使用 GPU

##### Copyright 2018 The TensorFlow Authors.

#@title 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
#
# https://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.

%cd ..
from set_env import temp_dir

/media/pc/data/lxw/ai/d2py/doc/libs/tf-chaos/guide

使用 GPU#

在 TensorFlow.org 上查看 在 Google Colab 中运行 在 GitHub 上查看源代码 下载笔记本

无需更改任何代码，TensorFlow 代码以及 tf.keras 模型就可以在单个 GPU 上透明运行。

注：使用 tf.config.list_physical_devices('GPU') 可以确认 TensorFlow 使用的是 GPU。

在一台或多台机器上，要顺利地在多个 GPU 上运行，最简单的方法是使用分布策略。

本指南适用于已尝试这些方法，但发现需要对 TensorFlow 使用 GPU 的方式进行精细控制的用户。要了解如何为单 GPU 和多 GPU 情景调试性能问题，请参阅优化 TensorFlow GPU 性能指南。

设置#

确保已安装最新的 TensorFlow GPU 版本。

import tensorflow as tf
print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))

Num GPUs Available:  2

概述#

TensorFlow 支持在各种类型的设备上执行计算，包括 CPU 和 GPU。我们使用字符串标识符来表示这些设备，例如：

  • "/device:CPU:0"：机器的 CPU。

  • "/GPU:0"：TensorFlow 可见的机器上第一个 GPU 的速记表示法。

  • "/job:localhost/replica:0/task:0/device:GPU:1"：TensorFlow 可见的机器上第二个 GPU 的完全限定名称。

如果一个 TensorFlow 运算同时有 CPU 和 GPU 实现，则在默认情况下，分配运算时会优先使用 GPU 设备。例如，tf.matmul 同时有 CPU 和 GPU 内核，在具有 CPU:0 和 GPU:0 设备的系统上，将选择 GPU:0 设备来运行 tf.matmul，除非明确要求在另一个设备上运行。

如果一个 TensorFlow 运算没有相应的 GPU 实现，则该运算将回退到 CPU 设备。例如，由于 tf.cast 只有一个 CPU 内核，在具有 CPU:0 和 GPU:0 设备的系统上，即使请求在 GPU:0 设备上运行 tf.cast，也会选择 CPU:0 设备来运行该运算。

记录设备放置#

为了找出将运算和张量分配到的目标设备，请将 tf.debugging.set_log_device_placement(True) 放在程序的第一行。启用设备放置记录将导致任何张量分配或运算被打印。

tf.debugging.set_log_device_placement(True)

# Create some tensors
a = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
b = tf.constant([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])
c = tf.matmul(a, b)

print(c)

input: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
_EagerConst: (_EagerConst): /job:localhost/replica:0/task:0/device:GPU:0
output_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
a: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
b: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
MatMul: (MatMul): /job:localhost/replica:0/task:0/device:GPU:0
product_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op MatMul in device /job:localhost/replica:0/task:0/device:GPU:0
tf.Tensor(
[[22. 28.]
 [49. 64.]], shape=(2, 2), dtype=float32)

以上代码将打印 MatMul 运算在 GPU:0 上执行的指示。

手动设备放置#

如果您希望在自己选择的设备上执行特定运算，而不是在自动选择的设备上执行，则可以使用 with tf.device 创建设备上下文。创建完成后，该上下文中的所有运算都会在同一指定设备上运行。

tf.debugging.set_log_device_placement(True)

# Place tensors on the CPU
with tf.device('/CPU:0'):
  a = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
  b = tf.constant([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])

c = tf.matmul(a, b)
print(c)

Executing op MatMul in device /job:localhost/replica:0/task:0/device:GPU:0
tf.Tensor(
[[22. 28.]
 [49. 64.]], shape=(2, 2), dtype=float32)

现在，您会看到已将 a 和 b 分配给 CPU:0。由于没有为 MatMul 运算明确指定设备，TensorFlow 运行时将根据运算和可用的设备选择一个设备（本例中为 GPU:0），并且在需要时会自动在设备之间复制张量。

限制 GPU 内存增长#

默认情况下，TensorFlow 会映射进程可见的所有 GPU（取决于 CUDA_VISIBLE_DEVICES）的几乎全部内存。这是为了减少内存碎片，更有效地利用设备上相对宝贵的 GPU 内存资源。为了将 TensorFlow 限制为使用一组特定的 GPU，我们使用 tf.config.set_visible_devices 方法。

gpus = tf.config.list_physical_devices('GPU')
if gpus:
  # Restrict TensorFlow to only use the first GPU
  try:
    tf.config.set_visible_devices(gpus[0], 'GPU')
    logical_gpus = tf.config.list_logical_devices('GPU')
    print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPU")
  except RuntimeError as e:
    # Visible devices must be set before GPUs have been initialized
    print(e)

Visible devices cannot be modified after being initialized

在某些情况下，我们希望进程最好只分配可用内存的一个子集，或者仅在进程需要时才增加内存使用量。TensorFlow 为此提供了两种控制方法。

第一个选项是通过调用 tf.config.experimental.set_memory_growth 来开启内存增长。此选项会尝试根据运行时分配的需求分配尽可能充足的 GPU 内存：首先分配非常少的内存，随着程序的运行，需要的 GPU 内存逐渐增多，于是扩展 TensorFlow 进程的 GPU 内存区域。内存不会被释放，因为这样会产生内存碎片。要关闭特定 GPU 的内存增长，请在分配任何张量或执行任何运算之前使用以下代码。

gpus = tf.config.list_physical_devices('GPU')
if gpus:
  try:
    # Currently, memory growth needs to be the same across GPUs
    for gpu in gpus:
      tf.config.experimental.set_memory_growth(gpu, True)
    logical_gpus = tf.config.list_logical_devices('GPU')
    print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
  except RuntimeError as e:
    # Memory growth must be set before GPUs have been initialized
    print(e)

Physical devices cannot be modified after being initialized

第二个启用此选项的方式是将环境变量 TF_FORCE_GPU_ALLOW_GROWTH 设置为 true。这是一个特定于平台的配置。

第二种方法是使用 tf.config.set_logical_device_configuration 配置虚拟 GPU 设备，并且设置可在 GPU 上分配多少总内存的硬性限制。

gpus = tf.config.list_physical_devices('GPU')
if gpus:
  # Restrict TensorFlow to only allocate 1GB of memory on the first GPU
  try:
    tf.config.set_logical_device_configuration(
        gpus[0],
        [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1024)])
    logical_gpus = tf.config.list_logical_devices('GPU')
    print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
  except RuntimeError as e:
    # Virtual devices must be set before GPUs have been initialized
    print(e)

Virtual devices cannot be modified after being initialized

这在要真正限制可供 TensorFlow 进程使用的 GPU 内存量时非常有用。在本地开发中，与其他应用（如工作站 GUI）共享 GPU 时，这是常见做法。

使用多 GPU 系统上的单个 GPU#

如果系统上有多个 GPU，则默认情况下会选择具有最小 ID 的 GPU。如果希望在不同的 GPU 上运行，则需要明确指定需要的 GPU：

tf.debugging.set_log_device_placement(True)

try:
  # Specify an invalid GPU device
  with tf.device('/device:GPU:2'):
    a = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
    b = tf.constant([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])
    c = tf.matmul(a, b)
except RuntimeError as e:
  print(e)

Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op MatMul in device /job:localhost/replica:0/task:0/device:GPU:0

如果指定的设备不存在，则会引发 RuntimeError 错误：.../device:GPU:2 unknown device。

当指定的设备不存在时，如果希望 TensorFlow 自动选择存在且支持的设备来执行运算，可以调用 tf.config.set_soft_device_placement(True)。

tf.config.set_soft_device_placement(True)
tf.debugging.set_log_device_placement(True)

# Creates some tensors
a = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
b = tf.constant([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])
c = tf.matmul(a, b)

print(c)

Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op MatMul in device /job:localhost/replica:0/task:0/device:GPU:0
tf.Tensor(
[[22. 28.]
 [49. 64.]], shape=(2, 2), dtype=float32)

使用多个 GPU#

为多个 GPU 开发的模型可使用额外的资源进行扩展。如果在具有单个 GPU 的系统上进行开发，可以使用虚拟设备模拟多个 GPU。这样，无需额外的资源，您就可以轻松对多 GPU 设置进行测试。

gpus = tf.config.list_physical_devices('GPU')
if gpus:
  # Create 2 virtual GPUs with 1GB memory each
  try:
    tf.config.set_logical_device_configuration(
        gpus[0],
        [tf.config.LogicalDeviceConfiguration(memory_limit=1024),
         tf.config.LogicalDeviceConfiguration(memory_limit=1024)])
    logical_gpus = tf.config.list_logical_devices('GPU')
    print(len(gpus), "Physical GPU,", len(logical_gpus), "Logical GPUs")
  except RuntimeError as e:
    # Virtual devices must be set before GPUs have been initialized
    print(e)

Virtual devices cannot be modified after being initialized

建立可供运行时使用的多个逻辑 GPU 后，可以通过 tf.distribute.Strategy 或手动放置来利用多个 GPU。

使用 tf.distribute.Strategy#

使用多个 GPU 的最佳做法是使用 tf.distribute.Strategy。下面是一个简单示例：

tf.debugging.set_log_device_placement(True)
gpus = tf.config.list_logical_devices('GPU')
strategy = tf.distribute.MirroredStrategy(gpus)
with strategy.scope():
  inputs = tf.keras.layers.Input(shape=(1,))
  predictions = tf.keras.layers.Dense(1)(inputs)
  model = tf.keras.models.Model(inputs=inputs, outputs=predictions)
  model.compile(loss='mse',
                optimizer=tf.keras.optimizers.SGD(learning_rate=0.2))

Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:0
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
value: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
AssignVariableOp: (AssignVariableOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
input: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
_EagerConst: (_EagerConst): /job:localhost/replica:0/task:0/device:GPU:1
output_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:1
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:1
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
value: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
AssignVariableOp: (AssignVariableOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:1
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
input: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
_EagerConst: (_EagerConst): /job:localhost/replica:0/task:0/device:GPU:0
output_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
x: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
y: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
FloorMod: (FloorMod): /job:localhost/replica:0/task:0/device:GPU:0
z_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op FloorMod in device /job:localhost/replica:0/task:0/device:GPU:0
x: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
Cast: (Cast): /job:localhost/replica:0/task:0/device:GPU:0
y_RetVal: (_DeviceRetval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op Cast in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
input: (_Arg): /job:localhost/replica:0/task:0/device:CPU:0
_EagerConst: (_EagerConst): /job:localhost/replica:0/task:0/device:GPU:0
output_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
seed: (_Arg): /job:localhost/replica:0/task:0/device:CPU:0
StatelessRandomGetKeyCounter: (StatelessRandomGetKeyCounter): /job:localhost/replica:0/task:0/device:GPU:0
key_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
counter_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op StatelessRandomGetKeyCounter in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
shape: (_DeviceArg): /job:localhost/replica:0/task:0/device:CPU:0
key: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
counter: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
alg: (_DeviceArg): /job:localhost/replica:0/task:0/device:CPU:0
StatelessRandomUniformV2: (StatelessRandomUniformV2): /job:localhost/replica:0/task:0/device:GPU:0
output_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op StatelessRandomUniformV2 in device /job:localhost/replica:0/task:0/device:GPU:0
x: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
y: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
Sub: (Sub): /job:localhost/replica:0/task:0/device:GPU:0
z_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op Sub in device /job:localhost/replica:0/task:0/device:GPU:0
x: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
y: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
Mul: (Mul): /job:localhost/replica:0/task:0/device:GPU:0
z_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op Mul in device /job:localhost/replica:0/task:0/device:GPU:0
x: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
y: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
AddV2: (AddV2): /job:localhost/replica:0/task:0/device:GPU:0
z_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op AddV2 in device /job:localhost/replica:0/task:0/device:GPU:0
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:0
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
value: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
AssignVariableOp: (AssignVariableOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
ReadVariableOp: (ReadVariableOp): /job:localhost/replica:0/task:0/device:GPU:0
value_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op ReadVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
input: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
Identity: (Identity): /job:localhost/replica:0/task:0/device:GPU:1
output_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
Executing op Identity in device /job:localhost/replica:0/task:0/device:GPU:1
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:1
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
value: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
AssignVariableOp: (AssignVariableOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:1
NoOp: (NoOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
dims: (_DeviceArg): /job:localhost/replica:0/task:0/device:CPU:0
value: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
Fill: (Fill): /job:localhost/replica:0/task:0/device:GPU:0
output_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op Fill in device /job:localhost/replica:0/task:0/device:GPU:0
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:0
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
value: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
AssignVariableOp: (AssignVariableOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
ReadVariableOp: (ReadVariableOp): /job:localhost/replica:0/task:0/device:GPU:0
value_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op ReadVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op Identity in device /job:localhost/replica:0/task:0/device:GPU:1
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:1
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
value: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
AssignVariableOp: (AssignVariableOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:1
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:0
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
value: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
AssignVariableOp: (AssignVariableOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
ReadVariableOp: (ReadVariableOp): /job:localhost/replica:0/task:0/device:GPU:0
value_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op ReadVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
input: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
Identity: (Identity): /job:localhost/replica:0/task:0/device:GPU:1
output_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
Executing op Identity in device /job:localhost/replica:0/task:0/device:GPU:1
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:1
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
value: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
AssignVariableOp: (AssignVariableOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:1
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
resource: (_Arg): /job:localhost/replica:0/task:0/device:GPU:0
ReadVariableOp: (ReadVariableOp): /job:localhost/replica:0/task:0/device:GPU:0
value_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
Executing op ReadVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op Identity in device /job:localhost/replica:0/task:0/device:GPU:1
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:1
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:1
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op Fill in device /job:localhost/replica:0/task:0/device:GPU:0
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op ReadVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op Identity in device /job:localhost/replica:0/task:0/device:GPU:1
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:1
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:1
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op Fill in device /job:localhost/replica:0/task:0/device:GPU:0
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:0
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op ReadVariableOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op Identity in device /job:localhost/replica:0/task:0/device:GPU:1
resource_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
VarHandleOp: (VarHandleOp): /job:localhost/replica:0/task:0/device:GPU:1
Executing op VarHandleOp in device /job:localhost/replica:0/task:0/device:GPU:1
Executing op AssignVariableOp in device /job:localhost/replica:0/task:0/device:GPU:1
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op NoOp in device /job:localhost/replica:0/task:0/device:GPU:0

此程序会在每个 GPU 上运行模型的一个副本，并将输入数据拆分到每个 GPU 上，也就是所谓的“数据并行”。

有关分布策略的详细信息，请查阅此处的指南。

手动放置#

tf.distribute.Strategy 通过跨设备复制计算在后台运行。您可以通过在每个 GPU 上构建模型来手动实现复制。例如：

tf.debugging.set_log_device_placement(True)

gpus = tf.config.list_logical_devices('GPU')
if gpus:
  # Replicate your computation on multiple GPUs
  c = []
  for gpu in gpus:
    with tf.device(gpu.name):
      a = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
      b = tf.constant([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])
      c.append(tf.matmul(a, b))

  with tf.device('/CPU:0'):
    matmul_sum = tf.add_n(c)

  print(matmul_sum)

Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op MatMul in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:1
Executing op _EagerConst in device /job:localhost/replica:0/task:0/device:GPU:1
a: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
b: (_Arg): /job:localhost/replica:0/task:0/device:GPU:1
MatMul: (MatMul): /job:localhost/replica:0/task:0/device:GPU:1
product_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:1
Executing op MatMul in device /job:localhost/replica:0/task:0/device:GPU:1
inputs_0: (_Arg): /job:localhost/replica:0/task:0/device:CPU:0
inputs_1: (_Arg): /job:localhost/replica:0/task:0/device:CPU:0
AddN: (AddN): /job:localhost/replica:0/task:0/device:CPU:0
sum_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:CPU:0
Executing op AddN in device /job:localhost/replica:0/task:0/device:CPU:0
tf.Tensor(
[[ 44.  56.]
 [ 98. 128.]], shape=(2, 2), dtype=float32)