launch¶
使用 python -m paddle.distributed.launch
方法启动分布式训练任务。
使用方法¶
python -m paddle.distributed.launch [-h] [--master MASTER] [--rank RANK]
[--log_level LOG_LEVEL] [--nnodes NNODES]
[--nproc_per_node NPROC_PER_NODE] [--log_dir LOG_DIR]
[--run_mode RUN_MODE] [--job_id JOB_ID] [--devices DEVICES]
[--host HOST] [--servers SERVERS] [--trainers TRAINERS]
[--trainer_num TRAINER_NUM] [--server_num SERVER_NUM]
[--gloo_port GLOO_PORT] [--with_gloo WITH_GLOO]
[--max_restart MAX_RESTART] [--elastic_level ELASTIC_LEVEL]
[--elastic_timeout ELASTIC_TIMEOUT]
training_script ...
基础参数¶
--master
: 主节点, 支持缺省 http:// 和 etcd://, 默认缺省 http://。例如--master=127.0.0.1:8080
. 默认值--master=None
.
--rank
: 节点序号, 可以通过主节点进行分配。默认值--rank=-1
.
--log_level
: 日志级别, 可选值为 CRITICAL/ERROR/WARNING/INFO/DEBUG/NOTSET, 不区分大小写. 0 号节点的日志默认不输出到标准输出,需要开启输出请使用 debug 模式。默认值--log_level=INFO
.
--nnodes
: 节点数量,支持区间设定以开启弹性模式,比如--nnodes=2:3
. 默认值--nnodes=1
.
--nproc_per_node
: 每个节点启动的进程数,在 GPU 训练中,应该小于等于系统的 GPU 数量。例如--nproc_per_node=8
--log_dir
: 日志输出目录。例如--log_dir=output_dir
。默认值--log_dir=log
。
--run_mode
: 启动任务的运行模式,可选有 collective/ps/ps-heter。例如--run_mode=ps
。默认值--run_mode=collective
。
--job_id
: 任务唯一标识,缺省将使用 default,会影响日志命名。例如--job_id=job1
. 默认值--job_id=default
.
--devices
: 节点上的加速卡设备,支持 gpu/xpu/npu/mlu。例如--devices=0,1,2,3
,这会启动 4 个进程,每个进程绑定到 1 个设备上。
training_script
: 需要运行的任务脚本,例如traing.py
。
training_script_args
:training_script
的输入参数,与普通起任务时输入的参数一样,例如--lr=0.1
。
Collective 参数¶
--ips
: [DEPRECATED] 需要运行分布式环境的节点 IP 地址,例如--ips=192.168.0.16,192.168.0.17
。 单机默认值是--ips=127.0.0.1
。
Parameter-Server 参数¶
--servers
: 多机分布式任务中,指定参数服务器服务节点的IP和端口,例如--servers="192.168.0.16:6170,192.168.0.17:6170"
。
--trainers
: 多机分布式任务中,指定参数服务器训练节点的IP和端口,也可只指定IP,例如--trainers="192.168.0.16:6171,192.168.0.16:6172,192.168.0.17:6171,192.168.0.17:6172"
。
--workers
: [DEPRECATED] 同 trainers。
--heter_workers
: 在异构集群中启动分布式任务,指定参数服务器异构训练节点的IP和端口,例如--heter_workers="192.168.0.16:6172,192.168.0.17:6172"
。
--trainer_num
: 指定参数服务器训练节点的个数。
--worker_num
: [DEPRECATED] 同 trainer_num。
--server_num
: 指定参数服务器服务节点的个数。
--heter_worker_num
: 在异构集群中启动单机模拟分布式任务, 指定参数服务器异构训练节点的个数。
--gloo_port
: 参数服务器模式中,用 Gloo 启动时设置的连接端口。同 http_port. Default--gloo_port=6767
.
--with_gloo
: 是否使用 gloo. 默认值--with_gloo=0
.
Elastic 参数¶
--max_restart
: 最大重启次数. 默认值--max_restart=3
.
--elastic_level
: 弹性级别设置,-1: 不开启, 0: 错误节点退出, 1: 节点内重启. 默认值--elastic_level=-1
.
--elastic_timeout
: 弹性超时时间,经过该时间达到最小节点数即开启训练。默认值--elastic_timeout=30
.
返回¶
None
代码示例零 (主节点, ip/port 自动识别)¶
# For training on multi node, run the following command in one of the nodes
python -m paddle.distributed.launch --nnodes 2 train.py
# Then the following info will be print
# Copy the following command to other nodes to run.
# --------------------------------------------------------------------------------
# python -m paddle.distributed.launch --master 10.0.0.1:38714 --nnodes 2 train.py
# --------------------------------------------------------------------------------
# Follow the instruction above and paste the command in other nodes can launch a multi nodes training job.
# There are two ways to launch a job with the same command for multi nodes training
# 1) using the following command in every nodes, make sure the ip is one of the training node and the port is available on that node
# python -m paddle.distributed.launch --master 10.0.0.1:38714 --nnodes 2 train.py
# 2) using the following command in every nodes with a independent etcd service
# python -m paddle.distributed.launch --master etcd://10.0.0.1:2379 --nnodes 2 train.py
# This functionality works will for both collective and ps mode and even with other arguments.
代码示例一 (collective, 单机)¶
# For training on single node using 4 gpus.
python -m paddle.distributed.launch --gpus=0,1,2,3 train.py --lr=0.01
代码示例二 (collective, 多机)¶
# The parameters of --gpus and --ips must be consistent in each node.
# For training on multiple nodes, e.g., 192.168.0.16, 192.168.0.17
# On 192.168.0.16:
python -m paddle.distributed.launch --gpus=0,1,2,3 --ips=192.168.0.16,192.168.0.17 train.py --lr=0.01
# On 192.168.0.17:
python -m paddle.distributed.launch --gpus=0,1,2,3 --ips=192.168.0.16,192.168.0.17 train.py --lr=0.01
代码示例三 (ps, cpu, 单机)¶
# To simulate distributed environment using single node, e.g., 2 servers and 4 workers.
python -m paddle.distributed.launch --server_num=2 --worker_num=4 train.py --lr=0.01
代码示例四 (ps, cpu, 多机)¶
# For training on multiple nodes, e.g., 192.168.0.16, 192.168.0.17 where each node with 1 server and 2 workers.
# On 192.168.0.16:
python -m paddle.distributed.launch --servers="192.168.0.16:6170,192.168.0.17:6170" --workers="192.168.0.16:6171,192.168.0.16:6172,192.168.0.17:6171,192.168.0.17:6172" train.py --lr=0.01
# On 192.168.0.17:
python -m paddle.distributed.launch --servers="192.168.0.16:6170,192.168.0.17:6170" --workers="192.168.0.16:6171,192.168.0.16:6172,192.168.0.17:6171,192.168.0.17:6172" train.py --lr=0.01
代码示例五 (ps, gpu, 单机)¶
# To simulate distributed environment using single node, e.g., 2 servers and 4 workers, each worker use single gpu.
export CUDA_VISIBLE_DEVICES=0,1,2,3
python -m paddle.distributed.launch --server_num=2 --worker_num=4 train.py --lr=0.01
代码示例六 (ps, gpu, 多机)¶
# For training on multiple nodes, e.g., 192.168.0.16, 192.168.0.17 where each node with 1 server and 2 workers.
# On 192.168.0.16:
export CUDA_VISIBLE_DEVICES=0,1
python -m paddle.distributed.launch --servers="192.168.0.16:6170,192.168.0.17:6170" --workers="192.168.0.16:6171,192.168.0.16:6172,192.168.0.17:6171,192.168.0.17:6172" train.py --lr=0.01
# On 192.168.0.17:
export CUDA_VISIBLE_DEVICES=0,1
python -m paddle.distributed.launch --servers="192.168.0.16:6170,192.168.0.17:6170" --workers="192.168.0.16:6171,192.168.0.16:6172,192.168.0.17:6171,192.168.0.17:6172" train.py --lr=0.01
代码示例七 (ps-heter, cpu + gpu, 单机)¶
# To simulate distributed environment using single node, e.g., 2 servers and 4 workers, two workers use gpu, two workers use cpu.
export CUDA_VISIBLE_DEVICES=0,1
python -m paddle.distributed.launch --server_num=2 --worker_num=2 --heter_worker_num=2 train.py --lr=0.01
代码示例八 (ps-heter, cpu + gpu, 多机)¶
# For training on multiple nodes, e.g., 192.168.0.16, 192.168.0.17 where each node with 1 server, 1 gpu worker, 1 cpu worker.
# On 192.168.0.16:
export CUDA_VISIBLE_DEVICES=0
python -m paddle.distributed.launch --servers="192.168.0.16:6170,192.168.0.17:6170" --workers="192.168.0.16:6171,192.168.0.17:6171" --heter_workers="192.168.0.16:6172,192.168.0.17:6172" train.py --lr=0.01
# On 192.168.0.17:
export CUDA_VISIBLE_DEVICES=0
python -m paddle.distributed.launch --servers="192.168.0.16:6170,192.168.0.17:6170" --workers="192.168.0.16:6171,192.168.0.17:6171" --heter_workers="192.168.0.16:6172,192.168.0.17:6172" train.py --lr=0.01
代码示例九 (elastic)¶
# With the following command, the job will begin to run immediately if 4 nodes are ready,
# or it will run after elastic_timeout if only 2 or 3 nodes ready
python -m paddle.distributed.launch --master etcd://10.0.0.1:2379 --nnodes 2:4 train.py
# once the number of nodes changes between 2:4 during training, the strategy holds