fuel-quantized 0.36.7

Backend-agnostic quantized (ggml/gguf block) numerics for fuel: BlockQX types, GgmlType trait, scalar + SIMD reference impls, CPU DynQuantizedStorage adapter.
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

// SPDX-License-Identifier: MIT OR Apache-2.0
//! Backend-agnostic quantized (ggml/gguf block-format) numerics.
//!
//! This crate owns the format-side of fuel's quantization stack:
//!
//! - [`k_quants`] — `BlockQX` block types, the [`GgmlType`](k_quants::GgmlType)
//!   trait, and scalar reference impls of every fmt fuel supports.
//! - [`utils`] — quantization-time helpers shared by k_quants impls.
//! - [`cpu`] — CPU-side adapters that bridge `Vec<BlockQX>` to the
//!   backend-agnostic [`fuel_backend_contract::quantized::DynQuantizedStorage`]
//!   trait, plus the `QuantizedDeviceKernels` impl on
//!   `fuel_cpu_backend::CpuBackendDevice`.
//!
//! Per-backend (CUDA, Metal, future Vulkan) quantized fast paths live in
//! their own crates and implement the same trait pair from
//! `fuel-core-types`.

pub mod awq_repack;
pub mod cpu;
pub mod dequant;
pub mod iq4;
pub mod iq_grid_types;
pub mod iq_grids;
pub mod k_quants;
pub mod tq_mx_types;
pub mod utils;

pub use cpu::{QuantizedType, cpu_from_data, cpu_zeros};
pub use dequant::dequant_ggml_bytes;
pub use k_quants::{
    BlockQ2K, BlockQ3K, BlockQ4_0, BlockQ4_1, BlockQ4K, BlockQ5_0, BlockQ5_1, BlockQ5K, BlockQ6K,
    BlockQ8_0, BlockQ8_1, BlockQ8K, GgmlType, K_SCALE_SIZE, QK_K, QK4_0, QK4_1, QK5_0, QK5_1,
    QK8_0, QK8_1, blocks_as_bytes, blocks_from_bytes_cow, matmul, matmul_f16,
};

// Re-export GgmlDType so downstream callers can write
// `fuel_quantized::GgmlDType` without naming fuel-core-types.
pub use fuel_backend_contract::quantized::{DynQuantizedStorage, QuantizedDeviceKernels};
pub use fuel_ir::quantized::GgmlDType;