IntelPython / dpnp / 42707614666

84%

push

github

Derive `dpnp.linalg.LinAlgError` from `numpy.linalg.LinAlgError` (#3184)

This PR makes `dpnp.linalg.LinAlgError` a subclass of
`numpy.linalg.LinAlgError`.

Before, it derived only from `ValueError`, so code written for NumPy
that handles a linear algebra failure with `except
numpy.linalg.LinAlgError` (or `scipy.linalg.LinAlgError`, the same
class) did not catch the errors dpnp raises, for example for a singular
matrix in `dpnp.linalg.inv` or a matrix that is not positive definite in
`dpnp.linalg.cholesky`. Libraries that dispatch through the Array API
rely on that `except` clause, for instance to fall back to a
least-squares solution when a factorization fails. CuPy uses NumPy's
class directly.

The LAPACK extension now registers its `LinAlgError` with
`numpy.linalg.LinAlgError` as the base class, so it covers both the
errors thrown from C++ and those raised in Python. It is still a
`ValueError`, and `except dpnp.linalg.LinAlgError` works as before.

1792 of 2998 branches covered (59.77%)

Branch coverage included in aggregate %.

3 of 3 new or added lines in 1 file covered. (100.0%)

29617 of 34206 relevant lines covered (86.58%)

39164.7 hits per line