Browse verified engine recipes, stand one up on your own GPU.
engine-room is the action half over a verified recipe layer: pick a recipe, then provision, launch, measure, and roll back — all on your own rig. Resolve against the live hardware, materialize the pinned artifacts, gate on a real performance baseline, and undo every irreversible step newest-first.
Detect the rig (no DB needed)
er rig
Browse the catalog
er list
Resolve — no side effects
er preflight <slug>
Plan it (dry-run by default)
er provision <slug>
Knowledge and action, split clean
A verified recipe says what to build; engine-room stands it up correctly on your rig and proves it works.
Polymorphic recipes
Four kinds, each measured on its own terms: launchable-server (tok/s, it/s), batch-producer (output + wall-clock), modifier (a measured delta), router-fleet (per-upstream health).
Verified + resolvable
The recipe layer (tensor-engine-knowledge/engines.db) is the trusted, sourced input — read-only, never written. preflight resolves it against the live rig with a capability + resolvability check before anything runs.
Reproducible + validated by design
Model-keyed baselines with a compatibility band, a correctness gate before any performance claim. Reproducibility is honestly partial today — content-addressed vendor pins are Goal 1, not an overclaim.
Stdlib-only + named compensators
Zero third-party deps in the core (sqlite3 + subprocess). Every irreversible step is recorded on a ledger before it runs and has a named, newest-first, identity-verified undo.
Usage
Python >=3.10, stdlib-only core. Clone gives you the executor; point er at a separately-provided recipe DB for the catalog.
Install (editable) — gives the er command
pip install -e .
# or run without installing:
# python -m executor.cli rigPoint at the recipe layer
# tensor-engine-knowledge/engines.db is a separate artifact
export ER_RECIPES_DB=/path/to/engines.db
# or pass --db on any command Browse + resolve (read-only)
er rig # detected rig fingerprint (no DB needed)
er list # list recipes (--all for everything)
er show # one recipe: baselines, artifacts, compensators
er preflight # resolve + capability check, no side effects Provision — dry-run by default
er provision # full plan + every compensator (no side effects)
er provision --execute --model # live: materialize -> launch -> measure
er status # the ledger: provisioned instances
er teardown # roll back, compensators newest-first