---

name: benchmarks-maintenance description: "Construct safe Petals benchmark command templates, choose tiny private-swarm maintenance smokes, route focused native checks, and interpret benchmark health signals without launching expensive workflows by default." disable-model-invocation: true metadata: disco-role: operating license: MIT

Petals Benchmarks and Maintenance

Use this sub-skill for Petals benchmark command construction, tiny smoke benchmarks, focused native maintenance checks, or CI-style private CPU swarm upkeep. Prefer safe command templates and checklists; do not launch benchmarks, model downloads, training-like loops, DHT peers, or server processes unless the user explicitly approves runtime, network, cache, and cleanup constraints.

Read references/benchmarking.md, references/native-smoke-tests.md, and references/troubleshooting.md. Use python scripts/build_benchmark_command.py --help to print commands. The builder targets the bundled runner scripts/run_petals_benchmark.py, which should be executed only after approval.

Route production server launch design to server-swarms, client generation to client-inference, prompt-tuning APIs to prompt-tuning, and low-level block correctness to distributed-blocks.