What it is
Miles is a Python framework for RL post-training of LLMs and vision-language models, pairing SGLang for rollout generation with Megatron-LM, FSDP2, or torchtitan for training. It supports fully asynchronous rollout and training, fast in-loop weight updates, low-precision training such as FP8 and NVFP4, LoRA, and fault recovery when inference engines fail. ML engineering teams use it to run GRPO, PPO, SFT, and on-policy distillation on large MoE models across NVIDIA and AMD GPU clusters, including training coding and computer-use agents in sandboxed environments.