trunk/13948faea12f1eb3a6df72f181a51f3e65dbf309: [MPS] Orthonormalize null-space U columns in svd_jacobi (#196139)
- PyTorch: 590 events in the last 90 days
- PyTorch: 584th Release in the last 90 days
- Previous: earlier the same day · trunk/44e77b112c9c76adc06256a90870a5f607575a1c: [dynamo] Type runtime predicate inputs as object (#196175)
What happened
The native Jacobi SVD forms U as (A V)_j / sigma_j, which collapses to zero for sigma_j at the noise floor, so rank-deficient inputs got a non-orthonormal U (U^H U != I) even though A = U S V^H still held. Replace those null columns with an orthonormal completion of the emitted columns (simd-group-parallel Gram-Schmidt of canonical vectors, like gesdd). Full-rank inputs are unaffected since the completion loop has z…
Summary assembled by rule from the sources below