PyTorch fixes silent sparse-matrix errors from strided COO indices

A transposed coordinate list could move values to the wrong row in CSR and BSR conversion, and the CPU path could write past its output.

By · Published

Two transparent grids of orange cubes sit on a desk; an arrow points to a grid with its bottom row filled, beside a code screen and logo mug.

A transposed coordinate list could move values to the wrong row in CSR and BSR conversion, and the CPU path could write past its output.