Strips show each layer's activations (64 bins); teal = positive, amber = negative, brightness = magnitude within that layer. Click any layer.
Wherever a script or engineer picks ranks, threads, GPU or precision by rule of thumb.
model.observe(...) runs the job once and measures it; model.decide(state) returns a choice, a confidence and every option's probability.
Above your threshold, apply the change automatically. Below it, hand the suggestion and alternatives to an engineer. STOP_TUNING means it thinks the config is already as good as it gets.
model.tune(...) loops decide → run → measure on this machine and keeps only changes that really made the job faster.
| Jev | HPC-DS1 |
|---|---|
jev.decide(context, options) | model.decide(state, candidates) |
| context: text | state: hardware, kernel and size, config, measured telemetry |
| options: your labels | candidates: any subset of the 9 actions (default: all valid) |
| choice, confidence, probabilities | same, plus stop probability and expected speedup |
| escalate below a threshold | should_escalate; tune(on_escalate=...) |
| deterministic | deterministic (same state → same output) |
Replayed over the real measurements of the held-out problem sizes, starting from every measured configuration. "% of best config" compares the final runtime with the fastest configuration measured for that problem.
Regret = how much slower the best runtime reachable after the chosen move is than the best reachable at all (geometric mean over states). Calibration error above 0.12 (project target) is flagged.