For this PR, I was thinking that the different between two 32 bit floats can only be resolve up to np.finfo(np.float32).resolution=1e-6 precisely, while 64 floats can resolve up to np.finfo(np.float64).resolution=1e-15. With the original tol=1e-10, float64 can resolve differences in that range, and thus continue to find a better solution, while float32 would not be able to resolve. (This is all a hypothesis, I can not reproduce the error on my local machine with the dependencies from the CI).

Looking at the docs for sparse.linalg.cg the tol is treated like a rtol, which means the "correct" tol would depend on the norm of y.

I’m going to see if this changes things on the CI for macPython by specifically pulling this branch.

This pass on travis ci on osx.

Here is the diff of the scikit-learn-wheels version that was used to run the travis ci instances.

Thanks @thomasjpfan