TurboQuant scheme (02d723a:docs/design/turboquant.md): rotation-based KV
quantization, a distinct axis from the uniform/per-layer affine kvQuant
above (composes with it in principle; v1 ships as a standalone CLI-only
runtime lever, same class as uniform kvBits — see
packages/inference/src/generation/autoregressive.ts maybeQuantizeKv). Valid
kBits: {2,4,5,8}; vBits: {2,3,4,5,8} (02d723a:docs/design/turboquant.md's
supported Lloyd-Max tables).
TurboQuant scheme (
02d723a:docs/design/turboquant.md): rotation-based KV quantization, a distinct axis from the uniform/per-layer affine kvQuant above (composes with it in principle; v1 ships as a standalone CLI-only runtime lever, same class as uniform kvBits — see packages/inference/src/generation/autoregressive.ts maybeQuantizeKv). Valid kBits: {2,4,5,8}; vBits: {2,3,4,5,8} (02d723a:docs/design/turboquant.md's supported Lloyd-Max tables).