install cudf on databricks

Question

I am trying to use cudf on databricks.

I started following https://medium.com/rapids-ai/rapids-can-now-be-accessed-on-databricks-unified-analytics-platform-666e42284bd1. But the init script link is broken.

Then, I followed this link (https://github.com/rapidsai/spark-examples/blob/master/getting-started-guides/csp/databricks/databricks.md#start-a-databricks-cluster) which install the cudf jars on the cluster. Still I could not import cudf.

I also tried: %sh conda install -c rapidsai -c nvidia -c numba -c conda-forge cudf=0.13 python=3.7 cudatoolkit=10.1 which also failed with a long error ending with:

     active environment : /databricks/python
    active env location : /databricks/python
            shell level : 2
       user config file : /root/.condarc
 populated config files : /databricks/conda/.condarc
          conda version : 4.8.2
    conda-build version : not installed
         python version : 3.7.6.final.0
       virtual packages : __cuda=10.2
                          __glibc=2.27
       base environment : /databricks/conda  (writable)
           channel URLs : https://conda.anaconda.org/nvidia/linux-64
                          https://conda.anaconda.org/nvidia/noarch
                          https://conda.anaconda.org/rapidsai/linux-64
                          https://conda.anaconda.org/rapidsai/noarch
                          https://conda.anaconda.org/numba/linux-64
                          https://conda.anaconda.org/numba/noarch
                          https://conda.anaconda.org/conda-forge/linux-64
                          https://conda.anaconda.org/conda-forge/noarch
                          https://conda.anaconda.org/pytorch/linux-64
                          https://conda.anaconda.org/pytorch/noarch
                          https://repo.anaconda.com/pkgs/main/linux-64
                          https://repo.anaconda.com/pkgs/main/noarch
                          https://repo.anaconda.com/pkgs/r/linux-64
                          https://repo.anaconda.com/pkgs/r/noarch
          package cache : /databricks/python/pkgs
                          /local_disk0/conda/pkgs
       envs directories : /databricks/conda/envs
                          /root/.conda/envs
               platform : linux-64
             user-agent : conda/4.8.2 requests/2.22.0 CPython/3.7.6 Linux/4.4.0-1114-aws ubuntu/18.04.5 glibc/2.27
                UID:GID : 0:0
             netrc file : None
           offline mode : False


An unexpected error has occurred. Conda has prepared the above report.

Upload successful.

Any idea how to use cudf on a databricks cluster ?

Answer 1

I think the OP want to use python with cudf. If so, that is not covered in the documentation.

But I tried to add below into the generate-init-script.ipynb to make it work:

#Use mamba to install packages to speed up conda resolve time
base=$(conda info --base)
conda create -y -n mamba -c conda-forge mamba
pip uninstall -y pyarrow
${base}/envs/mamba/bin/mamba remove -y c-ares zstd libprotobuf pandas
${base}/envs/mamba/bin/mamba install -y "pyarrow=1.0.1" -c "conda-forge"
${base}/envs/mamba/bin/mamba install -y -c "rapidsai" -c "nvidia" -c "conda-forge" -c "defaults" "cudf=0.18" "cudatoolkit=10.1"
conda env remove -n mamba

Note: Change the cudf version and cudatoolkit according to your env.

Answer 2

I remember helping write that blog a while ago :). It's out of date now.

Karthik and team made some great updates since with spark-rapids. Here is the newest implementation of RAPIDs with databricks in spark: https://nvidia.github.io/spark-rapids/docs/get-started/getting-started-databricks.html. That should get you running with the newest version of cudf.

I'll ask that someone add a disclaimer with that like on that specific blog, so that others don't get confused either. Thanks for alerting us through this question!

Answer 3

Perhaps you need cudatoolkit=10.2? You have virtual packages : __cuda=10.2 in that report.

I am investigating install issues on a databricks GPU cluster (different issue though) and noted that the version of CUDA was 10.2 and not the 10.1 that I expected.