Disable multi-gpu jobs for now

[ydshieh]

What does this PR do?

Disable multi-gpu jobs for model jobs in daily CI.

Context

Since 2 weeks, we can't receive the CI reports as the workflow runs failed to finish. There were remaining jobs waiting for runners. After a deeper look, they were all multi-gpu jobs.

@glegendre01 mentioned he is working on new runners, but there are still some /mnt/cache issues.

In the meantime, it's best for us to disable those multi-gpu jobs so we can still have some reports regarding single-gpu jobs.

example run page: https://github.com/huggingface/transformers/actions/runs/10576188096

I will revert #33136 before merge.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[amyeroberts]

@ydshieh Do we know why they're not running? e.g. is it OOM, running too long, etc.?

[ydshieh]

No. Some jobs could not find runners (or after waiting quite long) to run even if I didn't see any other jobs occupy those specified runner labels. (So it's more infra stuff.)

(some jobs indeed take long to run the tests - counting after once getting a runner, but that's another issue)

Link to an internal thread

[ydshieh]

Hi @amyeroberts I think we will need @glegendre01 to investigate this when he has time. In the meantime, it's probably better if we can have at least the single gpu jobs running and having their reports. There are indeed something we have to fix (some jobs take 3 hours to run and even expire in 6 hours etc.)