TorchTitan

Tracked

A PyTorch-native platform for generative model pretraining and distributed optimization.

Author PyTorch Open Sourced 2023-12-13 Last Commit Unknown

Overview

TorchTitan provides a production-ready PyTorch-native solution for LLM pretraining, offering multi-dimensional parallelism and practical training recipes.

Key features

  • FSDP2 and various parallelism strategies.
  • Training recipes and deployment guidance.

Use cases

  • Pretraining large generative models and distributed experiments.

Technical highlights

  • Close integration with PyTorch distributed primitives and tooling.