pytorch/torchtitan
PyTorch-native platform for training generative AI models
- Chart
- Not on a chart right now
- Stars
- 5.9k
What it is
A PyTorch-native platform for large-scale training that takes models from pretraining through supervised fine-tuning to reinforcement learning in one codebase. It composes FSDP2, tensor, pipeline, context, and expert parallelism without requiring model rewrites.
Who it is for
ML researchers and engineers building or experimenting with large-scale generative AI training pipelines.
Why it is trending
Released v0.4.0 on October 28, 2026, it stands out for combining composable N-D parallelism, Python recipe configs, and an integrated pretraining-to-RL stack via TitanRL.
Main features
- Composable FSDP2, TP, PP, CP, and EP parallelism via DeviceMesh and DTensor
- Unified pretraining-to-RL stack with TitanRL sharing model definitions with vLLM
- Python recipe-based configuration and component overrides
- MXFP8 and NVFP4 low-precision training with distributed checkpointing
Summarized on November 19, 2026 from the README and repository data. It can be wrong; check the source before relying on it. Report a wrong summary