⬆️ Update for TensorFlow 2.5+2.6

README.md

@@ -25,12 +25,42 @@ GPU available: True
 
 Results
 =======
-As of 2020-09, the only combinations that are working:
+As of 2021-10, the only combinations that are working:
 
 * TensorFlow 1.15.3 using CUDA Toolkit 10.0
-* TensorFlow 2.3.0 using CUDA Toolkit 10.1
+* (TensorFlow 2.3.0 using CUDA Toolkit 10.1)
+* TensorFlow 2.4.3 using CUDA Toolkit 11.0
+* TensorFlow 2.5.1 & TensorFlow 2.6.0 using CUDA Toolkit 11.1 & 11.2.1
 
-This is only for pip-installable TensorFlow, not self-compiled nor Anaconda.
+This is only for pip-installable TensorFlow from PyPI, not self-compiled nor
-We also did not test other TensorFlow versions. Note that these are the
+Anaconda. Note that these are the CUDA *Toolkit* versions, not the CUDA version
-CUDA *Toolkit* versions, not the CUDA version the driver supports (reported
+the driver supports (reported by `nvidia-smi`).
-by `nvidia-smi`).
+
+Full log `run-docker-compatibility-matrix`
+------------------------------------------
+~~~
+tensorflow-gpu 1.15 nvidia/cuda:10.0-cudnn7-runtime-ubuntu18.04 True
+tensorflow-gpu 1.15 nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04 False
+tensorflow-gpu 1.15 nvidia/cuda:10.2-cudnn7-runtime-ubuntu18.04 False
+tensorflow-gpu 1.15 nvidia/cuda:11.0-cudnn8-runtime-ubuntu18.04 False
+tensorflow-gpu 1.15 nvidia/cuda:11.1-cudnn8-runtime-ubuntu18.04 False
+tensorflow-gpu 1.15 nvidia/cuda:11.2.1-cudnn8-runtime-ubuntu18.04 False
+tensorflow 2.4.3 nvidia/cuda:10.0-cudnn7-runtime-ubuntu18.04 False
+tensorflow 2.4.3 nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04 False
+tensorflow 2.4.3 nvidia/cuda:10.2-cudnn7-runtime-ubuntu18.04 False
+tensorflow 2.4.3 nvidia/cuda:11.0-cudnn8-runtime-ubuntu18.04 True
+tensorflow 2.4.3 nvidia/cuda:11.1-cudnn8-runtime-ubuntu18.04 False
+tensorflow 2.4.3 nvidia/cuda:11.2.1-cudnn8-runtime-ubuntu18.04 False
+tensorflow 2.5.1 nvidia/cuda:10.0-cudnn7-runtime-ubuntu18.04 False
+tensorflow 2.5.1 nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04 False
+tensorflow 2.5.1 nvidia/cuda:10.2-cudnn7-runtime-ubuntu18.04 False
+tensorflow 2.5.1 nvidia/cuda:11.0-cudnn8-runtime-ubuntu18.04 False
+tensorflow 2.5.1 nvidia/cuda:11.1-cudnn8-runtime-ubuntu18.04 True
+tensorflow 2.5.1 nvidia/cuda:11.2.1-cudnn8-runtime-ubuntu18.04 True
+tensorflow 2.6.0 nvidia/cuda:10.0-cudnn7-runtime-ubuntu18.04 False
+tensorflow 2.6.0 nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04 False
+tensorflow 2.6.0 nvidia/cuda:10.2-cudnn7-runtime-ubuntu18.04 False
+tensorflow 2.6.0 nvidia/cuda:11.0-cudnn8-runtime-ubuntu18.04 False
+tensorflow 2.6.0 nvidia/cuda:11.1-cudnn8-runtime-ubuntu18.04 True
+tensorflow 2.6.0 nvidia/cuda:11.2.1-cudnn8-runtime-ubuntu18.04 True
+~~~

assets/Dockerfile

@@ -2,7 +2,9 @@ ARG BASE_IMAGE
 
 FROM $BASE_IMAGE
 
-ARG tf
+ARG tensorflow_package
+ARG tensorflow_version
+ARG test_nvidia_options
 
 RUN apt-get update && \
     apt-get install -y python3 python3-distutils curl &&\
@@ -11,9 +13,9 @@ RUN apt-get update && \
     curl -sSL https://bootstrap.pypa.io/get-pip.py -o get-pip.py && \
     python3 get-pip.py && \
     rm -f get-pip.py
-COPY assets/requirements-$tf.txt /tmp
+RUN pip install --no-cache-dir "$tensorflow_package == $tensorflow_version"
-RUN pip install --no-cache-dir -r /tmp/requirements-$tf.txt
 
 COPY test-nvidia /usr/bin
 
-CMD ["/usr/bin/test-nvidia"]
+ENV test_nvidia_options $test_nvidia_options
+CMD /usr/bin/test-nvidia $test_nvidia_options

assets/requirements-tf1.txt

@@ -1 +0,0 @@
-tensorflow-gpu == 1.15.*

assets/requirements-tf2.txt

@@ -1 +0,0 @@
-tensorflow == 2.*

run-docker-compatibility-matrix

@@ -1,17 +1,29 @@
 #!/bin/sh
 set -e
 
-for tf in tf1 tf2; do
+for tensorflow_version in 1.15 2.4.3 2.5.1 2.6.0; do
   #  Note: CUDA 11.0 only with CUDNN 8
   for BASE_IMAGE in \
     nvidia/cuda:10.0-cudnn7-runtime-ubuntu18.04 \
     nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04 \
     nvidia/cuda:10.2-cudnn7-runtime-ubuntu18.04 \
     nvidia/cuda:11.0-cudnn8-runtime-ubuntu18.04 \
+    nvidia/cuda:11.1-cudnn8-runtime-ubuntu18.04 \
+    nvidia/cuda:11.2.1-cudnn8-runtime-ubuntu18.04 \
   ; do
-    echo "== $tf $BASE_IMAGE"
+
+    # According to the docs at https://www.tensorflow.org/install/gpu, we should
+    # use different package names depending on the major version of TF.
+    if echo $tensorflow_version | grep -q ^1; then
+      tensorflow_package=tensorflow-gpu
+    else
+      tensorflow_package=tensorflow
+    fi
+
+
+    echo -n "$tensorflow_package $tensorflow_version $BASE_IMAGE "
     work_dir=`dirname $0`
-    image_id=`docker build -q --build-arg tf=$tf --build-arg BASE_IMAGE=$BASE_IMAGE -f assets/Dockerfile $work_dir`
+    image_id=`docker build -q --build-arg test_nvidia_options="--quiet" --build-arg tensorflow_package=$tensorflow_package --build-arg tensorflow_version=$tensorflow_version --build-arg BASE_IMAGE=$BASE_IMAGE -f assets/Dockerfile $work_dir`
     docker run --gpus all -it --rm -e TF_CPP_MIN_LOG_LEVEL=2 $image_id
     docker rmi $image_id >/dev/null || true
   done

test-nvidia

@@ -1,14 +1,24 @@
 #!/usr/bin/python3
 import os
+import argparse
 
 
-os.system('nvidia-smi -L')
+parser = argparse.ArgumentParser(description="Test if GPU is available in TensorFlow")
+parser.add_argument('--quiet', dest='quiet', action='store_true')
+parser.set_defaults(quiet=False)
+
+args = parser.parse_args()
+
+
+if not args.quiet:
+    os.system('nvidia-smi -L')
 
 
 if not 'TF_CPP_MIN_LOG_LEVEL' in os.environ:
     os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'  # '1' means >= WARN
 import tensorflow as tf
-print('TensorFlow', tf.__version__)
+if not args.quiet:
+    print('TensorFlow', tf.__version__)
 with tf.compat.v1.Session() as sess:
     hello = tf.constant('Hello, TensorFlow!')
     result = sess.run(hello)
@@ -17,6 +27,12 @@ with tf.compat.v1.Session() as sess:
 if hasattr(tf.config, 'list_physical_devices'):
     # TensorFlow 2
     is_gpu_available = len(tf.config.list_physical_devices('GPU')) > 0
+    if not args.quiet:
         print('GPU available:', is_gpu_available)
+    else:
+        print(is_gpu_available)
 else:
+    if not args.quiet:
         print('GPU available:', tf.test.is_gpu_available(cuda_only=True))
+    else:
+        print(tf.test.is_gpu_available(cuda_only=True))