Installing dask-hip#

2026-10-01

2 min read time

Applies to Linux

You can install dask-hip via AMD PyPI as described below. This is recommended for users of the package. For developers interested in modifying or contributing to the open-source dask-hip component, see the Build instructions.

See System requirements for information regarding supported operating systems, ROCm versions, and AMD GPUs before installing dask-hip.

ROCm component prerequisites#

dask-hip requires the amdsmi Python package, which is distributed with ROCm (not via AMD PyPI). Install it from the ROCm installation before installing dask-hip. See AMD SMI documentation for more information.

cd /opt/rocm/share/amd_smi
pip install .

Install dask-hip via AMD PyPI#

Packaged versions of dask-hip and its dependencies are distributed via AMD PyPI. This section describes how to install dask-hip via this package index.

Create and activate a Conda environment with a compatible Python version, such as 3.11 or 3.12 as shown below. For more information on compatible Python versions, see System requirements.

conda create --name dask-hip python=3.12 # Specify your Python version
conda activate dask-hip

Alternatively, create a Python virtual environment:

python3 -m venv dask-hip-env
source dask-hip-env/bin/activate

dask-hip can then be installed into either environment using pip and the AMD PyPI URL:

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

hip-ucxx support#

To install the amd-distributed-hipucxx package for high-performance UCX communication (ROCm-IPC, InfiniBand):

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

The hip-ucxx packages depend on amd-libucx, which brings its own ROCm-enabled UCX, so no separate UCX installation is needed.

Note

The amd-libucx dependency is a source distribution and compiles a ROCm-enabled UCX during pip install. Install its build prerequisites first (on Ubuntu):

sudo apt install make gcc g++ patchelf wget \
    rdma-core libibverbs-dev librdmacm-dev libnuma-dev

pynvml compatibility#

dask-hip relies on the pynvml compatibility shim provided by the hip-python-interop package, which maps pynvml API calls to their AMD SMI equivalents. This allows upstream packages such as dask.distributed to query GPU information transparently on AMD hardware without modification.