Cuda out of memory with plenty left

Hey, I’ve created a detailed issue around the topic on git, moving it to here.

Could be an issue with torchrl

If that comes from torchrl: can you tell us more about your replay buffer, specifically how you build it?
What other components of torchrl / tensordict are you using?
Does it replicate if you upgrade to torchrl 0.3.0?

I’m using a custom implemented PER GitHub - rlcode/per: Prioritized Experience Replay (PER) implementation in PyTorch

I’ve made sure to delete the tensors from memory in the implementation.

Alright, not sure what you use from tensordict/torchrl in the code then. If you think the problem can come from there feel free to ping me

1 Like