Update runs-on in workflow files

[ydshieh]

What does this PR do?

I also updated the labels on runners to have t4 and daily-ci, push-ci etc.

[LysandreJik]

Do we need the t4? If we want to move to V100s or better we'll need to update again. As long as it's nvidia-gpu it should suffice, no?

[mfuntowicz]

Maybe we can keep the t4 tag on the runner enrollment to identity which hardware it is running on and, in the future, if you end up having multiple flavours like we'll have for amd [mi210, mi250, mi300] you can easily target them (or not).

See how we would like to approach it for AMD gpu flavours: https://github.com/huggingface/transformers/pull/26346/files#diff-0517e7d6ca0b03d887391fd62602e2891e13f279d32d464e3f73c6e864b8152fR48

Within the workflow I agree, at the moment, it might not bring any additional constraints.

[ydshieh]

Technically no need t4, it's there just to keep more info. explicitly.

[LysandreJik]

Also is there a difference between the past-ci and daily-ci? Or is it just that you'd like to have runners dedicated to these tasks?

[mfuntowicz]

Also is there a difference between the past-ci and daily-ci? Or is it just that you'd like to have runners dedicated to these tasks?

It's a typo for push-ci no?

[LysandreJik]

I see runners with the following tags on nvidia's side: push-ci, daily-ci, past-ci, doctest-ci

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[ydshieh]

Also is there a difference between the past-ci and daily-ci? Or is it just that you'd like to have runners dedicated to these tasks?

So far, each CI workflow is running in a dedicated runner by using tags/labels like single-gpu-docker (daily CI has it) or docker-gpu (push CI has it). However I find this is difficult to maintain (they are all single gpu and docker but with differnt kind of label names to mean it).

Using explicit daily-ci, push-ci, past-ci etc. makes things clear and easy to understand IMO. But the machines are all of the same.

[ydshieh]

Talking to @glegendre01 we decide to keep these extra labels (push-ci etc.) at this moment, and will remove them just before we switch to AWS with the auto-scaling runners setting.

[glegendre01]

Yes, i'm not ready to deploy new runners on transformers now (focusing on 2 others topics). But will be able to handle it next week. at this moment we will remove extra labels

[LysandreJik]

Thanks!