Releasing GPU memory after deleting model

1

I’m experiencing some trouble with the GPU memory not being released after deleting a model.

The pseudo-code looks something like this:

for _ in range(5):
    data = get_data()
    model = MyModule() ### PyTorch model
    results = model(data)
    del model
    torch.cuda.empty_cache()

The model occupies around 4 GB of GPU memory and in the second iteration this code crashes on my 8GB GPU with the error

RuntimeError: CUDA out of memory. Tried to allocate 3.73 GiB (GPU 0; 7.93 GiB total capacity; 3.73 GiB already allocated; 3.54

I’m probably misunderstanding something here, but I thought the del operation together with empty_cache() would free up the memory.

Thanks for any input. :slight_smile:

1 Like
2

Hi,

The thing is that results most likely requires gradients. And these gradients would require most of the model parameters to be computed. And so the model cannot be freed yet :slight_smile:

1 Like
3

:man_facepalming: That was indeed it.
Thanks for your prompt reply! :slight_smile: