hf2q 0.1.3

Pure Rust CLI for converting HuggingFace models to hardware-optimized formats and serving them over an OpenAI-compatible API on Apple Silicon
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

//! Quantization module — pure-Rust GGML kernel ports for convert-v2.
//!
//! Post-P6 (ADR-033 §P6) the legacy two-pass quantize/calibrate/quality
//! stack is gone; only the [`ggml_quants`] subdir survives. It hosts
//! the pure-Rust kernel ports for every standard ggml ftype the new
//! convert-v2 pipeline emits, plus the APEX algorithmic tiers.
//!
//! Callers should import directly from `crate::quantize::ggml_quants`
//! (e.g. `quantize::ggml_quants::quantizer::Quantizer`,
//! `quantize::ggml_quants::LlamaFtype`,
//! `quantize::ggml_quants::apex::ApexPolicy`).
//!
//! The `imatrix` submodule implements ADR-033 §Pi — in-tree imatrix
//! generation + `.imatrix.gguf` read/write. Phase A (corpus loader,
//! accumulator, gguf writer/loader, file-based `--imatrix <path>`)
//! AND Phase B (forward-pass driver wired through the convert
//! orchestrator, exposed as `--imatrix-corpus <name>`) BOTH
//! SHIPPED 2026-05-19. Stage 3.0 (Gemma 4) + Stage 3b.4 (Qwen 3.5/3.6
//! MoE) are the driver-supported arches; other arches consume
//! pre-computed `.imatrix.gguf` via `--imatrix <path>`.

pub mod ggml_quants;
pub mod imatrix;