Enable rccl and rccl-tests for more amdgpu targets

[ScottTodd]

Motivation

RCCL builds have historically been too slow to support for all targets, see

• [Issue] RCCL slow build times (6-10 hours to compile for some GPU families) #2130
• [RCCL] Parallel build cleaned up v1 rocm-systems#4795
• Don't exclude strix halo in build #4661

Build time has improved recently. Let's see if it's improved enough to re-enable these targets.

Test Plan

Multi-arch CI on this PR, watch comm-libs stage

• Build should succeed
• Build time should be reasonable (not a bottleneck compared to other stages)
• Tests should run (and probably pass?)
• Inspect the configure/build logs and uploaded artifacts (look for what was actually enabled, what kpack files are included, etc.)

Test Result

• comm-libs build stage succeeded (logs here)
• comm-libs build time: 1h43m, math-libs gfx94X build time: 1h41m
• RCCL tests passed on Linux gfx942 (logs here)
• RCCL tests passed on Linux gfx950 (logs here)
• RCCL artifacts were split for each architecture, e.g. rccl_lib_gfx1151.tar.zst which includes .kpack/rccl_lib_gfx1151.kpack

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[ScottTodd]

Ah... I thought the test:rccl label would run just RCCL tests, but it triggered "full" tests for all components :/ https://github.com/ROCm/TheRock/actions/runs/25134138451?pr=4935

cc @geomin12

[marbre]

Ah... I thought the test:rccl label would run just RCCL tests, but it triggered "full" tests for all components :/ https://github.com/ROCm/TheRock/actions/runs/25134138451?pr=4935

cc @geomin12

It should or rather did work for me before. Is this a different label eventually?

[alex-breslow-amd]

Thanks for doing this Scott! :-)

[ScottTodd]

comm-libs build time: 1h43m, math-libs gfx94X build time: 1h41m

Note that this was for a build with all targets, which is what we'll see in release builds and opt-in on PRs. Most local builds and workflow runs on PRs and postsubmit will use a more limited list of targets with faster build time.

[fjankovi]

Thanks for this. We definitely need rccl for gfx1100/gfx1201 :)