cuml-installation-notes

This page summarizes 3 possible ways to install the GPU-accelerated ML libraries offered by cuML.

Conda / Docker

See https://rapids.ai/start.html#get-rapids.

Conda and Docker are at the moment the two officially supported release channels for cuML. When installing cuML via Docker, RAPIDS components will be installed from published Conda packages to the 'rapids' Conda environment (e.g., see here).

Build from source without Conda and without multi-GPU support

• See this section for system dependencies of cuML libraries.

• Follow this page to install CUDA Toolkit 11.0+.

• Follow this link to install libnccl.

• Follow this link to install ucx.

• And then run the following:

git clone https://github.com/rapidsai/cuml.git
cd cuml
# git checkout branch-21.08
cd cpp
mkdir build
cd build
cmake \
  -DSINGLEGPU=ON \
  -DBUILD_CUML_TESTS=OFF \
  -DBUILD_PRIMS_TESTS=OFF \
  -DBUILD_CUML_EXAMPLES=OFF \
  -DBUILD_CUML_BENCH=OFF \
  -DBUILD_CUML_PRIMS_BENCH=OFF \
  -DCMAKE_INSTALL_PREFIX=/usr \
  ..
make -j $(nproc)

to build cuML libraries from source.

NOTE: even though in principle this should work, at least on a release branch of the cuml repo, this, unlike Conda/Docker, is not an officially supported release channel of cuml (yet) AFAIK. So, your mileage may vary. Also, the build could take a non-negligible amount of time, even on an 8-core machine with modern CPUs.

EDIT

It looks like for cuML 21.10 or above, the spdlog header files (which are required by rmm and possibly other cuML dependencies) will not be installed by default when you run make install. As a temp workaround for this, you can do the following:

• Go to https://github.com/rapidsai/rapids-cmake, switch to branch-<the cuML version you want to install>.
• See the required spdlog version in the rapids-cmake/cpm/versions.json config file (e.g., something like this: https://github.com/rapidsai/rapids-cmake/blob/e7a79a669f3ccec524f99c50f518c0879c33f19a/rapids-cmake/cpm/versions.json#L21).
• Build and install that required version of spdlog from source using the commit tagged v<required spdlog version> from the https://github.com/gabime/spdlog.git repo.

Build from source without Conda and with multi-GPU support

Now this is even further away from the beaten paths and appears somewhat brittle, but at least it has worked for the branch-21.08 branch of the cuML repo.

To build cuML successfully with multi-GPU support, one will need a library named libcumlprims ("prims" meaning primitives). Because the author of this note was unable to find any source code for libcumlprims on the internet, the only viable alternative so far seems to be downloading and installing a pre- built copy of libcumlprims instead of building it from source. So,

# NOTE: you should choose the most recent version of libcumlprims that matches
# the version of CUDA you intend to use. Looks like currently only x86-64 is
# supported.
#
# Also, the `| sudo tar -xjvf - -C /usr` part is probably not something you
# should run directly in production!!
wget -O - https://anaconda.org/nvidia/libcumlprims/21.06.00/download/linux-64/libcumlprims-21.06.00-cuda11.2_gfda2e6c_0.tar.bz2 | sudo tar -xjvf - -C /usr

and then, similar build instructions as above except for removal of the -DSINGLEGPU=ON flag:

git clone https://github.com/rapidsai/cuml.git
cd cuml
# git checkout branch-21.08
mkdir build
cd build
cmake \
  -DBUILD_CUML_TESTS=OFF \
  -DBUILD_PRIMS_TESTS=OFF \
  -DBUILD_CUML_EXAMPLES=OFF \
  -DBUILD_CUML_BENCH=OFF \
  -DBUILD_CUML_PRIMS_BENCH=OFF \
  -DCMAKE_INSTALL_PREFIX=/usr \
  ..
make -j $(nproc)

Also, see comment above re: how to intall possibly missing spdlog header files if you are working with cuML 21.10 or above.