Installing and running rocOpt with ROCm 7.2.3 enabled

Installing and running rocOpt with ROCm 7.2.3 enabled#

2026-10-13

5 min read time

Applies to Linux

To run rocOpt with ROCm 7.2.3 enabled, you have two options to set up the Docker container:

  • Use a prebuilt Docker image that contains rocOpt and the required dependencies (recommended)

  • Build from source

Compatibility Matrix#

rocOpt version

Ubuntu

ROCm version (build → runtime)

Python version

PyTorch version

AMD Instinct GPU

1.0.0

24.04

7.1.1 (build) → 7.2.3 (runtime)

3.13

N/A

MI300X, MI355X

Note

ROCm version compatibility: This release of rocOpt is built against ROCm 7.1.1 and is intended to run on ROCm 7.2.3. This configuration serves as a workaround for a build issue identified in ROCm 7.2.3.

Option 1: Use a prebuilt Docker image#

The prebuilt docker image contains a fully configured rocOpt installation and all required dependencies pre-installed. The image can be found at https://hub.docker.com/r/rocm/rocopt/tags.

  1. Pull the Docker image.

    docker pull rocopt:rocopt-1.0.0.amd0_rocm7.2.3.ubuntu24.04_py3.13

  2. Start a Docker container using this image.

    docker run -it --privileged \
  --rm \
  --device=/dev/kfd \
  --device=/dev/dri \
  --group-add video \
  --cap-add=SYS_PTRACE \
  --security-opt seccomp=unconfined \
  --ipc=host \
  rocopt:rocopt-1.0.0.amd0_rocm7.2.3.ubuntu24.04_py3.13

Note

rocOpt container images have third-party vulnerabilities as explained in Known vulnerability exceptions and customer mitigations. Please review that content for additional information.

Option 2: Build from source#

  1. Clone the rocOpt repository.

    git clone https://github.com/AMD-Ecosystem/rocopt.git
cd rocopt

  2. Build the Docker image.

    docker build --file dockerfile.rocm --tag rocopt-rocm .

    This will create the rocopt-rocm image and install rocOpt along with its required dependencies (hipRAFT, hipMM, rocThrust, hipCUB, rocPRIM, rapids-cmake). The build stage compiles rocOpt against ROCm 7.1.1 (to avoid the ROCm 7.2.x MI300X codegen regression), while the resulting runtime image runs on the ROCm 7.2.3 runtime.

  3. Launch a container based on the image.

    docker run -it --privileged \
  --rm \
  --device=/dev/kfd \
  --device=/dev/dri \
  --group-add video \
  --cap-add=SYS_PTRACE \
  --security-opt seccomp=unconfined \
  --ipc=host \
  rocopt-rocm

  4. Inside the container, build the rocOpt C++ library.

    USE_ROCM=1 ./build.sh

Option 3: Installing the rocOpt wheels in a ROCm container#

Prerequisites#

  • An AMD GPU host with ROCm kernel support (/dev/kfd and /dev/dri present). Validated on AMD Instinct MI300X and MI355X.

  • Docker (or Podman).

Build steps#

  1. Start a ROCm-enabled Ubuntu 24.04 container

    docker run -it --rm \
    --device=/dev/kfd --device=/dev/dri \
    --group-add video \
    --security-opt seccomp=unconfined \
    rocm/dev-ubuntu-24.04:7.2.3 bash

  2. Create a Python 3.13 virtual environment

    The rocopt and rocopt_mps_parser wheels are only published as cp313-cp313, so Python 3.13 is required:

    apt-get update
apt-get install -y software-properties-common
add-apt-repository -y ppa:deadsnakes/ppa
apt-get update
apt-get install -y python3.13 python3.13-venv

python3.13 -m venv /opt/venv
source /opt/venv/bin/activate

  3. Install plain runtime dependencies

    pip install \
   numpy pandas pyyaml numba pyarrow fastapi \
   jsonref==1.1.0 msgpack==1.1.0 msgpack-numpy==0.4.8 \
   psutil "uvicorn==0.34.*" requests

  4. Install the ROCm math/FFT system packages

    libcuopt.so and cupy both need the ROCm math/FFT library suite at dlopen time; the base image only ships the core HIP/HSA runtime:

    apt install -y hipblas hipsparse hiprand hipsolver hipfft rocrand libtbb12

  5. Install the RAPIDS-DS dependency chain

    pip install --extra-index-url=https://pypi.amd.com/rocm-7.2.3/simple/ \
    rocm-llvm-python==7.2.3.562.46 \
    hip-python==7.2.3.562.43 \
    hip-python-as-cuda==7.2.3.562.43 \
    numba-hip==0.1.6 \
    amd-cupy==13.5.1 \
    amd-rocmds-logger==1.0.0 \
    amd-libhipmm==4.0.0 \
    amd-hipmm==4.0.0 \
    amd-libraft==1.0.0 \
    amd-pylibraft==1.0.0 \
    amd-libhipdf==3.0.0 \
    amd-pylibhipdf==3.0.0 \
    amd-hipdf==3.0.0

  6. Install rocopt

    pip install --extra-index-url=https://pypi.amd.com/rocm-7.2.3/simple/ \
    rocopt==1.0.0 \
    librocopt==1.0.0 \
    rocopt-server==1.0.0 \
    rocopt-sh-client==1.0.0 \
    rocopt-mps-parser==1.0.0

pip install requests  # rocopt_sh_client needs it, not pulled in by the above

  7. Symlink libcuopt.so into place

    ln -s /opt/venv/lib/python3.13/site-packages/libcuopt/libcuopt/lib64 \
      /opt/venv/lib/python3.13/site-packages/libcuopt/lib64

  8. Set required environment variables

    export ROCM_HOME=/opt/rocm
export LD_LIBRARY_PATH=/opt/venv/lib/python3.13/site-packages/libcuopt/lib64:/opt/venv/lib/python3.13/site-packages/librmm/lib64:/opt/venv/lib/python3.13/site-packages/rapids_logger/lib64:/opt/venv/lib/python3.13/site-packages/cuopt_mps_parser:/opt/venv/lib/python3.13/site-packages/cuopt/linear_programming/data_model${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}

    ROCM_HOME points cupy at the ROCm installation. LD_LIBRARY_PATH adds the shared library directories the compiled extensions load from at runtime. Persist both into /opt/venv/bin/activate so they survive new shells.

Test your installation#

After starting the container, verify your installation is working correctly. The exact verification steps vary by use case — common checks include:

  • Confirming the rocOpt Python package is installed and importable:

    python -c "import cuopt; print(cuopt.__version__)"

Note

The cuDSS-based barrier (interior-point) LP solver is not available on ROCm. Use the PDLP solver for large LPs and the dual-simplex solver for smaller problems.

Contents