Skill UI

TorchTitan delivers PyTorch-native distributed LLM pretraining with composable 4D parallelism (FSDP2, TP, PP, CP), Float8, torch.compile, and checkpointing so you can scale Llama 3.1, DeepSeek V3, or custom models across 8 to 512+ GPUs.

Torchforge RL Training documents Meta's PyTorch-native torchforge library, which separates algorithms from infrastructure so researchers can explore GRPO/SFT losses, manage Monarch actors, tap TorchTitan and vLLM for scalable distributed runs, and monitor KL/loss curves.