agilerl
Use AgileRL for reinforcement learning workflows: classical RL training, evolutionary HPO, evolvable networks, multi-agent PettingZoo training, offline/bandit data, and LLM fine-tuning.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- agilerl
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use AgileRL for reinforcement learning workflows: classical RL training, evolutionary HPO, evolvable networks, multi-agent PettingZoo training, offline/bandit data, and LLM fine-tuning.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/benchmarking.md1.8 KB
- 参考文档references/package-overview.md2.9 KB
- 参考文档references/repo-provenance.md1.0 KB
- 参考文档references/repo-routing-metadata.json513 B
- 参考文档references/troubleshooting.md2.8 KB
- 脚本scripts/check_agilerl_install.py2.8 KB
03 SKILL.md 正文
AgileRL
Use this repo skill when a task involves AgileRL, agilerl, evolutionary hyperparameter optimization for RL, Gymnasium/PettingZoo training with AgileRL algorithms, AgileRL evolvable modules/networks, offline RL datasets, contextual bandits, or AgileRL LLM post-training.
Quick Start
- Install:
pip install agilerl. - Optional LLM stack:
pip install "agilerl[llm]"only when the task uses GRPO/CISPO/GSPO, DPO/SFT, vLLM, PEFT, DeepSpeed, or transformer datasets. - Import check:
python -c "import agilerl, importlib.metadata as md; print(md.version('agilerl'))". - Read
references/package-overview.mdfor the algorithm family map,references/troubleshooting.mdfor cross-cutting dependency/backend issues, andreferences/repo-provenance.mdwhen checking whether this skill is stale. - Run
scripts/check_agilerl_install.py --helpfor a safe local capability probe; it does not train, download models, or require the source checkout.
Route By Task
| User task | Read |
|---|---|
Train PPO, DQN, RainbowDQN, DDPG, or TD3 on Gymnasium; wire create_population, train_on_policy, train_off_policy, replay/rollout buffers, or distributed classical training |
sub-skills/training-workflows/SKILL.md |
| Configure tournament selection, mutations, mutable RL hyperparameters, architecture mutation probabilities, or population evolution | sub-skills/hpo-and-mutation/SKILL.md |
Build custom EvolvableModule/EvolvableNetwork objects, net_config, MLP/CNN/LSTM/MultiInput/SimBa configs, actors, critics, or DummyEvolvable wrappers |
sub-skills/evolvable-modules/SKILL.md |
Use PettingZoo multi-agent training, MADDPG, MATD3, IPPO, AsyncPettingZooVecEnv, agent id grouping, or AgileRL wrappers |
sub-skills/multi-agent-and-wrappers/SKILL.md |
Use offline datasets, HDF5/Minari conversion, replay filling, CQL/ILQL, contextual bandits, NeuralUCB, NeuralTS, or BanditEnv |
sub-skills/offline-bandits-data/SKILL.md |
Plan LLM fine-tuning/post-training with GRPO, CISPO, GSPO, LLM PPO, REINFORCE, SFT, DPO, vLLM, DeepSpeed, PEFT, or AgileRL LLM envs |
sub-skills/llm-fine-tuning/SKILL.md |
| Compare benchmark scripts or adapt performance runs safely | references/benchmarking.md |
Common AgileRL Building Blocks
- Algorithms:
PPO,DQN,RainbowDQN,DDPG,TD3,CQN,ILQL,NeuralUCB,NeuralTS,MADDPG,MATD3,IPPO,GRPO,CISPO,GSPO, LLM PPO, REINFORCE,SFT, andDPO. - Population setup usually starts with
agilerl.utils.utils.create_population(...)and a Gymnasium or PettingZoo observation/action space. - Evolution usually combines
TournamentSelection(...)andMutations(...); useHyperparameterConfigandRLParameterfor mutable algorithm attributes. - Network configuration usually uses
net_configwithencoder_config,head_config, and optional recurrent/CNN/MultiInput fields. - Training helpers return trained populations and fitness metrics; full training can be long, so prefer import/config/smoke checks before launching runs.
Safe Validation Pattern
- Check installation and optional dependency availability with
scripts/check_agilerl_install.py. - Validate spaces, algorithm names,
INIT_HP,net_config, and HPO configs with the nearest sub-skill helper script. - Run tiny constructor/import checks before full training.
- Treat demos, tutorials, benchmark launchers, and original repo tests as evidence only; do not depend on the AgileRL source checkout from this generated skill.
- For long training, LLM, GPU, vLLM, DeepSpeed, Minari remote, UCI, or benchmark tasks, document required hardware/data/network first and ask before expensive execution.
Troubleshooting First Reads
- Install/import, Python version, missing optional extras, Torch/Gymnasium/PettingZoo/JAX/W&B issues:
references/troubleshooting.md. - Training shape/config errors:
sub-skills/training-workflows/references/troubleshooting.md. - Mutation/HPO mistakes:
sub-skills/hpo-and-mutation/references/troubleshooting.md. - Network architecture/config mistakes:
sub-skills/evolvable-modules/references/troubleshooting.md. - Multi-agent env/wrapper mistakes:
sub-skills/multi-agent-and-wrappers/references/troubleshooting.md. - Offline/bandit data schema mistakes:
sub-skills/offline-bandits-data/references/troubleshooting.md. - LLM optional dependency and backend mistakes:
sub-skills/llm-fine-tuning/references/troubleshooting.md.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).