I don't think it's plausible that we would add an openmp dependency to the project. But if the algorithm is "embarrassingly parallel" using joblib, that's another story.

actually, I think it's very plausible. It's just a question of when and how ;)

Note that #6448 not only provides parallelization of multi-class linear regression, but also manages the memory better than a naive parallelization would, since liblinear creates unnecessary copies, which become highly problematic on large datasets.

I've been using the patch (6448) for a long time now, and it's robust.

@mannby I think there is issues if people try to also use joblib because it currently doesn't support nested parallelism. I'm not sure what the right issue is to track that. maybe joblib/joblib#256 or #3754? @ogrisel and @GaelVaroquaux know more ;)

@alexhenrie In case you might be interested in this. We can now use OpenMP in the code base (although there are still some open questions about how to control the number of threads). A PR would be welcome.

Their patch is very large though, I was hoping that could be reduced a bit.

also cc @jeremiedbb

Hi, would like to reconsider that one.
I see that sklearn has its own liblinear fork. Is it following upstream changes or not?
Because there seems to be a not-so-new multicore liblinear https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/multicore-liblinear/
So n_jobs could simply send the parameter to that one, instead of being used only in the multiclass case. Is that right?

Context: we were having memory problems on a large dataset (99M samples) when using LogisticRegression (with any solver, liblinear or saga). So we used sklearn's dump_svmlight_file and then multicore liblinear, and that worked fine.

Hi, is there still no way to parallelize liblinear?