File size: 7,654 Bytes
079c32c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 |
from typing import Callable, Tuple, List, Any, Union
from easydict import EasyDict
import os
import numpy as np
import torch
import torch.distributed as dist
from .default_helper import error_wrapper
# from .slurm_helper import get_master_addr
def get_rank() -> int:
"""
Overview:
Get the rank of current process in total world_size
"""
# return int(os.environ.get('SLURM_PROCID', 0))
return error_wrapper(dist.get_rank, 0)()
def get_world_size() -> int:
"""
Overview:
Get the world_size(total process number in data parallel training)
"""
# return int(os.environ.get('SLURM_NTASKS', 1))
return error_wrapper(dist.get_world_size, 1)()
broadcast = dist.broadcast
allgather = dist.all_gather
broadcast_object_list = dist.broadcast_object_list
def allreduce(x: torch.Tensor) -> None:
"""
Overview:
All reduce the tensor ``x`` in the world
Arguments:
- x (:obj:`torch.Tensor`): the tensor to be reduced
"""
dist.all_reduce(x)
x.div_(get_world_size())
def allreduce_async(name: str, x: torch.Tensor) -> None:
"""
Overview:
All reduce the tensor ``x`` in the world asynchronously
Arguments:
- name (:obj:`str`): the name of the tensor
- x (:obj:`torch.Tensor`): the tensor to be reduced
"""
x.div_(get_world_size())
dist.all_reduce(x, async_op=True)
def reduce_data(x: Union[int, float, torch.Tensor], dst: int) -> Union[int, float, torch.Tensor]:
"""
Overview:
Reduce the tensor ``x`` to the destination process ``dst``
Arguments:
- x (:obj:`Union[int, float, torch.Tensor]`): the tensor to be reduced
- dst (:obj:`int`): the destination process
"""
if np.isscalar(x):
x_tensor = torch.as_tensor([x]).cuda()
dist.reduce(x_tensor, dst)
return x_tensor.item()
elif isinstance(x, torch.Tensor):
dist.reduce(x, dst)
return x
else:
raise TypeError("not supported type: {}".format(type(x)))
def allreduce_data(x: Union[int, float, torch.Tensor], op: str) -> Union[int, float, torch.Tensor]:
"""
Overview:
All reduce the tensor ``x`` in the world
Arguments:
- x (:obj:`Union[int, float, torch.Tensor]`): the tensor to be reduced
- op (:obj:`str`): the operation to perform on data, support ``['sum', 'avg']``
"""
assert op in ['sum', 'avg'], op
if np.isscalar(x):
x_tensor = torch.as_tensor([x]).cuda()
dist.all_reduce(x_tensor)
if op == 'avg':
x_tensor.div_(get_world_size())
return x_tensor.item()
elif isinstance(x, torch.Tensor):
dist.all_reduce(x)
if op == 'avg':
x.div_(get_world_size())
return x
else:
raise TypeError("not supported type: {}".format(type(x)))
synchronize = torch.cuda.synchronize
def get_group(group_size: int) -> List:
"""
Overview:
Get the group segmentation of ``group_size`` each group
Arguments:
- group_size (:obj:`int`) the ``group_size``
"""
rank = get_rank()
world_size = get_world_size()
if group_size is None:
group_size = world_size
assert (world_size % group_size == 0)
return simple_group_split(world_size, rank, world_size // group_size)
def dist_mode(func: Callable) -> Callable:
"""
Overview:
Wrap the function so that in can init and finalize automatically before each call
Arguments:
- func (:obj:`Callable`): the function to be wrapped
"""
def wrapper(*args, **kwargs):
dist_init()
func(*args, **kwargs)
dist_finalize()
return wrapper
def dist_init(backend: str = 'nccl',
addr: str = None,
port: str = None,
rank: int = None,
world_size: int = None) -> Tuple[int, int]:
"""
Overview:
Initialize the distributed training setting
Arguments:
- backend (:obj:`str`): The backend of the distributed training, support ``['nccl', 'gloo']``
- addr (:obj:`str`): The address of the master node
- port (:obj:`str`): The port of the master node
- rank (:obj:`int`): The rank of current process
- world_size (:obj:`int`): The total number of processes
"""
assert backend in ['nccl', 'gloo'], backend
os.environ['MASTER_ADDR'] = addr or os.environ.get('MASTER_ADDR', "localhost")
os.environ['MASTER_PORT'] = port or os.environ.get('MASTER_PORT', "10314") # hard-code
if rank is None:
local_id = os.environ.get('SLURM_LOCALID', os.environ.get('RANK', None))
if local_id is None:
raise RuntimeError("please indicate rank explicitly in dist_init method")
else:
rank = int(local_id)
if world_size is None:
ntasks = os.environ.get('SLURM_NTASKS', os.environ.get('WORLD_SIZE', None))
if ntasks is None:
raise RuntimeError("please indicate world_size explicitly in dist_init method")
else:
world_size = int(ntasks)
dist.init_process_group(backend=backend, rank=rank, world_size=world_size)
num_gpus = torch.cuda.device_count()
torch.cuda.set_device(rank % num_gpus)
world_size = get_world_size()
rank = get_rank()
return rank, world_size
def dist_finalize() -> None:
"""
Overview:
Finalize distributed training resources
"""
# This operation usually hangs out so we ignore it temporally.
# dist.destroy_process_group()
pass
class DDPContext:
"""
Overview:
A context manager for ``linklink`` distribution
Interfaces:
``__init__``, ``__enter__``, ``__exit__``
"""
def __init__(self) -> None:
"""
Overview:
Initialize the ``DDPContext``
"""
pass
def __enter__(self) -> None:
"""
Overview:
Initialize ``linklink`` distribution
"""
dist_init()
def __exit__(self, *args, **kwargs) -> Any:
"""
Overview:
Finalize ``linklink`` distribution
"""
dist_finalize()
def simple_group_split(world_size: int, rank: int, num_groups: int) -> List:
"""
Overview:
Split the group according to ``worldsize``, ``rank`` and ``num_groups``
Arguments:
- world_size (:obj:`int`): The world size
- rank (:obj:`int`): The rank
- num_groups (:obj:`int`): The number of groups
.. note::
With faulty input, raise ``array split does not result in an equal division``
"""
groups = []
rank_list = np.split(np.arange(world_size), num_groups)
rank_list = [list(map(int, x)) for x in rank_list]
for i in range(num_groups):
groups.append(dist.new_group(rank_list[i]))
group_size = world_size // num_groups
return groups[rank // group_size]
def to_ddp_config(cfg: EasyDict) -> EasyDict:
"""
Overview:
Convert the config to ddp config
Arguments:
- cfg (:obj:`EasyDict`): The config to be converted
"""
w = get_world_size()
if 'batch_size' in cfg.policy:
cfg.policy.batch_size = int(np.ceil(cfg.policy.batch_size / w))
if 'batch_size' in cfg.policy.learn:
cfg.policy.learn.batch_size = int(np.ceil(cfg.policy.learn.batch_size / w))
if 'n_sample' in cfg.policy.collect:
cfg.policy.collect.n_sample = int(np.ceil(cfg.policy.collect.n_sample / w))
if 'n_episode' in cfg.policy.collect:
cfg.policy.collect.n_episode = int(np.ceil(cfg.policy.collect.n_episode / w))
return cfg
|