Add lora test workflow (WIP)

[ltoniazzi]

Context

Partially solving Issue #8708.

Running the tests introduced in PR #8957 in a new git workflow (inspired by .github/workflows/server.yml).

Todo

• Test if workflow runs successfully.
• Should I add a test for the feature in server : add lora hotswap endpoint #8857 too?
• Do we want to refactor test models storage in HF to 1 architecture per model repo? This is an example of HuggingFace approach.

Checks

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[ltoniazzi]

@ngxson Thanks updated!

By the way, should I add a test run for the llama-server features added in #8857 as well?

[ltoniazzi]

Should I remove this line and the cmake dependency above?

[ltoniazzi]

Or should I use it to add lora tests that involve the server?

[ltoniazzi]

@ngxson Hey 😃 ! Are you able to take a look a this PR next week?

[ngxson]

Sorry for the delay. This is not my current priority, but I'll see if I have time next week.

[ltoniazzi]

@ngxson do you have time to look at this, or can you point me to another maintainer that can help here?

[ngxson]

I think we don't really need this anymore, as this is not a very important part and running test script locally works fine in most cases (we don't have much changes related to lora anyway)

We also don't want to add too much workflows to the CI. Every 1-2 minutes added to the CI will make it slow overtime, let alone the build step of this script that can easily take 5 minutes of the CI. So I think it's better not to continue implementing this for now.

[ltoniazzi]

@ngxson Yes no problem!

Should we merge only the modification in the test script to clone from a specific commit, so there is not risk in with updates to the HF test models repo? (See updated diff)

[ngxson]

Yes that looks good. Btw next time you can use add a new PR (if you want). The current PR can be kept as draft for reference in the future. For open-source projects, it's common to close a PR or to keep a non-functioning PR as draft.

[ltoniazzi]

Yes, let's do that! Closed as not planning to implement the CI workflow