MCMC with parallel chains get stuck in jupyter notebook on Ubuntu

[neerajprad]

As reported by @activatedgeek. MCMC with multiprocessing seems to work fine on mac, and even Ubuntu when run from the terminal, but gets stuck in Ubuntu when run in the notebook environment. I am opening this issue so that we can investigate further if this is a common issue faced by other users.

Jupyter version:

jupyter-client==5.2.4
jupyter-console==6.0.0
jupyter-core==4.4.0
jupyterlab==1.2.5
jupyterlab-server==1.0.6

[fehiepsi]

I believe this is related to these issues: pytorch/pytorch#17680 and pytorch/pytorch#20375.

Not related but I put here some observations in a terminal for reference:

• Without torch.multiprocessing.set_start_method("spawn") (or forkserver) line, baseball example will fail with the default fork method due to this issue.
• Under spawn method, the default sharing strategy file_descriptor in Linux will lead to bad value(s) in fds_in_keep issue. This does not happen with forkserver.

[neerajprad]

I think #20375 is the reason why this fails - the default on Ubuntu and python 3.8 across all platforms is "spawn" context which causes issues. I can replicate it in mac too if I set mp_context="spawn". I suppose one solution would be to change the default depending on the notebook environment.

@activatedgeek - In the meantime, to get around this issue, you could try to use mp_context='forkserver' when running in the notebook. Let us know if that works, or still causes issues.

[fehiepsi]

It might be better to add a note/warning to MCMC class so that users know about these gross of PyTorch multiprocessing. WDYT? I worry about forkserver might not be available on some platform.

[activatedgeek]

Thanks for the info @fehiepsi

@neerajprad The issue seems to still exist with mp_context='forkserver'.

[mennowitteveen]

Yes, thanks @fehiepsi for the info and all your great help in general.

I also bumped into this issue (on CentOS 7.9.2009), and I actually do not need the chains to run in parallel.

I need multiple chains for investigating convergence (In this case computing r_hat statistics).

[AndrewFalkowski]

Wondering if a fix for this ever emerged? num_chains>1 still fails when run in a jupyter notebook.

[fehiepsi]

Hi @AndrewFalkowski I think the issue has not been fixed in Pyro. I'm not sure what is the current status of the related PyTorch issues mentioned above.

[AndrewFalkowski]

Ah, bummer, I typically like to build models in jupyter notebooks before migrating to a proper script, guess I'll have to rework my workflow a little. Thanks!