Miles v0.1 Paper: Full-Stack Production Post-Training with LoRA and Diffusion Support
aigclink · x · 2026-09-13
The paper (arXiv:2609.08368) and GitHub repo for Miles v0.1 are now public. It is a full-stack, production-ready system for frontier post-training, building on slime's clean design and organizing every stage of the RL loop around one principle: components should be verified, clean, and customizable. Accuracy, efficiency, reliability, and scalability are first-class goals.
Architecture highlights
- Rollout engines built on SGLang; trainer with two backends: NVIDIA Megatron-LM and PyTorch FSDP.
- Three weight-synchronization transports for different deployment topologies.
- Beyond full-parameter RL, it supports LoRA RL, on-policy distillation, supervised fine-tuning, and true-on-policy rollout-training alignment, extending the same architecture to diffusion models.
- End-to-end case study: fully asynchronous agentic RL on GLM-5.2 744B-A40B for terminal-use coding tasks, on 64 NVIDIA GB300 GPUs, with a median step time of 263 seconds over the first 30 steps.
Code at github.com/radixark/miles; project site miles.radixark.com.
Related event: SGLang Team Open-Sources Miles v0.1 Production Post-Training Framework(2 posts)→
More from Infra
- MacBook Pro M4 ComfyUI test: 3-min images, but a 2-sec video takes 9+ hours — Reaperman-13 · 2026-09-13
- Cloud agents beat local Mac scaling, argues altryne amid parallel agent experiments — altryne · 2026-09-13
- Sell an RTX 5090 for a Mac Studio M5 Ultra 96GB? Weighing 1.8 vs 1.2 TB/s Bandwidth for Local AI — unchikuso · 2026-09-13
- Specialized Apple Silicon stacks beat LM Studio by 2x with native MTP speculative decoding — AccBalanced · 2026-09-13
- CUDA-accelerated Minecraft worldgen gets fast, verified bit-accurate against Java — gandamu_ml · 2026-09-13
- Why someone thinks Huawei should build a desktop inference box with 1TB of VRAM — AIFlow_ML · 2026-09-13