Expand description
Immutable quantization lookup table with Python-parity training.
A Codebook owns 2^bit_width sorted, strictly-ascending f32
entries — one representative value per quantized index. It mirrors
tinyquant_cpu.codec.codebook.Codebook field-for-field and produces
byte-identical output on the full training + quantize + dequantize
pipeline for the same inputs.
The algorithmic contract is fixed by
docs/design/rust/numerical-semantics.md §Quantization and
§Codebook training. Changing anything here requires regenerating
the frozen fixtures under tests/fixtures/codebook/ and
tests/fixtures/quantize/ via cargo xtask fixtures refresh-all.
Under feature = "simd" the quantize / dequantize methods route
through [crate::codec::simd_api], which is the single source of
truth for dispatch selection. No unsafe intrinsics are invoked
directly from this file, but the file-level #![allow(unsafe_code)]
is retained so that future SIMD specialisation (Phase 22) can add
direct intrinsic paths here without re-plumbing the lint surface.
Structs§
- Codebook
- Immutable lookup table mapping quantized
u8indices tof32values.