[feat] Add lru_cache to import_utils calls that did not previously have it

[tisles]

Fixes #1576

Adding @cache to the functions in import_utils results in a constant time for LORA loading, regardless of the length of python path. Have verified this using a python path with 50,000 directories - previously the lora load time was 259 seconds, now it's down to a constant time of 1.3 seconds 👍🏻

With some flamegraphs to prove! Before:

After:

[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.

[BenjaminBossan]

Thanks a lot for the PR and testing the solution, great that it solves the issue. Just out of curiosity, is there a specific reason for setting maxsize=None? From my understanding, just using the default should work equally well here.

[tisles]

Thanks a lot for the PR and testing the solution, great that it solves the issue. Just out of curiosity, is there a specific reason for setting maxsize=None? From my understanding, just using the default should work equally well here.

Yep, good point - you're right. I've updated it to remove the None setting.

[BenjaminBossan]

Thanks for identifying the issue and providing a fix, LGTM.