PyTorch 2.14.1 has been released, fixing linear algebra regressions and silent correctness issues on MPS with functions such as torch.linalg.lstsq and torch.linalg.svd. These include incorrect solutions for complex batched underdetermined systems, as well as inaccurate small singular values and non-orthogonal U for rank-deficient and ill-conditioned inputs. The CUDA 13.2 Linux binary has also been updated to CUDA 13.2.2, resolving two issues that caused corrupted register values: NVFP4 matrix multiplication ignoring tensor-level scaling and nested thread divergence.
Source:PyTorch 正式版本发布 · Read Original
Leave a Reply