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.DataLoaderwith 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