File size: 5,409 Bytes
9c6594c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
#
# This source code is licensed under the BSD license found in the
# LICENSE file in the root directory of this source tree.

# Copyright 2019 Kakao Brain
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Multithreading in pipeline parallelism."""
from contextlib import contextmanager
from queue import Queue
import sys
from threading import Thread
from types import TracebackType
from typing import TYPE_CHECKING, Callable, Dict, Generator, List, Optional, Tuple, Type, Union, cast

import torch

from .microbatch import Batch
from .stream import AbstractStream, use_device, use_stream

__all__: List[str] = []


ExcInfo = Tuple[Type[BaseException], BaseException, TracebackType]

# Queue is generic only in stubs.
# https://mypy.readthedocs.io/en/latest/common_issues.html#using-classes-that-are-generic-in-stubs-but-not-at-runtime
if TYPE_CHECKING:
    InQueue = Queue[Optional["Task"]]
    OutQueue = Queue[Tuple[bool, Union[Tuple["Task", Batch], ExcInfo, None]]]
else:
    InQueue = Queue
    OutQueue = Queue


class Task:
    """A task represents how to compute a micro-batch on a partition.

    It consists of two parts: :meth:`compute` and :meth:`finalize`.
    :meth:`compute` should be executed in worker threads concurrently.
    :meth:`finalize` should be executed after when worker threads complete to
    execute :meth:`compute`.

    :meth:`compute` might be boosted by worker threads. Because it produces
    several CUDA API calls by user code. In PyTorch, parallel CUDA API calls
    are not serialized through GIL. So more than one CUDA API call can be
    produced at the same time.

    """

    def __init__(
        self,
        stream: Optional[AbstractStream],
        *,
        compute: Callable[[], Batch],
        finalize: Optional[Callable[[Batch], None]],
    ) -> None:
        self.stream = stream
        self._compute = compute
        self._finalize = finalize
        self._grad_enabled = torch.is_grad_enabled()

    def compute(self) -> Batch:
        with use_stream(self.stream), torch.set_grad_enabled(self._grad_enabled):
            return self._compute()

    def finalize(self, batch: Batch) -> None:
        if self._finalize is None:
            return
        with use_stream(self.stream), torch.set_grad_enabled(self._grad_enabled):
            self._finalize(batch)


def worker(in_queue: InQueue, out_queue: OutQueue, device: torch.device) -> None:
    """The main loop of a worker thread."""
    with use_device(device):
        while True:
            task = in_queue.get()

            if task is None:
                break

            try:
                batch = task.compute()
            except Exception:
                exc_info = cast(ExcInfo, sys.exc_info())
                out_queue.put((False, exc_info))
                continue

            out_queue.put((True, (task, batch)))

    done = (False, None)
    out_queue.put(done)


def create_workers(
    devices: List[torch.device],
) -> Tuple[List[InQueue], List[OutQueue]]:
    """Spawns worker threads. A worker thread is bound to a device."""
    in_queues: List[InQueue] = []
    out_queues: List[OutQueue] = []

    # Spawn workers.
    workers: Dict[torch.device, Tuple[InQueue, OutQueue]] = {}

    def normalize_device(device: torch.device) -> torch.device:
        if device.type == "cuda" and device.index is None:
            return torch.device("cuda", index=torch.cuda.current_device())

        if device.type == "cpu" and device.index is not None:
            return torch.device("cpu")

        return device

    for device in devices:
        device = normalize_device(device)

        try:
            in_queue, out_queue = workers[device]
        except KeyError:
            in_queue = Queue()
            out_queue = Queue()
            workers[device] = (in_queue, out_queue)

            t = Thread(
                target=worker,
                args=(in_queue, out_queue, device),
                daemon=True,
            )
            t.start()

        in_queues.append(in_queue)
        out_queues.append(out_queue)

    return (in_queues, out_queues)


def join_workers(in_queues: List[InQueue], out_queues: List[OutQueue]) -> None:
    # Close workers.
    for in_queue in set(in_queues):
        in_queue.put(None)

    # Join running workers.
    running = set(out_queues)
    while running:
        out_queue = running.pop()
        ok, payload = out_queue.get()

        done = (False, None)
        if (ok, payload) == done:
            continue

        running.add(out_queue)


@contextmanager
def spawn_workers(
    devices: List[torch.device],
) -> Generator[Tuple[List[InQueue], List[OutQueue]], None, None]:
    try:
        (in_queues, out_queues) = create_workers(devices)
        yield (in_queues, out_queues)
    finally:
        join_workers(in_queues, out_queues)