ferrum-quantization 0.7.3

//! GGUF (GGML Universal Format) reader.
//!
//! Phase 1A scope: parse the file header, expose metadata + tensor descriptors,
//! and load individual tensors as candle `QTensor` (which already handles
//! dequant for every K-quant variant on CPU / Metal / CUDA).
//!
//! Phase 1A added the `GgufFile` reader. Phase 1B added `GgufLinear<B>`.
//! Phase 1C (this commit) adds `GgufLoader<B>` — implements `WeightLoader<B>`
//! against ferrum's HuggingFace-style tensor names by translating to GGUF's
//! `blk.{i}.attn_q.weight` shorthand and handling `qkv_proj` / `gate_up_proj`
//! fusion on the fly.
//!
//! ## Why wrap candle instead of writing a parser from scratch
//!
//! `candle_core::quantized::gguf_file::Content` already implements the full
//! GGUF v1/v2/v3 spec, including all current GGML K-quant variants and
//! Metal/CUDA/CPU dequant kernels. Re-implementing that for ferrum would be
//! 3-5 weeks of work duplicating well-tested code. Instead this module
//! provides a small adapter that:
//!
//!   1. Adds an `mmap`-backed `open(path)` constructor (candle's API takes
//!      a generic `Read + Seek` and pushes file handling to the caller).
//!   2. Provides typed metadata accessors keyed by string (`metadata_string`,
//!      `metadata_u32`, …) so callers don't pattern-match on `Value` everywhere.
//!   3. Documents the GGUF metadata key conventions ferrum relies on
//!      (`general.architecture`, `<arch>.block_count`, …) in one place.

pub mod file;
pub mod linear;
pub mod loader;
pub mod names;

pub use file::GgufFile;
pub use linear::{linear_from_qtensor, GgufLinear};
pub use loader::GgufLoader;
pub use names::{ferrum_to_gguf, gate_up_split_parts, qkv_split_parts};

// Re-exports — callers can import these from `ferrum_quantization::gguf` rather
// than reaching into `candle_core::quantized::*` directly. Keeps the dep
// surface explicit and lets us swap in a native parser later without churning
// downstream call sites.
pub use candle_core::quantized::gguf_file::{TensorInfo, Value, ValueType};
pub use candle_core::quantized::{GgmlDType, QTensor};