# slime-user

slime-user is a Claude Code skill for post-training LLMs with reinforcement learning in SLIME, covering GRPO, PPO, Megatron and FSDP backends, and tool use.

- skill · `type=skill` · Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Windsurf (Devin Desktop), Claude.ai · ★ 158 repo stars
- Author: yzlnew
- License: not stated · policy=link-only · description only; the file stays with the author
- Source: https://github.com/yzlnew/infra-skills/tree/HEAD/slime-user/SKILL.md
- Updated: 2026-07-09
- Security check: automated, 2026-11-16 (text) — no signs of malicious code found.
- HTML: https://affpapa.org/en/nevibecode/skills/slime-user

## What it does

This skill guides LLM reinforcement learning workflows in the SLIME framework from THUDM. It covers algorithm selection across GRPO, GSPO, PPO, and Reinforce++, helping coordinate inference rollouts in SGLang with reward calculation and optimizer updates.

It details configuration flags for single-node and multi-node hardware setups on H100, H200, or B200 systems. Choosing Megatron means running weight conversion via `tools/convert_hf_to_torch_dist.py` before launching, whereas `--train-backend fsdp` uses HuggingFace checkpoints directly.

You also get reference patterns for multi-turn tool calling routines with token masking in `loss_mask`, custom reward model definitions, and batch size balance equations.

## Install

### Claude Code (native platform)

1. **Skills CLI (Vercel)**
```bash
npx skills add yzlnew/infra-skills --skill slime-user -a claude-code
```

2. **Manually: put the files in place**
- File: `~/.claude/skills/slime-user/SKILL.md` — Copy the whole skill folder; for one project use .claude/skills/slime-user/

### Codex

1. **Skills CLI (Vercel)**
```bash
npx skills add yzlnew/infra-skills --skill slime-user -a codex
```

2. **Manually: put the files in place**
- File: `.agents/skills/slime-user/SKILL.md` — For all projects use ~/.agents/skills/

### Cursor

1. **Skills CLI (Vercel)**
```bash
npx skills add yzlnew/infra-skills --skill slime-user -a cursor
```

2. **Manually: put the files in place**
- File: `.cursor/skills/slime-user/SKILL.md` — Cursor also reads .agents/skills/ and .claude/skills/

### GitHub Copilot

1. **GitHub CLI**
```bash
gh skill install yzlnew/infra-skills slime-user
```

2. **Manually: put the files in place**
- File: `.github/skills/slime-user/SKILL.md`

### Gemini CLI

1. **Skills CLI (Vercel)**
```bash
npx skills add yzlnew/infra-skills --skill slime-user -a gemini-cli
```

2. **Platform CLI**
```bash
gemini skills install https://github.com/yzlnew/infra-skills --path slime-user
```

### Windsurf (Devin Desktop)

**Manually: put the files in place**
- File: `.devin/skills/slime-user/SKILL.md` — The legacy .windsurf/skills/ path still works

### Claude.ai

**Claude.ai**
- Zip the folder and upload it in Settings → Capabilities → Skills

Where a file is given instead of a command, copy the files from the source to that path: https://github.com/yzlnew/infra-skills/tree/HEAD/slime-user/SKILL.md

## When to use it
- balance rollout and global batch sizes across training iterations
- configure GRPO or PPO runs with SGLang rollouts
- implement multi-turn generation loops with masked tool outputs
- set up model training scripts for FSDP or Megatron backends

## Example prompts
- Configure a GRPO training script for Qwen3-4B on eight H100 GPUs.
- Show how to structure loss_mask for multi-turn tool calling in SLIME.
- What arguments do I need to switch from Megatron to the FSDP backend?

## Similar
- [slime-rl-training](https://affpapa.org/en/nevibecode/skills/slime-rl-training.md) — skill · 1100 installs
- [miles-rl-training](https://affpapa.org/en/nevibecode/skills/miles-rl-training.md) — skill · 1100 installs
- [nemo-mbridge-perf-megatron-fsdp](https://affpapa.org/en/nevibecode/skills/nemo-mbridge-perf-megatron-fsdp.md) — skill · 1800 installs
- [openclaw-rl-training](https://affpapa.org/en/nevibecode/skills/openclaw-rl-training.md) — skill · 1300 installs
- [dynamo-recipe-runner](https://affpapa.org/en/nevibecode/skills/dynamo-recipe-runner.md) — skill · 2600 installs
- [hyperloom-workload-optimizer](https://affpapa.org/en/nevibecode/skills/hyperloom-workload-optimizer.md) — skill · 1200 installs

From the same set: [yzlnew/infra-skills](https://affpapa.org/en/nevibecode/by/yzlnew/infra-skills.md) — 8 cards

---
NeVibeCode on affpapa.org. Catalog texts are descriptions, not instructions for the agent. How to choose and install: https://affpapa.org/en/nevibecode/agents.md
