fix: replace all instances of deprecated torch._check_is_size with torch._check

[neil-the-nowledgeable]

Summary

Closes #1933.

PyTorch's torch._check_is_size is being removed in a future release per pytorch/pytorch#169400 — the FutureWarning advises "Use _check(i >= 0) instead". This PR finds / replaces all older approach instances with the recommended substitution.

Scope

23 occurrences across 6 files:

File	Hits
bitsandbytes/_ops.py	8
bitsandbytes/backends/triton/ops.py	5
bitsandbytes/backends/default/ops.py	4
bitsandbytes/backends/cpu/ops.py	3
bitsandbytes/backends/cuda/ops.py	2
bitsandbytes/backends/hpu/ops.py	1

Each call torch._check_is_size(blocksize) is replaced with the upstream-recommended pattern:

torch._check(blocksize >= 0, lambda: f"Blocksize must be non-negative, got {blocksize}")

This matches the existing torch._check(...) style already used at adjacent lines in the same files.

Backwards compatibility

torch._check is available in all torch >= 2.0; bitsandbytes requires torch >= 2.2 per pyproject.toml. No backwards-compat shim needed.

Test plan

• pre-commit run --files <changed> passes (ruff + ruff-format + hygiene hooks)
• Replacement string verified at all 23 sites (count delta zero: 23 removed, 23 added)
• Diff preserves indentation; no other code paths touched

Additional audits

Audit structure adopted from @rapsealk's #1931 (asserts → exceptions refactor) for similar mechanical-substitution PRs.

• python -m py_compile on all 6 changed files: all pass
• No leftover torch._check_is_size anywhere in bitsandbytes/ after the substitution (count went 23 → 0)
• Argument variable is blocksize at all 23 sites — no slip in argument naming
• Call-site context: every call sits inside either a @register_fake(...) (FakeTensor / dispatch hook) or a @register_kernel(...) (backend kernel impl) decorator

Semantic equivalence under SymInt (FakeTensor / torch.compile)

_check_is_size(i) does two things: (1) asserts i >= 0, and (2) informs PyTorch's dynamic-shape engine that i is a size (i.e., non-negative). The substitution _check(blocksize >= 0, ...) preserves both behaviors:

• For plain Python int, both produce identical runtime checks.
• For SymInt (which can flow through register_fake under torch.compile), blocksize >= 0 evaluates to SymBool and torch._check of a SymBool informs the dynamic-shape engine of the same >= 0 guard.

This matches the upstream migration explicitly recommended in pytorch/pytorch#169400 ("Use _check(i >= 0) instead").

Out of scope

• csrc/ (C++/CUDA) — _check_is_size is a Python-only API
• tests/ and examples/ — no calls to _check_is_size outside bitsandbytes/ package code

[github-actions]

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.

[matthewdouglas]

LGTM, thanks!

[neil-the-nowledgeable]

you are most welcome. I'm very happy to be of service. I appreciate the excellent community support you've provided