Version compatibility between TF and NVIDIA drivers?

TensorFlow General Discussion
1

Hi, I’m having troubles to make Tensorflow 2.14 on my system (Fedora 38 with nvidia drivers version 535).

Whenever I import tensorflow from Python 3.9 I get this error:

2023-10-19 19:25:33.188414: E tensorflow/compiler/xla/stream_executor/cuda/cuda_dnn.cc:9342] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
2023-10-19 19:25:33.188446: E tensorflow/compiler/xla/stream_executor/cuda/cuda_fft.cc:609] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
2023-10-19 19:25:33.188473: E tensorflow/compiler/xla/stream_executor/cuda/cuda_blas.cc:1518] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
2023-10-19 19:25:33.193901: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.

I have tried installing CUDATOOLKIT 11.2 and CUdNN 8.1 with pip on my conda environment, ensuring that the PATH to the libraries is correct and files exist.

I have also tried with pip install tensorflow[and-cuda] that supossedly installs a bundled version of tensorflow with cuda

I reinstalled drivers, environments, deleted caches, followed every step from all the forums I found, and nothing seems to work. I hope this problem may be common in the tensorflow comunity and you can help me with this.

I’m guessing that maybe the nvidia driver version has something to do but I cannot find any information.

Thanks in advance!

3

Hi @Eduardo

I believe the message you get is a warning. not a error message, and it relates to your CPU, not to your GPU.

My understanding is that official Tensorflow builds do not resort CPU extensions such AVX2, FMA, etc… (that bump CPU performance) so that to be compatible with as many CPU as possible. You probably have to build yourself the library including everything that is required if you want to get the most of your CPU using Tensorflow.

I wouldn’t think the performance will be degraded too much as I trust Tensorflow manages well ops and send (most of) them appropriately to the GPU.

In short, you can safely disregard this warning.

If seeing it annoys you, add a command like this one at the beginning of your code:

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

Hopefully above isn’t misleading.