anamnesis 0.4.3

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
25
26
27
28
29
30
31
32
33
34
35

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

//! Precision recovery (dequantization) — built on [`parse`](mod@crate::parse).
//!
//! Each submodule handles one quantization family. All operations take raw byte
//! slices from the parsed `.safetensors` file and produce dequantized output as
//! raw `BF16` bytes suitable for writing back to a standard `.safetensors` file.

#[cfg(feature = "awq")]
pub mod awq;
#[cfg(feature = "bnb")]
pub mod bnb;
pub mod fp8;
#[cfg(feature = "gguf")]
pub mod gguf;
#[cfg(feature = "gptq")]
pub mod gptq;
#[cfg(feature = "pth")]
pub mod pth;
#[cfg(any(feature = "gptq", feature = "awq"))]
mod quant_utils;

#[cfg(feature = "awq")]
pub use awq::dequantize_awq_to_bf16;
#[cfg(feature = "bnb")]
pub use bnb::{dequantize_bnb4_to_bf16, dequantize_bnb_int8_to_bf16};
pub use fp8::{
    dequantize_fp8_to_bf16, dequantize_per_channel_fp8_to_bf16, dequantize_per_tensor_fp8_to_bf16,
};
#[cfg(feature = "gguf")]
pub use gguf::{dequantize_gguf_blocks_to_bf16, dequantize_gguf_to_bf16};
#[cfg(feature = "gptq")]
pub use gptq::dequantize_gptq_to_bf16;
#[cfg(feature = "pth")]
pub use pth::{pth_to_safetensors, pth_to_safetensors_bytes};