Module packing 

Trit packing/unpacking and FP32→ternary quantization.

Cold (5 trits/byte, base-3) is the disk-friendly format. Hot (4 trits/byte, 2-bpw) is the SIMD-friendly format.

Functions

cold_to_hot

Convert cold-packed trits to hot-packed trits.

pack_cold

Pack trits (i8 ∈ {-1, 0, +1}) into cold 5-trits-per-byte format.

pack_hot

Pack trits into hot 2-bpw format (4 trits per byte).

quantize

Quantize a FP32 vector to ternary trits using BitNet absmean scaling.

unpack_cold

Unpack cold 5-trits-per-byte format back to i8 trits.

unpack_hot

Unpack hot 2-bpw format back to i8 trits.