anamnesis 0.5.0

Parse any tensor format, recover any precision — framework-agnostic FP8/GPTQ/AWQ/BnB dequantization, NPZ parsing, and PyTorch .pth conversion for Rust
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

// SPDX-License-Identifier: MIT OR Apache-2.0

//! Precision compression (encoding) — the inverse of [`remember`](mod@crate::remember).
//!
//! Each submodule handles one quantization family on the encode side. All
//! operations take raw `BF16` bytes (the output dtype of every
//! [`remember`](mod@crate::remember) kernel) and produce raw quantized
//! bytes suitable for writing back to a `.safetensors` file alongside
//! per-block / per-row metadata (absmax, `SCB`, …).
//!
//! Phase 5 ships `bnb` only (`NF4` / `FP4` / `INT8`, requires the
//! `bnb` feature). Subsequent encode-kernel families (`FP8`, `GGUF`
//! legacy / `K-quants` / `IQ` / `TQ` / `MXFP4`) land in Phase 7.5 and
//! reuse the [`round_trip`] validation harness introduced here.

#[cfg(feature = "bnb")]
pub mod bnb;
pub mod round_trip;

#[cfg(feature = "bnb")]
pub use bnb::{
    encode_bnb4, encode_bnb4_compute_absmax, encode_bnb4_double_quant, encode_bnb_int8,
    encode_bnb_int8_compute_scb, FP4_CODEBOOK, NF4_CODEBOOK,
};