GPU Memory Allocation Issue with POT==0.9.1

atong01 / conditional-flow-matching

TorchCFM: a Conditional Flow Matching library

https://arxiv.org/abs/2302.00482

MIT License

1.25k stars 101 forks source link

GPU Memory Allocation Issue with POT==0.9.1 #57

Open atong01 opened 1 year ago

atong01 commented 1 year ago

With POT 0.9.1 GPU memory is allocated on import in combination with deep learning packages see https://github.com/PythonOT/POT/issues/523 https://github.com/PythonOT/POT/issues/516

Recommend downgrading to POT < 0.9.1 for now.

RoyJames commented 1 year ago

I just verified that this is fixed by https://github.com/PythonOT/POT/pull/520 when I install it from their master. Previously nearly all GPU memory would be all pre-allocated and now I can see the true usage.

kilianFatras commented 1 year ago

Oh that’s interesting! Thank you for your interest and checking this issue with us. We have been busy with the 1.0.4 release and we have not checked it for a while. @atong01 would you mind checking as well on the code that was broken?