Module ternary 

BitNet b1.58 ternary quantization codec.

Encodes FP32 vectors into ternary trits {-1, 0, +1} using per-vector mean-absolute-value (absmean) scaling. Two storage layouts are supported:

• Cold (QuantMode::TernaryPacked): 5 trits/byte via base-3 packing (1.6 bpw). Optimal for disk; decompressed on page-in.
• Hot (QuantMode::TernarySimd): 2 bpw, 4 trits/byte, suitable for direct SIMD load. VectorCodec::encode produces hot format by default.

Re-exports

pub use codec::TernaryCodec;

pub use codec::TernaryQuantized;

pub use codec::TernaryQuery;

pub use packing::cold_to_hot;

pub use packing::pack_cold;

pub use packing::pack_hot;

pub use packing::quantize;

pub use packing::unpack_cold;

pub use packing::unpack_hot;

pub use simd::ternary_dot;

Modules

codec

TernaryCodec — BitNet b1.58 ternary VectorCodec implementation.

packing

Trit packing/unpacking and FP32→ternary quantization.

simd

SIMD ternary dot product with runtime CPU-feature dispatch.