Re: ROCM support in ucx, mpi
Hi Alastair,
On 2023-11-15 05:37, Alastair McKinstry wrote:
Recent releases of UCX and MPI support ROCM. I would like to enable
that capability.
Can the ROCm team please advise? I do not have amd locally (mostly
devel on arm64) so I would like to know how I would enable testing -
is there CI/CD for this hardware available?
I am working with a couple universities to ensure there are a variety of
AMD GPU servers available for testing. There should be some
announcements in the upcoming months. For the moment, the AMD GPU
servers for the ROCm team's CI are all hosted in private residences,
which may not be set up to allow outside access. However, as Christian
mentioned, they will run autopkgtests on a variety of AMD GPUs and
publish the results.
AMD donated some GPUs last year and Jonathan Carter was setting up a
machine for Debian Developers to use [1]. You could check with him to
see if that machine is available. Additionally, I think Jonathan may
still have a spare Radeon RX 6800 that has not yet been claimed, so that
may be an option if you have an otherwise suitable local machine. On
Debian, the entire ROCm stack has been built for arm64 but AFAIK nobody
has tested it on anything but amd64.
Sincerely,
Cory Bloor
[1]: https://lists.debian.org/debian-ai/2022/11/msg00021.html
Reply to: