CI: setup-dependent pip cache

[gante]

What does this PR do?

This PR makes two changes to the way we cache our pip dependencies in the add-model-like.yml GH actions workflow:

1. The name of the cache depends on the hash of setup.py;
2. We do not restore the cache from partial name matches.

(this pattern exists in one of our CI files, github-torch-hub.yml , here)

Together, these changes will make us start from a fresh environment whenever we change setup.py. Having a stale cache was causing us to have dependency problems (e.g. old, incompatible protobuf version), and potentially miss dependency issues from fresh installs.

If you agree, I will also port these changes to model-templates.yml and update_metdata.yml, which have the same pattern/issue. EDIT: ported.

[sgugger]

Thanks for fixing! Pinging @LysandreJik also on this.

[LysandreJik]

Ok for me!

[HuggingFaceDocBuilderDev]

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

[gante]

Cool, porting the changes to the other two files as well then 👍

[ydshieh]

I learn this thing from you in this PR, @gante!

LGTM.

[ydshieh]

@gante: to learn more from you: how you figured out the cause for the error you mentioned:

https://github.com/huggingface/transformers/runs/6007067240?check_suite_focus=true

If it was me, I don't even know if I could figure this out!

[gante]

@ydshieh It definitely helps that I had this exact issue (stale CI caches) in my previous role :)

To pin the error to this issue, I reran the failing CI workflow locally, from a fresh virtual env. Since it ran without issues, I had a look at the .yml file, and saw that it had a cache for pip. Then I went on to see that pip install -e .[dev] was doing in the failing CI file, and I noticed that it had error messages due to incompatible package versions, which I did not have locally -- because an old version was cached.