nir-rs

Pure-Rust implementation of the Neuromorphic Intermediate Representation (NIR)

Typed NIR graphs in Rust, with opt-in HDF5 .nir read/write that interoperates with the official Python reference. The graph model has no system dependencies; only the hdf5 feature links native libhdf5.

NIR is to spiking neural networks what ONNX is to conventional nets (or GGUF to LLMs): a framework-agnostic graph format so models can move between simulators and hardware without being rewritten.

Why nir-rs?

• Official NIR is primarily Python (neuromorphs/NIR)
• This crate is a pure-Rust graph model with the same wire types and HDF5 layout
• Suitable for embedded, server, and tooling pipelines that should not embed a Python runtime
• Opt-in HDF5 I/O and Serde for debug serialization — enable only what you need

Upstream

• Spec / reference: github.com/neuromorphs/NIR
• Primitives docs: neuroir.org
• Paper: Nature Communications (2024) (DOI 10.1038/s41467-024-52259-9) — cite

Wire compatibility: HDF5 node type strings must match the Python IR (CubaLIF, Conv2d, SumPool2d, …), not informal aliases (CurrLIF, Convolution, …).

Scope

This crate provides:

• The NIR graph model and standard node types
• Reading and writing .nir (HDF5) files
• Round-trip fidelity checks and structural validation
• An idiomatic Rust API (NirGraph, closed NirNode enum, tensors, errors)

This crate does not provide:

• SNN training or simulation
• Mapping graphs onto specific neuromorphic hardware
• Framework-specific importers/exporters (those belong in the tools that produce or consume NIR)

Status

Release notes and the upstream compatibility matrix:

Compatibility claims are fixture-backed (tests/fixtures/).

Install

[dependencies]
nir-rs = "0.4.2"

HDF5 .nir I/O (needs a system libhdf5, or a static build — see File I/O):

[dependencies]
nir-rs = { version = "0.4.2", features = ["hdf5"] }

Debug Serde (JSON / RON / etc.; not a wire standard):

[dependencies]
nir-rs = { version = "0.4.2", features = ["serde"] }

From git — pin a release tag (same tree as the matching crates.io release once the tag exists):

nir-rs = { git = "https://github.com/Limen-Neural/nir-rs", tag = "v0.4.2" }

For unreleased work on the default branch:

nir-rs = { git = "https://github.com/Limen-Neural/nir-rs", branch = "main" }

Quick start

use nir_rs::nodes::{Input, Output};
use nir_rs::{NirGraph, NirNode};

fn main() -> nir_rs::Result<()> {
    let mut g = NirGraph::new();
    g.insert_node(
        "input",
        NirNode::Input(Input {
            shape: vec![4],
            metadata: Default::default(),
        }),
    )?;
    g.insert_node(
        "output",
        NirNode::Output(Output {
            shape: vec![4],
            metadata: Default::default(),
        }),
    )?;
    g.add_edge("input", "output");
    g.validate_structure()?;
    Ok(())
}

File I/O

.nir is the official NIR interchange format: an HDF5 container whose layout is fixed by upstream. Files written here load in Python nir.read, and files written by nir.write load here.

fn main() -> nir_rs::Result<()> {
    let graph = nir_rs::io::read("model.nir")?;
    for (name, node) in &graph.nodes {
        println!("{name}: {}", node.type_name());
    }
    nir_rs::io::write("copy.nir", &graph)?;
    Ok(())
}

I/O is behind the opt-in hdf5 feature, which links native libhdf5. Without that feature the crate has no system dependencies:

Without the feature, io::read / io::write still exist and return NirError::Unimplemented, so downstream code compiles either way.

Round-trip fidelity is graph-level, not byte-level: node names and types, ordered edges, and parameter values are preserved; HDF5 group order and chunk layout may differ from h5py. In-memory dtypes (f32, f64, i64, bool) round-trip exactly; narrower on-disk integers widen to i64 on read. Absent optional fields (v_reset, w_in) use the same defaults as Python so graphs match nir.read in memory.

Example: load, inspect, save a LIF graph

cargo run --example load_inspect_lif --features hdf5
# optional paths:
cargo run --example load_inspect_lif --features hdf5 -- model.nir copy.nir

Default input is tests/fixtures/lif_norse.nir; default output is a PID-qualified file in the system temp directory.

Docker

Images ship a Rust 1.97 + libhdf5 environment with the crate sources and the load_inspect_lif example binary (not an SNN simulator).

docker pull ghcr.io/limen-neural/nir-rs:0.4.2
docker pull ghcr.io/limen-neural/nir-rs:latest

docker run --rm ghcr.io/limen-neural/nir-rs:latest rustc --version
docker run --rm ghcr.io/limen-neural/nir-rs:latest load_inspect_lif

The same tags may also appear on Docker Hub; prefer GHCR for a stable, documented image path.

Local image:

docker build -t nir-rs:local .

Debug serialization

The opt-in serde feature implements Serialize / Deserialize for the graph model. It is independent of hdf5:

[dependencies]
nir-rs = { version = "0.4.2", features = ["serde"] }
serde_json = "1"

JSON is debug/test output, not a NIR interchange standard. Use HDF5 .nir for Python and hardware tooling. JSON cannot represent NaN/infinities faithfully.

Toolchain

Rust 1.97.1 — rust-toolchain.toml, package.rust-version, and CI (Linux / macOS / Windows) all pin the same version.

Develop

cargo fmt --check
cargo test                 # graph model only — no libhdf5
cargo test --features serde
cargo test --all-features  # + HDF5 fixtures / round-trip
cargo clippy --all-targets --all-features -- -D warnings
cargo doc --no-deps --all-features

Wire compatibility is checked against real Python-written .nir fixtures under tests/fixtures/ (BSD-3; see that directory's README). No Python interpreter is required for default builds, tests, or CI.

API and fidelity details: COMPATIBILITY.md, docs.rs/nir-rs.

Citing NIR

If you use NIR in your work (including via this crate), please cite the Nature Communications paper:

@article{NIR2024,
    title={Neuromorphic intermediate representation: A unified instruction set for interoperable brain-inspired computing},
    author={Pedersen, Jens E. and Abreu, Steven and Jobst, Matthias and Lenz, Gregor and Fra, Vittorio and Bauer, Felix Christian and Muir, Dylan Richard and Zhou, Peng and Vogginger, Bernhard and Heckel, Kade and Urgese, Gianvito and Shankar, Sadasivan and Stewart, Terrence C. and Sheik, Sadique and Eshraghian, Jason K.},
    rights={2024 The Author(s)},
    DOI={10.1038/s41467-024-52259-9},
    number={1},
    journal={Nature Communications},
    volume={15},
    year={2024},
    month=sep,
    pages={8122},
}

Machine-readable form: CITATION.cff (GitHub “Cite this repository”).

Acknowledgements

NIR was originally conceived at the Telluride Neuromorphic Workshop 2023 by the authors below (alphabetical order), as listed by upstream neuromorphs/NIR:

• Steven Abreu
• Felix Bauer
• Jason Eshraghian
• Matthias Jobst
• Gregor Lenz
• Jens Egholm Pedersen
• Sadique Sheik
• Peng Zhou

This crate is an independent pure-Rust implementation of that IR; it is not the official Python reference package.

License

Dual-licensed under either:

• Apache License, Version 2.0 (LICENSE-APACHE-2.0 or https://www.apache.org/licenses/LICENSE-2.0)
• MIT License (LICENSE-MIT or https://opensource.org/licenses/MIT)

at your option.