libffcv/ffcv

Converts datasets into a binary `.beton` format with on-the-fly decoding pipelines that leverage JPEG decompression and augmentation on GPU, eliminating I/O bottlenecks common in PyTorch dataloaders. Provides a drop-in `Loader` replacement compatible with existing training loops while supporting configurable field types, custom transforms, and multi-worker parallel loading. Includes pre-tuned training recipes for ImageNet and CIFAR-10 that achieve state-of-the-art throughput and convergence speed on single and multi-GPU setups.

2,986 stars and 422 monthly downloads. Used by 1 other package. No commits in the last 6 months. Available on PyPI.