We currently have to hardcoded those instance types in the gcp_catalog.py file. And for other older versions of GPUs, we will attach to n1 instances based on the CPUs/memory requested by the user.