All A/B, secondary, and thermal configs (excludes the fit memory-probes). Throughput is mean ± population sd over repeats. This is the table the charts derive from.
| arm | method | attn | seq | bs | rank | pack | tok/s real | peak GB | phase | status |
|---|
The raw per-run records (full config + environment manifest + all
metrics, one JSON object per run) live in data/runs.jsonl in the repo — the committed ground
truth. Phases: eager-vs-aotriton-* = primary A/B, secondary-qwen3-4b = RQ2 sweep,
run = the thermal long run. The RQ3 fit_probe records are diagnostic memory
probes and are charted on the Memory page.