[BUG] RMM 24.10 build script fails to build python bindings

[ccoulombe]

Describe the bug
Building from source v24.10 fails with cmake not finding RMMConfig.

Steps/Code to reproduce bug

git clone --recurse-submodules https://github.com/rapidsai/rmm.git --branch v24.10.00
cd rmm
module load python/3.11 cuda/12
virtualenv ~/ENV && source ~/ENV/bin/activate
pip install --upgrade pip
pip install 'rapids-build-backend>=0.3.0,<0.4.0.dev0' "scikit-build-core[pyproject]>=0.10.0"
bash build.sh

returns :

    
        Could not find a package configuration file provided by "rmm" (requested
        version 24.10.00) with any of the following names:
      
          rmmConfig.cmake
          rmm-config.cmake

Expected behavior
Build librmm, install librmm, build rmm, install rmm, successfully. The script should set where to find RMMConfig

Environment details (please complete the following information):

• Environment location: Bare-metal
• Method of RMM install: source
• Please run and attach the output of the rmm/print_env.sh script to gather relevant environment details
  Relevent env. details:

***CMake***
/cvmfs/soft.computecanada.ca/custom/bin/cmake
cmake version 3.31.0

CMake suite maintained and supported by Kitware (kitware.com/cmake).

***g++***
/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/x86_64-pc-linux-gnu/gcc-bin/12/g++
g++ (Gentoo 12.3.1_p20230526 p2) 12.3.1 20230526
Copyright (C) 2022 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.


***nvcc***
/cvmfs/soft.computecanada.ca/easybuild/software/2023/x86-64-v3/Core/cudacore/12.2.2/bin/nvcc
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Tue_Aug_15_22:02:13_PDT_2023
Cuda compilation tools, release 12.2, V12.2.140
Build cuda_12.2.r12.2/compiler.33191640_0

***Python***
/tmp/coulombc/ENV/bin/python
Python 3.11.5

print_env.sh: line 58: type: conda: not found
***pip packages***
/tmp/coulombc/ENV/bin/pip
Package                          Version
-------------------------------- -----------------------
attrs                            24.2.0+computecanada
jsonschema                       4.23.0+computecanada
jsonschema_specifications        2024.10.1+computecanada
packaging                        24.2+computecanada
pathspec                         0.12.1+computecanada
pip                              24.2+computecanada
PyYAML                           6.0.2+computecanada
rapids-build-backend             0.3.2
rapids-dependency-file-generator 1.16.0
referencing                      0.35.1+computecanada
rpds_py                          0.20.0+computecanada
scikit_build_core                0.10.7+computecanada
setuptools                       75.2.0
tomlkit                          0.13.2
wheel                            0.44.0

[ccoulombe]

Specifying CMAKE_PREFIX_PATH helped/solved it:

bash build.sh rmm --cmake-args=\"-DCMAKE_PREFIX_PATH=$PWD/build/install/lib64/cmake\"

[bdice]

Thanks for reporting @ccoulombe. Can you give a little more detail on the CMAKE_PREFIX_PATH? Do you think there is a flaw in RMM's build logic that could be corrected or is it something needed for your particular system? We test build.sh in some of our builds, and I would expect the steps you followed to work. I will try to reproduce your steps on my system (outside of a conda environment) to verify that things work as expected.

[vyasr]

From a quick glance, the build.sh script was designed for use in a conda environment and is making some invalid assumptions. In our CI we only use this script when building in a conda environment, and as a result the INSTALL_PREFIX is set to the CONDA_PREFIX, which is always on CMake's default search paths. When you use build.sh outside of conda the line I linked above is installing to $LIBRMM_BUILD_DIR/install, which is not part of the default search path. The solution would be to add -DCMAKE_PREFIX_PATH=${INSTALL_PREFIX}; to this line.