Using refine_regularization=0. breaks jit compile in EqualityConstrainedQP about jaxopt HOT 2 CLOSED

Nice catch !

The issue is that the jax.scipy.sparse.linalg API uses A, b, x0 signature whereas the jaxopt.linear_solve API uses matvec, b, ridge, init so we cannot write code that supports both. We have to chose between the canonical Jaxopt way (with init), but unfortunately it is not exposed publicly... or the jax.scipy.sparse.linalg that the user knows about, but is incompatible... Upsettingly, neither keyword nor positional arguments can do the trick.

Hopefully in the next iterations of the library we will have a well defined API.

from jaxopt.

One possibility could be to add a * between positional and kwargs for the jaxopt solvers? This would at least remove error ambiguity.

from jaxopt.