Execution providers
Learn how to enable execution providers to leverage hardware acceleration.
Execution providers (EPs) enable ONNX Runtime to execute ONNX graphs with hardware acceleration. If you have specialized hardware like a GPU or NPU, execution providers can provide a massive performance boost to your ort
applications. For more information on the intricacies of execution providers, see the ONNX Runtime docs.
ONNX Runtime must be compiled with support for each execution provider. pyke provides precompiled binaries for some of the most common EPs, so you wonβt need to compile ONNX Runtime from source. Below is a table showing available EPs, their support in ort
, and their binary availability status.
EP | Supported | Binaries | Static linking |
---|---|---|---|
NVIDIA CUDA | π’ | π’ | β |
NVIDIA TensorRT | π’ | π’ | β |
Microsoft DirectML | π’ | π’ | π’ |
Apple CoreML | π’ | π’ | π’ |
AMD ROCm | π’ | π’ | β |
Intel OpenVINO | π’ | β | β |
Intel oneDNN | π’ | β | β |
XNNPACK | π’ | π’ | π’ |
Qualcomm QNN | π’ | β | β |
Huawei CANN | π’ | β | β |
Android NNAPI | π’ | β | β |
Apache TVM | π’ | β | β |
Arm ACL | π’ | β | β |
ArmNN | π’ | β | β |
AMD MIGraphX | β | β | β |
AMD Vitis AI | β | β | β |
Microsoft Azure | β | β | β |
Rockchip RKNPU | β | β | β |
Some EPs supported by ONNX Runtime are not supported by ort
due to a lack of hardware for testing. If your preferred EP is missing support and youβve got the hardware, please open an issue!
Registering execution providers
To use an execution provider with ort
, youβll need to enable its respective Cargo feature, e.g. the cuda
feature to use CUDA, or the coreml
feature to use CoreML.
[dependencies]
ort = { version = "2.0", features = [ "cuda" ] }
See Cargo features for the full list of features.
In order to configure sessions to use certain execution providers, you must register them when creating an environment or session. You can do this via the SessionBuilder::with_execution_providers
method. For example, to register the CUDA execution provider for a session:
use ort::{CUDAExecutionProvider, Session};
fn main() -> anyhow::Result<()> {
let session = Session::builder()?
.with_execution_providers([CUDAExecutionProvider::default().build()])?
.commit_from_file("model.onnx")?;
Ok(())
}
You can, of course, specify multiple execution providers. ort
will register all EPs specified, in order. If an EP does not support a certain operator in a graph, it will fall back to the next successfully registered EP, or to the CPU if all else fails.
use ort::{CoreMLExecutionProvider, CUDAExecutionProvider, DirectMLExecutionProvider, TensorRTExecutionProvider, Session};
fn main() -> anyhow::Result<()> {
let session = Session::builder()?
.with_execution_providers([
// Prefer TensorRT over CUDA.
TensorRTExecutionProvider::default().build(),
CUDAExecutionProvider::default().build(),
// Use DirectML on Windows if NVIDIA EPs are not available
DirectMLExecutionProvider::default().build(),
// Or use ANE on Apple platforms
CoreMLExecutionProvider::default().build()
])?
.commit_from_file("model.onnx")?;
Ok(())
}
Configuring EPs
EPs have configuration options to control behavior or increase performance. Each XXXExecutionProvider
struct returns a builder with configuration methods. See the API reference for the EP structs for more information on which options are supported and what they do.
use ort::{CoreMLExecutionProvider, Session};
fn main() -> anyhow::Result<()> {
let session = Session::builder()?
.with_execution_providers([
CoreMLExecutionProvider::default()
// this model uses control flow operators, so enable CoreML on subgraphs too
.with_subgraphs()
// only use the ANE as the CoreML CPU implementation is super slow for this model
.with_ane_only()
.build()
])?
.commit_from_file("model.onnx")?;
Ok(())
}
Fallback behavior
ort
will silently fail and fall back to executing on the CPU if all execution providers fail to register. In many cases, though, youβll want to show the user an error message when an EP fails to register, or outright abort the process.
To receive these registration errors, instead use ExecutionProvider::register
to register an execution provider:
use ort::{CUDAExecutionProvider, ExecutionProvider, Session};
fn main() -> anyhow::Result<()> {
let builder = Session::builder()?;
let cuda = CUDAExecutionProvider::default();
if cuda.register(&builder).is_err() {
eprintln!("Failed to register CUDA!");
std::process::exit(1);
}
let session = builder.commit_from_file("model.onnx")?;
Ok(())
}
You can also check whether ONNX Runtime is even compiled with support for the execution provider with the is_available
method.
use ort::{CoreMLExecutionProvider, ExecutionProvider, Session};
fn main() -> anyhow::Result<()> {
let builder = Session::builder()?;
let coreml = CoreMLExecutionProvider::default();
if !coreml.is_available() {
eprintln!("Please compile ONNX Runtime with CoreML!");
std::process::exit(1);
}
// Note that even though ONNX Runtime was compiled with CoreML, registration could still fail!
coreml.register(&builder)?;
let session = builder.commit_from_file("model.onnx")?;
Ok(())
}
Global defaults
You can configure ort
to attempt to register a list of execution providers for all sessions created in an environment.
use ort::{CUDAExecutionProvider, Session};
fn main() -> anyhow::Result<()> {
ort::init()
.with_execution_providers([CUDAExecutionProvider::default().build()])
.commit()?;
let session = Session::builder()?.commit_from_file("model.onnx")?;
// The session will attempt to register the CUDA EP
// since we configured the environment default.
Ok(())
}
ort::init
must come before you create any sessions, otherwise the configuration will not take effect!
Sessions configured with their own execution providers will extend the execution provider defaults, rather than overriding them.
Troubleshooting
If it seems like the execution provider is not registering properly, or you are not getting acceptable performance, see the Troubleshooting: Performance page for more information on how to debug any EP issues.
Notes
CoreML
Statically linking to CoreML (the default behavior when using downloaded binaries + the coreml
Cargo feature) requires an additional Rust flag in order to link properly. Youβll need to provide the flag -C link-arg=-fapple-link-rtlib
to rustc
. You can do this via an entry in .cargo/config.toml
, in a build script, or in an environment variable.
See Configuration: Hierarchical structure for more information on where the configuration file can be placed.
[target.aarch64-apple-darwin]
rustflags = ["-Clink-arg=-fapple-link-rtlib"]
[target.x86_64-apple-darwin]
rustflags = ["-Clink-arg=-fapple-link-rtlib"]