Does SGDRegressor work with accept_large_sparse=True? If so we should just make it the default, there is no need to add an estimator parameter for this in general.

Please add a test with a 64bit sparse matrix. I don't think it's this simple, but I'd happily be proven wrong

I think this PR partially addresses #11355

@jnothman @cmarmo @rth sorry for going AWOL on this -- I can pick this up during this week! is this still something we'd want to work on?

I do not think SGDRegressor can support large sparse matrices out of the box. Currently this PR, lets accept_large_sparse through, but once it gets into the Cython code here:

The dtype of x_ind_ptr is fixed to int, which is usually 32 bits. Also SequentialDataset, would also need to be updated to template out the pointer which also uses int* for the x_ind_ptr: https://github.com/scikit-learn/scikit-learn/blob/main/sklearn/utils/_seq_dataset.pxd.tp

TLDR: I think this is a hard task that involves going through the Cython files and making sure that int64 indices work with templating or fused types.

@thomasjpfan makes sense -- I guess it would be one thing if we'd just need to make x_ind_ptr a 64bit, but looks like there's a lot more to do.

admittedly I don't have a ton of time to work on this so happy to keep slowly playing with this PR for a bit ... but I don't think it's something I have a lot of time to dedicate to closing out, so feel free to close this

makes sense -- I guess it would be one thing if we'd just need to make x_ind_ptr a 64bit, but looks like there's a lot more to do.

Most of the complication comes from needing to support 32bit and 64bit indices.

but I don't think it's something I have a lot of time to dedicate to closing out, so feel free to close this

I'm okay with closing. Thank you for working on this issue.