[zupengwang]

The CUDA implementation of int8_vectorwise_quant passes the input pointer directly to a kernel that assumes contiguous rows. Transposed inputs and row/column slices therefore produce incorrect quantized values and row scales without raising an error. This also affects int8_double_quant and FP16 inputs to Linear8bitLt.

Make the input contiguous before calling the CUDA kernel. Add regression coverage for transposes, strided slices with nonzero storage offsets, thresholds 0 and 6, both quantization functions, and Linear8bitLt outputs and input gradients. The CUDA vectorwise tests also run torch.library.opcheck.

Minimal reproducer on the unpatched CUDA backend:

import torch
from bitsandbytes.functional import int8_vectorwise_quant

torch.manual_seed(0)
x = torch.randn(64, 8, device="cuda", dtype=torch.float16).t()
actual, scales, _ = int8_vectorwise_quant(x)
expected, expected_scales, _ = int8_vectorwise_quant(x.contiguous())
torch.testing.assert_close(actual, expected, rtol=0, atol=0)
torch.testing.assert_close(scales, expected_scales, rtol=0, atol=0)

Validation against main 833649043474794b8fe7a4136e0c40faf077b2e0, using a source build on RTX 4090 (SM89), Python 3.12.14, PyTorch 2.13.0+cu130, and CUDA 13.0:

• All 20 new CUDA regression cases fail before the fix and pass after it, using the same native library.
• BNB_TEST_DEVICE=cuda python -m pytest tests/test_ops.py::TestLLMInt8Ops tests/test_linear8bitlt.py -q -ra: 84 passed, 2 existing xfails for torch.compile reduce-overhead mode. This total includes the 20 new cases.
• The two new tests with BNB_TEST_DEVICE=cpu: 14 passed, 6 skipped because double quantization is CUDA/ROCm-only. opcheck in the new test is scoped to CUDA; the default CPU backend has a separate existing output-stride mismatch with its fake implementation for transposed inputs.
• Ruff 0.14.3 lint/format checks and git diff --check passed.

These are local hardware results; upstream CI and other accelerators have not been run.