tensorflow: Could not create cudnn handle: CUDNN_STATUS_NOT_INITIALIZED

TensorFlow: 1.10.1 cuda: 9.0 cudnn: cudnn7.1 for cuda9.0 python3 Ubuntu 18.04 nvidia-driver: 390

Try running cnns error got:

Could not create cudnn handle: CUDNN_STATUS_NOT_INITIALIZED
2018-09-04 23:13:52.571542: E tensorflow/stream_executor/cuda/cuda_dnn.cc:360] Possibly insufficient driver version: 390.77.0

I think the cuda version are already the newest version, and my cudnn and cuda are OK. Why still got this problem? The error message throw from cuda_dnn.cc should be more specific what’s going on.

About this issue

Original URL
State: closed
Created 6 years ago
Reactions: 3
Comments: 16 (6 by maintainers)

Most upvoted comments

Yeah, in my case I ran TF 1.13 using CUDA 9.0 with CuDNN for CUDA 10.0 (since I was testing new TF 2.0). Downgrade sudo apt install libcudnn7=7.5.0.56-1+cuda9.0 solved it.

bzamecnik on Mar 8, 2019

Is it really the driver version being the problem?

No, the problem was an incompatible Cudnn version (7.1.2) being installed by creating an Anaconda environment which I had to manually overwrite with the previously installed system’s version (7.3.1).

saskra on Oct 17, 2018

I found my problem is caused by installing the wrong version of CUDNN. @jinfagang could you try nvidia-smi and dpkg -l | grep cudnn and paste the output here?

byronyi on Sep 27, 2018

Is this still an issue ?

Harshini-Gadige on Oct 31, 2018