How much GPU memory are required for inference if it's 10M context?