RuntimeError: CUDA error: out of memory

[Arong-gong]

I run the run_classifier_TABSA.py by two 1080ti GPU.
However, I get the RuntimeError every time.
In the original Pytorch code of BERT, as we know, 12G GPU memory is enough.
for this situation,
How much GPU memory should I use? Or what should I do?

[HSLCY]

Hello, since your batch size is 24, four 1080ti GPUs are needed.
The following picture is our running situation:

For less GPU memory, you can try smaller batch size.
By the way, one of the GPUs will occupy more memory than others.

[Arong-gong]

Thanks for your response.