrations with HuggingFace, Weights & Biases, DeepSpeed, and TIMM, industrial strength distributed training, streaming dataloaders, and more.
We provide a functional interface so you can enjoy our speedups in your existing training loop. Use Composer in your codebase to keep updated as we continue to roll out more efficient implementations and kernels.
Try out the Composer speedup methods in our notebooks, also available on Colab, and get started in minutes.
Let's build more efficient machine learning together.