File size: 7,836 Bytes
b84549f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 |
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
# This file mainly comes from
# https://github.com/facebookresearch/detectron2/blob/master/detectron2/utils/comm.py
# Copyright (c) Facebook, Inc. and its affiliates.
# Copyright (c) Megvii Inc. All rights reserved.
"""
This file contains primitives for multi-gpu communication.
This is useful when doing distributed training.
"""
import functools
import os
import pickle
import time
from contextlib import contextmanager
from loguru import logger
import numpy as np
import torch
from torch import distributed as dist
__all__ = [
"get_num_devices",
"wait_for_the_master",
"is_main_process",
"synchronize",
"get_world_size",
"get_rank",
"get_local_rank",
"get_local_size",
"time_synchronized",
"gather",
"all_gather",
]
_LOCAL_PROCESS_GROUP = None
def get_num_devices():
gpu_list = os.getenv('CUDA_VISIBLE_DEVICES', None)
if gpu_list is not None:
return len(gpu_list.split(','))
else:
devices_list_info = os.popen("nvidia-smi -L")
devices_list_info = devices_list_info.read().strip().split("\n")
return len(devices_list_info)
@contextmanager
def wait_for_the_master(local_rank: int):
"""
Make all processes waiting for the master to do some task.
"""
if local_rank > 0:
dist.barrier()
yield
if local_rank == 0:
if not dist.is_available():
return
if not dist.is_initialized():
return
else:
dist.barrier()
def synchronize():
"""
Helper function to synchronize (barrier) among all processes when using distributed training
"""
if not dist.is_available():
return
if not dist.is_initialized():
return
world_size = dist.get_world_size()
if world_size == 1:
return
dist.barrier()
def get_world_size() -> int:
if not dist.is_available():
return 1
if not dist.is_initialized():
return 1
return dist.get_world_size()
def get_rank() -> int:
if not dist.is_available():
return 0
if not dist.is_initialized():
return 0
return dist.get_rank()
def get_local_rank() -> int:
"""
Returns:
The rank of the current process within the local (per-machine) process group.
"""
if _LOCAL_PROCESS_GROUP is None:
return get_rank()
if not dist.is_available():
return 0
if not dist.is_initialized():
return 0
return dist.get_rank(group=_LOCAL_PROCESS_GROUP)
def get_local_size() -> int:
"""
Returns:
The size of the per-machine process group, i.e. the number of processes per machine.
"""
if not dist.is_available():
return 1
if not dist.is_initialized():
return 1
return dist.get_world_size(group=_LOCAL_PROCESS_GROUP)
def is_main_process() -> bool:
return get_rank() == 0
@functools.lru_cache()
def _get_global_gloo_group():
"""
Return a process group based on gloo backend, containing all the ranks
The result is cached.
"""
if dist.get_backend() == "nccl":
return dist.new_group(backend="gloo")
else:
return dist.group.WORLD
def _serialize_to_tensor(data, group):
backend = dist.get_backend(group)
assert backend in ["gloo", "nccl"]
device = torch.device("cpu" if backend == "gloo" else "cuda")
buffer = pickle.dumps(data)
if len(buffer) > 1024 ** 3:
logger.warning(
"Rank {} trying to all-gather {:.2f} GB of data on device {}".format(
get_rank(), len(buffer) / (1024 ** 3), device
)
)
storage = torch.ByteStorage.from_buffer(buffer)
tensor = torch.ByteTensor(storage).to(device=device)
return tensor
def _pad_to_largest_tensor(tensor, group):
"""
Returns:
list[int]: size of the tensor, on each rank
Tensor: padded tensor that has the max size
"""
world_size = dist.get_world_size(group=group)
assert (
world_size >= 1
), "comm.gather/all_gather must be called from ranks within the given group!"
local_size = torch.tensor([tensor.numel()], dtype=torch.int64, device=tensor.device)
size_list = [
torch.zeros([1], dtype=torch.int64, device=tensor.device)
for _ in range(world_size)
]
dist.all_gather(size_list, local_size, group=group)
size_list = [int(size.item()) for size in size_list]
max_size = max(size_list)
# we pad the tensor because torch all_gather does not support
# gathering tensors of different shapes
if local_size != max_size:
padding = torch.zeros(
(max_size - local_size,), dtype=torch.uint8, device=tensor.device
)
tensor = torch.cat((tensor, padding), dim=0)
return size_list, tensor
def all_gather(data, group=None):
"""
Run all_gather on arbitrary picklable data (not necessarily tensors).
Args:
data: any picklable object
group: a torch process group. By default, will use a group which
contains all ranks on gloo backend.
Returns:
list[data]: list of data gathered from each rank
"""
if get_world_size() == 1:
return [data]
if group is None:
group = _get_global_gloo_group()
if dist.get_world_size(group) == 1:
return [data]
tensor = _serialize_to_tensor(data, group)
size_list, tensor = _pad_to_largest_tensor(tensor, group)
max_size = max(size_list)
# receiving Tensor from all ranks
tensor_list = [
torch.empty((max_size,), dtype=torch.uint8, device=tensor.device)
for _ in size_list
]
dist.all_gather(tensor_list, tensor, group=group)
data_list = []
for size, tensor in zip(size_list, tensor_list):
buffer = tensor.cpu().numpy().tobytes()[:size]
data_list.append(pickle.loads(buffer))
return data_list
def gather(data, dst=0, group=None):
"""
Run gather on arbitrary picklable data (not necessarily tensors).
Args:
data: any picklable object
dst (int): destination rank
group: a torch process group. By default, will use a group which
contains all ranks on gloo backend.
Returns:
list[data]: on dst, a list of data gathered from each rank. Otherwise,
an empty list.
"""
if get_world_size() == 1:
return [data]
if group is None:
group = _get_global_gloo_group()
if dist.get_world_size(group=group) == 1:
return [data]
rank = dist.get_rank(group=group)
tensor = _serialize_to_tensor(data, group)
size_list, tensor = _pad_to_largest_tensor(tensor, group)
# receiving Tensor from all ranks
if rank == dst:
max_size = max(size_list)
tensor_list = [
torch.empty((max_size,), dtype=torch.uint8, device=tensor.device)
for _ in size_list
]
dist.gather(tensor, tensor_list, dst=dst, group=group)
data_list = []
for size, tensor in zip(size_list, tensor_list):
buffer = tensor.cpu().numpy().tobytes()[:size]
data_list.append(pickle.loads(buffer))
return data_list
else:
dist.gather(tensor, [], dst=dst, group=group)
return []
def shared_random_seed():
"""
Returns:
int: a random number that is the same across all workers.
If workers need a shared RNG, they can use this shared seed to
create one.
All workers must call this function, otherwise it will deadlock.
"""
ints = np.random.randint(2 ** 31)
all_ints = all_gather(ints)
return all_ints[0]
def time_synchronized():
"""pytorch-accurate time"""
if torch.cuda.is_available():
torch.cuda.synchronize()
return time.time()
|