ci : use hf-jobs-cpu-xl runner in server sanitize workflow

[ggerganov]

Overview

Update the Server (sanitize) workflow to use the hf-jobs-cpu-xl runner instead of hf-jobs-cpu-upgrade.

Additional information

Should fix https://github.com/ggml-org/llama.cpp/actions/runs/35821670631/job/107054637988#step:9:7781

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp

[ggerganov]

@CISC Seems like hf-jobs-cpu-xl is not recognized?

[CISC]

@CISC Seems like hf-jobs-cpu-xl is not recognized?

I see it in the README, but not sure where that name originated from, it's not listed here:
https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/_space_api.py#L68

[ggerganov]

Maybe we can move these to the new AMD runners.

[CISC]

Maybe we can move these to the new AMD runners.

A little awkward as we have to set up a container, but doable.

[ggerganov]

@CISC Seems like hf-jobs-cpu-xl is not recognized?

I see it in the README, but not sure where that name originated from, it's not listed here: https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/_space_api.py#L68

@abidlabs Does it make sense to add the cpu-xl hardware preset to the hf-jobs?

[abidlabs]

Ah this is a bug in the dispatcher, because in flavors.py we get the labels from SpaceHardware (which doesn't include cpu-xl or cpu-performance) instead of JobHardware I'll switch the map to JobHardware and then if you redeploy the ggml-org dispatcher, then this PR should pick up a runner as is I believe

[abidlabs]

Done: huggingface/jobs-actions#20

[ggerganov]

Thanks, it works now!

[CISC]

The jobs seem to hang and time out...

[ggerganov]

My guess is that running sanitized server builds in parallel starves the machine from resources. Let's try to disable the pytest workers for this workflow. I'll open a PR.