using python3+gpu runtime only first concurrent notebook gets python3, others get python2 about colabtools HOT 4 CLOSED

First off, thanks for the outstanding bug report. 👍

I can't seem to repro this issue -- I've got two notebooks open right now that are both set to py3+gpu, and forcing a restart of my backend (*) still gets me py3+gpu with both:

A few questions to try and dig a little more:

• is it possible to share two notebooks that repro the issue with [email protected]?
• just to confirm, how are you opening the second notebook? (eg from recent, directly navigating, something else completely?)

from colabtools.

So I tried reproducing it on a fresh VM with two notebooks starting from scratch and everything worked fine at first just like for you. Then I added some code I used in my full notebook to fetch dependencies and any copies I made after running that would get python version 2 again.

Basically the responsible code grabs a Pipfile and companion Pipfile.lock from https://github.com/lefant/dockerfiles-lefant/blob/master/python3-keras/ and then uses !pip3 install pipenv && pipenv install --three --deploy --system to install all the packages. I noticed now that these packages include ipython and jupyter, so that might be very relevant (originally I use that Pipfile to build a docker container that has its own jupyter server). I will try with only those / without those packages next, but first I have to wait until my VM goes away... ;)

notebook links (shared with [email protected] and available for public viewing):
https://colab.research.google.com/notebook#fileId=1IsB6qOqB7N18DwNYXEGzfkozKMTlRTDH
https://colab.research.google.com/notebook#fileId=1fzxvf5LP3RELfkE8cfxrRRbL2gliwCpG

from colabtools.

Once I stop installing jupyter, ipython and tensorflow from Pipfile everythings works fine on a fresh VM. I am not sure how exactly installing a custom version of jupyter and ipython messes things up, but then I do not have a use case for it, so I am closing this issue. Sorry for the noise.

from colabtools.

Thanks for digging through -- I'm betting what happened is that some step of the jupyter install overwrote the path for the python3 kernelspec.

from colabtools.