Benjamin Bossan

You could make the `dtype` an argument to the `nn.Module` and cast the whole module to that dtype. Then you can grid search over the dtype. Here is an example:...

> I agree this is a workaround. What I imagined as a better solution is that **init**() of NeuralNet can handle this if a corresponding parameter is assigned instead of...

Thanks for the report and reproducer. When I ran your snippet, I did not encounter any error. What version of skorch and torch are you using? I'm using the latest...

Thanks for the info. This issue should be fixed on the master branch of skorch via #1092. Could you please install skorch from source and check if the error goes...

I didn't have time yet for a deep dive into the FSDP issue you reported in #2833, hopefully I'll have time tomorrow. Once that is fixed, I plan to take...

@ojh31 Given the upcoming transformers v5 release and the work associated with that, I didn't have time week to investigate your issue (which can potentially take quite some time because...

I finally had time to look at this issue. I couldn't quite get your script to run, but I took the existing `sft/train.py` script and dropped this into the `main`...

Thanks for confirming. I agree it would be nice if it worked with FSDP but would think that in the grand scheme of things, it's not a significant factor. LMK...

I have no strong (pun not intended) opinion on this matter. As is, the existing type annotations are more type hints for convenience but not really checked for correctness, which...

> Personally, I don't really like working with such libraries that get mixed up with all your types. If existing type annotations are incorrect, that's certainly not on purpose. We...