torchrec
TorchRec is a PyTorch library designed for optimizing large-scale recommender systems through advanced sparsity and parallelism techniques. It facilitates efficient multi-device training with GPU-sharded embedding tables. Features include diverse sharding strategies, automated sharding plans, and kernels optimized by FBGEMM. TorchRec also supports quantization for precision reduction in training and inference, offering practical applications like DLRM. Used by platforms such as Twitter and Meta, TorchRec is a key tool for modern recommendation system development.
