AOTInductor: Ahead-Of-Time Compilation for Torch.Export-ed Models


AOTInductor and its related features are in prototype status and are subject to backwards compatibility breaking changes.

AOTInductor is a specialized version of TorchInductor , designed to process exported PyTorch models, optimize them, and produce shared libraries as well as other relevant artifacts. These compiled artifacts are specifically crafted for deployment in non-Python environments, which are frequently employed for inference deployments on the server side.

In this tutorial, you will gain insight into the process of taking a PyTorch model, exporting it, compiling it into a shared library, and conducting model predictions using C++.

Model Compilation

Using AOTInductor, you can still author the model in Python. The following example demonstrates how to invoke aot_compile to transform the model into a shared library.

This API uses torch.export to capture the model into a computational graph, and then uses TorchInductor to generate a .so which can be run in a non-Python environment. For comprehensive details on the torch._export.aot_compile API, you can refer to the code here. For more details on torch.export, you can refer to the torch.export docs.


If you have a CUDA-enabled device on your machine and you installed PyTorch with CUDA support, the following code will compile the model into a shared library for CUDA execution. Otherwise, the compiled artifact will run on CPU. For better performance during CPU inference, it is suggested to enable freezing by setting export TORCHINDUCTOR_FREEZING=1 before running the Python script below.

import os
import torch

class Model(torch.nn.Module):
    def __init__(self):
        self.fc1 = torch.nn.Linear(10, 16)
        self.relu = torch.nn.ReLU()
        self.fc2 = torch.nn.Linear(16, 1)
        self.sigmoid = torch.nn.Sigmoid()

    def forward(self, x):
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        x = self.sigmoid(x)
        return x

with torch.no_grad():
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model = Model().to(device=device)
    example_inputs=(torch.randn(8, 10, device=device),)
    batch_dim = torch.export.Dim("batch", min=1, max=1024)
    so_path = torch._export.aot_compile(
        # Specify the first dimension of the input x as dynamic
        dynamic_shapes={"x": {0: batch_dim}},
        # Specify the generated shared library path
        options={"aot_inductor.output_path": os.path.join(os.getcwd(), "")},

In this illustrative example, the Dim parameter is employed to designate the first dimension of the input variable “x” as dynamic. Notably, the path and name of the compiled library remain unspecified, resulting in the shared library being stored in a temporary directory. To access this path from the C++ side, we save it to a file for later retrieval within the C++ code.

Inference in C++

Next, we use the following C++ file inference.cpp to load the shared library generated in the previous step, enabling us to conduct model predictions directly within a C++ environment.


The following code snippet assumes your system has a CUDA-enabled device and your model was compiled to run on CUDA as shown previously. In the absence of a GPU, it’s necessary to make these adjustments in order to run it on a CPU: 1. Change model_container_runner_cuda.h to model_container_runner_cpu.h 2. Change AOTIModelContainerRunnerCuda to AOTIModelContainerRunnerCpu 3. Change at::kCUDA to at::kCPU

#include <iostream>
#include <vector>

#include <torch/torch.h>
#include <torch/csrc/inductor/aoti_runner/model_container_runner_cuda.h>

int main() {
    c10::InferenceMode mode;

    torch::inductor::AOTIModelContainerRunnerCuda runner("");
    std::vector<torch::Tensor> inputs = {torch::randn({8, 10}, at::kCUDA)};
    std::vector<torch::Tensor> outputs =;
    std::cout << "Result from the first inference:"<< std::endl;
    std::cout << outputs[0] << std::endl;

    // The second inference uses a different batch size and it works because we
    // specified that dimension as dynamic when compiling
    std::cout << "Result from the second inference:"<< std::endl;
    std::vector<torch::Tensor> inputs2 = {torch::randn({2, 10}, at::kCUDA)};
    std::cout <<[0] << std::endl;

    return 0;

For building the C++ file, you can make use of the provided CMakeLists.txt file, which automates the process of invoking python for AOT compilation of the model and compiling inference.cpp into an executable binary named aoti_example.

cmake_minimum_required(VERSION 3.18 FATAL_ERROR)

find_package(Torch REQUIRED)

add_executable(aoti_example inference.cpp


target_link_libraries(aoti_example "${TORCH_LIBRARIES}")
set_property(TARGET aoti_example PROPERTY CXX_STANDARD 17)

Provided the directory structure resembles the following, you can execute the subsequent commands to construct the binary. It is essential to note that the CMAKE_PREFIX_PATH variable is crucial for CMake to locate the LibTorch library, and it should be set to an absolute path. Please be mindful that your path may vary from the one illustrated in this example.

$ mkdir build
$ cd build
$ CMAKE_PREFIX_PATH=/path/to/python/install/site-packages/torch/share/cmake cmake ..
$ cmake --build . --config Release

After the aoti_example binary has been generated in the build directory, executing it will display results akin to the following:

$ ./aoti_example
Result from the first inference:
[ CUDAFloatType{8,1} ]
Result from the second inference:
[ CUDAFloatType{2,1} ]


Access comprehensive developer documentation for PyTorch

View Docs


Get in-depth tutorials for beginners and advanced developers

View Tutorials


Find development resources and get your questions answered

View Resources