Tensorflow 不使用 GPU，发现 xla_gpu 不是 gpu

Question

我刚开始探索 AI，从未使用过 Tensorflow，即使 Linux 对我来说也是新手。

我之前安装了 NVIDIA 驱动程序 430。它带有 CUDA 10.1

由于 Tensorflow-gpu 1.14 不支持 CUDA 10.1，我卸载了 CUDA 10.1 并下载了 CUDA 10.0

cuda_10.0.130_410.48_linux.run

安装后我就跑了

nvcc --version nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2018 NVIDIA Corporation Built on Sat_Aug_25_21:08:01_CDT_2018 Cuda compilation tools, release 10.0, V10.0.130 当我尝试在 Jupyter Notebook 中使用 GPU 时，代码仍然不起作用

import tensorflow as tf
with tf.device('/gpu:0'):
    a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')
    b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')
    c = tf.matmul(a, b)

with tf.Session() as sess:
    print (sess.run(c))

错误：

---------------------------------------------------------------------------
InvalidArgumentError                      Traceback (most recent call last)
~/anaconda3/lib/python3.7/site-packages/tensorflow/python/client/session.py in _do_call(self, fn, *args)
   1355     try:
-> 1356       return fn(*args)
   1357     except errors.OpError as e:

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/client/session.py in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata)
   1338       # Ensure any changes to the graph are reflected in the runtime.
-> 1339       self._extend_graph()
   1340       return self._call_tf_sessionrun(

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/client/session.py in _extend_graph(self)
   1373     with self._graph._session_run_lock():  # pylint: disable=protected-access
-> 1374       tf_session.ExtendSession(self._session)
   1375 

InvalidArgumentError: Cannot assign a device for operation MatMul: {{node MatMul}}was explicitly assigned to /device:GPU:0 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 ]. Make sure the device specification refers to a valid device.
     [[MatMul]]

During handling of the above exception, another exception occurred:

InvalidArgumentError                      Traceback (most recent call last)
<ipython-input-19-3a5be606bcc9> in <module>
      6 
      7 with tf.Session() as sess:
----> 8     print (sess.run(c))

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata)
    948     try:
    949       result = self._run(None, fetches, feed_dict, options_ptr,
--> 950                          run_metadata_ptr)
    951       if run_metadata:
    952         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)
   1171     if final_fetches or final_targets or (handle and feed_dict_tensor):
   1172       results = self._do_run(handle, final_targets, final_fetches,
-> 1173                              feed_dict_tensor, options, run_metadata)
   1174     else:
   1175       results = []

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/client/session.py in _do_run(self, handle, target_list, fetch_list, feed_dict, options, run_metadata)
   1348     if handle is None:
   1349       return self._do_call(_run_fn, feeds, fetches, targets, options,
-> 1350                            run_metadata)
   1351     else:
   1352       return self._do_call(_prun_fn, handle, feeds, fetches)

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/client/session.py in _do_call(self, fn, *args)
   1368           pass
   1369       message = error_interpolation.interpolate(message, self._graph)
-> 1370       raise type(e)(node_def, op, message)
   1371 
   1372   def _extend_graph(self):

InvalidArgumentError: Cannot assign a device for operation MatMul: node MatMul (defined at <ipython-input-9-b145a02709f7>:5) was explicitly assigned to /device:GPU:0 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 ]. Make sure the device specification refers to a valid device.
     [[MatMul]]

Errors may have originated from an input operation.
Input Source operations connected to node MatMul:
 b (defined at <ipython-input-9-b145a02709f7>:4)    
 a (defined at <ipython-input-9-b145a02709f7>:3)

但是，如果我从终端用 Python 运行这段代码，它就可以工作。我可以看到输出

[[22. 28.] [49. 64.]]

Answer 1

您需要确保安装了适当的CUDA 和CuDNN 版本。

• 您可以通过以下链接的建议检查您的CuDNN 版本：How to verify CuDNN installation?
  • 或在 linux 机器上运行 cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2
• 您可以在此处查看CUDA 版本：xcat.docs
  • nvcc -V
  • 或通过运行nvidia-smi
• 在此处阅读xla_gpus：tensorflow xla 和此处：github xla_gpu issue
  • xla 由 tensorflow 制作，比标准 tensorflow 更快。
  • 我不确定为什么没有CuDNN 的CUDA 会调用gpus xla_gpus。 Nvidia gpus 需要 CUDA 和 CuDNN 才能与 Tensorflow 正常工作，因此看起来 tensorflow 正在尝试使用自己的库在 GPU 上进行计算。但是，我不太确定。