anamnesis 0.7.4

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
36
37
38
39
40
41
42
43
44
45
46
47
48
49

// 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 = "npz")]
pub mod npz;
pub mod output;
#[cfg(feature = "pth")]
pub mod pth;
#[cfg(any(feature = "gptq", feature = "awq"))]
pub(crate) mod quant_utils;

#[cfg(feature = "awq")]
pub use awq::{dequantize_awq, dequantize_awq_to_bf16};
#[cfg(feature = "bnb")]
pub use bnb::{
    dequantize_bnb_int8, dequantize_bnb_int8_to_bf16, dequantize_bnb4,
    dequantize_bnb4_double_quant, dequantize_bnb4_double_quant_to_bf16, dequantize_bnb4_to_bf16,
};
pub use fp8::{
    dequantize_fp8, dequantize_fp8_to_bf16, dequantize_per_channel_fp8,
    dequantize_per_channel_fp8_to_bf16, dequantize_per_tensor_fp8,
    dequantize_per_tensor_fp8_to_bf16,
};
#[cfg(feature = "gguf")]
pub use gguf::{
    dequantize_gguf, dequantize_gguf_blocks, dequantize_gguf_blocks_to_bf16,
    dequantize_gguf_to_bf16,
};
#[cfg(feature = "gptq")]
pub use gptq::{dequantize_gptq, dequantize_gptq_to_bf16};
#[cfg(feature = "npz")]
pub use npz::{npz_to_safetensors, npz_to_safetensors_bytes};
pub use output::{Bf16Out, F16Out, F32Out, OutputElement};
#[cfg(feature = "pth")]
pub use pth::{pth_to_safetensors, pth_to_safetensors_bytes};