In that case you have two options either to reduce the memory footprint of the model or use distributed deployment techniques