[NDTensorsCUDAExt] Fix QR-based SVD for some rectangular matrices

[kmp5VT]

Description

CUDA's QRAlgorithm fails when an (N x M) matrix has N < M. Here I fix this issue by catching the problem in the SVD dispatch and transposing the matrices and the resulting SVD.

TODO

• Fix the QR SVD algorithm
• Make sure all the unit tests still pass

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (66f5d39) 85.31% compared to head (c89e33a) 54.54%.
Report is 7 commits behind head on main.

❗ Current head c89e33a differs from pull request most recent head 4e3b304. Consider uploading reports for the commit 4e3b304 to get more accurate results

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1229       +/-   ##
===========================================
- Coverage   85.31%   54.54%   -30.78%     
===========================================
  Files          89       88        -1     
  Lines        8445     8392       -53     
===========================================
- Hits         7205     4577     -2628     
- Misses       1240     3815     +2575     

see 37 files with indirect coverage changes

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mtfishman]

I see, it looks like other code is sending alg keyword arguments into this CUDA SVD function which only make sense for the CPU SVD. Maybe it is an issue with picking good defaults for alg which are aware of the backend?

[mtfishman]

Maybe need to overload default_svd_alg for CuArrays.

[mtfishman]

Looks good, thanks!