[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: ROCm Enabled gloo



Hi Mo,

On 2025-07-01 08:55, M. Zhou wrote:
-------------------------------------------------------------------------

CMake Warning at cmake/Hip.cmake:22 (find_package):
   Could not find a configuration file for package "HIP" that is compatible
   with requested version "1.0".
The following configuration files were considered but not accepted: /usr/lib/aarch64-linux-gnu/cmake/hip/hip-config.cmake, version: 5.7.0
     /lib/aarch64-linux-gnu/cmake/hip/hip-config.cmake, version: 5.7.0
Call Stack (most recent call first):
   cmake/Dependencies.cmake:140 (include)
   CMakeLists.txt:111 (include)

The gloo library is using the old hip_add_library functionality provided by /usr/lib/*/cmake/hip/FindHIP.cmake. I'm not sure exactly why it's not being found. It's probably either because they're specifying version 1.0 when they do the find_package call (which doesn't make much sense to me), or because the location of FindHIP.cmake needs to be added to the CMAKE_MODULE_PATH for the build.

In any case, I recall patching this in a few places when I was trying to build PyTorch upstream a while back. I didn't keep my patches around, but it should be an easy fix. Spaarsh, if you have any trouble with gloo-rocm, please let me know and I'd be happy to lend a hand.

Over the long-term, I might recommend that upstream move to using enable_language(HIP), given that they're already using enable_language(CUDA). The hip_add_library functions in FindHIP.cmake predate the existence of proper HIP support in CMake and there's better ways of doing things now.

Sincerely,
Cory Bloor


Reply to: