Crate executorch

Expand description

Bindings for ExecuTorch - On-device AI across mobile, embedded and edge for PyTorch.

Provides a high-level Rust API for executing PyTorch models on mobile, embedded and edge devices using the ExecuTorch library, specifically the C++ API. PyTorch models are created and exported in Python, and then loaded and executed on-device using the ExecuTorch library.

The following example create a simple model in Python, exports it, and then executes it in Rust:

Create a model in Python and export it:

import torch
from executorch.exir import to_edge
from torch.export import export

class Add(torch.nn.Module):
    def __init__(self):
        super(Add, self).__init__()

    def forward(self, x: torch.Tensor, y: torch.Tensor):
        return x + y


aten_dialect = export(Add(), (torch.ones(1), torch.ones(1)))
edge_program = to_edge(aten_dialect)
executorch_program = edge_program.to_executorch()
with open("model.pte", "wb") as file:
    file.write(executorch_program.buffer)

Execute the model in Rust:

use executorch::evalue::IntoEValue;
use executorch::module::Module;
use executorch::tensor_ptr;
use ndarray::array;

let mut module = Module::new("model.pte", None);

let (tensor1, tensor2) = (tensor_ptr![1.0_f32], tensor_ptr![1.0_f32]);
let inputs = [tensor1.into_evalue(), tensor2.into_evalue()];

let outputs = module.forward(&[inputs]).unwrap();
assert_eq!(outputs.len(), 1);
let output = outputs.into_iter().next().unwrap();
let output = output.as_tensor().into_typed::<f32>();

println!("Output tensor computed: {:?}", output);
assert_eq!(array![2.0], output.as_array());

§Cargo Features

data-loader: Includes additional structs in the data_loader module for loading data. Without this feature the only available data loader is BufferDataLoader. The libextension_data_loader.a static library is required, compile C++ executorch with EXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON.
module: Includes the module API, a high-level API for loading and executing PyTorch models. It is an alternative to the lower-level Program API, which is more suitable for embedded systems. The libextension_module_static.a static library is required, compile C++ executorch with EXECUTORCH_BUILD_EXTENSION_MODULE=ON. Also includes the std feature.
tensor-ptr: Includes the tensor::TensorPtr struct, a smart pointer for tensors that manage the lifetime of the tensor object alongside the lifetimes of the data buffer and additional metadata. The extension_tensor.a static library is required, compile C++ executorch with EXECUTORCH_BUILD_EXTENSION_TENSOR=ON. Also includes the std feature.
etdump Includes the ETDumpGen struct, an implementation of an EventTracer, used for debugging and profiling. The libetdump.a static library is required, compile C++ executorch with EXECUTORCH_BUILD_DEVTOOLS=ON and EXECUTORCH_ENABLE_EVENT_TRACER=ON. In addition, the flatcc (or flatcc_d) library is required, available at {CPP_EXECUTORCH_DIR}/third-party/flatcc/lib/, and should be linked by the user.
ndarray: Conversions between executorch tensors and ndarray arrays. Adds a dependency to the ndarray crate. This feature is enabled by default.
f16: Adds a dependency to the half crate, which provides a fully capable f16 and bf16 types. Without this feature enabled, both of these types are available with a simple conversions to/from u16 only. Note that this only affect input/output tensors, the internal computations always have the capability to operate on such scalars.
num-complex: Adds a dependency to the num-complex crate, which provides a fully capable complex number type. Without this feature enabled, complex numbers are available as a simple struct with two public fields without any operations. Note that this only affect input/output tensors, the internal computations always have the capability to operate on such scalars.
std: Enable the standard library. This feature is enabled by default, but can be disabled to build executorch in a no_std environment. See the examples/no_std example. Also includes the alloc feature. NOTE: no_std is still WIP, see https://github.com/pytorch/executorch/issues/4561
alloc: Enable allocations. When this feature is disabled, all methods that require allocations will not be compiled. This feature is enabled by the std feature, which is enabled by default. Its possible to enable this feature without the std feature, and the allocations will be done using the alloc crate, that requires a global allocator to be set.

By default the std and ndarray features are enabled.

§Build

To use the library you must compile the C++ executorch library yourself, as there are many configurations that determines which modules, backends, and operations are supported. See the executorch-sys crate for more info.

§Embedded Systems

The library is designed to be used both in std and no_std environments. The no_std environment is useful for embedded systems, where the standard library is not available. The alloc feature can be used to provide an alternative to the standard library’s allocator, but it is possible to use the library without allocations at all. Due to some difference between Cpp and Rust, it is not trivial to provide such API, and the interface may feel more verbose. See the memory::Storage struct for stack allocations of Cpp objects, and the examples/no_std example.

§API Stability

The C++ API is still in Beta, and this Rust lib will continue to change with it. Currently the supported executorch version is 0.5.0.

Re-exports§

pub use ndarray;ndarray
pub use half;half
pub use num_complex;num-complex

Modules§

data_loader: Data loaders for loading execution plans (models) from a data source.
evalue: Module for EValue and related types.
event_tracer: Event tracers for debugging and profiling.
memory: Memory management classes.
modulemodule: A higher-level for simple execution of programs.
platform: Platform abstraction layer to allow individual platform libraries to override symbols in ExecuTorch.
program: Lower-level API for loading and executing ExecuTorch programs.
scalar: Custom scalar types that can be used in tensors.
tensor: Tensor struct is an input or output tensor to an executorch program.
util: Utility functions and types.

Macros§

storage: A macro to create a pinned Storage object(s).
tensor_ptrtensor-ptr and ndarray: A short syntax for creating a TensorPtr.

Enums§

CError: Error codes returned by the Cpp library.
Error: ExecuTorch Error type.

Crate executorchCopy item path