Fix not load dynamic library for Tensorflow GPU

I want to use my GPU for Tensorflow.

I tried this Could not load dynamic library 'cudart64_101.dll' on tensorflow CPU-only installation

Unfortunately, I keep getting an error Could not load dynamic library 'cudart64_110.dll'; dlerror: cudart64_110.dll not found. How can I fix this? Python-version: 3.8.3, CUDA 10.1

2020-11-03 12:30:28.832014: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'cudart64_110.dll'; dlerror: cudart64_110.dll not found
2020-11-03 12:30:28.832688: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'cublas64_11.dll'; dlerror: cublas64_11.dll not found
2020-11-03 12:30:28.833342: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'cublasLt64_11.dll'; dlerror: cublasLt64_11.dll not found
2020-11-03 12:30:28.833994: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'cufft64_10.dll'; dlerror: cufft64_10.dll not found
2020-11-03 12:30:28.834645: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'curand64_10.dll'; dlerror: curand64_10.dll not found
2020-11-03 12:30:28.835297: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'cusolver64_10.dll'; dlerror: cusolver64_10.dll not found
2020-11-03 12:30:28.835948: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'cusparse64_11.dll'; dlerror: cusparse64_11.dll not found
2020-11-03 12:30:28.836594: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'cudnn64_8.dll'; dlerror: cudnn64_8.dll not found
2020-11-03 12:30:28.836789: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1761] 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...
2020-11-03 12:30:28.837575: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2020-11-03 12:30:28.838495: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1265] Device interconnect StreamExecutor with strength 1 edge matrix:
2020-11-03 12:30:28.838708: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1271]      
2020-11-03 12:30:28.838831: I tensorflow/compiler/jit/xla_gpu_device.cc:99] Not creating XLA devices, tf_xla_enable_xla_devices not set

Solution 1:

Solution

I would suggest you to use conda (Ananconda/Miniconda) to create a separate environment and install tensorflow-gpu, cudnn and cudatoolkit. Miniconda has a much smaller footprint than Anaconda. I would suggest you to install Miniconda if you do not have conda already.

Quick Installation

# Quick and dirty: with channel specification
conda create -n tf_gpu_env python=3.8 
conda activate tf_gpu_env
conda install tensorflow-gpu -c anaconda
conda install cudnn -c conda-forge 
conda install cudatoolkit -c anaconda

Check that Tensorflow is using the GPU

For more details see this: https://www.tensorflow.org/guide/gpu

# Sanity check for validating 
# visibility of the GPUs to TF
import tensorflow as tf
print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))

Easily Reproducible Installation with Environment File: environment.yml

Although you can quickly create a conda (Ananconda/Miniconda) environment as shown earlier, it is much more desirable to make the installation process as reproducible as possible: enters the environment.yml file.

Save the environment file at the root of your repository (or project) and run the following command to install all the packages in an isolated conda environment (here I have named it tfgpu_env) using the environment file. The top part of the environment file contains some useful commands.

conda env create -f environment.yml

Save the following as environment.yml under your repository. And consider pinning the following three libraries:

An example:

  • tensorflow-gpu=2.4
  • cudnn=8
  • cudatoolkit=11
## filename: environment.yml

## Environment File Definition

name: tfgpu_env # tensorflow-gpu environment
channels:
  - conda-forge
  - anaconda
  - default
dependencies:
  - python=3.8
  ## Core Necessities
  - numpy # -c conda-forge, anaconda
  - pandas # -c conda-forge, anaconda
  - tabulate # -c conda-forge, anaconda  # necessary for df.to_markdown() in pandas
  - scipy # -c conda-forge, anaconda
  - matplotlib # -c conda-forge, anaconda
  ## Jupyter Related
  - jupyter # -c anaconda, conda-forge
  - jupyterlab # -c anaconda, conda-forge
  - jupyter_dashboards # -c conda-forge  (see: https://medium.com/plotly/introducing-jupyterdash-811f1f57c02e)
  - jupyter_contrib_nbextensions # -c conda-forge
  ## Progressbar
  - tqdm # -c conda-forge, anaconda
  ## Machine Learning
  - tensorflow-gpu # -c anaconda | version: 2.4.1 (linux), 2.3.0 (windows)
  # - tensorflow # -c anaconda | version: 2.2.0 (linux), 2.1.0 (windows)
  - cudnn # -c conda-forge | version: 8.1.0.77 (linux/windows)
  #       # -c anaconda | version: 7.6.5 (linux/windows)
  - cudatoolkit # -c conda-forge | version: 8.1.0.77 (linux/windows)
  #             # -c anaconda | version: 11.0.221 (linux/windows)
  - scikit-learn # -c conda-forge, anaconda
  ## Hyperparameter Optimization
  - optuna # -c conda-forge # works for pytorch, tf/keras, mxnet, scikit-learn, xgboost, lightgbm
  - keras-tuner # -c conda-forge
  ## Image Processing
  - opencv # -c conda-forge, anaconda
  - imageio # -c anaconda, conda-forge
  ## Image Augmentation
  - albumentations # -c conda-forge
  - imgaug # -c conda-forge
  ## Code Linting
  - pylint # -c conda-forge, anaconda
  - autopep8 # -c conda-forge, anaconda
  ## Installations with pip
  - pip:
    ## Web App Framework
    # - Flask-Testing
    - streamlit # https://docs.streamlit.io/en/stable/troubleshooting/clean-install.html

Useful Instructions

You might as well copy and paste the following instructions in the environment file itself, to keep them handy.

# Instruction:
#-----------------------------------------------------------
#
## For an environment installed locally (under: ./.venv)
# mkdir -p .venv && cd .venv
# conda env create --prefix . -f ../environment.yml
## For Updating local environment
# cd .venv
# conda env update --prefix . -f ../environment.yml  --prune
#
## For an environment installed globally
## with a name: fav_env 
# NOTE: The env-name is stored inside the 
#       environment.yml file.
# conda env create -f environment.yml
## For Updating global environment from env-file
# conda env update -f ./environment.yml  --prune
#
## Update conda itself
# conda update -n base -c defaults conda
#
## Creating a global environment in one-line: py37, py38
# conda create -n py37 python=3.7
# conda create -n py38 python=3.8
#
### In each of the envs: base, py37, py38
# conda install jupyter jupyterlab numpy scipy pandas matplotlib scikit-learn scikit-image tqdm plotly imageio requests pylint autopep8 tabulate opencv
#
## Export a platform independent copy of an environment
#  conda env export --from-history > path/to/environment.yml
### Make exports directory (if not present already) and export
# $targetDir = conda_exports
# mkdir ./$targetDir
# conda env export --from-history > ./$targetDir/exported_environment.yml

References

  1. How to manage packages using Conda.

  2. https://docs.anaconda.com/anaconda/user-guide/tasks/tensorflow/

  3. https://www.tensorflow.org/guide/gpu

  4. https://www.fastwebhost.in/blog/how-to-find-if-linux-is-running-on-32-bit-or-64-bit/

  5. Miniconda installtion: https://docs.conda.io/en/latest/miniconda.html

Solution 2:

Well, you can see that your Tensorflow installation is looking for Cuda libraries of version 11, 10, while you have 10.1. So in order to fix this, install the proper Cuda version. Why is it looking for 3 different versions, I have no idea. But you can find valid combinations of Cuda, Tensorflow, and CUDNN here.

EDIT: Removed 8 from the Cuda version, Tensorflow is actually looking for CUDNN version 8. So don't forget to install CUDNN as well (my guess is that you are installing the latest version of Tensorflow -> that's why is it looking for the latest Cuda and CUDNN releases.)