Hi team - A customer is getting a memory issue us...
# squads
n
Hi team - A customer is getting a memory issue using
neuralforecast
. The customer claims that it works ok on CPU but not on GPU (memory issue - see trace attached). Any ideas what might be going on here? Ref: https://app.intercom.com/a/inbox/j7y9c2ep/inbox/conversation/905
Copy code
Hello! I'm trying to work your library on a dataset of around 6MB and using CPU.
It works even on a 90MB file using CPU.

When I go to a Mac Mini M4 (using GPU) the following error occurs:
Invalid buffer size: 29.20 GB.

How can it be that the SAME notebook has completely a different order of size and can't be run on a GPU? Do you have any suggestions?
👀 1
o
No clue, avoid using Mac? 😉 But seriously, from the trace it is the normal point where the memory gets addressed when creating the windows, so it seems it's just a really large window. Can they share the data so that we can reproduce?
n
I have asked the to post on the repo instead of intercom. Let's see if they can provide it.
o
Thanks. From their initial post it could well be that the windows exceeds 30 GB, and their GPU RAM may be lower. As Mac M4 Mini has unified RAM, you can first suggest to monitor the ram use when using CPU, and make sure the addressable unified RAM for GPU is at least as big.
I think about 2/3 or 75% of RAM is by default available for GPU use in Mac unified RAM architectures. So it could be that they exceed this when running on GPU, but not when running on CPU
So maybe ask first to provide their maximum RAM use as percentage of their total RAM when using CPU when running this.
n
Thanks. I will respond accordingly.