目录
在前面的文章之中,我们已经学习了PyTorch 分布式的基本模块,介绍了官方的几个例子,我们接下来会介绍PyTorch的弹性训练,本文是第二篇,重点关注的是如何启动弹性训练,并且可以对系统总体架构有所了解。
弹性训练系列文章如下:
[源码解析] PyTorch 分布式之弹性训练(1) --- 总体思路
为了更好的说明(这个说明可能在后面文章也会出现,因为太重要了),我们先总述一下TE 最重要的 Agent 和 Rendezvous 两个概念。
rendezvous
实现 worker 间的相互发现(比如把状态上报到KVStore),成员变动时候基于 rendezvous
进行变更同步等等。我们从源码中取出示意图看看,大家先有一个总体概念。
我们知道,PET是从 PyTorch v1.9 合并进来的,因为合并了弹性训练,所以分布式启动的方式有了很大的改变。
V1.9 之前是使用 torch/distributed/launch.py 进行启动,比如:
python -m torch.distributed.launch --nproc_per_node=NUM_GPUS_YOU_HAVE
--nnodes=2 --node_rank=0 --master_addr="192.168.1.1"
--master_port=1234 YOUR_TRAINING_SCRIPT.py (--arg1 --arg2 --arg3
and all other arguments of your training script)
此处参数含义是:
nnodes
:是参与训练的节点数目。nproc_per_node
:每个节点上运行的进程数目。node_rank
:当前节点标识符。master_addr
和 master_port
是 master 监听的地址和端口。当运行时,torch.distributed.launch
会设置一些环境变量,包括 world_size
,master_addr
和 master_port
等等。然后在当前机器上创建 nproc_per_node
个进程,这些进程构成了一个本地组。如果一共有 NODE_SIZE
个机器参与训练,则一共有 NODE_SIZE * TRAINERS_PER_NODE
个进程。如果想启动一个分布式训练任务,则需要在所有的机器上执行相关命令。
PyTorch 1.9 使用 torch/distributed/run.py 进行启动。如果依然采用 torch/distributed/launch.py,其实其内部已经透传给 run.py,具体参见代码:
def main(args=None):
logger.warn(
"The module torch.distributed.launch is deprecated "
"and going to be removed in future."
"Migrate to torch.distributed.run"
)
args = parse_args(args)
run(args)
torch.distributed.run
是之前torch.distributed.launch
的一个超集,提供如下新功能:
RANK
和 WORLD_SIZE
是自动分配的。为了使用弹性训练,用户代码也需要做一些修改,如果用户的训练脚本已经支持 torch.distributed.launch ,则只需要修改几处就可以使用torch.distributed.run
:
rdzv_backend
和 rdzv_endpoint
。对于大多数用户来说,这其实就是“c10d”(参见“rendezvous“)。其实这就替代了之前的MASTER_ADDR 和 MASTER_PORT。use_env
参数已被删除。请从 LOCAL_RANK 环境变量中获取local_rank (例如,os.environ["LOCAL_RANK"]
)。load_checkpoint(path)
和 save_checkpoint(path)
逻辑,即手动处理Checkpoint。因为当worker失败时,我们将使用最近的checkpoint来恢复现场,重启所有worker。下面是一个训练脚本的示例,该脚本在每个epoch上设置检查点,因此在失败时最差也只是会丢失一个epoch的训练成果。
def main():
args = parse_args(sys.argv[1:])
state = load_checkpoint(args.checkpoint_path)
initialize(state)
# torch.distributed.run ensure that this will work
# by exporting all the env vars needed to initialize the process group
torch.distributed.init_process_group(backend=args.backend)
for i in range(state.epoch, state.total_num_epochs)
for batch in iter(state.dataset)
train(batch, state.model)
state.epoch += 1
save_checkpoint(state)
所以,我们接下来看看在新模式之下,如何分布式启动。
部署一般按照如下方式。
--rdzv_endpoint
传递给启动程序脚本)当使用作业/群集管理器时,多节点作业的入口点命令应为 launcher。
我们首先通过几个例子来看看如何启动分布式训练。
单节点多worker的启动方式如下,其实就是Standalone 模式,这是分布式模式的一种特例,具体就是针对单机多 Worker 提供了一些便利设置。
python -m torch.distributed.run
--standalone
--nnodes=1
--nproc_per_node=$NUM_TRAINERS
YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)
如下是容错方式启动,固定数目workers,没有弹性训练。 --nproc_per_node=$NUM_TRAINERS 一般是 单节点上GPU 个数。
python -m torch.distributed.run
--nnodes=$NUM_NODES
--nproc_per_node=$NUM_TRAINERS
--rdzv_id=$JOB_ID
--rdzv_backend=c10d
--rdzv_endpoint=$HOST_NODE_ADDR
YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)
HOST_NODE_ADDR
, 的格式是: : ,指定了 C10d rendezvous 后端所运行的节点地址和端口,这个节点可以是训练集群中任意节点,但是最好找一个高带宽的节点。
下面是弹性训练,弹性区间为 (min=1
, max=4
)。通过指定rdzv参数,可以实现多机训练,具备容错与弹性能力。
在多台机器上分别执行以下命令启动:最小节点数为MIN_SIZE,最大为MAX_SIZE,利用etcd服务实现一致性和信息同步。
python -m torch.distributed.run
--nnodes=1:4
--nproc_per_node=$NUM_TRAINERS
--rdzv_id=$JOB_ID
--rdzv_backend=c10d
--rdzv_endpoint=$HOST_NODE_ADDR
YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)
HOST_NODE_ADDR
, 的格式是: : ,指定了 C10d rendezvous 后端所运行的节点地址和端口,这个节点可以是训练集群中任意节点,但是最好找一个高带宽的节点。
关于 rendezvous backend,有几点说明:
对于多节点训练,需要指定:
--rdzv_id
: 一个唯一的 job id,在参与job的所有节点之间共享。--rdzv_backend
: torch.distributed.elastic.rendezvous.RendezvousHandler
的一个实现。 (--rdzv_backend
默认是static模式,不支持容错和弹性伸缩)--rdzv_endpoint
: rendezvous backend 所运行的 endpoint,通常格式为:host:port
。就是取代了之前的 master address / port 设置。目前,以下几种后端可以直接使用,c10d
(推荐), etcd-v2
, and etcd
(legacy) 。为了使用 etcd-v2
或者 etcd
,需要搭建一个 v2
api开启的 etcd server (即. --enable-v2
)。
既然以上启动都是用 torch/distributed/run.py,所以我们仔细分析一下这个脚本,该脚本提供三个功能:
RANK
and WORLD_SIZE
;启动脚本中,一些参数定义如下:
Node
- 物理实例或容器;映射到与 job manager 所协调的单元。Worker
- 分布式训练环境中的worker。WorkerGroup
- 执行相同功能的一组worker(例如trainers)。LocalWorkerGroup
- 在同一节点上运行的工作组中的workers子集。 节点
运行 LOCAL_WORLD_SIZE
个workers,这些 workers 组成LocalWorkerGroup
。LocalWorkerGroups
组成WorkerGroups
。RANK
- 工作组中worker的rank,是全局rank,可以认为是一个全局GPU资源列表。 RANK
和LOCAL_RANK
的稳定性做任何假设和依赖编码。LOCAL_RANK
- 本地工作组中,某个worker 的 rank,可以认为是当前节点上的GPU资源列表。GROUP_RANK
- worker group的rank。介于0和“最大节点数”之间的数字。如果每个节点运行一个单一工作组,那GROUP_RANK
就是这个节点的rank。ROLE_RANK
- 对于具有相同角色worker来说,他们之间共享的rank,角色在“WorkerSpec”中被指定。WORLD_SIZE
- 工作组中worker的总数。因为节点会加入/离开,所以WORLD_SIZE
会变化,不能依赖 WORLD_SIZE
的稳定性进行编码。LOCAL_WORLD_SIZE
- 本地工作组的大小,即本地运行的worker数目,等于在torch.distributed.run
运行时候指定的--nproc_per_node
。目前,torch/distributed/run.py 仅支持同构的 LOCAL_WORLD_SIZE
。也就是说,假设所有节点运行相同数量的本地工作者(每个角色)。ROLE_WORLD_SIZE
- 具有同样角色的workers总数,在 WorkerSpec
之中被指定。rdzv_id
- 用户定义的id,用于唯一标识作业的工作组。这个id在每个节点加入特定工作组时候使用。rdzv_backend
-rendezvous 的后端(例如“c10d”)。这通常是一个强一致性的键值存储。rdzv_endpoint
- rendezvous 后端端点;通常以“<host>:<port>
”的形式出现。run_id
: 用户定义的id,它唯一地标识分布式应用程序的一个实例。它通常映射到作业id并用于允许节点加入正确的分布式应用程序。TORCHELASTIC_RUN_ID
- 与 rendezvous run_id
相等,即唯一的job id。TORCHELASTIC_RESTART_COUNT
- 迄今为止,工作组重启的次数。TORCHELASTIC_MAX_RESTARTS
- 配置的最大重启数目。为了更好的理解上面的参数,我们选取部分相关函数/变量看看。
这两个变量是动态生成的,所以从 state 之中取出。
rank, world_size = self._get_world()
def _get_world(self) -> Tuple[int, int]:
state = self._state_holder.state
return state.participants[self._this_node], len(state.participants)
该全局变量存储了每个 group 的 global rank 到 local rank 映射信息。
# Process group's global rank to local rank mapping
_pg_group_ranks: Dict[ProcessGroup, Dict[int, int]] = {}
其赋值举例如下:
# Create the global rank to group rank mapping
_pg_group_ranks[pg] = {
global_rank: group_rank
for group_rank, global_rank in enumerate(ranks)
}
我们可以利用 global rank 从 _pg_group_ranks 之中提取对应的 local rank。
def _get_group_rank(group: ProcessGroup, rank):
"""
Helper that gets a given group's local rank in the group from a given global
rank.
"""
if group is GroupMember.WORLD:
raise RuntimeError("group.WORLD does not have local rank to global "
"rank mapping")
if group not in _pg_group_ranks:
raise RuntimeError("The given group does not exist")
try:
group_rank = _pg_group_ranks[group][rank]
except KeyError:
raise RuntimeError(f"The global rank {rank} is not part of the group {group}") from None
return group_rank
我们可以利用一个 group 的 local rank 获取到其 gloabl rank。
def _get_global_rank(group, group_rank):
"""
Helper that gets a given group's global rank from a given local rank in the
group.
"""
if group is GroupMember.WORLD:
raise RuntimeError("group.WORLD does not have local rank to global "
"rank mapping")
group_rank_map = _pg_group_ranks[group]
for rank, grp_rank in group_rank_map.items():
if grp_rank == group_rank:
return rank
raise RuntimeError("The group rank is not part of the group")
我们可以 _get_group_size 获取到某一个group 的大小。
def _get_group_size(group):
"""
Helper that gets a given group's world size.
"""
if group is GroupMember.WORLD or group is None:
default_pg = _get_default_group()
return default_pg.size()
if group not in _pg_group_ranks:
raise RuntimeError("The given group does not exist")
return len(_pg_group_ranks[group])
这个变量可以得到每个node之上支持多少个进程。
def determine_local_world_size(nproc_per_node: str):
try:
logging.info(f"Using nproc_per_node={nproc_per_node}.")
return int(nproc_per_node)
except ValueError:
if nproc_per_node == "cpu":
num_proc = os.cpu_count()
device_type = "cpu"
elif nproc_per_node == "gpu":
if not torch.cuda.is_available():
raise ValueError("Cuda is not available.")
device_type = "gpu"
num_proc = torch.cuda.device_count()
elif nproc_per_node == "auto":
if torch.cuda.is_available():
num_proc = torch.cuda.device_count()
device_type = "gpu"
else:
num_proc = os.cpu_count()
device_type = "cpu"
else:
raise ValueError(f"Unsupported nproc_per_node value: {nproc_per_node}")
)
return num_proc
脚本入口主要代码如下,可以看到,其调用到了 elastic_launch 来完成功能,所以我们下一节就要顺藤摸瓜来看看这个函数。
from torch.distributed.launcher.api import LaunchConfig, elastic_launch
def run(args):
if args.standalone: # 有两种模式:Standalone 模式和分布式模式,这里要判断一下
args.rdzv_backend = "c10d"
args.rdzv_endpoint = "localhost:29400"
args.rdzv_id = str(uuid.uuid4())
log.info(
f"\n**************************************\n"
f"Rendezvous info:\n"
f"--rdzv_backend={args.rdzv_backend} "
f"--rdzv_endpoint={args.rdzv_endpoint} "
f"--rdzv_id={args.rdzv_id}\n"
f"**************************************\n"
)
config, cmd, cmd_args = config_from_args(args)
elastic_launch(
config=config,
entrypoint=cmd,
)(*cmd_args)
def main(args=None):
args = parse_args(args)
run(args)
if __name__ == "__main__":
logging.basicConfig(
level=logging.INFO, format="[%(levelname)s] %(asctime)s %(module)s: %(message)s"
)
main()
我们下面就从 elastic_launch 开始,看看在单节点上如何启动运行。我们首先给出一个总体示意图,图上是两个节点,每个节点有一个 agent,agent下面是一个 worker group,组下面是4个worker。
我们再从源码中找一个例子来看看,这里只是设置了两个workers。
import uuid
import torch
from torch.distributed.launcher.api import LaunchConfig, elastic_launch
def worker_fn(t1, t2):
return torch.add(t1, t2)
def main():
t1 = torch.rand((3,3), requires_grad=True)
t2 = torch.rand((3, 3), requires_grad=True)
config = LaunchConfig(
min_nodes=2,
max_nodes=4,
nproc_per_node=1,
run_id=str(uuid.uuid4()),
role="trainer",
rdzv_endpoint="localhost:29400",
rdzv_backend="c10d",
max_restarts=1,
monitor_interval=1,
start_method="spawn",
)
outputs = elastic_launch(config, worker_fn)(t1, t2)
if __name__ == '__main__':
main()
输出如下,可以看到有两个 worker 进程 和一个 agent 进程。
{"name": "torchelastic.worker.status.SUCCEEDED", "source": "WORKER", "timestamp": 0, "metadata": {"run_id": "7fbf85fe-b8b3-462e-887e-8121e3062e0b", "global_rank": 0, "group_rank": 0, "worker_id": "12172", "role": "trainer", "hostname": "DESKTOP-0GO3RPO", "state": "SUCCEEDED", "total_run_time": 31, "rdzv_backend": "c10d", "raw_error": null, "metadata": "{\"group_world_size\": 1, \"entry_point\": \"worker_fn\", \"local_rank\": [0], \"role_rank\": [0], \"role_world_size\": [2]}", "agent_restarts": 0}}
{"name": "torchelastic.worker.status.SUCCEEDED", "source": "WORKER", "timestamp": 0, "metadata": {"run_id": "7fbf85fe-b8b3-462e-887e-8121e3062e0b", "global_rank": 1, "group_rank": 0, "worker_id": "3276", "role": "trainer", "hostname": "DESKTOP-0GO3RPO", "state": "SUCCEEDED", "total_run_time": 31, "rdzv_backend": "c10d", "raw_error": null, "metadata": "{\"group_world_size\": 1, \"entry_point\": \"worker_fn\", \"local_rank\": [1], \"role_rank\": [1], \"role_world_size\": [2]}", "agent_restarts": 0}}
{"name": "torchelastic.worker.status.SUCCEEDED", "source": "AGENT", "timestamp": 0, "metadata": {"run_id": "7fbf85fe-b8b3-462e-887e-8121e3062e0b", "global_rank": null, "group_rank": 0, "worker_id": null, "role": "trainer", "hostname": "DESKTOP-0GO3RPO", "state": "SUCCEEDED", "total_run_time": 31, "rdzv_backend": "c10d", "raw_error": null, "metadata": "{\"group_world_size\": 1, \"entry_point\": \"worker_fn\"}", "agent_restarts": 0}}
顺着代码我们深入挖掘一下。elastic_launch 的作用就是启动一个 torchelastic agent,然后通过这个 agent来调用用户程序入口,agent 会启动 worker 进行训练,并且管理 worker 生命周期。
class elastic_launch:
"""
Launches an torchelastic agent on the container that invoked the entrypoint.
1. Pass the ``entrypoint`` arguments as non ``kwargs`` (e.g. no named parameters)/
``entrypoint`` can be a function or a command.
2. The return value is a map of each worker's output mapped
by their respective global rank.
"""
def __init__(
self,
config: LaunchConfig,
entrypoint: Union[Callable, str, None],
):
self._config = config
self._entrypoint = entrypoint
def __call__(self, *args, **kwargs):
return launch_agent(self._config, self._entrypoint, list(args)) # 内部会调用用户程序
launch_agent 启动了一个 LocalElasticAgent,调用了其 run 方法。
@record
def launch_agent(
config: LaunchConfig,
entrypoint: Union[Callable, str, None],
args: List[Any],
) -> Dict[int, Any]:
if not config.run_id:
run_id = str(uuid.uuid4().int)
config.run_id = run_id
entrypoint_name = _get_entrypoint_name(entrypoint, args)
rdzv_parameters = RendezvousParameters(
backend=config.rdzv_backend,
endpoint=config.rdzv_endpoint,
run_id=config.run_id,
min_nodes=config.min_nodes,
max_nodes=config.max_nodes,
**config.rdzv_configs,
)
agent = None
rdzv_handler = rdzv_registry.get_rendezvous_handler(rdzv_parameters)
master_addr, master_port = _get_addr_and_port(rdzv_parameters)
try:
spec = WorkerSpec( # 1. 得到spec
role=config.role,
local_world_size=config.nproc_per_node,
entrypoint=entrypoint,
args=tuple(args),
rdzv_handler=rdzv_handler, # RendezvousHandler
max_restarts=config.max_restarts,
monitor_interval=config.monitor_interval,
redirects=config.redirects,
tee=config.tee,
master_addr=master_addr,
master_port=master_port,
)
cfg = metrics.MetricsConfig(config.metrics_cfg) if config.metrics_cfg else None
metrics.initialize_metrics(cfg)
agent = LocalElasticAgent( # 2. 构建代理
spec=spec, start_method=config.start_method, log_dir=config.log_dir
)
result = agent.run() # 3. 启动代理
events.record(agent.get_agent_status_event(WorkerState.SUCCEEDED))
if result.is_failed():
# ChildFailedError is treated specially by @record
# if the error files for the failed children exist
# @record will copy the first error (root cause)
# to the error file of the launcher process.
raise ChildFailedError(
name=entrypoint_name,
failures=result.failures,
)
else:
return result.return_values
except ChildFailedError:
raise
except Exception:
if agent:
events.record(agent.get_agent_status_event(WorkerState.FAILED))
else:
events.record(_construct_event(config))
raise
finally:
rdzv_handler.shutdown()
这里有几个关键点:
WorkerSpec :这是配置信息,里面包含了代理所需要的某些全局信息,比如 RendezvousHandler,role,entry(用户函数)。
spec = {WorkerSpec}
args = {tuple: 2} (tensor, tensor)
fn = {NoneType} None
local_world_size = {int} 1
master_addr = {NoneType} None
master_port = {NoneType} None
max_restarts = {int} 1
monitor_interval = {int} 1
rdzv_handler = {DynamicRendezvousHandler}
redirects = {Std} Std.NONE
role = {str} 'trainer'
tee = {Std} Std.NONE
entry = worker_fn
代理会从这里提取各种所需信息。比如_start_workers 会从中获取 store。
use_agent_store = spec.rdzv_handler.get_backend() == "static"
此时逻辑为:
+--------------------------+ +---------------------------------------------------+
|LocalElasticAgent | | WorkerSpec |
| | | |
| WorkerSpec +--------------> | rdzv_handler = {DynamicRendezvousHandler} --------+
| | | | |
| rdzv_run_id | | entry = worker_fn | |
| | | | |
| store | | role = {str} 'trainer' | |
| | | | |
| | +---------------------------------------------------+ |
| | |
| | |
| | |
| | |
| | +-----------------------------------------+ |
+--------------------------+ |DynamicRendezvousHandler | |
| | |
| | |
| _settings: RendezvousSettings | <---+
| |
| _store: Store |
| |
| _state_holder: _RendezvousStateHolder |
| |
| _op_executor: _RendezvousOpExecutor |
| |
+-----------------------------------------+
WorkerGroup 代表了一个工作组。WorkerGroup 作为一个整体来管理多个 workers,进行批量处理。
class WorkerGroup:
"""
Represents the set of ``Worker`` instances for the given ``WorkerSpec``
managed by ``ElasticAgent``. Whether the worker group contains cross
instance workers or not depends on the implementation of the agent.
"""
__slots__ = ["spec", "workers", "store", "group_rank", "group_world_size", "state"]
def __init__(self, spec: WorkerSpec):
self.spec = spec
self.workers = [Worker(local_rank=i) for i in range(self.spec.local_world_size)]
# assigned after rdzv
self.store = None
self.group_rank = None
self.group_world_size = None
self.state = WorkerState.INIT
在SimpleElasticAgent 初始化之中,会建立一个 WorkerGroup。
class SimpleElasticAgent(ElasticAgent):
"""
An ``ElasticAgent`` that manages workers (``WorkerGroup``)
for a single ``WorkerSpec`` (e.g. one particular type of worker role).
"""
def __init__(self, spec: WorkerSpec, exit_barrier_timeout: float = 300):
self._worker_group = WorkerGroup(spec)
self._remaining_restarts = self._worker_group.spec.max_restarts
self._store = None
self._exit_barrier_timeout = exit_barrier_timeout
self._total_execution_time = 0
具体如下:
+-----------------------------+ +------------------------------------------------+
| LocalElasticAgent | | WorkerSpec |
| | | |
| +------------------------+ | | rdzv_handler = {DynamicRendezvousHandler} -------+
| |WorkerGroup | | | | |
| | spec +--------------> | entry = worker_fn | |
| | workers | | | | |
| | store | | | role = {str} 'trainer' | |
| | group_rank | | | | |
| | group_world_size | | +------------------------------------------------+ |
| | | | |
| +------------------------+ | |
| | |
| rdzv_run_id | |
| store | +-----------------------------------------+ |
| | |DynamicRendezvousHandler | |
+-----------------------------+ | | |
| | |
| _settings: RendezvousSettings | <--+
| |
| _store: Store |
| |
| _state_holder: _RendezvousStateHolder |
| |
| _op_executor: _RendezvousOpExecutor |
| |
+-----------------------------------------+
SimpleElasticAgent 是 LocalElasticAgent 的基类,所以会先运行到WorkerSpec.run 方法这里,run方法则调用了 _invoke_run。
@prof
def run(self, role: str = DEFAULT_ROLE) -> RunResult:
start_time = time.monotonic()
try:
result = self._invoke_run(role) # 调用
self._total_execution_time = int(time.monotonic() - start_time)
self._record_metrics(result)
self._record_worker_events(result)
return result
finally:
# record the execution time in case there were any exceptions during run.
self._total_execution_time = int(time.monotonic() - start_time)
self._shutdown()
代理在 invoke_run 之中做如下操作:
def _invoke_run(self, role: str = DEFAULT_ROLE) -> RunResult:
# NOTE: currently only works for a single role
spec = self._worker_group.spec
role = spec.role
self._initialize_workers(self._worker_group) # 启动worker
monitor_interval = spec.monitor_interval
rdzv_handler = spec.rdzv_handler
while True:
assert self._worker_group.state != WorkerState.INIT
# 定期监控
time.sleep(monitor_interval)
# 监控客户程序运行情况
run_result = self._monitor_workers(self._worker_group) # 得到进程运行结果
state = run_result.state
self._worker_group.state = state
put_metric(f"workers.{role}.remaining_restarts", self._remaining_restarts)
put_metric(f"workers.{role}.{state.name.lower()}", 1)
if state == WorkerState.SUCCEEDED:
# 程序正常结束
self._exit_barrier()
return run_result
elif state in {WorkerState.UNHEALTHY, WorkerState.FAILED}:
# 程序出错
if self._remaining_restarts > 0: # 重试
self._remaining_restarts -= 1
self._restart_workers(self._worker_group)
else:
self._stop_workers(self._worker_group) # 重试次数达到,结束workers
self._worker_group.state = WorkerState.FAILED
self._exit_barrier()
return run_result
elif state == WorkerState.HEALTHY:
# 节点成员关系有变化,比如scale up,就会有新节点waiting
# membership changes do not count as retries
num_nodes_waiting = rdzv_handler.num_nodes_waiting()
group_rank = self._worker_group.group_rank
# 如果有新的节点在waiting,就重启所有workers
if num_nodes_waiting > 0:
self._restart_workers(self._worker_group)
else:
raise Exception(f"[{role}] Worker group in {state.name} state")
于是最终逻辑如下:
+----------------------------------------------+
| LocalElasticAgent |
| | +---------------------------------------------------+
| rdzv_run_id | | WorkerSpec |
| | | |
| store +------------------------+ | | rdzv_handler = {DynamicRendezvousHandler} +-------+
| |WorkerGroup | | | | |
| _pcontext | spec +------------> | entry = worker_fn | |
| | workers | | | | |
| | store | | | role = {str} 'trainer' | |
| | group_rank | | | | |
| | group_world_size | | +---------------------------------------------------+ |
| | | | |
| +------------------------+ | |
| +----------------------------------------+ | |
| | _invoke_run | | |
| | | | +-----------------------------------------+ |
| | _initialize_workers +------------------------+ |DynamicRendezvousHandler | |
| | | | | | | |
| | | | | | | |
| | while True: | | | | _settings: RendezvousSettings | <---+
| | _monitor_workers(_worker_group) | | | | |
| | + | | | | _store: Store |
| | | _pcontext.wait | | | | |
| | | | | | | _state_holder: _RendezvousStateHolder |
| +----------------------------------------+ | | | |
| | | | | _op_executor: _RendezvousOpExecutor |
+----------------------------------------------+ | | |
| | +-----------------------------------------+
| |
v v
+-------------------------------------------------+
| +------------+ +------------+ +------------+ |
| |Process | |Process | |Process | |
| | | | | | | |
| | work_fn | | work_fn | | work_fn | |
| | | | | | | |
| +------------+ +------------+ +------------+ |
+-------------------------------------------------+
手机如下:
至此,脚本如何启动和单体流程我们分析完毕,下一篇我们来具体分析代理。
PyTorch Elastic源码阅读(