Remove upstream partial transformers stubs

[Avasam]

Upstreamed at huggingface/transformers#38335

I'm not entirely certain if it's worth mentioning in the Upstreamed libraries section given how tiny the stub was. Up to you.

[isaac-mcfadyen]

Hey 👋

I'm a user of Pylance and would like to +1 the request for putting in the Upstreamed libraries section of the README here.

I am currently hitting this old issue (microsoft/pylance-release#1258) which was resolved by adding type stubs in this package.

I'm guessing that they've been removed from this repo and added to Transformers but a new version of Transformers with the stubs hasn't been published yet (meaning there's probably a small limbo period where that old issue will occur), so it might be good to mention just in case anyone else runs into this in the next few days.

[Avasam]

I could've kept this as draft until next transformers release since some type information is lost atm.

Right now as a quick fix you could install these stubs on the commit right before the partial transformers stubs were removed: pip install git+https://github.com/microsoft/python-type-stubs@aaa9f03e47df046345e6d17cea5a0ae1c53d1af7 if you really see a difference.

That would also cause pyright and mypy to find the stubs. Not just Pylance.

Looking at microsoft/pylance-release#1258 (comment) , I could not replicate those numbers, at most I got ~800ms
Similarly concerning huggingface/transformers#38335 (comment)

(Unless I just restarted the language server, but then, partial stubs or not, the initial run does take a lot more time, indeed 4-6s, because with from transformers import AutoTokenizer instead of from transformers.models.auto.tokenization_auto import AutoTokenizer analysis has to be run on a lot more files due to all the re-exports before pyright/pylance even finds the AutoTokenizer symbol)

[isaac-mcfadyen]

Interesting, perhaps my issue is something else then?

I also saw the initial ~3000ms at server reload, and then it looks like the language server is getting killed because of a timeout:

2025-05-31 14:02:42.911 [info] [Info  - 2:02:42 PM] (1492239) BG: [BG(1)] Long operation: binding: file:///data/PROJECT/.venv/lib/python3.12/site-packages/transformers/models/__init__.py (3385ms)
2025-05-31 14:02:42.917 [info] [Info  - 2:02:42 PM] (1492239) BG: [BG(1)] Long operation: binding: file:///data/PROJECT.venv/lib/python3.12/site-packages/transformers/__init__.py (3456ms)
2025-05-31 14:02:48.043 [info] (Client) The existing extension didn't exit within 10 seconds. New instance will start, but you might encounter issues.

I assumed that was the causing issues, because completions often hang for a while and syntax highlighting is sluggish (adding/changing a line sometimes takes upwards of ~2s to rehighlight), but maybe not? I'll try these stubs from the old commit and see if that works.

UPDATE: I think that works! I'm still seeing a tiny bit of lag for syntax highlighting, but more like ~1/2s instead of the ~2s from before!

[debonte]

@isaac-mcfadyen, for what it's worth, we haven't shipped a release of Pylance with this change yet.

[isaac-mcfadyen]

Oh that's definitely odd, my issue must be unrelated then... Sorry for hijacking this PR 😄