torchrec

Why we included this project

TorchRec is the library Meta uses to train and serve its own production recommendation models, so if you are building a large-scale recommender in PyTorch, it is worth a close look. Plain PyTorch lacks the sparsity and parallelism primitives recommendation workloads need, and TorchRec fills that gap, most notably by sharding very large embedding tables across many GPUs with table-wise, row-wise, or column-wise partitioning. A built-in planner generates optimized sharding plans automatically, and pipelined training overlaps data loading, inter-device communication, and compute to keep GPUs busy. It also bundles FBGEMM kernels, quantization for reduced-precision training and inference, common RecSys modules, and example datasets like Criteo click logs. It is the foundation of Meta's DLRM and Twitter's Algorithm ML, which is a strong signal that it scales well beyond toy setups.