File size: 5,165 Bytes
ac1cc14
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os

import torch
import socket

try:
    import horovod.torch as hvd
except ImportError:
    hvd = None


def is_global_master(args):
    return args.rank == 0


def is_local_master(args):
    return args.local_rank == 0


def is_master(args, local=False):
    return is_local_master(args) if local else is_global_master(args)


def is_using_horovod():
    # NOTE w/ horovod run, OMPI vars should be set, but w/ SLURM PMI vars will be set
    # Differentiating between horovod and DDP use via SLURM may not be possible, so horovod arg still required...
    ompi_vars = ["OMPI_COMM_WORLD_RANK", "OMPI_COMM_WORLD_SIZE"]
    pmi_vars = ["PMI_RANK", "PMI_SIZE"]
    if all([var in os.environ for var in ompi_vars]) or all(
        [var in os.environ for var in pmi_vars]
    ):
        return True
    else:
        return False


def is_using_distributed():
    if "WORLD_SIZE" in os.environ:
        return int(os.environ["WORLD_SIZE"]) > 1
    if "SLURM_NTASKS" in os.environ:
        return int(os.environ["SLURM_NTASKS"]) > 1
    return False


def world_info_from_env():
    local_rank = 0
    for v in (
        "SLURM_LOCALID",
        "MPI_LOCALRANKID",
        "OMPI_COMM_WORLD_LOCAL_RANK",
        "LOCAL_RANK",
    ):
        if v in os.environ:
            local_rank = int(os.environ[v])
            break
    global_rank = 0
    for v in ("SLURM_PROCID", "PMI_RANK", "OMPI_COMM_WORLD_RANK", "RANK"):
        if v in os.environ:
            global_rank = int(os.environ[v])
            break
    world_size = 1
    for v in ("SLURM_NTASKS", "PMI_SIZE", "OMPI_COMM_WORLD_SIZE", "WORLD_SIZE"):
        if v in os.environ:
            world_size = int(os.environ[v])
            break

    return local_rank, global_rank, world_size


def init_distributed_device(args):
    # Distributed training = training on more than one GPU.
    # Works in both single and multi-node scenarios.
    args.distributed = False
    args.world_size = 1
    args.rank = 0  # global rank
    args.local_rank = 0
    if args.horovod:
        assert hvd is not None, "Horovod is not installed"
        hvd.init()
        world_size = int(os.environ["OMPI_COMM_WORLD_SIZE"])
        world_rank = int(os.environ["OMPI_COMM_WORLD_RANK"])
        local_rank = int(os.environ["OMPI_COMM_WORLD_LOCAL_RANK"])
        args.local_rank = local_rank
        args.rank = world_rank
        args.world_size = world_size
        # args.local_rank = int(hvd.local_rank())
        # args.rank = hvd.rank()
        # args.world_size = hvd.size()
        args.distributed = True
        os.environ["LOCAL_RANK"] = str(args.local_rank)
        os.environ["RANK"] = str(args.rank)
        os.environ["WORLD_SIZE"] = str(args.world_size)
        print(
            f"Distributed training: local_rank={args.local_rank}, "
            f"rank={args.rank}, world_size={args.world_size}, "
            f"hostname={socket.gethostname()}, pid={os.getpid()}"
        )
    elif is_using_distributed():
        if "SLURM_PROCID" in os.environ:
            # DDP via SLURM
            args.local_rank, args.rank, args.world_size = world_info_from_env()
            # SLURM var -> torch.distributed vars in case needed
            os.environ["LOCAL_RANK"] = str(args.local_rank)
            os.environ["RANK"] = str(args.rank)
            os.environ["WORLD_SIZE"] = str(args.world_size)
            torch.distributed.init_process_group(
                backend=args.dist_backend,
                init_method=args.dist_url,
                world_size=args.world_size,
                rank=args.rank,
            )
        elif "OMPI_COMM_WORLD_SIZE" in os.environ:  # using Summit cluster
            world_size = int(os.environ["OMPI_COMM_WORLD_SIZE"])
            world_rank = int(os.environ["OMPI_COMM_WORLD_RANK"])
            local_rank = int(os.environ["OMPI_COMM_WORLD_LOCAL_RANK"])
            args.local_rank = local_rank
            args.rank = world_rank
            args.world_size = world_size
            torch.distributed.init_process_group(
                backend=args.dist_backend,
                init_method=args.dist_url,
                world_size=args.world_size,
                rank=args.rank,
            )
        else:
            # DDP via torchrun, torch.distributed.launch
            args.local_rank, _, _ = world_info_from_env()
            torch.distributed.init_process_group(
                backend=args.dist_backend, init_method=args.dist_url
            )
            args.world_size = torch.distributed.get_world_size()
            args.rank = torch.distributed.get_rank()
        args.distributed = True
        print(
            f"Distributed training: local_rank={args.local_rank}, "
            f"rank={args.rank}, world_size={args.world_size}, "
            f"hostname={socket.gethostname()}, pid={os.getpid()}"
        )

    if torch.cuda.is_available():
        if args.distributed and not args.no_set_device_rank:
            device = "cuda:%d" % args.local_rank
        else:
            device = "cuda:0"
        torch.cuda.set_device(device)
    else:
        device = "cpu"
    args.device = device
    device = torch.device(device)
    return device