PaperSpace out of memory or killed by signal

Part 1 (2020)
1

Im writing this post just in case other find either of these two errors while training models:
[Errno 12] Cannot allocate memory

DataLoader worker (pid 63) is killed by signal: Killed.

For me the solution was getting on to a machine that has a GPU.

Hope that was of any help!

2 Likes
2

Hi @Salazar. I am getting the same error. I don’t quite follow your suggested solution . DO you mean I need to use a non_GPU instance? How can model training work without a GPU?

3

Ahh sorry yes thats confusing - I found that the reason why I was getting that error was because I was on a machine with no GPU. So once I switched to a machine that has a gpu it worked.

1 Like
4

Thanks, that solved the problem for me too.

Mysteriously I had set up Paperspace with a GPU and successfully trained a model before, but it then somehow reverted to a CPU-only instance.

5

one magic line which you should have at the top of every notebook:
assert torch.cuda.is_available(), "GPU not available"

2 Likes