steveseguin/ml-bottleneck

refresh-evidence

Refresh the community benchmark snapshot (data/localmaxxing-snapshot.js) from the public Localmaxxing API, understand how gold cases are selected and what the rows mean, triage a failing weekly refresh, and map new models/hardware so their runs become…

لغة النص الأصلي: الإنجليزية

calibrate-engine

Re-fit or re-anchor the planner's decode/prefill physics (FRAMEWORK_PROFILES, LAYER_OVERHEAD_SCALES, device kernelOverheadScale) against the gold-case corpus, triage outliers and roofline violations, and re-pin the exact-value regression tests. Use when…

لغة النص الأصلي: الإنجليزية

add-hardware

Add or update a device template in DEVICE_TEMPLATES (engine.js) — GPU, Apple/AMD unified-memory system, CPU, NVMe tier, or accelerator — with official peak specs, the right backend overhead scale, default runtime, picker group, and evidence mapping. Use when…

لغة النص الأصلي: الإنجليزية

add-model

Add or update an LLM preset in MODEL_PRESETS (engine.js) from its official config.json so the planner's physics (KV cache, FLOPs, fixed per-layer overhead) are right for it; wires the picker, the evidence rules, and the tests. Use whenever a user asks to…

لغة النص الأصلي: الإنجليزية