radixark/miles
Miles is an enterprise-facing reinforcement learning framework for LLM and VLM post-training, forked from and co-evolving with slime.
Python · ★ 2,615 · Fork 447 · 在 GitHub 查看
在手机上浏览源码、订阅每日 AI 趋势日报与项目架构深度解读:获取 CopoHub 客户端(iOS & HarmonyOS)→
GitHub Trending 上榜记录
radixark/miles 已在 GitHub Trending 每日榜单上榜 2 天,最高排名第 14,首次上榜 2026-09-04,最近一次上榜 2026-09-05。
| 日期 | 排名 | Star | 当日增长 |
|---|---|---|---|
| 2026-09-05 | #14 | 2,615 | +64 |
| 2026-09-04 | #14 | 2,540 | +55 |
同期上榜的项目
- mattpocock/skills – Skills for Real Engineers. Straight from my .agents directory.
- affaan-m/ECC – The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
- DietrichGebert/ponytail – Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
- NousResearch/hermes-agent – The agent that grows with you
- fmtlib/fmt – A modern formatting library
- anthropics/skills – Public repository for Agent Skills
Miles 项目深度剖析与商业评估报告
Miles 项目深度剖析与商业评估报告
报告日期:2026-09-04 | 分析对象:radixark/miles v0.1 | 性质:企业级 LLM/VLM 后训练强化学习框架
---
1. 核心定位与商业落地价值
1.1 根本痛点洞察
Miles 所解决的并非"能不能做 RL",而是 "RL 从研究原型到生产系统之间那道巨大的工程鸿沟" 。当前开源 RL 框架(如 TRL、OpenRLHF)在单机或小规模集群上表现尚可,但一旦触及以下场景便集体失效:
万卡级分布式训练:Megatron-LM 的 TP/PP/CP/EP 多维并行与 SGLang rollout 集群的协同调度 万亿参数(1T+)模型的在线权重热更新:传统 checkpoint 保存/加载路径在参数规模下耗时以分钟计,导致 RL 训练中 rollout 与 training 之间产生无法容忍的"气泡" MoE 模型的训练/推理一致性:专家路由在 rollout(SGLang)与 training(Megatron)之间天然不一致,直接导致 loss 震荡甚至发散 多轮 Agentic 任务的长序列管理:需要维持跨轮次的 KV cache、工具调用状态与 reward 追踪
Miles 的战略定位是 "RL 训练的操作系统层" ,而非又一个算法库——它提供的是 RL 生产化所需的全部基础设施,包括调度、容错、精度控制与观测性。
在 README 中展示上榜徽章
[](https://github.fq6825.top/repo/radixark/miles)