radixark/miles

Health v2 · maintenance only

Adoption not part of the score

2263 stars · 403 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Miles is an open-source, enterprise-grade reinforcement learning framework for large-scale LLM and VLM post-training, forked from and co-evolving with slime. It pairs SGLang for high-throughput rollout generation with Megatron-LM (or PyTorch FSDP2) for scalable training, supporting trillion-parameter models with features like fully async RL, P2P weight updates, and low-precision (MXFP8/NVFP4) training.

Use cases

• run GRPO reinforcement learning post-training on large language models
• train frontier-scale MoE models like DeepSeek-V4 with RL
• fine-tune LLMs with SFT, RLHF, or on-policy distillation
• run agentic RL with multi-turn tool-use environments
• post-train multimodal VLMs with reinforcement learning
• train LoRA adapters on trillion-parameter models
• run low-precision RL training with FP8 or FP4 on Blackwell GPUs

When to choose

• you need production-grade RL post-training at frontier/trillion-parameter scale
• you want day-0 recipes for new frontier open models
• you need fast agentic multi-turn rollout with SGLang and async RL
• you want both Megatron-LM and FSDP training backends with LoRA support
• you run multi-node NVIDIA or AMD GPU clusters

When to avoid

• you only need simple fine-tuning of small models on a single GPU
• you want a lightweight RLHF library without distributed training infrastructure
• you lack access to multi-GPU clusters or high-end accelerators
• you need a framework with broad community support outside the SGLang/Megatron ecosystem

Facets

framework · maturity active

machine-learning llm-training reinforcement-learning gpu-computing machine-learning large-language-models reinforcement-learning gpu-computing python cloud post-training rlhf grpo sglang megatron-lm fsdp lora moe agentic-rl on-policy-distillation sft vlm low-precision-training ai-agents linux gpu docker

Member repositories

For agents

markdown · JSON · MCP: product_card(name="radixark/miles")