srtctl · git:20260914.911a12b · 2026-09-14 · sha256 26f8c2acf8f9a435
srtctl git:20260914.911a12bA
Immutable. This exact content is served forever at /api/v1/blob/26f8c2acf8f9a435.
--- name: srtctl description: Run srt-slurm (srtctl) inference benchmarks on a Slurm cluster. Use when setting up a checkout, writing or migrating a recipe, submitting a job, or reading its results. --- # srtctl `srtctl` turns a YAML recipe into a Slurm job that launches inference workers, a frontend, the services they need, a metrics scraper and a benchmark client, then writes logs and results to `outputs/<job_id>/`. Work from the srt-slurm checkout that holds the cluster's `srtslurm.yaml`; the docs below are paths in that checkout. ## Invariants - `srtctl dry-run -f <recipe>` before `srtctl apply`. A recipe that fails dry-run fails on the cluster. - Recipes are `schema: 2`. Start from `examples/`, never from scratch. Block-style YAML only. - Read `srtslurm.yaml` for model aliases and cluster defaults instead of guessing. Ask the user where model weights live before adding `model_paths`; never download weights unless asked. - A benchmark that "succeeded" with an implausible `Total generated tokens` did not succeed. ## Where to look | Task | Read | |---|---| | Fresh checkout: install, `make setup`, first `srtslurm.yaml` | `docs/installation.md` (Run Setup, Configure srtslurm.yaml) | | Every `srtslurm.yaml` key, including `preflight`, `default_mounts`, aliases | `docs/config-reference.md` (Cluster Config Fields) | | Pick a starting recipe by engine, frontend and topology | `examples/README.md` (Matrix) | | Any recipe field | `docs/config-reference.md`, or the MCP `explain_field` tool | | Prefill/decode split, `nodes: colocate`, GPU fit | `docs/config-reference.md` (roles, Colocating decode on the prefill nodes) | | Dynamo install, `dynamo.source`, sidecar mode | `docs/config-reference.md` (dynamo, Native sidecar mode) | | etcd, NATS, exporters, Mooncake, declared sidecars | `docs/services.md` (Implicit Services) | | Move a v1 recipe (`backend:`, `infra:`, `resources.*_nodes`) to schema 2 | `docs/cli.md` (srtctl migrate), `docs/legacy-v1.md` for the key mapping | | `apply` flags: `--set`, `--serve-only`, `--tags`, `--json`, `--no-preflight` | `docs/cli.md` (srtctl apply) | | Sweeps and override files | `docs/sweeps.md`, `docs/overrides.md` | | What a job wrote and how to read it | `docs/monitoring.md` (Log Structure, benchmark.out) | | Metrics and the per-run dashboard | `docs/component-dashboard.md` | ## MCP `srtctl-mcp` offers the schema tools (`schema_summary`, `explain_field`, `validate_config`, `resolve_config`, `get_config_reference`) anywhere, and the job tools (`submit_job`, `dry_run`, `job_status`, `job_logs`, `list_jobs`, `cancel_job`) when it runs on a login node inside the checkout. See `docs/README.md`.