My WSL2 instance disappears within seconds of launching a GPU-based LLM program in python

Question

After over a year of problem free GPU training in WSL2, now when I launch my GPU-based python scripts, the WSL instance dies within a few seconds, and seems to leave no events or logs behind as it dies...

I have reinstalled a brand new Ubuntu instance with only the minimal installs required to faciliate my GPU-based LLM use case, and this instance dies in exactly the same way as my original instance did.

Unless I can resolve this, I will need to switch to running Linux bare metal rather than Windows/WSL2... I contacted MSFT support, demonstrated the issue, and was reommended to ask a question here as they were unable to help.

I am running a slightly modified version of FastChat (https://github.com/lm-sys/FastChat)"github.com") using a fine-tuned model based on Facebooks LLama2 7B state model, and simply running the model with test prompts. This causes the WSL2 instance to die dependably every time.

Answer 1

Hello there,

Given your wsl -l -v output, it looks like the docker-desktop-data instance got set as the default when you uninstalled Ubuntu. That's not actually a bootable instance, since it has no /init in it.

That hopefully will explain why the wsl command is exiting. Just running wsl.exe (from PowerShell, CMD, or the Start Menu) is attempting to launch docker-desktop-data and then immediately exits.

You should be able to fix it by executing wsl --set-default Ubuntu from PowerShell or CMD. If not, we can dig deeper.

Hope this resolves your Query !!

--If the reply is helpful, please Upvote and Accept it as an answer--