[Mazukiri]

Summary

• Convert inputs to contiguous in the CUDA backend for int8_vectorwise_quant, int8_linear_matmul and int8_mm_dequant before their raw pointers go to the kernels and cuBLASLt. This is the same approach Handle non-contiguous tensors in quantize/dequantize ops #1859 took for the blockwise and 4-bit ops.
• Without it, strided views are read as if they were dense. Examples are hidden_states[:, 0] in BERT-style poolers and classification heads, and hidden_states[:, -1:, :] with batch > 1. Every row except the first comes from the wrong address, and nothing reports an error. Linear8bitLt with fp16 activations reaches this through MatMul8bitLt. The repro and the effect on real models are in LLM.int8(): int8 CUDA ops ignore input strides, so Linear8bitLt silently returns wrong rows for non-contiguous inputs like hidden_states[:, 0] #2130.
• .contiguous() returns the tensor itself when it is already contiguous, so the usual path does not change.

Fixes #2130

Test plan

• New tests in TestNonContiguousInputs (tests/test_ops.py):
  • test_int8_vectorwise_quant_non_contiguous (threshold 0.0 and 6.0, with an outlier column)
  • test_int8_linear_matmul_non_contiguous
  • test_int8_mm_dequant_non_contiguous
• New test test_linear8bitlt_non_contiguous_input (tests/test_linear8bitlt.py), covering h[:, 0] and h[:, -1:, :] with threshold 0.0 and 6.0.
• NVIDIA L4 (sm_89), CUDA 13.0, torch 2.14.1, built from main @ a616886 with -DCOMPUTE_CAPABILITY=89: the 6 new CUDA test cases fail without the change and pass with it. With the change, TestNonContiguousInputs plus the new Linear8bitLt test give 84 passed, 36 skipped.
• Existing int8 tests on the same L4: tests/test_ops.py, tests/test_linear8bitlt.py, tests/test_functional.py and tests/test_autograd.py, selected with -k "int8 or Int8 or matmullt or linear8bit or NonContiguous or non_contiguous or LLMInt8 or linear_no_igemmlt". With the change: 1338 passed, 52 skipped, 2 xfailed. Without it: 1332 passed, and the only failures are the 6 new CUDA test cases.
• CPU: the new tests pass with and without the change, because the CPU path is not affected.
• ruff check and ruff format --check (ruff 0.14.x) are clean.
• End to end on the same L4 with main @ a616886. Models were loaded with load_in_8bit=True and dtype=torch.float16, and a batch of 8 was compared with running the same inputs one at a time:
  • distilbert-base-uncased-finetuned-sst-2-english: before the change, every sentence in the batch is labelled positive (agreement 0.50); after it, the labels match (agreement 1.00).
  • cross-encoder/ms-marco-MiniLM-L6-v2: the top-3 goes from [4, 0, 3] back to [0, 2, 6].

🤖 Generated with Claude Code