Per-layer bits from the loaded model's quant specs — the authoritative,
always-present source for by_bits scaling (each adapted linear carries its
own QuantSpec, so this reflects the real mixed-precision assignment without
needing the optiq_metadata.json sidecar). Keyed by the same module paths
resolveRanks iterates, so it drops straight into opts.bitsMap.
Per-layer bits from the loaded model's quant specs — the authoritative, always-present source for
by_bitsscaling (each adapted linear carries its ownQuantSpec, so this reflects the real mixed-precision assignment without needing the optiq_metadata.json sidecar). Keyed by the same module pathsresolveRanksiterates, so it drops straight intoopts.bitsMap.