Tensorflow can't use GPU. tf.test.is_gpu_available() show GPU but cannot use

Question

I have Ubuntu 18.04. Python 3.7.3, Tensorflow 2.0.0

here's my cuda version:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Nov__3_21:07:56_CDT_2017
Cuda compilation tools, release 9.1, V9.1.85

My computer is UX430UQ, graphic card is GeForce 940MX

Here's the output from nvidia-smi:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.87.01    Driver Version: 418.87.01    CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce 940MX       On   | 00000000:01:00.0 Off |                  N/A |
| N/A   45C    P0    N/A /  N/A |    283MiB /  2004MiB |      9%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0      1014      G   /usr/lib/xorg/Xorg                            24MiB |
|    0      1164      G   /usr/bin/gnome-shell                          47MiB |
|    0      1440      G   /usr/lib/xorg/Xorg                           123MiB |
|    0      1615      G   /usr/bin/gnome-shell                          84MiB |
+-----------------------------------------------------------------------------+

Here's the output when I run sudo apt-get install cuda:

Reading package lists...
Building dependency tree...
Reading state information...
cuda is already the newest version (10.1.243-1).
0 upgraded, 0 newly installed, 0 to remove and 138 not upgraded.

Here's the output when I run tf.test.is_gpu_available()

2019-10-08 21:04:37.186069: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:1006] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero

2019-10-08 21:04:37.188434: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1618] Found device 0 with properties:

name: GeForce 940MX major: 5 minor: 0 memoryClockRate(GHz): 1.2415

pciBusID: 0000:01:00.0

2019-10-08 21:04:37.188863: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libcudart.so.10.0'; dlerror: libcudart.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-8.0/lib64

2019-10-08 21:04:37.189156: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libcublas.so.10.0'; dlerror: libcublas.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-8.0/lib64

2019-10-08 21:04:37.189426: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libcufft.so.10.0'; dlerror: libcufft.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-8.0/lib64

2019-10-08 21:04:37.189687: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libcurand.so.10.0'; dlerror: libcurand.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-8.0/lib64

2019-10-08 21:04:37.189946: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libcusolver.so.10.0'; dlerror: libcusolver.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-8.0/lib64

2019-10-08 21:04:37.190202: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libcusparse.so.10.0'; dlerror: libcusparse.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-8.0/lib64

2019-10-08 21:04:37.190236: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudnn.so.7

2019-10-08 21:04:37.190244: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1641] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu for how to download and setup the required libraries for your platform.

Skipping registering GPU devices...

2019-10-08 21:04:37.190261: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1159] Device interconnect StreamExecutor with strength 1 edge matrix:

2019-10-08 21:04:37.190268: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1165] 0

2019-10-08 21:04:37.190276: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1178] 0: N

Let's see: 1 Your installed TF wants CUDA 10.0: Could not load dynamic library 'libcudart.so.10.0', 2. Your installed CUDA version appears to be 9.1: Cuda compilation tools, release 9.1, V9.1.85, and 3. your LD_LIBRARY_PATH is pointing to CUDA 8.0: LD_LIBRARY_PATH: /usr/local/cuda-8.0/lib64 Much of this can probably be sorted out if you do a careful job of installing CUDA. The CUDA 10.0 linux install guide is here. You should start by installing CUDA 10.0 correctly. — Robert Crovella
Hi, the nvidia-smi shows that I already installed Cuda 10.1 but somehow the nvcc - - version is 9.1.85. I just include the output of nvidia-smi as an edit — yew onn
nvidia-smi doesn't tell you the installed CUDA version the way you think it does, see here. Anyway, if you have actually installed CUDA 10.1 properly, then you just need to set the PATH and LD_LIBRARY_PATH variables correctly to use it, which is covered in the CUDA linux install guide I already linked (step 7). However your TF is expecting CUDA 10.0 and you can't use CUDA 10.1 as a substitute/replacement for CUDA 10.0 for TF. — Robert Crovella
I just did the two steps: 1) export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}} and 2) export PATH=/usr/local/cuda-10.0/bin${PATH:+:${PATH}}. The output for echo $LD_LIBRARY_PATH is /usr/local/cuda-10.1/lib64 but the tf.test.is_gpu_available() is still saying LD_LIBRARY_PATH: /usr/local/cuda-8.0/lib64. Also, it says cuda is already the newest version (10.1.243-1). when I run sudo apt-get install cuda. (I just edited this into the question). What should I do about the CUDA 10.1 vs CUDA 10.0 for TF? — yew onn
After running export PATH=/usr/local/cuda-10.1/bin:/usr/local/cuda-10.1/NsightCompute-2019.1${PATH:+:${PATH}} , nvcc --version now shows Cuda compilation tools, release 10.1, V10.1.243. However, running tf.test.is_gpu_available() is still saying LD_LIBRARY_PATH: /usr/local/cuda-8.0/lib64 even when the output of echo $LD_LIBRARY_PATH is /usr/local/cuda-10.0/lib64 — yew onn

DachuanZhao DachuanZhao · Accepted Answer · 2020-11-26T01:38:28

You should use cuda10 and cudnn7.4 referring to this web

Tensorflow can't use GPU. tf.test.is_gpu_available() show GPU but cannot use

1 Answers