File size: 6,851 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 |
# 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.
"""Tracks skip tensors on a thread."""
from contextlib import contextmanager
import threading
from typing import Dict, Generator, List, Optional, Tuple
from torch import Tensor
from ..checkpoint import is_checkpointing
from ..dependency import fork, join
from ..microbatch import Batch
from ..stream import AbstractStream
from .layout import SkipLayout
from .namespace import Namespace
from .portal import Portal
__all__: List[str] = []
class SkipTracker:
"""Tracks saved skip tensors.
It will update the given micro-batch in place. This is because when it
manipulates the underlying skip tensors, the current micro-batch also has
to be connected with the skip tensors.
One thread has one skip tracker. Call :func:`current_skip_tracker` to get
the skip tracker on the current thread.
"""
def __init__(self) -> None:
self.tensors: Dict[Tuple[Namespace, str], Optional[Tensor]] = {}
def save(self, batch: Batch, ns: Namespace, name: str, tensor: Optional[Tensor]) -> None:
self.tensors[(ns, name)] = tensor
def load(self, batch: Batch, ns: Namespace, name: str) -> Optional[Tensor]:
return self.tensors.pop((ns, name))
def copy(
self,
batch: Batch,
prev_stream: AbstractStream,
next_stream: AbstractStream,
ns: Namespace,
name: str,
) -> None:
raise TypeError("copy is not supported for non-portal skip tensors")
@property
def index(self) -> int:
return 0
class SkipTrackerThroughPotals(SkipTracker):
"""Tracks saved skip tensors through portals. The skip tensors will be
hidden in portals so that the autograd engine does not need to track them.
This tracker is only used when the training or evaluating module is wrapped
with :class:`torchpipe.Pipe`.
"""
def __init__(self, skip_layout: SkipLayout, index: int) -> None:
super().__init__()
self.skip_layout = skip_layout
self.portals: Dict[Tuple[Namespace, str], Portal] = {}
self.__index = index
@property
def index(self) -> int:
return self.__index
def save(self, batch: Batch, ns: Namespace, name: str, tensor: Optional[Tensor]) -> None:
"""Saves the stashed skip tensor in a portal. The portal is then
connected to the given micro-batch with :class:`Join`.
"""
if not self.skip_layout.requires_copy(ns, name):
super().save(batch, ns, name, tensor)
return
# See [Tensor Life of Portal] at Portal.put_tensor() to understand the
# below tensor_life values. Here are the selected events which retrieve
# the tensor in portal:
#
# 1. [x] blue()
# ...
# 6. [x] PortalOrange.forward
# ...
# 8. [x] PortalOrange.forward (recomputed)
# ...
# 11. [x] blue() (recomputed)
#
if (ns, name) not in self.portals:
if is_checkpointing():
# Under checkpointing, the tensor used by the first
# PortalOrange should be alive in the portal. This tensor will
# be used again by the second PortalOrange during the
# recomputation.
tensor_life = 3 # Delete at [8. PortalOrange.forward (recomputed)]
else:
tensor_life = 2 # Delete at [6. PortalOrange.forward]
assert batch.index == self.index
portal = Portal(tensor, tensor_life, batch.index)
portal.ns_name = (ns, name)
self.portals[(ns, name)] = portal
else:
# Under recomputation, the portal already exists.
portal = self.portals[(ns, name)]
# The existing tensor life already became 0. It should be reset as
# 1 to delete the tensor after the second PortalBlue immediately.
tensor_life = 1 # Delete at [11. blue() (recomputed)]
portal.put_tensor(tensor, tensor_life)
phony = portal.blue()
batch[0] = join(batch[0], phony)
def load(self, batch: Batch, ns: Namespace, name: str) -> Optional[Tensor]:
"""Loads a skip tensor from the corresponding portal to pop. The given
micro-batch is connected to the portal with :class:`Fork`.
"""
if not self.skip_layout.requires_copy(ns, name):
tensor = super().load(batch, ns, name)
return tensor
portal = self.portals[(ns, name)]
batch[0], phony = fork(batch[0])
tensor = portal.orange(phony)
return tensor
def copy(
self,
batch: Batch,
prev_stream: AbstractStream,
next_stream: AbstractStream,
ns: Namespace,
name: str,
) -> None:
"""Copies the skip tensor in the corresponding portal. The given
micro-batch and the portal will be tied with :class:`Fork` and
:class:`Join`.
"""
assert self.skip_layout.requires_copy(ns, name)
batch[0], phony = fork(batch[0])
portal = self.portals[(ns, name)]
phony = portal.copy(prev_stream, next_stream, phony)
batch[0] = join(batch[0], phony)
class ThreadLocal(threading.local):
def __init__(self) -> None:
self.skip_tracker: Optional[SkipTracker] = None
thread_local = ThreadLocal()
@contextmanager
def use_skip_tracker(skip_tracker: SkipTracker) -> Generator[None, None, None]:
"""Registers the given skip tracker on the current thread within a
context::
with use_skip_tracker(my_skip_tracker):
...
"""
orig = thread_local.skip_tracker
thread_local.skip_tracker = skip_tracker
try:
yield
finally:
thread_local.skip_tracker = orig
def current_skip_tracker() -> SkipTracker:
"""Gets the skip tracker on the current thread."""
skip_tracker = thread_local.skip_tracker
if skip_tracker is None:
skip_tracker = SkipTracker()
thread_local.skip_tracker = skip_tracker
return skip_tracker
|