Nail in edge case of torch dtype being overriden permantly in the case of an error #35845
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| @@ -245,6 +246,25 @@ def set_zero3_state(): | |||
| _is_ds_init_called = False | |||
|
|
|||
|
|
|||
| def restore_default_torch_dtype(func): | |||
There was a problem hiding this comment.
I think this approach should generally work. It would fail if there was ever a case where calling from_pretrained should purposefully change the default dtype for the rest of the process. I'm not sure if such a use case exist, just wanted to highlight it.
Co-authored-by: Benjamin Bossan <BenjaminBossan@users.noreply.github.com>
ArthurZucker
approved these changes
Feb 3, 2025
|
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. |
MekkCyber
pushed a commit
that referenced
this pull request
Feb 7, 2025
…e of an error (#35845) * Nail in edge case of torch dtype * Rm unused func * Apply suggestions from code review Co-authored-by: Benjamin Bossan <BenjaminBossan@users.noreply.github.com> * Refactor tests to only mock what we need, don't introduce injection functions * SetUp/TearDown * Do super --------- Co-authored-by: Benjamin Bossan <BenjaminBossan@users.noreply.github.com>
elvircrn
pushed a commit
to elvircrn/transformers
that referenced
this pull request
Feb 13, 2025
…e of an error (huggingface#35845) * Nail in edge case of torch dtype * Rm unused func * Apply suggestions from code review Co-authored-by: Benjamin Bossan <BenjaminBossan@users.noreply.github.com> * Refactor tests to only mock what we need, don't introduce injection functions * SetUp/TearDown * Do super --------- Co-authored-by: Benjamin Bossan <BenjaminBossan@users.noreply.github.com>
sbucaille
pushed a commit
to sbucaille/transformers
that referenced
this pull request
Feb 16, 2025
…e of an error (huggingface#35845) * Nail in edge case of torch dtype * Rm unused func * Apply suggestions from code review Co-authored-by: Benjamin Bossan <BenjaminBossan@users.noreply.github.com> * Refactor tests to only mock what we need, don't introduce injection functions * SetUp/TearDown * Do super --------- Co-authored-by: Benjamin Bossan <BenjaminBossan@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
@BenjaminBossan and I have been trying to hunt down an issue causing some headache over the last few days, as to why some tests were failing if done in specific instances.
What we cornered it to, is if a user:
from_pretrainedor_from_config), with adtypeother than the torch default dtypeWhat will occur is the torch dtype will permantly change, which is why some users may "discover" this issue and it goes away after a full restart (a likely scenario that could happen). What we currently do doesn't catch if we have an exception, this does, and is used as an additive
This PR introduces a safety decorator called
restore_default_torch_dtypewhich will restore whatever torch's original dtype was prior to calling the function (since we restore it anyways at the end).This PR also adds a new
test_injectionfunction. The aim is for this function to be mocked at critical points where we need testing, and we cannot easily get to it (e.g. in this case, there was only one function we could mock infrom_pretrainedand zero functions we could mock in_from_configto make sure this fix worked!)Fixes # (issue)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
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.
@SunMarc @BenjaminBossan @ArthurZucker