GitHub Workflow Status Crates.io (latest)

Wonnx is a GPU-accelerated ONNX inference run-time written 100% in Rust, ready for the web.

Supported Platforms (enabled by `wgpu`)

API | Windows | Linux & Android | macOS & iOS | ----- | ----------------------------- | ------------------ | ------------------ | Vulkan | ✅ | ✅ | | Metal | | | ✅ | DX12 | ✅ (W10 only) | | | DX11 | :construction: | | | GLES3 | | :ok: | |

:whitecheckmark: = First Class Support — :ok: = Best Effort Support — :construction: = Unsupported, but support in progress

Getting started

From the command line

Ensure your system supports either Vulkan, Metal or DX12 for access to the GPU. Then either download a binary release, or install Rust and run cargo install --git https://github.com/webonnx/wonnx.git wonnx-cli to install the CLI.

The CLI tool (nnx) provides a convenient interface for tinkering with models (see the README for more information):

bash nnx info ./data/models/opt-squeeze.onnx nnx infer ./data/models/opt-squeeze.onnx -i data=./data/images/pelican.jpeg --labels ./data/models/squeeze-labels.txt --top 3

From Rust

Add the wonnx crate as dependency (cargo add wonnx if you have cargo-add). Then, see the examples for usage examples, or browse the API docs.

From Python

bash pip install wonnx

And then, to use:

python from wonnx import Session session = Session.from_path( "../data/models/single_relu.onnx" ) inputs = {"x": [-1.0, 2.0]} assert session.run(inputs) == {"y": [0.0, 2.0]}

Then run python3 with the above Python code!

For more details on the Python package including build instructions, see wonnx-py.

In the browser, using WebGPU + WebAssembly

bash npm install @webonnx/wonnx-wasm

And then, on the client side:

````js import init, { Session, Input } from "@webonnx/wonnx-wasm";

// Check for WebGPU availability first: if(navigator.gpu) { .. } await init(); const session = await Session.fromBytes(modelBytes /* Uint8Array containing the ONNX file */); const input = new Input(); input.insert("x", [13.0, -37.0]); const result = await session.run(input); // This will be an object where the keys are the names of the model outputs and the values are arrays of numbers. session.free(); input.free(); ````

The package @webonnx/wonnx-wasm provides an interface to WONNX, which is included as WebAssembly module and will use the browser's WebGPU implementation. See wonnx-wasm-example for a more complete usage example involving a bundler.

For more details on the JS/WASM package including build instructions, see wonnx-wasm.

For development

To work on wonnx itself, follow the following steps:

Install Rust
Install Vulkan, Metal, or DX12 for the GPU API.
git clone this repo.

bash git clone https://github.com/webonnx/wonnx.git

Then, you're all set! You can run one of the included examples through cargo:

bash cargo run --example squeeze --release

Running other models

To run an onnx model, first simplify it with nnx prepare (substitute with cargo run -- prepare when inside this repo):

bash nnx prepare -i ./some-model.onnx ./some-model-prepared.onnx

To specify dynamic dimension parameters, add e.g. --set batch_size=1.

You can also use an external tool, such as onnx-simplifier, with the command:

```bash

pip install -U pip && pip install onnx-simplifier

python -m onnxsim mnist-8.onnx opt-mnist.onnx ```

Then you can run it using the CLI (see README or programmatically, following the examples in the examples folder. To run an example:

bash cargo run --example mnist --release

Tested models

Squeezenet
MNIST
BERT

GPU selection

Except when running in WebAssembly, you may set the following environment variables to influence GPU selection by WGPU:

WGPU_ADAPTER_NAME with a substring of the name of the adapter you want to use (e.g. 1080 will match NVIDIA GeForce 1080ti).
WGPU_BACKEND with a comma separated list of the backends you want to use (vulkan, metal, dx12, dx11, or gl).
WGPU_POWER_PREFERENCE with the power preference to choose when a specific adapter name isn't specified (high or low)

Contribution: On implementing a new Operator

Contribution are very much welcomed even without large experience in DL, WGSL, or Rust. I hope that, this project can be a sandbox for all of us to learn more about those technologies beyond this project initial scope.

To implement an operator all you have to do is: 1. Add a new matching pattern in compiler.rs 2. Retrieve its attributes values using the get_attribute function: Rust let alpha = get_attribute("alpha", Some(1.0), node); // or without default value let alpha = get_attribute::<f32>("alpha", None, node); 3. Add any variable you want to use in the WGSL shader using context. 4. Write a new WGSL template in the templates folder.

Available types are in structs.wgsl but you can also generate new ones within your templates. 5. Respect the binding layout that each entry is incremented by 1 starting from 0, with input first and output last. If the number of binding is above 4. Increment the binding group. You can change the input within sequencer.rs 6. Write the logic.

There is default variables in the context: - {{ i_lens[0] }}: the length of the input 0. This also work for output: {{ o_lens[0] }} and other input {{ i_lens[1] }} - {{ i_shape[0] }}: the array of dimensions of input 0. To get the first dimension of the array, just use: {{ i_shape[0][0] }} - {{ i_chunks[0] }}: the size of the chunks of each dimensions of input 0. By default, each variable is represented as a long array of values where to get to specific values you have to move by chunks. Those chunks are represented within this variable. To get the size of the chunks of the first dimensions use: {{ i_chunks[0][0] }}. - {{ op_type }} the op type as some op_type like activation are using the same template.

Test it using the utils function and place it in the tests folder. The test can look as follows: ```Rust
[test]

fn testmatmulsquare_matrix() { // USER INPUT

let n = 16;
let mut input_data = HashMap::new();

let data_a = ndarray::Array2::eye(n);
let mut data_b = ndarray::Array2::<f32>::zeros((n, n));
data_b[[0, 0]] = 0.2;
data_b[[0, 1]] = 0.5;

let sum = data_a.dot(&data_b);

input_data.insert("A".to_string(), data_a.as_slice().unwrap());
input_data.insert("B".to_string(), data_b.as_slice().unwrap());

let n = n as i64;
let model = model(graph(
    vec![tensor("A", &[n, n]), tensor("B", &[n, n])],
    vec![tensor("C", &[n, n])],
    vec![],
    vec![],
    vec![node(vec!["A", "B"], vec!["C"], "MatMul", "MatMul", vec![])],
));

let session =
    pollster::block_on(wonnx::Session::from_model(model)).expect("Session did not create");

let result = pollster::block_on(session.run(input_data)).unwrap();

// Note: it is better to use a method that compares floats with a tolerance to account for differences
// between implementations; see `wonnx/tests/common/mod.rs` for an example.
assert_eq!((&result["C"]).try_into().unwrap(),sum.as_slice().unwrap());

} ```

Check out tera documentation for other templating operation: https://tera.netlify.app/docs/

If at any point you want to do optimisation of several node you can do it within sequencer.rs.

Supported Operators (ref ONNX IR)

Known limitations

The Clip, Resize, Reshape, Split, Pad and ReduceSum ops accept (typically optional) secondary inputs to set various parameters (i.e. axis). These inputs are only supported if they are supplied as initializer tensors (i.e. do not depend on inputs and are not outputs of other ops), because wonnx pre-compiles all operations to shaders in advance (and must know these parameters up front).
Internally 64-bit integers are not supported (the reason is they are not supported in the current version of WGSL); inputs and initializers with 64-bit scalars are converted to 32-bit values (possibly overflowing).
For MatMul and Gemm, the matrix dimensions must be divisible by 2, or the output matrix must be of size (1, N). Matrix multiplication only supports floats, not integers (this is a WebGPU/WGSL limitation).

Shape inference

WONNX needs to know the shape of input and output tensors for each operation in order to generate shader code for executing it. ONNX models however do not always contain this information for intermediate values. Shape inference is the process of deducing the shape of intermediate values from the shape of inputs and outputs and the characteristics of each operation.

WONNX supports a limited form of shape inference (the process of determining what the shapes are of the various nodes in a model's graph). Shape inference is available programmatically as well as through the CLI. Before shape inference can be performed, all dynamic dimension parameters need to be replaced with static values. Shape inference only infers output shapes from input shapes for specific supported ops (see the table above). Inference cannot succeed if the shape for any input of a node is not known. Nodes that already have fully defined shapes for their outputs are left unchanged (and the outputs are used for shape inference on nodes that use these outputs as inputs).

To perform shape inference using the CLI, run a command similar to this (here batch_size and sequence_length are dynamic dimension parameters; the -i flag enables shape inference):

bash nnx prepare model.onnx model-prepared.onnx --set batch_size=1 --set sequence_length=255 -i

To perform shape inference programmatically, use apply_dynamic_dimensions and infer_shapes from the wonnx_preprocessing::shape_inference module.

Constant folding

Some models contain subgraphs whose output can be determined statically, as they do not depend on the specific inputs provided during inference. WONNX can replace such constant intermediate values with static values ('constant folding'). This is supported in the following cases:

Output of nodes of the Constant op type (these are replaced with initializers)
Output of nodes of the Shape op type where the shape of the input is known (up front or during inference)
Output of nodes of which all inputs are constant (possibly after folding), and for which the operator is supported by WONNX.

Constant folding is performed as part of shape inference, unless disabled (from the CLI pass --no-fold-constants to disable). This is done in order to support models that dynamically calculate shapes using operators such as Shape/Squeeze/Unsqueeze depending on dynamically set dimension parameters (e.g. batch size).

Supported Platforms (enabled by wgpu)