Spaces:
Paused
Paused
File size: 7,237 Bytes
3f9c56c |
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 |
# Copyright (c) Facebook, Inc. and its affiliates.
"""
This file contains primitives for multi-gpu communication.
This is useful when doing distributed training.
"""
import functools
import numpy as np
import torch
import torch.distributed as dist
_LOCAL_PROCESS_GROUP = None
_MISSING_LOCAL_PG_ERROR = (
"Local process group is not yet created! Please use detectron2's `launch()` "
"to start processes and initialize pytorch process group. If you need to start "
"processes in other ways, please call comm.create_local_process_group("
"num_workers_per_machine) after calling torch.distributed.init_process_group()."
)
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()
@functools.lru_cache()
def create_local_process_group(num_workers_per_machine: int) -> None:
"""
Create a process group that contains ranks within the same machine.
Detectron2's launch() in engine/launch.py will call this function. If you start
workers without launch(), you'll have to also call this. Otherwise utilities
like `get_local_rank()` will not work.
This function contains a barrier. All processes must call it together.
Args:
num_workers_per_machine: the number of worker processes per machine. Typically
the number of GPUs.
"""
global _LOCAL_PROCESS_GROUP
assert _LOCAL_PROCESS_GROUP is None
assert get_world_size() % num_workers_per_machine == 0
num_machines = get_world_size() // num_workers_per_machine
machine_rank = get_rank() // num_workers_per_machine
for i in range(num_machines):
ranks_on_i = list(range(i * num_workers_per_machine, (i + 1) * num_workers_per_machine))
pg = dist.new_group(ranks_on_i)
if i == machine_rank:
_LOCAL_PROCESS_GROUP = pg
def get_local_process_group():
"""
Returns:
A torch process group which only includes processes that are on the same
machine as the current process. This group can be useful for communication
within a machine, e.g. a per-machine SyncBN.
"""
assert _LOCAL_PROCESS_GROUP is not None, _MISSING_LOCAL_PG_ERROR
return _LOCAL_PROCESS_GROUP
def get_local_rank() -> int:
"""
Returns:
The rank of the current process within the local (per-machine) process group.
"""
if not dist.is_available():
return 0
if not dist.is_initialized():
return 0
assert _LOCAL_PROCESS_GROUP is not None, _MISSING_LOCAL_PG_ERROR
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
assert _LOCAL_PROCESS_GROUP is not None, _MISSING_LOCAL_PG_ERROR
return dist.get_world_size(group=_LOCAL_PROCESS_GROUP)
def is_main_process() -> bool:
return get_rank() == 0
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
if dist.get_backend() == dist.Backend.NCCL:
# This argument is needed to avoid warnings.
# It's valid only for NCCL backend.
dist.barrier(device_ids=[torch.cuda.current_device()])
else:
dist.barrier()
@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 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() # use CPU group by default, to reduce GPU RAM usage.
world_size = dist.get_world_size(group)
if world_size == 1:
return [data]
output = [None for _ in range(world_size)]
dist.all_gather_object(output, data, group=group)
return output
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()
world_size = dist.get_world_size(group=group)
if world_size == 1:
return [data]
rank = dist.get_rank(group=group)
if rank == dst:
output = [None for _ in range(world_size)]
dist.gather_object(data, output, dst=dst, group=group)
return output
else:
dist.gather_object(data, None, 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 reduce_dict(input_dict, average=True):
"""
Reduce the values in the dictionary from all processes so that process with rank
0 has the reduced results.
Args:
input_dict (dict): inputs to be reduced. All the values must be scalar CUDA Tensor.
average (bool): whether to do average or sum
Returns:
a dict with the same keys as input_dict, after reduction.
"""
world_size = get_world_size()
if world_size < 2:
return input_dict
with torch.no_grad():
names = []
values = []
# sort the keys so that they are consistent across processes
for k in sorted(input_dict.keys()):
names.append(k)
values.append(input_dict[k])
values = torch.stack(values, dim=0)
dist.reduce(values, dst=0)
if dist.get_rank() == 0 and average:
# only main process gets accumulated, so only divide by
# world_size in this case
values /= world_size
reduced_dict = {k: v for k, v in zip(names, values)}
return reduced_dict
|