chore(typing): initial `ty` integration by tarekziade · Pull Request #43396 · huggingface/transformers

tarekziade · 2026-01-21T16:52:29Z

What does this PR do?

Initial ty integration. To avoid a gigantic, risky patch, let's start with a baby step where we add the tooling to make repo-check and activate it on a subset of the repo.

That gives us a human-readable patch, and allows us to get confortable with the tool before we expand it.

This patch:

adds ty to make repo-check and the CI
runs only against src/transformers/utils
has a large list of ignores in its configuration for now, that we will loosen later

Before submitting

This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
Did you read the contributor guideline,
Pull Request section?
Was this discussed/approved via a Github issue or the forum? Please add a link
to it if that's the case.
Did you make sure to update the documentation with your changes? Here are the
documentation guidelines, and
here are tips on formatting docstrings.
Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

HuggingFaceDocBuilderDev · 2026-01-21T17:02:03Z

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.

github-actions · 2026-01-22T10:13:39Z

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=43396&sha=6cce79

Cyrilvallez

ALright! Happy to start with ty! Asked quite a few questions in comments more for my understanding and get familiar with ty than anything else

Cyrilvallez · 2026-02-04T09:54:23Z

-    account_id = os.getenv("TRAINING_JOB_ARN").split(":")[4] if "TRAINING_JOB_ARN" in os.environ else None
+    training_job_arn = os.getenv("TRAINING_JOB_ARN")
+    account_id = training_job_arn.split(":")[4] if training_job_arn is not None else None


What's the reason of ty complaining on this one?

possibly-missing-attributebecause os.getenv can return None and ty can't know that the in os.environ check is similar. I think it's a pretty reasonable tradeoff, we remove that uncertainty by storing the getenv() value

Cyrilvallez · 2026-02-04T09:57:56Z

+        if hasattr(torch.backends, "fp32_precision"):
+            torch.backends.fp32_precision = precision_mode


It seem to require a lot of existence checks. Do you know how they are chosen to be required or not? Does ty base itself e.g. on your currently installed version of torch in the env? Because that may become an issue

In this very specific case ty can't infere the existence of this attributes:

At runtime, torch.backends module replaces itself with a GenericModule instance (line 122 in torch/backends/init.py)

GenericModule defines fp32_precision as a ContextProp descriptor (lines 116-119)

However, there's no .pyi stub file for torch.backends, so type checkers can't see that the attribute exists

Does ty base itself e.g. on your currently installed version of torch in the env?

yes it uses its stubs

tarekziade · 2026-02-04T14:24:44Z

ALright! Happy to start with ty! Asked quite a few questions in comments more for my understanding and get familiar with ty than anything else

Thanks for the first pass!

one remark about the assert is not None checks:

it's all the same pattern:

a class has an attribute it sets to None in its constructor
a method B makes the assumption is has another value than None, set by another method A

So it makes sense for a type checker to detect that one possible code execution pass is to call B with no previous call to A. Adding assertions like this is the simplest non intrusive fix. With no assertions, we need to refactor the class so calls to B can't happen before calls to A by design.

…alance with ty

…ts module

…hecks

github-actions · 2026-02-06T09:01:31Z

[For maintainers] Suggested jobs to run (before merge)

run-slow: dia, whisper

tarekziade · 2026-02-06T09:22:48Z

run-slow: dia, whisper

github-actions · 2026-02-06T09:23:58Z

This comment contains run-slow, running the specified jobs:

models: ["models/dia", "models/whisper"]
quantizations: []

github-actions · 2026-02-06T09:41:57Z

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	b98e83f4	merge commit
PR	3556795b	branch commit
main	0b2900dd	base commit

✅ No failing test specific to this PR 🎉 👏 !

tarekziade · 2026-02-20T07:38:52Z

supersed by #44167
(which is a branch in the same repo for convenency)

tarekziade force-pushed the tarekziade-ty branch from 6cce790 to e018467 Compare February 3, 2026 13:32

tarekziade requested a review from Cyrilvallez February 3, 2026 19:33

Cyrilvallez reviewed Feb 4, 2026

View reviewed changes

tarekziade force-pushed the tarekziade-ty branch from 24c0234 to 72b6d85 Compare February 5, 2026 08:01

tarekziade added 20 commits February 6, 2026 09:58

initial ty integration

da0da51

narrow ty check to utils and add more ignores

09efe27

circleci does not use the makefile, fix ty changes there

dac72f4

remove unecessary ignores

21bbc0a

removed a couple of ignores

3bed4d1

causal_mask can be a Tensor, BlockMask or None, lets be explicit

76f66f7

simplify to_json_string

5b07638

make it more readable - we know qc cannot be None here but its best b…

1b57a49

…alance with ty

explicitely assert training_tracker, do not silently ignore

7d65390

make _is_package_available return type unique, and fully private to i…

b634db3

…ts module

explicit contract at the base class level is nicer than duck typing c…

c226084

…hecks

add a comment about monky patching the logger

050a7a8

added ignore

448f487

better one

6b65a46

fixed test

2389e27

tweaks

064bcec

added a logger protocol

a71340a

removed some asserts

c998e27

just ignore type check on that one

5be4580

better descirption

06ebe02

tarekziade force-pushed the tarekziade-ty branch from 86ef683 to 06ebe02 Compare February 6, 2026 09:00

not needed anymore

74fd531

tarekziade added 2 commits February 6, 2026 10:09

yeah callable dont always have names

e043bb7

one more func name ficx

3556795

tarekziade closed this Feb 20, 2026

tarekziade deleted the tarekziade-ty branch February 20, 2026 07:40

		if hasattr(torch.backends, "fp32_precision"):
		torch.backends.fp32_precision = precision_mode

Conversation

tarekziade commented Jan 21, 2026

What does this PR do?

Before submitting

Who can review?

Uh oh!

HuggingFaceDocBuilderDev commented Jan 21, 2026

Uh oh!

github-actions bot commented Jan 22, 2026

Uh oh!

Cyrilvallez left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Cyrilvallez Feb 4, 2026

Choose a reason for hiding this comment

Uh oh!

tarekziade Feb 4, 2026

Choose a reason for hiding this comment

Uh oh!

Cyrilvallez Feb 4, 2026

Choose a reason for hiding this comment

Uh oh!

tarekziade Feb 4, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

tarekziade commented Feb 4, 2026

Uh oh!

github-actions bot commented Feb 6, 2026

Uh oh!

tarekziade commented Feb 6, 2026

Uh oh!

github-actions bot commented Feb 6, 2026

Uh oh!

github-actions bot commented Feb 6, 2026

CI Results

Commit Info

Uh oh!

tarekziade commented Feb 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

tarekziade commented Feb 20, 2026 •

edited

Loading