@pytorchbot retest this please

This will lose backward hooks, you'll need to serialize (and reconstitute them as well).

See also #10220, where we have a parallel code path specifically for the multiprocessing serialization case.

@elanmart Could you update the diff to include backward hooks like here? https://github.com/pytorch/pytorch/pull/10220/files#diff-0d553ed0e9ca41fdac21632b4d503a10R41

@ailzhang please excuse my late response, I was having a few days off. Yes, of course I'll do my best to fix that.

No rush. Just wondering if there is any progress on this?

@ssnl well yeah there's really no excuse, I lost track of this PR, which kept people waiting.

I'm not sure how the multiprocessing codepath should affect the solution here. Is sticking an extra classmethod to the nn.Parameter an acceptable solution?

@elanmart Adding a classmethod is fine by me, but it seems to break other things (see CI).

@ssnl the CI seems to run fine now, what do you think?

Hi @elanmart
Serialization of my model with dill works with pytorch 0.4.0, but stopped working in 0.4.1.
I just tried to serialize it with torch-nightly 1.0.0.dev20180930, but unsuccessfully so.
I hit the default recursion limit, and increasing it results in a segfault. Sounds like infinite recursion to me. Any ideas?
Thanks.

I just had a look at the initial issue, #9981.
The error I'm seeing on torch-nightly 1.0.0.dev20180930 is identical.
Either your patch is not in there, or it doesn't have the desired effect.

Argh, never mind. I was still on torch 0.4.1 because a dependency pulled it in.
It is really unfortunate that the nightly build of torch is not called torch, but torch-nightly...
that is a headache for package management.