services-cli-operations
Operate AReaL 2.0 service CLIs, service lifecycles, model registration, status/log/state inspection, Hermes online RL session/reward wiring, and safe command validation.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- services-cli-operations
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Operate AReaL 2.0 service CLIs, service lifecycles, model registration, status/log/state inspection, Hermes online RL session/reward wiring, and safe command validation.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/online-rl-service-recipes.md13.4 KB
- 参考文档references/service-cli-reference.md18.8 KB
- 参考文档references/troubleshooting.md12.0 KB
- 脚本scripts/check_service_cli.py35.7 KB
03 SKILL.md 正文
AReaL Services And CLI Operations
Use this sub-skill when the task is to operate AReaL 2.0 service surfaces rather than author a new training algorithm or backend. Covered surfaces are:
areal inf: inference gateway/router lifecycle, model registration, model listing, logs, status, state, and OpenAI-compatible inference endpoints.areal agent: agent-service gateway/router/worker/data-proxy lifecycle, session-affine agent calls, status, logs, and state.areal train: safe driver invocation shape for AReaL training scripts.- Service HTTP APIs: inference, agent, training, and weight-update gateway/router/worker/data-proxy contracts.
- Online RL service wiring: start a session, route external or agent traffic through the inference gateway, set rewards, export or refresh trajectories, and shut down cleanly.
Route boundaries
- If the user needs to write or repair an
AgentRunnable,AgentWorkflow,RolloutWorkflow, dataset loader, reward function, or framework adapter, route to../custom-data-rewards-workflows/SKILL.md. - If the service starts but SGLang/vLLM/FSDP/Megatron/Archon workers fail, GPUs are misallocated, weight sync hangs, CUDA/NCCL fails, or backend specs need cluster planning, route to
../distributed-engines-backends/SKILL.md. - This sub-skill may generate commands and validate them, but bundled scripts must not start services, training, model downloads, or credentialed provider calls.
Operating checklist
-
Identify the surface:
areal inf,areal agent,areal train, direct service module, online RL session/reward, or weight update. -
Read
references/service-cli-reference.mdfor command/API contracts, state locations, config precedence, and side effects. -
For online RL, Hermes, or self-evolution loops, read
references/online-rl-service-recipes.mdbefore composing commands. Keep admin keys, agent-service keys, andsk-sess-*session keys distinct. -
Validate command text before handing it to a user or scheduler:
python scripts/check_service_cli.py --command 'areal inf register --model-name qwen --backend sglang:d1 --model-path /models/qwen' python scripts/check_service_cli.py --config service.toml --config-type infThe checker is static: it uses shell parsing and TOML inspection only. It does not import AReaL, contact HTTP endpoints, start services, start training, or read model files.
-
Use
areal <group> status,ps, andlogsas the first live checks once the user has confirmed that live service inspection is allowed. Usereferences/troubleshooting.mdto map symptoms to fixes.
Safety defaults
- Treat every
run,register,deregister,stop,train run, directpython -m areal.v2.*service module, and weight-update request as side-effecting. - Never paste real API keys into generated examples; use shell variables such as
$INF_ADMIN_KEY,$AGENT_ADMIN_KEY, and$SESSION_API_KEY. - Do not use built-in demo keys for non-local services. For production or shared hosts, require user-provided secrets and explicit bind host/port decisions.
- Do not run native repo tests, example services, training jobs, or model downloads as part of this sub-skill. Classify those as native candidates or live operations requiring user approval.
Bundled references
references/service-cli-reference.md— CLI matrix, config/state/log layout, HTTP API summaries, model registration, direct module entrypoints.references/online-rl-service-recipes.md— online proxy and Hermes/session/reward recipes with concrete request payloads.references/troubleshooting.md— service, key, session, reward, state, port, and backend-boundary failure modes.scripts/check_service_cli.py— safe static validator for service CLI commands, TOML snippets, backend specs, and quoted engine/proxy args.
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).