Fix scale slicing for negative indices on blocked dims

[Kaif10]

_slice_scale_for_dimension maps a data-tensor slice onto the scale with floor/ceil division by the block size, but never resolves negative indices first. A negative end with |end| < block_size ceil-divides to 0, so the scale is sliced to length zero.

For per-row quantization the block spans the whole row, so any negative end on the last dim triggers it:

m = torch.nn.Sequential(torch.nn.Linear(128, 64, bias=False, dtype=torch.bfloat16))
quantize_(m, Int8WeightOnlyConfig(version=2, granularity=PerRow()))
w = m[0].weight          # scale (64, 1)

w[:, :-32].scale.shape   # torch.Size([64, 0])   <- no error at slice time
w[:, :96].scale.shape    # torch.Size([64, 1])
w[:, :-32].dequantize()  # RuntimeError: shape '[64, 1]' is invalid for input of size 0

Same with Float8WeightOnlyConfig() / Float8DynamicActivationFloat8WeightConfig(granularity=PerRow()), which fails later with ZeroDivisionError. The slice itself succeeds and returns a tensor with a plausible shape, so the error surfaces far from its cause.

Negative start happens to give the right block because floor division lands on the same block when the dim is block-aligned, and dim=0 is unaffected because a block size of 1 delegates to aten.slice, which handles negatives natively. Only the ceil-divided end breaks.

Fix

Resolve negative start/end against data_shape[dim] (clamped at 0, same as Python slice semantics) before the block arithmetic. Since Float8Tensor, Int8Tensor (scale and zero_point) and the prototype static float8 tensor all call this helper, they are all fixed.

Tests

• test_int8_tensor_cpu.py::test_slice_negative_indices: CPU. PerRow weight-only, PerGroup(32) weight-only and PerRow dynamic, over (None, -32), (-32, None), (-96, -32), (-1000, None) on both dims. Each negative slice is compared against the equivalent non-negative slice on qdata, scale and dequantize(). The PerRow + negative-end cases fail on main and pass with this change; the rest pass on both and guard the unaffected paths.
• test_float8_tensor.py::test_slice_negative_indices: the same check for Float8Tensor with PerRow, next to the existing test_slice (GPU-gated like the rest of that class). I validated its body on CPU by substituting Float8WeightOnlyConfig: it fails on main with a (64, 0) scale and passes with this change.

ruff check and ruff format are clean. test_int8_tensor_cpu.py gains 12 passing tests. Its existing compile=True cases fail identically before and after this change on my machine (no C++ compiler for Inductor on Windows), so they're unrelated.

Suggested labels: module: inference, topic: bug fix.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4958

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[Kaif10]

Could a maintainer approve the CI workflows for this first-time-contributor PR? Suggested label: module: inference.

[Kaif10]

@andrewor14 friendly ping. This is a 4-line fix to _slice_scale_for_dimension for negative indices, with tests that fail on main. CI is still waiting on workflow approval.