tuning-optimization
Build and validate Optuna hyperparameter optimization commands, study reuse, and report artifacts for RL Zoo.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- tuning-optimization
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Build and validate Optuna hyperparameter optimization commands, study reuse, and report artifacts for RL Zoo.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/optuna-study-artifacts.md3.4 KB
- 参考文档references/optuna-tuning.md6.8 KB
- 参考文档references/troubleshooting.md4.2 KB
- 脚本scripts/tuning_command_builder.py11.0 KB
03 SKILL.md 正文
tuning-optimization
Use this sub-skill when a task is specifically about RL Zoo Optuna hyperparameter optimization (HPO): starting -optimize / --optimize-hyperparameters, choosing sampler/pruner/budget flags, coordinating shared studies, loading a stored trial, or interpreting HPO report artifacts.
Route first
- Base training flags, run budgets, device/thread/vector-env choices, resume/checkpoint flags, and ordinary training lifecycle:
../training-cli/SKILL.md. - YAML/Python hyperparameter files,
-paramsoverride grammar, wrappers/callbacks, and config syntax:../config-hyperparams/SKILL.md. - Plotting or benchmarking the resulting files:
../plotting-benchmarking/SKILL.md. - Optional dependency and console-entry caveats:
../../references/install-and-environment.md.
Runtime entry points
Prefer installed-package commands:
python -m rl_zoo3.train ... --optimize-hyperparametersis the safest base command for HPO because it targets the train module directly.rl_zoo3 train ... --optimize-hyperparametersis equivalent when the console entry point and optional plot-import path are available.
Use the bundled non-executing helper to build and validate command strings:
python scripts/tuning_command_builder.py --help
The helper does not start training or optimization. It validates high-risk HPO combinations such as --pruner halving with too few jobs, --sampler auto without optunahub, unsupported HPO algorithms, and --trial-id without a reusable study.
References
references/optuna-tuning.md— HPO command patterns, flag semantics, lifecycle, and search-space notes.references/optuna-study-artifacts.md— report files, per-trial evaluation logs, storage reuse, and deprecated study-parser handling.references/troubleshooting.md— common HPO failures and fast fixes.
Operating boundaries
- HPO is distinct from ordinary training: always make the training timestep budget and Optuna trial budget explicit before proposing a real run.
- Treat search spaces as algorithm-specific; if an algorithm is not in the HPO sampler/converter map, do not claim it is tunable through RL Zoo HPO without extending that map.
- Do not copy or rely on deprecated source helper scripts for new workflows. Use storage plus
--study-name/--trial-idfor trial reuse.
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).