OptionalusableBytes: number
Explicit memory budget in bytes (admission control). When set it replaces the machine-derived usable ceiling (ram × WIRED_FRACTION) outright — the budget IS the usable envelope.
OptionalkvScheme: KvSchemeOptions
Active KV-quant scheme; a quantized cache holds more context in the same budget, so the solved ceiling (and admission) must bill it.
Bytes of
.experts.tensors (registry). MoE decode reads only top_k/num_experts of them per token; residency still needs all.