问题：Cannot assign a device for operation Variable

wzg2016

关注

原创最新推荐文章于 2026-09-04 16:21:14 发布 4.3k 阅读 AI 写同款· GEO 优化 ›

TensorFlow图模式性能优化实战：从原理到生产落地 TensorFlow图模式是一种基于静态计算图的执行范式，其核心原理是将Python张量运算编译为优化后的CUDA内核，通过OP融合、常量折叠和内存复用等编译器级优化显著降低GPU显存访问与调度开销。相比eager模式，图模式在计算密集型场景下可提升吞吐量2–4倍，并将GPU利用率推至98%以上，具备明确的技术价值：稳（服务SLA保障）、快（P99延迟压降）、省（云资源成本削减）。典型应用场景包括实时推荐排序、边缘端T4推理、长周期训练任务及高并发线上服务。本文聚焦TensorFlow图模式与性能提升两大热阅读详情

报错代码：

import tensorflow as tf
import numpy as np
import os
os.environ["CUDA_DEVICES_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1"

with tf.device("/gpu:0"): ## 与下面注释的命令等价
# with tf.device("/job:localhost/replica:0/task:0/device:GPU:0"):
    a = tf.Variable(0,dtype=tf.float32)

with tf.device("/gpu:1"):
    b = tf.constant(0, shape=[1], dtype=tf.float32)

config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.log_device_placement = True
with tf.Session(config=config) as sess:
    sess.run(tf.global_variables_initializer())
    c = sess.run(a)
    b_ = sess.run(b)
    print(b_,c)

报错：

tensorflow.python.framework.errors_impl.InvalidArgumentError: Cannot assign a device for operation Variable: node Variable (defined at /wzg_code/×××/multi_gpu_test.py:19) was explicitly assigned to /device:GPU:1 but available devices are [ /job:localhost/replica:0/task:0/device:CPU:0, /job:localhost/replica:0/task:0/device:XLA_CPU:0, /job:localhost/replica:0/task:0/device:XLA_GPU:0, /job:localhost/replica:0/task:0/device:XLA_GPU:1 ]. Make sure the device specification refers to a valid device.
	 [[Variable]]

当时考虑了将 tf.device("gpu:0") 的位置改成： tf.device(“/job:localhost/replica:0/task:0/device:XLA_GPU:0”），仍然报错，最后发现需要在sess的配置上进行修改，增加 config.allow_soft_placement=True,虽然不知道为什么，但是work了。更正后的代码如下：

import tensorflow as tf
import numpy as np
import os
os.environ["CUDA_DEVICES_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1"

with tf.device("/gpu:0"):
# with tf.device("/job:localhost/replica:0/task:0/device:GPU:0"):
    a = tf.Variable(0,dtype=tf.float32)

with tf.device("/gpu:1"):
    b = tf.constant(0, shape=[1], dtype=tf.float32)


#config = tf.ConfigProto()
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
config.log_device_placement = True
with tf.Session(config=config) as sess:
    sess.run(tf.global_variables_initializer())
    c = sess.run(a)
    b_ = sess.run(b)
    print(b_,c)

收录于