Remove bitsandbytes _check_is_size warning filter

[yashb98]

What does this PR do?

Removes the _check_is_size filterwarnings entry from pyproject.toml, which is the action item in #6447.

Fixes #6447

The entry was added because bitsandbytes called the deprecated torch._check_is_size in its CUDA quant ops. The upstream fix (bitsandbytes-foundation/bitsandbytes#1940) was merged but unreleased at the time, so the issue asked to remove the filter once a release containing it reached CI. bitsandbytes 0.50.0 is that release, and CI is already on it (#6546).

Verification

Ran the three tests named in the issue with the filter removed, on an aarch64 GB10, torch 2.13.0+cu130, capturing all warnings:

bitsandbytes	result	_check_is_size warnings
0.50.0	3 passed	0
0.49.2	3 passed	14, from bitsandbytes/backends/cuda/ops.py:213

So the filter is genuinely dead on 0.50.0, and the 0.49.2 row confirms the check is meaningful rather than vacuous: it does still catch the warning on the older release. Independently, _check_is_size appears 0 times anywhere in the installed 0.50.0 package and 6 times in 0.49.2.

Worth noting that filterwarnings does not start with error, so this cannot fail CI on an older bitsandbytes either. It would only restore log noise.

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? Please add a link to it if that's the case. Yes, Remove bitsandbytes _check_is_size warning filter once the minimum required bitsandbytes is >= 0.50.0 #6447, which states this removal as its action item.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests? No new tests, the three existing test_peft_with_quantization tests already cover it; results above.

AI writing disclosure

• No AI usage: the PR was written entirely by a human.
• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.
• AI-generated: the PR was mostly or fully generated by an AI tool.

Who can review?

Anyone once CI is green.

---

Note

Low Risk
Test-configuration-only change with no runtime or dependency logic; worst case on an older bitsandbytes is extra warning noise in logs, not CI failure.

Overview
Removes the pytest filterwarnings ignore for bitsandbytes’ deprecated torch._check_is_size FutureWarning from pyproject.toml, closing the follow-up in #6447.

That entry was only needed while CI still pulled bitsandbytes before the upstream fix shipped; with bitsandbytes 0.50.0 on CI, PEFT + quantization tests no longer emit that warning, so the filter is dead weight.

^{Reviewed by Cursor Bugbot for commit f232d09. Bugbot is set up for automated code reviews on this repo. Configure here.}

[albertvillanova]

Thanks for your contribution, @yashb98.

Unfortunately, I'm going to close this, and the reason is a wording bug in my issue #6447, not a problem with the PR.

#6447 says to remove the filter "once bitsandbytes ships the fix". That criterion is wrong, and this PR implements it faithfully. The correct criterion is: remove the filter once TRL's minimum required bitsandbytes version is >= the first release containing the fix (0.50.0).

pyproject.toml declares bitsandbytes with no lower bound, so every resolvable release is still a supported configuration, and filterwarnings is part of the declared test configuration for all of them, not a snapshot of what today's resolver installs in CI.

I am going to reword #6447 so the action item is unambiguous, and keep it open until the floor moves.

Apologies for the wasted round trip: the fault is in the issue, not in your work.

[yashb98]

No worries, and thanks for the explanation. The declared-config vs CI snapshot point makes sense. I read "ships the fix" as the trigger and didn't account for the missing lower bound. I'll watch #6447 and resubmit the same diff once the minimum required bitsandbytes moves to >= 0.50.0.