RadixArk's Miles Lands on the PyTorch Foundation Blog — a PyTorch-Native RL Post-Training Stack With a Two-Decorator API

PyTorch Foundation Blog·low signal

Miles composes SGLang for rollout, Megatron-LM for training, and Ray for orchestration behind a small pluggable trainer, exposing a decorator-based API where you annotate @rollout and @update on an ordinary PyTorch loop and the framework handles distribution and weight sync. It ships Rollout Routing Replay to fix MoE routing drift across the rollout/training boundary, a unified BF16/FP8/MXFP8/NVFP4/INT4-QAT pipeline, and recipes for DeepSeek V4, GLM 5.2, Qwen3.6, Kimi K2.6 and Nemotron 3 Ultra. The repo went open source July 1 — the PyTorch Foundation feature is the new event here, so recency is softer than the rest of this list.