TensorFlow : failed call to cuInit: CUDA_ERROR_NO_DEVICE

13
votes

My test :

import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
sess = tf.Session()`

Error :

c:\l\work\tensorflow-1.1.0\tensorflow\stream_executor\cuda\cuda_driver.cc:405]

failed call to cuInit: CUDA_ERROR_NO_DEVICE

-> but "/cpu:0" works fine

Config :

nvidia-smi : enter image description here

  • CUDA Version 9.1
  • tensorflow-1.1.0
  • Windows 10
  • cudnn64_7.dll (installed in C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.1\bin)
  • Only tensorflow-gpu is installed in my Conda environment

Why can't Tensorflow detect my GPU ?

2
windowstensorflowcudnn
If you're using Tensorflow 1.1 from the prebuild binaries, you need CUDA 8 and cuDNN 5.1. Did you build this yourself? – GPhilo
I have updated Tensorflow to v1.5. The error message disappeared but it is still using my CPU instead of my GPU. Do you what could be the reason? – Vincent Roye
You installed via pip "tensorflow" instead of "tensorflow-gpu" perhaps? – GPhilo
no, tensorflow-gpu is installed (only) – Vincent Roye
Now I'm on : - Tensorflow-gpu 1.6 - CUDA 9.0 - Cudnn 7.0 Still : failed call to cuInit: CUDA_ERROR_NO_DEVICE – Vincent Roye

2 Answers

26
votes

The issue was solved on GitHub. This error message will be shown if you set an invalid value for the CUDA_VISIBLE_DEVICES environment variable, e.g. when you only have a single GPU (which has ID 0) and set CUDA_VISIBLE_DEVICES=1 or CUDA_VISIBLE_DEVICES=2.

22
votes

The true answer is to add following line of code:

import os
os.environ['CUDA_VISIBLE_DEVICES'] = "0"