| 117 | |
| 118 | |
| 119 | def run(rank, world_size): |
| 120 | os.environ['MASTER_ADDR'] = 'localhost' |
| 121 | os.environ['MASTER_PORT'] = '29500' |
| 122 | |
| 123 | # Initialize the process group first |
| 124 | dist.init_process_group( |
| 125 | backend="gloo", |
| 126 | rank=rank, |
| 127 | world_size=world_size |
| 128 | ) |
| 129 | |
| 130 | options=rpc.TensorPipeRpcBackendOptions( |
| 131 | num_worker_threads=16, |
| 132 | rpc_timeout=60 |
| 133 | ) |
| 134 | if rank != 0: |
| 135 | rpc.init_rpc( |
| 136 | f"trainer{rank}", |
| 137 | rank=rank, |
| 138 | world_size=world_size, |
| 139 | rpc_backend_options=options |
| 140 | ) |
| 141 | # trainer passively waiting for ps to kick off training iterations |
| 142 | else: |
| 143 | rpc.init_rpc( |
| 144 | "ps", |
| 145 | rank=rank, |
| 146 | world_size=world_size, |
| 147 | rpc_backend_options=options |
| 148 | ) |
| 149 | run_ps([f"trainer{r}" for r in range(1, world_size)]) |
| 150 | |
| 151 | # block until all rpcs finish |
| 152 | rpc.shutdown() |
| 153 | dist.destroy_process_group() |
| 154 | |
| 155 | |
| 156 | if __name__=="__main__": |