Training a model from scratch. If i want to use 8 GPUs to train a model(a server),how to configure these flags's value? flags: num_clones worker_replicas num_ps_tasks