Handle non-contiguous inputs in CPU blockwise quantize/dequantize

[egeozkoc]

Summary

Fixes #1995. The CPU quantize_blockwise / dequantize_blockwise kernels pass raw pointers (get_ptr) to the native library, which assumes contiguous memory. Non-contiguous inputs (e.g. strided slices) were read in physical order and produced silently incorrect results — the same bug class as #1342 / #1690, which #1859 fixed for CUDA only.

Per the maintainer's guidance on #1995, this applies the same approach as the CUDA backend: convert the input with A = A.contiguous() at kernel entry.

Changes

• bitsandbytes/backends/cpu/ops.py: add A = A.contiguous() at the start of the CPU quantize_blockwise and dequantize_blockwise kernels (in the latter, before torch.empty_like so the output is contiguous as well).
• tests/test_ops.py: enable the previously CPU-skipped TestNonContiguousInputs blockwise tests.

Verification

Run locally on the CPU backend (with the compiled native library):

• The two enabled blockwise non-contiguous tests now pass — 18 cases (fp16/bf16/fp32 × blocksize 64/128/256); these were previously silently incorrect.
• All CPU blockwise tests pass with no regression on the contiguous path (.contiguous() is a no-op there): 289 passed, 0 failed.

Minimal repro before/after:

import torch
import bitsandbytes.functional as F

A = torch.randn(64, 128)
A_nc = A[::2, :]                       # non-contiguous
qn, sn = F.quantize_blockwise(A_nc, blocksize=64)
qc, sc = F.quantize_blockwise(A_nc.contiguous(), blocksize=64)
print(torch.equal(qn, qc))             # before: False, after: True

🤖 Generated with Claude 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.