A high-performance dataloader implemented in Rust that serves as a drop-in replacement for PyTorch’s DataLoader. Leverages Rust’s concurrency and memory safety to deliver faster data loading throughput for ML training pipelines.

Key Features

  • PyTorch-compatible API — swap out torch.utils.data.DataLoader with minimal code changes
  • Rust performance — multithreaded data loading and prefetching with zero-copy where possible
  • Seamless Python integration — exposes a Python interface while keeping the hot path in Rust

View on GitHub