Re: RFS: rccl/5.4.3-3~exp1 -- ROCm Communication Collectives Library

[Christian Kastner <ckk@debian.org>]

Hey Cory,

On 2024-03-14 05:36, Cordell Bloor wrote:
> I've added packages for tests and documentation to rccl. These tests
> require at least two GPUs to execute, so they're a bit different from
> the other libraries. However, aside from that, this is pretty standard
> 
> The rccl library also has performance tests in a separate repository
> [1]. The contents of that repository have not been packaged.

I initially postponed this because the two-GPU-requirement sounded
complicated, but it looks like I overthought it; it's really simple,
actually.

I assume this is still up-to-date? (I added a d/gbp.conf.)

Slightly tangential: What do you think about setting up a specific
worker configuration for multi-GPU tests, for example configuring
pinwheel as
  * amd64+gfx90a when one GPU is in use
  * amd64+gfx90a_x2 (or similar) when both GPUs are in use?

pinwheel/gfx90a is just one example, other configuration would of course
also work.

Best,
Christian