Ignore benign bitsandbytes _check_is_size FutureWarning in CI

[albertvillanova]

This PR silences a benign FutureWarning emitted in CI by the PEFT + quantization tests when a bitsandbytes-quantized model is loaded.

Motivation

test_peft_with_quantization in the DPO, KTO, and SFT trainer tests triggers a FutureWarning from bitsandbytes:

bitsandbytes/backends/cuda/ops.py:213: FutureWarning: _check_is_size will be removed in a future PyTorch release along with guard_size_oblivious. Use _check(i >= 0) instead.
  torch._check_is_size(blocksize)

The warning comes from bitsandbytes, which calls the deprecated torch._check_is_size in its CUDA quantization ops. TRL only triggers it indirectly by loading a quantized model; there is nothing actionable on the TRL side. The warning clutters CI.

Solution

The issue is already fixed upstream in bitsandbytes (bitsandbytes-foundation/bitsandbytes#1940), but that fix is unreleased (latest is 0.49.2, which predates it; it is only in the continuous-release_main build). Following the existing precedent in pyproject.toml (and the same approach as #6436), this filters the non-actionable upstream warning until a bitsandbytes release including the fix is available. Removal is tracked in #6447.

Changes

• Add a filterwarnings entry in pyproject.toml to ignore the _check_is_size FutureWarning from bitsandbytes

---

Note

Low Risk
Test/CI configuration only; no runtime or training behavior changes.

Overview
Adds a pytest filterwarnings ignore in pyproject.toml for the bitsandbytes FutureWarning about deprecated torch._check_is_size, which shows up when PEFT + quantization tests load 4-bit models.

This matches the existing pattern for other upstream noise (Liger, PyTorch JIT, etc.): documented with tracking issue #6447 and the upstream bitsandbytes fix PR, until a release includes the fix.

^{Reviewed by Cursor Bugbot for commit c72b8f4. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

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.