Module: aprender::format

Public module of the aprender-core crate.

Source

crates/aprender-core/src/format.rs or directory.

Example

use aprender::format::{ModelCard, AprConverter, ConvertOptions};
// See `cargo doc -p aprender-core --open` for full API reference.

Module summary

aprender::format is the model-file plumbing layer. It owns the .apr binary format (v1 and v2), bidirectional conversion to/from GGUF, SafeTensors, and ONNX, on-disk quantization, sharded checkpoints, golden-output references, diff/lint tooling, and the layout-contract enforcement that keeps row-major APR data safe from column-major imports. If a tensor lives on disk, something in this module handles it.

Key types

The submodules quantize, gguf, onnx, sharded, signing, and layout_contract each have their own focused public APIs — see the rustdoc.

Usage patterns

Pattern 1: Attach a model card to a checkpoint

use aprender::format::{ModelCard, TrainingDataInfo};

let card = ModelCard {
    name: "albor-370m-v1".to_string(),
    description: "370M-param distilled student.".to_string(),
    license: "Apache-2.0".to_string(),
    training_data: Some(TrainingDataInfo {
        dataset: "CSN-Python".to_string(),
        num_samples: 412_345,
        source: Some("https://huggingface.co/datasets/code_search_net".to_string()),
    }),
    citations: vec![],
    intended_use: Some("Code completion and small-scale generation.".to_string()),
};
println!("model card: {}", card.name);

Pattern 2: Diff two checkpoints tensor-by-tensor

use aprender::format::{DiffOptions, DiffCategory};
// AprConverter / diff_models work on real on-disk artefacts;
// see `apr diff a.apr b.apr` for the CLI surface that wraps these.
//
//     let opts = DiffOptions::default();
//     let report = aprender::format::diff_models(&path_a, &path_b, &opts)?;
//     for entry in &report.entries {
//         if entry.category == DiffCategory::Numerical {
//             println!("{}: max_abs={}", entry.tensor_name, entry.max_abs_diff);
//         }
//     }

See also

• serialization — the higher-level AprReader / AprWriter and SafeTensors helpers
• inspect — read-only inspection of model metadata + tensor stats
• models — model implementations that load weights from format
• primitives — the Matrix / Vector types that on-disk tensors deserialize into

Full API

Run cargo doc -p aprender-core --open for the rendered rustdoc, or browse docs.rs/aprender for the published version.