Workbench instance GPU availability

Google Cloud Build with AI Custom ML & MLOps
1

Hi,

I’m new to Vertex AI and trying to get started with a simple notebook + GPU to test out some basic functionality. Figured I’d spin up a Colab Enterprise Runtime or a Workbench Instance. However, whatever zone I try to spin either up in, I get this error:

instance-20240516-130652: The zone ‘projects/<my_project>/zones/europe-west1-b’ does not have enough resources available to fulfill the request. Try a different zone, or try again later.: Something went wrong. Sorry about that.

From what I understand, this has to do with the availability of GPUs in each zone. Going over each zone and trying to spin up an instance with my fingers crossed is not a workable/scalable solution.

How do others solve this?

Thanks!

2

Hi @DavidKroon ,

I understand it can be frustrating to navigate a new API. Let me help you find a solution to your GPU availability issue. Here are some steps that might help.

  • You can explore this document to see which regions have GPUs available
  • If europe-west1-b isn’t working, try spinning up your Workbench instance in another zone within Europe
  • Make sure your Workbench instance allows GPU usage and has enough GPU resources allocated
  • If your budget allows, you could change your machine type to one with more GPUs for your Workbench instance

I hope these steps help you resolve the issue.

3

Hi Isolatorio,

Thank you for your reply.

  • My workbench instance does allow GPU usage and has enough GPU resources allocated.

  • Regarding attaching more GPUs: if I can’t attach 1 GPU, why would I be able to attach more than 1? I could try to attach a different type of GPU, but I’d rather not pay a few x more for a larger machine that I don’t need.

  • My problem is not so much that I don’t know which areas offer GPUs, but rather that I can’t find a scalable/practical way of finding out whether the GPUs are actually available for use. As said in my original question, I can’t go over each zone that offers GPUs, try to spin one up, wait a couple of minutes, and simply hope that one is indeed available. For your reference, I tried 5 different zones across the globe, all of which failed. That is in my opinion not a scalable or practical approach every time I need to use a notebook with a GPU.

So my question is: how can I find out which zone has GPUs available (so not just offers GPUs but actually has GPUs available)?

Thanks,

David

P.S. New APIs are only frustrating rather than exciting when they don’t work as advertised.

4

I’m experiencing the same problem. I’m wondering why it’s so complicated to instantiate a service.