fix(broadcast_tensors): handle zero-size dims in broadcast shape computation

[cccxxttt]

Summary

_compute_broadcast_shape in broadcast_tensors.py computed each output dimension as max(dim_size, shape[i]). This is wrong for the broadcasting rule "a size of 1 is broadcastable": max() treats 0 < 1, so a zero-size dimension paired with a singleton (1) collapsed to 1 instead of staying 0. This violates PyTorch broadcasting semantics, where any non-1 size (including 0) wins over a 1, and only two distinct non-1 sizes are incompatible.

Concretely, torch.broadcast_shapes([0], [1]) should give (0,), but the old code returned (1,). This broke the shape=(0,) (empty-tensor) cases in test_smooth_l1_loss, where the loss inputs must broadcast to a zero-size output.

Fix

Resolve each dimension by taking the unique non-1 size:

• skip size-1 inputs (they are broadcastable and never determine the result),
• the first non-1 size wins,
• a second, different non-1 size raises a RuntimeError (matching torch.broadcast_shapes's "The size of tensor a (...) must match the size of tensor b (...) at non-singleton dimension" message),
• if every input is 1, the result is 1.

This now matches torch.broadcast_shapes behavior, including for zero-size dims:

# before
broadcast_shapes([0], [1])  -> (1,)   # wrong
# after
broadcast_shapes([0], [1])  -> (0,)   # correct
broadcast_shapes([0,3], [1,3]) -> (0,3)
broadcast_shapes([2], [3]) -> RuntimeError

Test

The shape=(0,) cases in tests/test_smooth_l1_loss.py (test_smooth_l1_loss and test_smooth_l1_loss_backward) were previously failing because of this; they now pass.

CUDA_VISIBLE_DEVICES=<n> python3 -m pytest tests/test_smooth_l1_loss.py -v
# 523 passed

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.