Torchtitan is a PyTorch-native platform for training generative AI models, developed in the open by the PyTorch team on GitHub. It serves as both a reference implementation and a working toolkit for large-scale language model training, showing how far you can get with native PyTorch features alone — no heavyweight external training framework required. The project has earned more than 5,800 stars on GitHub and is actively developed, with its latest updates in 2026.
Torchtitan is built to be readable, modular and hackable: models, parallelization strategies and training components are cleanly separated so researchers can swap pieces in and out and understand exactly what each part does. It demonstrates state-of-the-art distributed training techniques — fully sharded data parallelism, tensor, pipeline and context parallelism — in pure PyTorch, and it is the codebase the PyTorch team itself uses to validate new distributed features. Pre-training recipes for the Llama model family are included, along with detailed performance and memory profiling, so teams can reproduce large training runs and then adapt them to their own models and clusters. Recent work has extended it in directions such as multimodal training and reinforcement-learning post-training recipes, reflecting how LLM pipelines have evolved beyond plain pre-training. Because everything is plain PyTorch, debugging and experimentation use the tooling ML engineers already know.
Torchtitan is completely free and open source under the BSD-3-Clause license. You only pay for your own compute infrastructure.
Torchtitan is for ML engineers and researchers who train large language models and want a clean, modern PyTorch codebase to learn from, benchmark with, or build production training pipelines on.