Build Options

By default cuNumeric.jl will install cupynumeric with Conda.jl into a new conda environment stored inside the .julia/conda directory.

To customize the build process create a LocalPreferences.toml file in the same directory as the Project.toml where cuNumeric.jl is installed (or in the cuNumeric.jl git repo).

The LocalPreferences.toml defines 3 keys : mode, conda_jl_env, and user_env.

The mode option can be either:

• "local_env" : cuNumeric.jl will link against a local install of cupynumeric
• "conda_jl" : cuNumeric.jl will use Conda.jl to install cupynumeric

When mode is set to local_env the user_env option must be defined as the absolute path to the conda environment with cupynumeric installed (e.g., the value of CONDA_PREFIX when your environment is active). For me this path is: /home/emeitz/.conda/envs/cunumeric-gpu.

The conda_jl mode only needs to be set if you previously used a local environment and want to switch back to a Conda.jl install. The conda_jl_env key will set the name of the environment created by Conda.jl, by default this is just cupynumeric.

The LocalPreferences.toml might look like:

[cuNumeric]
mode = "local_env"
conda_jl_env = "cupynumeric"
user_env = "/home/emeitz/.conda/envs/cunumeric-gpu"

After setting the LocalPreferences.toml you must re-build cuNumeric.jl by running:

julia -e 'using Pkg; Pkg.build()'

Using Local CUDA

A common reason to link against an existing environment is to avoid cupynumeric re-installing the entire CUDA toolkit. To make your own conda environment built off an existing CUDA install run the following with whatever version of CUDA is on the LD_LIBRARY_PATH in place of 12.2.

conda create --name myenv 
conda activate myenv
CONDA_OVERRIDE_CUDA="12.2" \
  conda install -c conda-forge -c legate cupynumeric