onnx-rs

Zero-dependency ONNX model parser and encoder in Rust. Parses .onnx files (protobuf wire format) directly into typed Rust structs without relying on prost, protobuf, or any other crate.

Usage

use onnx_rs::ast::*;

// Parse
let bytes = std::fs::read("model.onnx").unwrap();
let model = onnx_rs::parse(&bytes).unwrap();

let graph = model.graph.as_ref().unwrap();
for node in &graph.node {
    match &node.op_type {
        OpType::Conv => println!("found conv: {}", node.name),
        OpType::Relu => println!("found relu: {}", node.name),
        op => println!("{}: {}", op, node.name),
    }
}

// Encode back to protobuf
let bytes = onnx_rs::encode(&model);
std::fs::write("output.onnx", &bytes).unwrap();

Features

• Zero dependencies — hand-rolled protobuf wire format decoder
• Typed AST — OpType enum with 170+ standard ONNX operators, Custom(String) fallback for vendor ops
• Full ONNX IR — Model, Graph, Node, TensorProto, TypeProto (Tensor/Sequence/Map/Optional/SparseTensor), Attribute, Function, TrainingInfo
• All 23 data types — including Float8 (e4m3fn, e5m2, etc.), Int4, Uint4, BFloat16
• Roundtrip encode/decode — parse() and encode() are fully symmetric

Benchmarks

Measured on Apple Silicon. onnx-rs vs the official C++ protobuf implementation (via Python onnx package). Parse only — file I/O excluded.

Model	Size	onnx-rs	C++ protobuf	Speedup
NLLB Decoder	1.7 GB	0.4ms	55.3ms	138x
NLLB Encoder	1.5 GB	0.2ms	29.8ms	149x
BGE Large	1.2 GB	0.3ms	55.7ms	186x
VGG-19	548 MB	8.2ms	8.5ms	1.0x
GPT-2	523 MB	0.3ms	8.3ms	28x
Swin Transformer	422 MB	2.6ms	8.2ms	3.1x
BERT-SQuAD	415 MB	0.1ms	12.6ms	126x
Donut Encoder	297 MB	2.2ms	6.1ms	2.8x
YOLOv4	245 MB	0.1ms	4.0ms	40x
ResNet-152	230 MB	3.9ms	4.0ms	1.0x
Mask R-CNN	169 MB	0.9ms	3.7ms	4.1x
Faster R-CNN	168 MB	0.7ms	3.5ms	5.0x
MiniLM-L6	86 MB	0.1ms	1.5ms	15x
EfficientNet	49 MB	0.1ms	0.9ms	9x
Inception v2	42 MB	0.1ms	0.9ms	9x

Models using raw_data (the default for modern exporters like PyTorch) are parsed near-instantly via zero-copy borrows. Models using float_data (legacy packed floats) use bulk memcpy on little-endian platforms.

All 15 models roundtrip cleanly: onnx_rs::parse -> onnx_rs::encode -> validated by onnx.checker.check_model() in the official C++ implementation.

Run benchmarks locally:

cargo bench

License

MIT