Install ONNX Runtime for ROCm

Install ONNX Runtime for ROCm#

ONNX Runtime accelerates machine learning inference using the AMD MIGraphX execution provider on ROCm-supported GPUs.

ROCm version

10.0.0

7.14.1

7.14.0

Installation method

pip

Prerequisites#

Install the ROCm Core SDK#

Install ONNX Runtime using pip#

Create and activate a virtual environment or activate an existing ROCm 10.0.0 environment.

Python 3.14

python3.14 -m venv .venv
source .venv/bin/activate

Python 3.12

python3.12 -m venv .venv
source .venv/bin/activate

Install MIGraphX using pip. See Install MIGraphX for installation instructions.

Install ONNX Runtime and the onnxruntime-ep-migraphx execution provider plugin.

python -m pip install --extra-index-url https://stable.repo.amd.com/rocm/onnxruntime/whl-next/ "onnxruntime-ep-migraphx==1.0.0+rocm10.0.0"

As a workaround due to packaging issues, create the required soname symlink and set library search paths. For example, if you installed MIGraphX using pip:
```
SP=$(python -c "import site; print(site.getsitepackages()[0])")
ln -sf $SP/onnxruntime/capi/libonnxruntime.so.1.29.0 $SP/onnxruntime/capi/libonnxruntime.so.1
export LD_LIBRARY_PATH=$SP/onnxruntime/capi:$SP/migraphx_libs:$LD_LIBRARY_PATH
```
Tip

The symlink only needs to be created once. For persistence, LD_LIBRARY_PATH should be set for each shell session or added to your shell startup script (e.g. ~/.bashrc).
Confirm ONNX Runtime is correctly installed and the MIGraphX execution provider is available.

The EP plugin must be explicitly registered using onnxruntime_ep_migraphx before querying available providers. migraphx must be imported first to initialize the ROCm runtime.
```
python -c "import migraphx, onnxruntime as ort, onnxruntime_ep_migraphx as m; [ort.register_execution_provider_library(n,p) for n,p in zip(m.get_ep_names(), m.get_library_paths())]; print(ort.get_available_providers())"
```
You should see MIGraphXExecutionProvider in the output:
```
['CPUExecutionProvider', 'MIGraphXExecutionProvider']
```

Create and activate a virtual environment or activate an existing ROCm 7.14.1 environment. To create a new Python 3.12 virtual environment:
```
python3.12 -m venv .venv
source .venv/bin/activate
```

Download and install the wheels.

python -m pip install https://rocm.frameworks.amd.com/whl-multi-arch/onnxruntime-migraphx/onnxruntime_migraphx-1.23.2%2Brocm7.14.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl

Verify your installation#

Download and extract the test binary zip file.

wget https://rocm.frameworks.amd.com/whl-multi-arch/onnxruntime-migraphx/onnxruntime_migraphx-1.23.2%2Brocm7.14.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.zip
unzip onnxruntime_migraphx-1.23.2+rocm7.14.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.zip
cd build/Linux/Release

Run the test suite.

./onnxruntime_test_all \
    --gtest_filter=-:CudaKernelTest.SoftmaxGrad_LargeTensor_LastAxis_Float16:CudaKernelTest.SoftmaxGrad_LargeTensor_LastAxis_Float16_NoPowerOfTwo:CudaKernelTest.SoftmaxGrad_LargeTensor_AllAxis_Float16:CudaKernelTest.SoftmaxGrad_LargeTensor_AllAxis_Float16_NoPowerOfTwo:CudaKernelTest.LogSoftmaxGrad_LargeTensor_LastAxis_Float16:CudaKernelTest.LogSoftmaxGrad_LargeTensor_LastAxis_Float16_NoPowerOfTwo:CudaKernelTest.LogSoftmaxGrad_LargeTensor_AllAxis_Float16:CudaKernelTest.LogSoftmaxGrad_LargeTensor_AllAxis_Float16_NoPowerOfTwo:ReductionOpTest.ReductionVariationTest:GatherOpTest.Gather_invalid_index_cpu:Scatter.InvalidIndex:GradientCheckerTest.AddGrad:GradientCheckerTest.SubGrad:GradientCheckerTest.MulGrad:GradientCheckerTest.DivGrad:NhwcTransformerTests*:QDQTransformerTests*Sample Output for Unit Test

Create and activate a virtual environment or activate an existing ROCm 7.14.0 environment. To create a new Python 3.12 virtual environment:
```
python3.12 -m venv .venv
source .venv/bin/activate
```