Installing MONAI on ROCm

Installing MONAI on ROCm#

2026-09-28

3 min read time

Applies to Linux

To install MONAI on ROCm, you have the following options:

System requirements#

ROCm version

Ubuntu version

Python version

AMD Instinct™ GPU

PyTorch for AMD ROCm version

Numpy version

10.0.0

24.04

3.12

MI300X, MI325X, or MI355X

ROCm-enabled PyTorch (ships with ROCm)

No earlier than 1.24 and no later than 2.4

For the complete list of dependencies, see the requirements.txt file.

Setting up the environment#

  1. Start a Docker container with the ROCm Ubuntu Docker image from Docker Hub:

    docker run --cap-add=SYS_PTRACE --ipc=host --privileged=true \
--shm-size=512GB --network=host --device=/dev/kfd \
--device=/dev/dri --group-add video -it \
-v $HOME:$HOME --name ${LOGNAME}_monai \
ubuntu:24.04

    MONAI on ROCm is installed in the Docker container. All subsequent commands must be run from within the Docker container.

  2. Install required system dependencies.

    apt-get update && \
apt-get install -y --no-install-recommends \
  python3-venv python3-pip python3-dev \
  build-essential git cmake ninja-build yasm \
  openssh-client \
  libgomp1 libstdc++-13-dev \
  libopenslide-dev libwebp-dev libzstd-dev && \
rm -rf /var/lib/apt/lists/*

  3. Create and activate the development environment.

    python3 -m venv /opt/venv
source /opt/venv/bin/activate
pip install --upgrade pip

  4. Install PyTorch and amd-hipcim for ROCm.

    pip install \
    --index-url https://stable.repo.amd.com/rocm/whl-next/ \
    "rocm[libraries,devel,device-gfx942,device-gfx950]==10.0.*" \
    "torch[device-gfx942,device-gfx950]" \
    "torchvision[device-gfx942,device-gfx950]" \
    torchaudio

pip install amd-hipcim \
    --extra-index-url=https://pypi.amd.com/rocm-10.0.0/simple/

rocm-sdk init

  5. Set environment variables.

    export ROCM_PATH=$(python3 -c "import _rocm_sdk_core, os; print(os.path.dirname(_rocm_sdk_core.__file__))")
export ROCM_HOME=$ROCM_PATH
export ROCM_LIBRARIES_PATH=$(python3 -c "import _rocm_sdk_libraries, os; print(os.path.dirname(_rocm_sdk_libraries.__file__))")
export PATH=$ROCM_PATH/bin:$PATH
export LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/lib/rocm_sysdeps/lib:$ROCM_PATH/lib/llvm/lib:$ROCM_LIBRARIES_PATH/lib:$LD_LIBRARY_PATH
export AMDGPU_TARGETS="gfx942"
export OMP_NUM_THREADS=1

    Note

    For MI300X and MI325X, set AMDGPU_TARGETS=gfx942. For MI355X, set AMDGPU_TARGETS=gfx950. Set OMP_NUM_THREADS=1 to suppress OpenMP warnings during setup.

Installing using package manager#

From within the Docker container, use these steps to install MONAI on ROCm from AMD PyPI.

  1. Install optional dependencies based on the workload.

    pip install ITK nibabel gdown tqdm lmdb psutil pandas einops mlflow \
            pynrrd clearml transformers pydicom fire ignite         \
            parameterized tensorboard pytorch-ignite onnx

  2. Install NumPy and MONAI on ROCm.

    pip install "numpy<2.5,>=1.24"

pip install --no-deps amd-monai \
    --index-url https://stable.repo.amd.com/rocm/whl-next/ \
    --extra-index-url=https://pypi.amd.com/rocm-10.0.0/simple/

Building from source#

To build MONAI on ROCm from source, follow the steps given in this section.

  1. Clone the MONAI on ROCm repository.

    git clone git@github.com:AMD-Ecosystem/monai.git
cd monai

  2. Set environment variables.

    The devel extra installed with rocm during environment setup provides the development packages the build needs.

    export ROCM_DEVEL_PATH=$(python3 -c "import _rocm_sdk_devel, os; print(os.path.dirname(_rocm_sdk_devel.__file__))")
export LIBRARY_PATH=$ROCM_DEVEL_PATH/lib:$ROCM_PATH/lib
export CPATH=$ROCM_DEVEL_PATH/include:/usr/lib/gcc/x86_64-linux-gnu/13/include:$CPATH
export PYTORCH_ROCM_ARCH=$AMDGPU_TARGETS

# Symlinks required by hipcc for GPU bitcode and unversioned .so stubs
mkdir -p $ROCM_PATH/amdgcn
ln -sf $ROCM_PATH/lib/llvm/amdgcn/bitcode $ROCM_PATH/amdgcn/bitcode

  3. Install development dependencies and build a wheel.

    pip install -r requirements-dev.txt -c amd-constraints.txt \
    --index-url https://stable.repo.amd.com/rocm/whl-next/ \
    --extra-index-url https://pypi.org/simple/ \
    --build-constraint amd-constraints.txt

BUILD_MONAI=1 FORCE_CUDA=1 python3 setup.py bdist_wheel
pip install --no-deps dist/amd_monai-*.whl

    The wheel file is generated under the dist directory.

Verify installation#

Verify the MONAI on ROCm installation. Run these commands from within the Docker container.

python3 -c "import monai; print(monai.__version__)"

import torch
import monai

print(f"MONAI version: {monai.__version__}")
print(f"PyTorch version: {torch.__version__}")
print(f"ROCm available: {torch.version.hip is not None}")
print(f"GPU available: {torch.cuda.is_available()}")
if torch.cuda.is_available():
    print(f"GPU: {torch.cuda.get_device_name(0)}")

pip show -v amd-monai
Contents