To launch a training run across multiple machines, run app.py with the arguments start_ip corresponding to the IP of a known machine, and self_ip corresponding to the IP by which other machines should communicate with the current machine. learning_rate specifies learning rate and model_name specifies a HuggingFace Transformers-supported model. gradient_acc_steps specifies the minimum number of gradients each machine should calculate before sharing the gradients with the others and updating their weights. You can also use port to specify the current machine's port and start_port to specify a known machine's port.