电脑基础 · 2023年4月3日

Pytorch - 弹性训练原理

Pytorch在1.9.0引入了torchrun,用其替代1.9.0以前版本的torch.distributed.launch。torchrun在torch.distributed.launch 功能的基础上主要新增了两个功能:

  • Failover: 当worker训练失败时,会自动重新启动所有worker继续进行训练;

  • Elastic: 可以动态增加或或删除node节点;

弹性训练代码同DDP代码编写的思路基本一致,只要在DDP代码上增加以下两点即可:

  • checkpoint处理:由于再每次增加或删除node时,会将所有worker kill掉,然后再重新启动所有worker进行训练。因此,在训练代码中要对训练的状态进行保存,以保证重启后能接着上次的状态继续训练。

  • 超参调解:由于node节点数的变化,会导致global batch size的变化,因此我们的learning rate一般也要做相应的调整,保证训练出的模质量不受影响。

代码见第二节 最下面

当编写完弹性训练代码后,我们可以使用torchrun来启动弹性训练任务:

  • --nnodes=1:3 :表示当前训练任务接受最少1个node,最多3个node参与分布式训练;

  • --nproc_per_node=4:表示每个node上节点有4个process

  • --max_restarts=3: worker group最大的重启次数;这里需要注意的是,node fail、node scale down和node scale up都会导致restart;

  • --rdzv_id=1:一个unique的job id,所有node均使用同一个job id;

  • --rdzv_backend: rendezvous的backend实现,默认支持c10d和etcd两种;rendezvous用于多个node之间的通信和协调;

  • --rdzv_endpoint:rendezvous的地址,应该为一个node的host ip和port;

torchrun \
    --nnodes=1:3\
    --nproc_per_node=4\
    --max_restarts=3\
    --rdzv_id=1\
    --rdzv_backend=c10d\
    --rdzv_endpoint="192.0.0.1:1234"\
    train_elastic.py

3 整体架构

Pytorch - 弹性训练原理

弹性调度的架构如上图所示,其中最关键角色为elastic agent。在每个Node上面都有一个elastic agent进程,其负责管理当前Node上面的所有workers。

当我们调用torchrun 命令启动弹性训练任务后:

  • 首先,elastic agent会触发rendezvous 流程; rendezvous的功能是在所有elastic agent间做协调和同步,该接口会一直阻塞直到至少min个elastic agent加入进来后返回;

  • 然后,elastic agent会启动当前Node的所有workers

  • 最后,elastic agent会监控当前Node上所有workers的运行状态,并根据workers的状态进行相应的处理(例如restart worker)

4 Elastic Agent

本小结,我们详细分析下Elastic Agent的实现。Elastic Agent在Pytorch代码中由以下对象构成:

  • Elastic Agent是抽象基类

  • SimpleElasticAgent提供了更完整的Agent接口,并且实现了部分接口

  • LocalElasticAgent则是实现剩余的接口

  • Pytorch - 弹性训练原理

Elastic Agent在代码中的调用逻辑如下:

  • torch.distributed.launcher.api:launch_agent() 弹性训练逻辑的入口;

    • 首先、会构建一个RendezvousParameters来描述Rendezvous调用时所需要的参数,例如min_nodes/max_nodes/endpoint等;

    • 然后、构建WorkerSpec描述当前Node上启动Wokers的信息, 例如max_restart/entrypoint等;

    • 再然后,构建LocalElasticAgent对象;

    • 最后,调用LocalElasticAgent的run接口启动当前node的workers进行弹性训练;

  • Elastic run接口主要由两个部分逻辑组成:

    • 若process group的状态为succeeded:调用_exit_barrier接口等待所有node上agent相应并退出

    • 若process group的状态为unhealthyfailed: 如果重试次数小于_remaining_restart则restart所有worker进程,否则stop所有worker,并退出;

    • 若process group的状态为healthy: 则判断当前是否有node等待加入,如果有则restart_worker;(注:restart worker的实现逻辑是先stop 所有worker,然后在调用_initialize_workers)

    • SimpleElasticAgent._initialize_workers:先调用_rendezvous等待至少min 个node加入,然后调用_start_workers接口在当前node上启动worker process

    • while loop monitor worker:while循环,监控上一步启动process的状态

    • Pytorch - 弹性训练原理

5 Rendezvous

5.1 基本概念

Pytorch中Rendezvous的实现涉及到很多概念,我们这里先把这些概念一一介绍下,然后再介绍Rendezvous的实现这样会清晰很多。

首先是_RendezvousState,每个ElasticAgent上都会存储一份_RendezvousState,并会在必要时进行彼此间的同步,_RendezvousState存储的内容如下:

  • round: The current round of the rendezvous.

  • complete: A boolean value indicating whether the current round of the rendezvous is complete.

  • deadline: The time at which the current round of the rendezvous will be considered complete if it is still waiting for nodes to join.

  • closed: A boolean value indicating whether the rendezvous is closed.

  • participants: A dictionary of the participants and their corresponding ranks.

  • wait_list:A set of nodes that are waiting to participate in the next round of the rendezvous.

  • last_heartbeats: A dictionary containing each node's last heartbeat time.

那_RendezvousState是如何在所有ElasticAgent间进行同步的呢,Pytorch中又提出了Store的概念,在Pytorch中有TCPStoreFileStoreHashStore三种类型,在弹性训练场景,默认使用TCPStore。

TCPStore的典型用法如下:

  • 其是一个典型的server-client架构,我们在process1上启动server,在proess2上启动client,通过TCPStore的set和get接口可以进行数据的设置和获取

  • 在Rendezvous实现中即是通过TCPStore来对_RendezvousState进行设置和获取的。

import torch.distributed as dist
from datetime import timedelta
# Run on process 1 (server)
server_store = dist.TCPStore("127.0.0.1", 1234, 2, True, timedelta(seconds=30))
# Run on process 2 (client)
client_store = dist.TCPStore("127.0.0.1", 1234, 2, False)
# Use any of the store methods from either the client or server after initialization
server_store.set("first_key", "first_value")
client_store.get("first_key")

Pytorch的Rendezvous实现中,通过C10dRendezvousBackend对TCPStore进行了封装,并提供了set_stateget_state接口,方便state的操作。(注:Pytorch中还提供了EtcdRendezvousBackend,该类型的RendezvousBackend通过Etcd来进行_RendezvousState的同步)。

C10dRendezvousBackend的主要实现如下,可以很清晰的看到get_state和set_state的实现,均是对store接口的调用.

class C10dRendezvousBackend(RendezvousBackend):
    def get_state(self) -> Optional[Tuple[bytes, Token]]:
        """See base class."""
        base64_state: bytes = self._call_store("get", self._key)
        return self._decode_state(base64_state)
    def set_state(
        self, state: bytes, token: Optional[Token] = None
    ) -> Optional[Tuple[bytes, Token, bool]]:
        """See base class."""
        base64_state_str: str = b64encode(state).decode()
        if token:
            # Shortcut if we know for sure that the token is not valid.
            if not isinstance(token, bytes):
                result = self.get_state()
                if result is not None:
                    tmp = *result, False
                    # Python 3.6 does not support tuple unpacking in return
                    # statements.
                    return tmp
                return None
            token = token.decode()
        else:
            token = self._NULL_SENTINEL
        base64_state: bytes = self._call_store("compare_set", self._key, token, base64_state_str)
        state_token_pair = self._decode_state(base64_state)
        if state_token_pair is None:
            return None
        new_state, new_token = state_token_pair
        # C10d Store's compare_set method does not offer an easy way to find out
        # whether our write attempt was successful. As a brute-force solution we
        # perform a bitwise comparison of our local state and the remote state.
        return new_state, new_token, new_state == state
    
    def _call_store(self, store_op: str, *args, **kwargs) -> Any:
        try:
            return getattr(self._store, store_op)(*args, **kwargs)
        except (ValueError, RuntimeError, TimeoutError) as exc:
            raise RendezvousConnectionError(
                "The connection to the C10d store has failed. See inner exception for details."
            ) from exc    

在RendezvousBackend的基础上,Pytorch提出了一个更偏向业务层面的概念**_RendezvousStateHolder**,其提供了_RendezvousState进行获取、同步、标记更新的接口,这些接口的实现均是调用RendezvousBackend的set_state和get_state完成的。

_RendezvousStateHolder的定义如下:

class _RendezvousStateHolder(ABC):
    """Holds the shared rendezvous state synced with other nodes."""
    def state(self) -> _RendezvousState:
        """Gets the local state."""
    def sync(self) -> Optional[bool]:
        """Reads or writes the latest state.
        Returns:
            A boolean value indicating whether the local state, in case marked
            as dirty, was successfully synced with other nodes.
        """
    def mark_dirty(self) -> None:
        """Marks the local state as dirty."""

Rendezvous的基础设置都准备好了,状态在 _RendezvousState中保存,状态的同步通过 _RendezvousStateHolder来完成,此时还差一项,就是Rendezvous state的是如何变更的。这个变更通过 _RendezvousXXXOp和 _RendezvousOpExecutor共同来完成。

Pytorch首先提供了_RendezvousExitOp/_RendezvousJoinOp/_RendezvousCloseOp/_RendezvousKeepAliveOp来对应ElasticAgent的退出、加入、Rendezvous关闭和心跳保保持四个操作。这些OP的实现逻辑是根据OP的类型和当前_RendezvousState的内容来决定来返回一个action,_RendezvousOpExecutor则执行对应的action。

例如_RendezvousExitOp 对应ElasticAgent的退出操作

  • 如果当前节点仍旧在participants列表中,则返回一个REMOVE_FROM_PARTICIPANTS,_RendezvousOpExecutor在接收到这个action后会执行_remove_from_participants逻辑;

  • 如果当前节点没有在participants列表中,返回FINISH,这个状态_RendezvousOpExecutor不会做任何操作;

class _RendezvousExitOp:
    """Represents a rendezvous exit operation."""
    def __call__(self, ctx: _RendezvousContext, deadline: float) -> _Action:
        if ctx.node in ctx.state.participants:
            if time.monotonic() > deadline:
                return _Action.ERROR_TIMEOUT
            return _Action.REMOVE_FROM_PARTICIPANTS
        return _Action.FINISH
    
 

_DistributedRendezvousOpExecutor的核心接口如下:

  • run提供了执行Rendezvous op的总入口

  • 其他接口则对应了Rendezvous op返回的action的实现。这些action的实现本质上都是对_RendezvousState内容的修改,例如_mark_rendezvous_closed是将_RendezvousState的close字段设置为了True。

class _DistributedRendezvousOpExecutor:
  def run(self, state_handler: Callable[[_RendezvousContext, float], _Action], deadline: float,) -> None:
  def _keep_alive(self) -> None:
  def _add_to_participants(self) 
  def _add_to_wait_list(self)
  def _remove_from_participants(self)
  def _remove_from_wait_list(self)
  def _mark_rendezvous_complete(self)
  def _mark_rendezvous_closed(self):
        self._state.closed = True

最后一个要介绍的概念是RendezvousHandler,其是Rendezvous系统最上层的对外接口,ElasticAgent通过该接口来在所有节点间进行协调。在Pytorch中提供了DynamicRendezvousHandler、EtcdRendezvousHandler和StaticTCPRendezvous三种实现,这里我们仅关注DynamicRendezvousHandler。

RendezvousHandler中最核心的接口是next_rendezvous,ElasticAgent会调用该接口来等待至少min个node的加入。他们实现我们后面再进行讲解。

上面介绍的这些概念,可以通过如下的关系图来进行描述。

Pytorch - 弹性训练原理

5.2 实现逻辑

在熟系完Rendezvous的基本概念后,我们现在可以来看其实现逻辑了。

首先,我们看DynamicRendezvousHandler.next_rendezvous的实现逻辑(注:ElasticAgent通过调用该接口实现的node间的协调)。DynamicRendezvousHandler.next_rendezvous 一共由5个步骤组成:

  • DynamicRendezvousHandler._stop_heartbeats():停止先TCPStore的心跳操作,通过调用定时器_PeriodicTimer的cancel接口实现;

  • Execute Exit OP:执行退出逻辑,如果当前node已经在participants中了,则先把当前节点从_RendezvousState的participants列表中删除;

  • Execute Join OP: 下图仅描述了一个常规的场景,源码中还有一些特殊情况需要处理;

    • 将自己加入到_RendezvousState的participants列表中;

    • 向TCPStore发起心跳,等待至少min个node加入;

    • 当_RendezvousState的participants的个数大于min时,mark rendezvous;

    • 此时,Join OP执行完成,返回给_RendezvousOpExecutor 个Finish action;

  • DynamicRendezvousHandler._start_heartbeats(): 开启心跳,这个逻辑通过_PeriodicTimer定期执行_RendezvousKeepAliveOp实现;_RendezvousKeepAliveOp的操作则是对_RendezvousState的last_heartbeats进行更新来实现;

  • DynamicRendezvousHandler._get_world():从_RendezvousState中获取当前rank和work_size信息;

  • Pytorch - 弹性训练原理

下面我们再看下Rendezvous的OP是如何执行的。上文提到OP是通过_DistributedRendezvousOpExecutor.run()接口统一来完成的。

  • 主流程包裹在while循环中,直到OP的action为finish方可退出循环;

  • 首先,会调用_BackendRendezvousStateHolder.sync()接口在所有node间进行_RendezvousState的同步;

    • 若当前node有内容需要更新,则调用C10dRendezvousBackend.set_state()来更新;若没有,则调用C10dRendezvousBackend.get_state()来获取最新的state;

    • 若获取了最新的state,则对当前node上存储的state进行更新;

  • 然后,调用当前需要执行的OP,OP接口会返回一个ACTION,_DistributedRendezvousOpExecutor则根据ACTION的内容执行keep_alive/add_to_participants/add_to_wait_list等操作;

  • Pytorch - 弹性训练原理

6 Failover

Failover分为两种情况:

  • ElasticAgent Process正常,但是worker process 出错

  • ElasticAgent Process 异常退出

6.1 Worker Fail

对于worker fail的场景,worker process的异常状态会被ElasticAgent捕获,实现逻辑在SimpleElasticAgent的_invoke_run接口中。

  • 该接口实现中会循环monitor 当前node上所有worker process的状态,如果process 异常,则会进行入UNHEALTHY/FAILED状态的处理流程。

  • 如果当前重试的次数小于_remain_restart,则会发起restart worker的流程

  • Pytorch - 弹性训练原理

restart worker的实现逻辑也很清晰: whaosoft aiot http://143ai.com

  • 先stop 点前node上所有worker

  • 然后重新走_initialize_workers逻辑来进行Rendezvous和start worker

    def _restart_workers(self, worker_group: WorkerGroup) -> None:
        """
        Restarts (stops, rendezvous, starts) all local workers in the group.
        """
        role = worker_group.spec.role
        log.info(f"[{role}] Stopping worker group")
        self._stop_workers(worker_group)
        worker_group.state = WorkerState.STOPPED
        self._initialize_workers(worker_group)

6.2 ElasticAgent Fail

首先,我们看下当一个node Fail掉后,弹性训练是如何运行的。这有两个node:node0和node1,开始node0和node1同时进行分布式训练,当训练到一定时间后,我们将node1 kill掉。

这是node1上的日志:

[763] epoch 14 (rank = 4, local_rank = 0) loss = 1.2388396263122559
[765] epoch 14 (rank = 6, local_rank = 2) loss = 1.4543075561523438
[766] epoch 14 (rank = 7, local_rank = 3) loss = 1.0290627479553223
[764] epoch 14 (rank = 5, local_rank = 1) loss = 1.1143463850021362
^CTraceback (most recent call last):
Traceback (most recent call last):
  File "/opt/conda/bin/torchrun", line 33, in <module>
    sys.exit(load_entry_point('torch==1.11.0', 'console_scripts', 'torchrun')())
  File "/opt/conda/lib/python3.8/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 345, in wrapper
    return f(*args, **kwargs)
  File "/opt/conda/lib/python3.8/site-packages/torch/distributed/run.py", line 724, in main
    run(args)
  File "/opt/conda/lib/python3.8/site-packages/torch/distributed/run.py", line 715, in run
    elastic_launch(
  File "/opt/conda/lib/python3.8/site-packages/torch/distributed/launcher/api.py", line 131, in __call__
    return launch_agent(self._config, self._entrypoint, list(args))
  File "/opt/conda/lib/python3.8/site-packages/torch/distributed/launcher/api.py", line 236, in launch_agent
    result = agent.run()
  File "/opt/conda/lib/python3.8/site-packages/torch/distributed/elastic/metrics/api.py", line 125, in wrapper
    result = f(*args, **kwargs)
  File "/opt/conda/lib/python3.8/site-packages/torch/distributed/elastic/agent/server/api.py", line 709, in run
    result = self._invoke_run(role)
  File "/opt/conda/lib/python3.8/site-packages/torch/distributed/elastic/agent/server/api.py", line 850, in _invoke_run
    time.sleep(monitor_interval)
  File "/opt/conda/lib/python3.8/site-packages/torch/distributed/elastic/multiprocessing/api.py", line 60, in _terminate_process_handler
    raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
torch.distributed.elastic.multiprocessing.api.SignalException: Process 759 got signal: 2

这是node0上的日志,我们可以得出以下结论:

  • 当Elastic Agent退出时,会导致其他存活的Elastic Agent中的process 运行失败;这是因为剩余process无法在正常进行collective communication了;

  • 存活的Elastic Agent会按照UNHEALTHY/FAILED的处理逻辑来重启本机的worker;若失败的Elastic Agent没有重启,则剩余的Elastic Agent重新构建worker group继续进行训练,若失败的Elastic Agent重新启动(例如kubernetes中job提供重启的机制),则会重新加入到整个训练任务中;

# 1) 此时node0和node1共同进行分布式训练
...
[11762] epoch 14 (rank = 2, local_rank = 2) loss = 1.1763713359832764                                                                                                          [702/1958]
[11760] epoch 14 (rank = 0, local_rank = 0) loss = 1.324049949645996
# 2) 此时node1被kill掉,因此当执行collective communication时,会报出异常
[E ProcessGroupNCCL.cpp:406] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete d
ata. To avoid this inconsistency, we are taking the entire process down.
terminate called after throwing an instance of 'std::runtime_error'
  what():  NCCL error: unhandled system error, NCCL version 21.0.3
ncclSystemError: System call (socket, malloc, munmap, etc) failed.
# 3)stop 其他三个process
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 11761 closing signal SIGTERM
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 11762 closing signal SIGTERM
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 11763 closing signal SIGTERM
ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: -6) local_rank: 0 (pid: 11760) of binary: /opt/conda/bin/python
# 4)重新走_initialize_workers逻辑
[11828] Initializing process group with: {'MASTER_ADDR': 'iZ2ze9q3ftqtxtqlkrk6tuZ', 'MASTER_PORT': '40539', 'WORLD_SIZE': '4', 'LOCAL_WORLD_SIZE': '4'}[11825] Initializing process group
 with: {'MASTER_ADDR': 'iZ2ze9q3ftqtxtqlkrk6tuZ', 'MASTER_PORT': '40539', 'WORLD_SIZE': '4', 'LOCAL_WORLD_SIZE': '4'}
[11826] Initializing process group with: {'MASTER_ADDR': 'iZ2ze9q3ftqtxtqlkrk6tuZ', 'MASTER_PORT': '40539', 'WORLD_SIZE': '4', 'LOCAL_WORLD_SIZE': '4'}
[11827] Initializing process group with: {'MASTER_ADDR': 'iZ2ze9q3ftqtxtqlkrk6tuZ', 'MASTER_PORT': '40539', 'WORLD_SIZE': '4', 'LOCAL_WORLD_SIZE': '4'}
[11827] (rank = 2, local_rank = 2) train worker starting...
[11828] (rank = 3, local_rank = 3) train worker starting...
[11825] (rank = 0, local_rank = 0) train worker starting...
[11826] (rank = 1, local_rank = 1) train worker starting...
# 5)node0 独自进行分布式训练
load checkpoint from checkpoint.ptload checkpoint from checkpoint.ptload checkpoint from checkpoint.ptload checkpoint from checkpoint.pt
[11826] epoch 14 (rank = 1, local_rank = 1) loss = 0.839302122592926
[11828] epoch 14 (rank = 3, local_rank = 3) loss = 0.8971960544586182
[11825] epoch 14 (rank = 0, local_rank = 0) loss = 1.3382269144058228

7 Scale Up/Down

Scale Down的可以理解为上文中Elastic Agent退出,但是没有重启的场景,因此这里不再赘述。

Scale UP这里要再介绍一下,Scale UP的流程仍旧可以用上图进行描述:

  • 当有新的节点加入时,由于当前Elastic已经建立一个的Rendezvous,其无法加入,所以当前Node会被加入到_RendezvousState的wait_list中

  • 当ElasticAgent和对应的worker process都正常运行时,monitor会返回Healthy的状态;此时,ElasticAgent会检查_RendezvousState的waiting list的node个数,发现waiting list大于0,则出发restart worker来发起新一轮的Rendezvous以将新的加入,这样新的Node加入到了worker group中;

Pytorch - 弹性训练原理

二 \ 代码----

著名物理学家,诺贝尔奖得主Richard Feynman办公室的黑板上写了:"What I cannot create, I do not understand."。在程序员界也经常有"show me the code"的口号。因此,我打算写一系列的分布式训练的文章,将以往抽象的分布式训练的概念以代码的形式展现出来,并保证每个代码可执行、可验证、可复现,并贡献出来源码让大家相互交流。

经过调研发现pytorch对于分布式训练做好很好的抽象且接口完善,因此本系列文章将以pytorch为主要框架进行,文章中的例子很多都来自pytorch的文档,并在此基础上进行了调试和扩充。

最后,由于分布式训练的理论介绍网络上已经很多了,理论部分的介绍不会是本系列文章的重点,我会将重点放在代码层面的介绍上面。

Pytorch - 分布式训练极简体验:https://zhuanlan.zhihu.com/p/477073906

Pytorch - 分布式通信原语(附源码):https://zhuanlan.zhihu.com/p/478953028

Pytorch - 手写allreduce分布式训练(附源码):https://zhuanlan.zhihu.com/p/482557067

Pytorch - 算子间并行极简实现(附源码):https://zhuanlan.zhihu.com/p/483640235

Pytorch - 多机多卡极简实现(附源码):https://zhuanlan.zhihu.com/p/486130584

1. 介绍

Pytorch在1.9.0引入了torchrun,用其替代1.9.0以前版本的torch.distributed.launch。torchrun在torch.distributed.launch 功能的基础上主要新增了两个功能:

  • Failover: 当worker训练失败时,会自动重新启动所有worker继续进行训练;

  • Elastic: 可以动态增加或或删除node节点,本文将通过一个例子说明Elastic Training应该如何使用;

本例中会先在Node0上启动4 GPU的worker group ,等其训练一段时间后,会在Node1上再启动4 GPU的workers,并与Node1上的workers构成一个新的worker group,最终构成一个2机8卡的分布式训练。

Pytorch - 弹性训练原理

2. 模型构建

一个简单的全连接模型神经网络模型

class ToyModel(nn.Module):
    def __init__(self):
        super(ToyModel, self).__init__()
        self.net1 = nn.Linear(10, 10)
        self.relu = nn.ReLU()
        self.net2 = nn.Linear(10, 5)
    def forward(self, x):
        return self.net2(self.relu(self.net1(x)))

3. checkpoint 处理

由于再每次增加或删除node时,会将所有worker kill掉,然后再重新启动所有worker进行训练。因此,在训练代码中要对训练的状态进行保存,以保证重启后能接着上次的状态继续训练。

需要保存的信息一般有如下内容:

  • model :模型的参数信息

  • optimizer :优化器的参数信心

  • epoch:当前执行到第几个epoch

save和load的代码如下所示

  • torch.save:利用python的pickle将python的object 进行序列化,并保存到本地文件;

  • torch.load : 将torch.save后的本地文件进行反序列化,并加载到内存中;

  • model.state_dict(): 存储了model 每个layer和其对应的param信息

  • optimizer.state_dict():存储了优化器的参数信信息

def save_checkpoint(epoch, model, optimizer, path):
    torch.save({
    "epoch": epoch,
    "model_state_dict": model.state_dict(),
    "optimize_state_dict": optimizer.state_dict(),
}, path)
def load_checkpoint(path):
    checkpoint = torch.load(path)
    return checkpoint

4. 训练代码

初始化逻辑如下:

  • 1~3行: 输出当前worker的关键环境变量,用于后面的结果展示

  • 5~8行:创建模型、优化器和损失函数

  • 10~12行:初始化参数信息

  • 14~19行:如果存在checkpoint,则加载checkpoint,并赋值给model、optimizer和firt_epoch

    local_rank = int(os.environ["LOCAL_RANK"])
    rank = int(os.environ["RANK"])
    print(f"[{os.getpid()}] (rank = {rank}, local_rank = {local_rank}) train worker starting...")
    
    model = ToyModel().cuda(local_rank)
    ddp_model = DDP(model, [local_rank])
    loss_fn = nn.MSELoss()
    optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)
    optimizer.zero_grad()
    max_epoch = 100
    first_epoch = 0
    ckp_path = "checkpoint.pt"
    
    if os.path.exists(ckp_path):
        print(f"load checkpoint from {ckp_path}")
        checkpoint = load_checkpoint(ckp_path)
        model.load_state_dict(checkpoint["model_state_dict"])
        optimizer.load_state_dict(checkpoint["optimize_state_dict"])
        first_epoch = checkpoint["epoch"]

训练逻辑:

  • 1行:epoch执行的次数为first_epoch到max_epoch,以便能够在worker被重启后继续原有的epoch继续训练;

  • 2行:为了展示动态添加node效果,这里添加sleep函数来降低训练的速度;

  • 3~8行:模型训练流程;

  • 9行:为了简单,文本每个epoch进行一次checkpoint保存;将当前的epoch,model和optimizer保存到checkpoint中;

    for i in range(first_epoch, max_epoch):
        time.sleep(1) # 为了展示动态添加node效果,这里添加sleep函数来降低训练的速度
        outputs = ddp_model(torch.randn(20, 10).to(local_rank))
        labels = torch.randn(20, 5).to(local_rank)
        loss = loss_fn(outputs, labels)
        loss.backward()
        print(f"[{os.getpid()}] epoch {i} (rank = {rank}, local_rank = {local_rank}) loss = {loss.item()}\n")
        optimizer.step()
        save_checkpoint(i, model, optimizer, ckp_path)

5. 启动方式

由于我们使用torchrun来启动多机多卡任务,无需使用spawn接口来启动多个进程(torchrun会负责将我们的python script启动为一个process),因此直接调用上文编写的train函数,并在前后分别添加DistributedDataParallel的初始化和效果函数即可。

下面代码描述了上文train接口的调用。

def run():
    env_dict = {
        key: os.environ[key]
        for key in ("MASTER_ADDR", "MASTER_PORT", "WORLD_SIZE", "LOCAL_WORLD_SIZE")
    }
    print(f"[{os.getpid()}] Initializing process group with: {env_dict}")
    dist.init_process_group(backend="nccl")
    train()
    dist.destroy_process_group()
if __name__ == "__main__":
    run()

本例中使用torchrun来执行多机多卡的分布式训练任务(注:torch.distributed.launch已经被pytorch淘汰了,尽量不要再使用)。启动脚本描述如下(注:node0和node1均通过该脚本进行启动)

  • --nnodes=1:3 :表示当前训练任务接受最少1个node,最多3个node参与分布式训练;

  • --nproc_per_node=4:表示每个node上节点有4个process

  • --max_restarts=3: worker group最大的重启次数;这里需要注意的是,node fail、node scale down和node scale up都会导致restart;

  • --rdzv_id=1:一个unique的job id,所有node均使用同一个job id;

  • --rdzv_backend: rendezvous的backend实现,默认支持c10d和etcd两种;rendezvous用于多个node之间的通信和协调;

  • --rdzv_endpoint:rendezvous的地址,应该为一个node的host ip和port;

torchrun \
    --nnodes=1:3\
    --nproc_per_node=4\
    --max_restarts=3\
    --rdzv_id=1\
    --rdzv_backend=c10d\
    --rdzv_endpoint="192.0.0.1:1234"\
    train_elastic.py

6. 结果分析

代码:BetterDL - train_elastic.py:https://github.com/tingshua-yts/BetterDL/blob/master/test/pytorch/DDP/train_elastic.py

运行环境: 2台4卡 v100机器

image: pytorch/pytorch:1.11.0-cuda11.3-cudnn8-runtime
gpu: v100

先在node0上执行执行启动脚本

torchrun \
    --nnodes=1:3\
    --nproc_per_node=4\
    --max_restarts=3\
    --rdzv_id=1\
    --rdzv_backend=c10d\
    --rdzv_endpoint="192.0.0.1:1234"\
    train_elastic.py

得到如下结果

  • 2~5行:当前启动的是单机4卡的训练任务,因此WORLD_SIZE为4, LOCAL_WORKD_SIZE也为4

  • 6~9行:共有4个rank参与了分布式训练,rank0~rank3

  • 10~18行: rank0~rank3 均从epoch=0开始训练

r/workspace/DDP# sh run_elastic.sh
[4031] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '44901', 'WORLD_SIZE': '4', 'LOCAL_WORLD_SIZE': '4'}
[4029] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '44901', 'WORLD_SIZE': '4', 'LOCAL_WORLD_SIZE': '4'}
[4030] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '44901', 'WORLD_SIZE': '4', 'LOCAL_WORLD_SIZE': '4'}
[4032] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '44901', 'WORLD_SIZE': '4', 'LOCAL_WORLD_SIZE': '4'}
[4029] (rank = 0, local_rank = 0) train worker starting...
[4030] (rank = 1, local_rank = 1) train worker starting...
[4032] (rank = 3, local_rank = 3) train worker starting...
[4031] (rank = 2, local_rank = 2) train worker starting...
[4101] epoch 0 (rank = 1, local_rank = 1) loss = 0.9288564920425415
[4103] epoch 0 (rank = 3, local_rank = 3) loss = 0.9711472988128662
[4102] epoch 0 (rank = 2, local_rank = 2) loss = 1.0727070569992065
[4100] epoch 0 (rank = 0, local_rank = 0) loss = 0.9402943253517151
[4100] epoch 1 (rank = 0, local_rank = 0) loss = 1.0327017307281494
[4101] epoch 1 (rank = 1, local_rank = 1) loss = 1.4485043287277222
[4103] epoch 1 (rank = 3, local_rank = 3) loss = 1.0959293842315674
[4102] epoch 1 (rank = 2, local_rank = 2) loss = 1.0669530630111694
...

在node1上执行与上面相同的脚本

torchrun \
    --nnodes=1:3\
    --nproc_per_node=4\
    --max_restarts=3\
    --rdzv_id=1\
    --rdzv_backend=c10d\
    --rdzv_endpoint="192.0.0.1:1234"\
    train_elastic.py

node1上结果如下:

  • 2~5行:由于添加node1,当前执行的是2机8卡的分布式训练任务,因此WORLD_SIZE=8, LOCAL_WORLD_SIZE=4

  • 6~9行:当前node1上workers的rank为rank4 ~rank7

  • 13~20行: 由于node1是在node0上work训练到epoch35的时候加入的,因此其接着epoch 35开始训练

/workspace/DDP# sh run_elastic.sh
[696] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '42913', 'WORLD_SIZE': '8', 'LOCAL_WORLD_SIZE': '4'}
[697] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '42913', 'WORLD_SIZE': '8', 'LOCAL_WORLD_SIZE': '4'}
[695] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '42913', 'WORLD_SIZE': '8', 'LOCAL_WORLD_SIZE': '4'}
[694] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '42913', 'WORLD_SIZE': '8', 'LOCAL_WORLD_SIZE': '4'}
[697] (rank = 7, local_rank = 3) train worker starting...
[695] (rank = 5, local_rank = 1) train worker starting...
[694] (rank = 4, local_rank = 0) train worker starting...
[696] (rank = 6, local_rank = 2) train worker starting...
load checkpoint from checkpoint.ptload checkpoint from checkpoint.pt
load checkpoint from checkpoint.pt
load checkpoint from checkpoint.pt
[697] epoch 35 (rank = 7, local_rank = 3) loss = 1.1888569593429565
[694] epoch 35 (rank = 4, local_rank = 0) loss = 0.8916441202163696
[695] epoch 35 (rank = 5, local_rank = 1) loss = 1.5685604810714722
[696] epoch 35 (rank = 6, local_rank = 2) loss = 1.11683189868927
[696] epoch 36 (rank = 6, local_rank = 2) loss = 1.3724170923233032
[694] epoch 36 (rank = 4, local_rank = 0) loss = 1.061527967453003
[695] epoch 36 (rank = 5, local_rank = 1) loss = 0.96876460313797
[697] epoch 36 (rank = 7, local_rank = 3) loss = 0.8060566782951355
...

node0上结果如下:

  • 6~9行: node0上的works在执行到epoch 35时,node1上执行了训练脚本,请求加入到训练任务中

  • 10~13行:所有workers重新启动,由于添加了node1,当前执行的是2机8卡的分布式训练任务,因此WORLD_SIZE=8, LOCAL_WORLD_SIZE=4

  • 14~17行:当前node1上works的rank为rank0~rank3

  • 18~21行:加载checkpoint

  • 22~30行:接着checkpoint中的model、optimizer和epoch继续训练

...
[4100] epoch 35 (rank = 0, local_rank = 0) loss = 1.0746158361434937
[4101] epoch 35 (rank = 1, local_rank = 1) loss = 1.1712706089019775
[4103] epoch 35 (rank = 3, local_rank = 3) loss = 1.1774182319641113
[4102] epoch 35 (rank = 2, local_rank = 2) loss = 1.0898035764694214
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 4100 closing signal SIGTERM
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 4101 closing signal SIGTERM
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 4102 closing signal SIGTERM
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 4103 closing signal SIGTERM
[4164] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '42913', 'WORLD_SIZE': '8', 'LOCAL_WORLD_SIZE': '4'}
[4165] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '42913', 'WORLD_SIZE': '8', 'LOCAL_WORLD_SIZE': '4'}
[4162] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '42913', 'WORLD_SIZE': '8', 'LOCAL_WORLD_SIZE': '4'}
[4163] Initializing process group with: {'MASTER_ADDR': '192.0.0.1', 'MASTER_PORT': '42913', 'WORLD_SIZE': '8', 'LOCAL_WORLD_SIZE': '4'}
[4162] (rank = 0, local_rank = 0) train worker starting...
[4163] (rank = 1, local_rank = 1) train worker starting...
[4164] (rank = 2, local_rank = 2) train worker starting...
[4165] (rank = 3, local_rank = 3) train worker starting...
load checkpoint from checkpoint.pt
load checkpoint from checkpoint.pt
load checkpoint from checkpoint.pt
load checkpoint from checkpoint.pt
[4165] epoch 35 (rank = 3, local_rank = 3) loss = 1.3437936305999756
[4162] epoch 35 (rank = 0, local_rank = 0) loss = 1.5693414211273193
[4163] epoch 35 (rank = 1, local_rank = 1) loss = 1.199862003326416
[4164] epoch 35 (rank = 2, local_rank = 2) loss = 1.0465545654296875
[4163] epoch 36 (rank = 1, local_rank = 1) loss = 0.9741991758346558
[4162] epoch 36 (rank = 0, local_rank = 0) loss = 1.3609280586242676
[4164] epoch 36 (rank = 2, local_rank = 2) loss = 0.9585908055305481
[4165] epoch 36 (rank = 3, local_rank = 3) loss = 0.9169824123382568
...