# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from __future__ import print_function import subprocess import os import sys import time import argparse default_envs = { "PADDLE_TRAINER_ENDPOINTS": "127.0.0.1:6170,127.0.0.1:6171,127.0.0.1:6172,127.0.0.1:6173,127.0.0.1:6174,127.0.0.1:6175,127.0.0.1:6176,127.0.0.1:6177", "LD_LIBRARY_PATH": os.getenv("LD_LIBRARY_PATH", ""), "PATH": os.getenv("PATH"), "LD_PRELOAD": os.getenv("LD_PRELOAD", ""), "PADDLE_TRAINERS_NUM": "8", "NCCL_DEBUG": "INFO", "GLOG_v": "0", "NCCL_SOCKET_IFNAME": "eth0", "NCCL_IB_GID_INDEX": "3", "NCCL_IB_RETRY_CNT": "0", "PYTHONPATH": os.getenv("PYTHONPATH", ""), } GPUS = 8 def start_procs(gpus, entrypoint, entrypoint_args, log_dir): procs = [] log_fns = [] os.system("mkdir -p %s" % log_dir) # ======== update parent envs ======= for k, v in os.environ.items(): if k.startswith("FLAGS_") or k.startswith("NCCL_") or \ k.startswith("GLOG_"): default_envs[k] = v # ======== for dist training ======= node_trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) current_ip = os.getenv("POD_IP", "127.0.0.1") trainer_ips = os.getenv("PADDLE_TRAINERS", current_ip).split(",") num_nodes = len(trainer_ips) all_nodes_devices_endpoints = "" for n in trainer_ips: for i in range(gpus): if all_nodes_devices_endpoints: all_nodes_devices_endpoints += "," all_nodes_devices_endpoints += "%s:617%d" % (n, i) nranks = num_nodes * gpus # ======== for dist training ======= for i in range(gpus): curr_env = {} curr_env.update(default_envs) curr_env.update({ "FLAGS_selected_gpus": "%d" % i, "PADDLE_TRAINER_ID": "%d" % (node_trainer_id * gpus + i), "PADDLE_CURRENT_ENDPOINT": "%s:617%d" % (current_ip, i), # nranks "PADDLE_TRAINERS_NUM": "%d" % nranks, "PADDLE_TRAINER_ENDPOINTS": all_nodes_devices_endpoints }) print("starting process ", i, entrypoint, entrypoint_args, curr_env) fn = open("%s/workerlog.%d" % (log_dir, i), "w") log_fns.append(fn) cmd = [sys.executable, "-u", entrypoint] + entrypoint_args procs.append(subprocess.Popen(cmd, stdout=fn, stderr=fn, env=curr_env)) for i in range(gpus): try: procs[i].communicate() procs[i].terminate() log_fns[i].close() except: pass def parse_args(): parser = argparse.ArgumentParser( description='''start paddle training using multi-process mode. NOTE: your train program ***must*** run as distributed nccl2 mode, see: http://www.paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html#permalink-8--nccl2- And your train program must read environment variables below in order to let different process init properly: FLAGS_selected_gpus PADDLE_TRAINER_ID PADDLE_CURRENT_ENDPOINT PADDLE_TRAINERS_NUM PADDLE_TRAINER_ENDPOINTS POD_IP (current node ip address, not needed for local training) ''') parser.add_argument( '--gpus', type=int, default=8, help='start number of processes for every gpu') parser.add_argument( '--log_dir', type=str, default="mylog", help='directory to put logs per process.') parser.add_argument( 'entrypoint_script', type=str, help="The entrypoint script to be launched in parallel," "followed by all the arguments for each process," "e.g. train.py --lr 0.1") parser.add_argument('entrypoint_args', nargs=argparse.REMAINDER) return parser.parse_args() def main(): args = parse_args() # launch multiple training process start_procs(args.gpus, args.entrypoint_script, args.entrypoint_args, args.log_dir) if __name__ == "__main__": main()