setup-and-cluster
Guides agents through RLinf installation target selection, environment probes, Ray cluster startup, Hydra cluster configuration, component placement, heterogeneous node groups, execution modes, and scheduler API orientation.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- setup-and-cluster
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Guides agents through RLinf installation target selection, environment probes, Ray cluster startup, Hydra cluster configuration, component placement, heterogeneous node groups, execution modes, and scheduler API orientation.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/configuration-and-placement.md6.6 KB
- 参考文档references/installation-and-cluster.md8.8 KB
- 参考文档references/scheduler-api.md6.4 KB
- 参考文档references/troubleshooting.md6.7 KB
- 脚本scripts/render_cluster_plan.py7.7 KB
- 脚本scripts/rlinf_env_probe.py9.9 KB
03 SKILL.md 正文
setup-and-cluster
Use this sub-skill when the task is to prepare, inspect, or explain RLinf runtime setup: choosing an install target, verifying that rlinf/Ray/Torch/backends import, starting or checking a Ray cluster, reading cluster: YAML, planning component placement, reasoning about heterogeneous node groups, or orienting around the scheduler API.
Do not use this sub-skill for task-specific training/evaluation commands, metrics/checkpoints/debug sessions, or adding new RLinf models/environments/algorithms. Route those to the sibling sub-skills embodied-workflows, reasoning-agent-workflows, operations-evaluation-debugging, or extension-development as appropriate.
Read these bundled references
- Installation and cluster readiness — install target choices, Docker-vs-UV decisions, safe probes, and single-/multi-node Ray procedures.
- Configuration and placement — Hydra basics,
cluster.num_nodes, short and node-group placement syntax, execution modes, and heterogeneous hardware planning. - Scheduler API mental model —
Cluster,Worker,WorkerGroup, placement strategies, channels, and how config becomes Ray actors. - Troubleshooting — missing Ray, stale Ray environment variables, network/NIC issues, placement over-allocation, optional accelerator packages, and Python dependency problems.
Use these bundled scripts
Run scripts from this sub-skill directory or by absolute path; they do not depend on the original RLinf checkout unless you explicitly pass --repo-root.
python scripts/rlinf_env_probe.py --help
python scripts/rlinf_env_probe.py --json
python scripts/render_cluster_plan.py --help
python scripts/render_cluster_plan.py path/to/config.yaml
scripts/rlinf_env_probe.pysafely probes Python, package imports, Ray command/status, relevant RLinf environment variables, and CUDA/Torch availability.scripts/render_cluster_plan.pyparses a YAML config without importing RLinf and summarizescluster.num_nodes,component_placement,node_groups,env_configs, and hardware blocks.
Operating rules
- Verify before advising: run the bundled environment probe,
ray status, or the placement-render script when the user gives an environment/config to inspect. - Treat installation as a planning decision unless the user explicitly authorized environment creation. This sub-skill describes install selectors but should not launch dependency installation by default.
- Set
RLINF_NODE_RANKbeforeray starton every node. If the value or Python environment changes after Ray starts, stop and restart Ray on that node. - Keep
cluster.num_nodesequal to the real joined node count and use zero-based node ranks. - For heterogeneous clusters, prefer explicit
node_groupsandcomponent_placementnode-group form; reserve short form for homogeneous accelerator layouts. - Never call
ray.init()before RLinf'sClusterin driver code; RLinf owns the Ray namespace and manager actors.
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).