GPU RAM requirements for the datacentric challenge?

Hi,

I was wondering whether the GPU RAM requirement of 8500 MBs is accurate, we tried to reproduce the fixed model, but found that a 12 GB GPU is not enough.

Thank you!

Hi aledelmo_clario,

That's interesting because the model was benchmarked on an RTX3060 with 12GB of VRAM. To meet the requirement of 8500MB of VRAM for batch_size=2, you need to use mixed precision. This can be achieved by using the configuration flag trainer: precision. For newer NVIDIA GPUs, this could be bf16-mixed, and for older ones, it needs to be 16. An exemplary configuration file can be found here. If the issue persists, please open an issue on GitHub, providing more details such as the command line output. Hope this helps.

Best, Jakob

Hi Jakob,

indeed, the issue happened with an older gpu, it works fine when using the correct 16 value.

Thanks for your help!