Does rocm xgboost work on linux?

[nipnipj]

Does rocm xgboost work on linux? Particularly archlinux.
How can I use my amd gpu with xgboost on archlinux?

[hliuca]

Hi @nipnipj yes it works on Linux. It needs ROCm to compile, https://rocm.docs.amd.com/projects/install-on-linux/en/latest/.

But I didn't try on archlinux, as it seems archlinux has not been supported by ROCm officially. Do you have docker? It will be easier to use ROCm on supported OS, such as https://hub.docker.com/r/rocm/dev-ubuntu-22.04/tags.

[nipnipj]

I read that xgboost does support Nvidia gpu's, but not AMD gpu's. Using xgboost with a Nvidia gpu on archlinux is easy, thanks to CUDA.
When instaling https://aur.archlinux.org/packages/opencl-amd. I can use my AMD gpu and run Stan on GPU. It seem that AMD gpu's are not currently supported by softtware like xgboost, tho.

[hliuca]

Hi @nipnipj xgboost supports ROCm now, here is the build instructions,
https://github.com/ROCm/xgboost/blob/master-rocm/README-ROCm.md

if you see issues, let me know please. Thank you.

[nipnipj]

I got the following error after cmake .. -DUSE_HIP=ON -DR_LIB=ON

CMake Error at CMakeLists.txt:273 (find_package):
  By not providing "Findrocthrust.cmake" in CMAKE_MODULE_PATH this project
  has asked CMake to find a package configuration file provided by
  "rocthrust", but CMake did not find one.

  Could not find a package configuration file provided by "rocthrust" with
  any of the following names:

    rocthrustConfig.cmake
    rocthrust-config.cmake

  Add the installation prefix of "rocthrust" to CMAKE_PREFIX_PATH or set
  "rocthrust_DIR" to a directory containing one of the above files.  If
  "rocthrust" provides a separate development package or SDK, be sure it has
  been installed.


-- Configuring incomplete, errors occurred!

[hliuca]

@nipnipj the reason is rocthrust isn't installed,

apt install rocthrust-dev

if it complains about a package, which can be found through,
apt list | grep some_missing_package

[nipnipj]

I was missing 2 packages

sudo pacman -S rocthrust
sudo pacman -S hipcub

Now it is building.

[hliuca]

excellent @nipnipj :-)

[nipnipj]

@hliuca It seem that it is working. Now I need to test it on R, I'll try version 2.0.3 since that's the version that works with tidymodels pacakge.

[nipnipj]

I confirm AMD gpu is now visible. Thanks.
Tho there is an error that make my r session close crash. No error message is displayed.

[nipnipj]

Not sure whether this is helpful here, but this is the error from R console:

:0:/long_pathname_so_that_rpms_can_package_the_debug_info/src/clr/hipamd/src/hip_global.cpp:117 : 4511938730d us:  Cannot find Symbol with name: _ZN7xgboost4tree20EvaluateSplitsKernelILi64EEEvjNS_6common4SpanIKNS0_19EvaluateSplitInputsELm18446744073709551615EEENS0_25EvaluateSplitSharedInputsENS3_IjLm18446744073709551615EEENS0_13TreeEvaluator14SplitEvaluatorINS0_16GPUTrainingParamEEENS3_INS0_20DeviceSplitCandidateELm18446744073709551615EEE
fish: Job 1, 'R' terminated by signal SIGABRT (Abort)

[hliuca]

What is your GPU? Instinct or navi? Thanks.

[nipnipj]

It is a RX 6650 XT. A navi one.

[hliuca]

I see. I work on Instinct GPU all the time and I don't have access to navi cards. Let me see if I can find a navi card and do some debug. I will update here. Thank you.

[hliuca]

Updates:
1). The root cause is Navi cards and MI GPUs have different architectures. Navi cards use wave32 (warp size 32), and MI GPUs use wave64 (warp size 64).
2). Not every ROCm works Radeon. We need a special ROCm version for Radeon GPUs, https://rocm.docs.amd.com/projects/radeon/en/latest/index.html
3). I have a fix now, and I will have to modify a few repos and branches. I will update here in the next days.

[hliuca]

@nipnipj temp fix,

1. Please use rocm 6.3.4 (official for radeon cards).
2. https://github.com/ROCm/xgboost/tree/release_2.0.3-rocm_navi

I need to test this branch and then merge to release.