⚡ Lightning Launch

Agentic RL, ready to run.

Bring your agent. Define your reward. Train with ms-swift.

中文 · Quickstart · Adapters · Rewards · Deployment · Why this project?

Lightning Launch connects existing agents to Agent Lightning + ms-swift/Megatron with a compact, inspectable workflow: task → agent/tool rollout → external reward → policy update → updated-policy rollout. Keep your agent's tool loop and choose a local process, container, or remote worker.

One-command learning demo / 一条命令训练并对比

On Linux with two NVIDIA GPUs (tested: 2 × RTX 3060 12 GB), CUDA 12.8 toolkit and compatible drivers, Python 3 and uv, run from the checkout:

bash scripts/run_learning_demo.sh .runs/hermes-mbpp

自动安装独立 GPU runtime 和固定版本 Hermes，下载模型和 MBPP，校验数据，完成 24 步原生 ms-swift GRPO、训练前后配对评估，并生成 .runs/hermes-mbpp/experiment/REPORT.md 和 JSON 明细。无需自行准备数据或修改 YAML。 安装和模型下载需要网络；运行目录包含模型缓存、工作区及 checkpoint，请预留磁盘空间。 已有兼容环境可通过 AGL_TRAIN_PYTHON、AGL_HERMES_ROOT、AGL_MODEL 复用。 每次运行使用新目录；完整设置见 demo guide。

Measured Hermes + Qwen3.5-0.8B results (24 GRPO steps):

Split	Tasks × samples	Assertion pass fraction: base → trained	Fully solved: base → trained
Train	8 × 2	22.92% → 37.50%	3/16 → 6/16
Validation	4 × 2	25.00% → 37.50%	2/8 → 3/8
Test subset	8 × 2	12.50% → 12.50%	2/16 → 2/16

80 次评估均完成，176 个保存的工作区重算 reward 无差异。训练集和验证集有提升， 测试子集持平。样本较少，断言对 agent 可见，结果用于展示可工作的学习流程， 不代表隐藏测试基准或广泛泛化；重复采样也不是新增独立题目。 完整实测记录 · 机器可读证据

CPU smoke check

git clone https://github.com/0KEAHA/lightning-launch.git
cd lightning-launch
python3.11 -m venv .venv && . .venv/bin/activate
pip install -e .
lightning-launch init my-agent-rl
lightning-launch demo --output .runs/demo

The demo runs a small CPU policy-gradient example to exercise rewards and audits. Real LLM training requires a compatible GPU runtime and model, downloaded separately. Follow the complete walkthrough, then:

lightning-launch doctor --config configs/quickstart.yaml
lightning-launch run --config configs/quickstart.yaml
lightning-launch rewards .runs/runs/YOUR_RUN

Public datasets + real agent products

lightning-launch recipe gsm8k ./gsm8k-opencode --agent opencode
lightning-launch recipe mbpp ./mbpp-hermes --agent hermes

Downloads, checksums, splits, evaluators and reference self-checks are automatic. Use --hardware two-gpu-shared for the tested small-GPU topology. Install the chosen product and GPU runtime, then evaluate or train with the generated config. See product setup, dataset service and real-data evidence. For measured learning effects, see the regression diagnosis and fixed-budget comparison.

Built for the integration work around RL

• Keep ms-swift. Native GRPO, Megatron parallelism, LoRA, policy synchronization and checkpoints through explicit Swift plugins.
• Bring an existing agent. Hermes, Codex, Claude Code and OpenCode adapters; LangGraph/LangChain, OpenAI Agents, AutoGen, smolagents, CrewAI and DeepSeek Harness SDK entries; custom commands and remote workers.
• Own your reward. Nine built-in scorers, correctness-gated cost bonuses, versioned composition, Python/command/HTTP extensions, component audits and paired policy comparisons.
• See what actually trained. Original token IDs/logprobs, tool feedback masks, complete-group validation, Store readback and acceptance reports.
• Keep the checkout small. Source, examples, documentation and tests. No Swift images, model weights, GPU runtime bundles or private deployment configuration.

“Adapter included” is not a claim that every release, provider feature or deployment has been tested. Validation separates real GPU training, installed SDK/tool probes and unverified combinations.

Choose your path

You have…	Start here
No prepared dataset	Public data recipes
A compatible Swift environment	GPU quickstart
An existing agent framework	Adapter contract and examples
A custom scorer, test runner or reward service	Reward interface
Containers, remote services or Kubernetes jobs	Deployment contract
Questions about upstream differences	Contribution and attribution

Relationship to Agent Lightning

Agent Lightning already provides real-harness training and proxy-based integration. Our contribution is the ms-swift/Megatron backend integration and ready-to-run operating layer: adapters, trajectory replay, reward management, portable tasks, isolated deployment and testable evidence. We do not claim a new RL algorithm, benchmark superiority or official Microsoft/DeepSeek affiliation. See the source-backed comparison.

Contribute

Start with CONTRIBUTING.md. Add a real tool-roundtrip test for a new adapter, preserve exact trajectory and reward semantics, and document tested versions. MIT license; upstream notices are retained in NOTICE.