pytorch releases

Details

torch.linalg.lstsq: Fixed an issue where incorrect solutions were returned for complex batched underdetermined systems on MPS.
torch.linalg.svd: Resolved issues on MPS where U was not orthogonal or small singular values were inaccurately computed for rank-deficient and ill-conditioned inputs.
CUDA 13.2.2 Update: Updated to NVIDIA CUDA 13.2.2 binaries to resolve the following two critical issues:
- cuBLAS: Fixed an issue where cublasLtMatmul() ignored tensor-wide scaling during NVFP4 matrix multiplication.
- Compiler: Fixed an issue where register values were corrupted or left in a stale state due to reconvergence failures in kernels with nested thread branching.

torch.linalg.svd, torch.linalg.svdvals, torch.linalg.lstsq: Fixed an issue where inputs with more than 8192 elements failed on MPS due to Metal pipeline state errors.
torch.svd(out=): Fixed an internal assertion failure when processing complex inputs on MPS.

torch.linalg.lstsq: Fixed an issue where incorrect solutions were returned for complex batched underdetermined systems on MPS.
torch.linalg.svd: Resolved issues on MPS where U was not orthogonal or small singular values were inaccurately computed for rank-deficient and ill-conditioned inputs.
CUDA 13.2.2 Update: Updated to NVIDIA CUDA 13.2.2 binaries to resolve the following two critical issues:
- cuBLAS: Fixed an issue where cublasLtMatmul() ignored tensor-wide scaling during NVFP4 matrix multiplication.
- Compiler: Fixed an issue where register values were corrupted or left in a stale state due to reconvergence failures in kernels with nested thread branching.

torch.linalg.svd, torch.linalg.svdvals, torch.linalg.lstsq: Fixed an issue where inputs with more than 8192 elements failed on MPS due to Metal pipeline state errors.
torch.svd(out=): Fixed an internal assertion failure when processing complex inputs on MPS.

pytorch