Alternative backends
Since ONNX Runtime  is written in C++, linking troubles often arise when attempting to use it in a Rust project - especially with WASM. v2.0.0-rc.10
of ort
introduced support for alternative backends — that is, ONNX executors that do not use ONNX Runtime.
As the Rust ML scene has evolved, many exciting new inference engines supporting ONNX models have popped up, like 🤗 Hugging Face’s candle
, Burn , and tract
. These libraries, being written in pure Rust (minus some GPU kernels) play much nicer when it comes to linking, and often support any platform Rust’s standard library does. They’re also, of course, memory safe and 🦀blazingly🔥fast🚀!
Internally, alternative backend implementations are simply glue code between these libraries and the ONNX Runtime C API. Because they implement the same API as ONNX Runtime, using them in ort
is as simple as adding one line of code!
Using an alternative backend
Alternative backends are experimental, and are constantly changing and growing — use them at your own risk!
We may not be able to provide the same level of support for different backends as we do with ONNX Runtime.
Install the alternative backend
We’ll use ort-tract
for this example.
[dependencies]
ort-tract = "0.1.0+0.21"
...
Enable the alternative-backend
feature
This instructs ort
to not try to download/link to ONNX Runtime.
[dependencies.ort]
version = "=2.0.0-rc.10"
default-features = false # Disables the `download-binaries` feature since we don't need it
features = [
"alternative-backend"
]
Initialize the backend
Use ort::set_api
to use the crate’s API implementation (replacing ort_tract
with whichever backend crate you choose to use):
fn main() {
// This should run as early in your application as possible - before you ever use `ort`!
ort::set_api(ort_tract::api());
}
Done!
Be sure to check each backend’s docs page to see which APIs are and are not implemented.
Available backends
ort
currently has the following backends:
ort-candle
, based on 🤗 Hugging Facecandle
- đź”· Supports: CPU, CUDA (though not available via
ort-candle
right now), WebAssembly - ⚠️ Limited operator support; though most transformer models have good support.
- đź”· Supports: CPU, CUDA (though not available via
ort-tract
, based ontract
- đź”· Supports: CPU, WebAssembly
- ✅ Great operator support