Fairseq

Meta AI's PyTorch-based sequence modeling toolkit for training high-performance Transformers and language models.

Fairseq (Facebook AI Research Sequence-to-Sequence Toolkit) delivers the core infrastructure for training custom translation, summarization, and speech models. It features optimized implementations for industry benchmarks: RoBERTa, wav2vec 2.0, and BART. The framework scales seamlessly across multi-GPU setups (via distributed data parallel) to handle massive datasets like WMT or LibriSpeech. Developers use its modular API to iterate on neural architectures quickly without rewriting low-level training loops.

https://github.com/facebookresearch/fairseq