Update dockerfile

deepray/core/base_trainer.py

@@ -256,7 +256,7 @@ def __init__(
       self.global_batch_size *= get_world_size()
       learning_rate *= get_world_size()
 
-    # TODO: fuhailin 
+    # TODO: fuhailin
     # if isinstance(optimizer, optimizers.Optimizer):
     self.optimizer = optimizer
     # else:

tools/build_base_container.sh

@@ -4,11 +4,13 @@ set -x -e
 
 PY_VERSION=${1:-"3.8"}
 TF_VERSION=${2:-"2.9.3"}
+CUDA_VERSION=${3:-"11.6.2"}
 
 docker build \
     -f tools/docker/base_container.Dockerfile \
-    --build-arg TF_VERSION=2.9.3 \
-    --build-arg TF_PACKAGE=tensorflow \
-    --build-arg PY_VERSION=$PY_VERSION \
+    --build-arg CUDA_VERSION=${CUDA_VERSION} \
+    --build-arg TF_VERSION=${TF_VERSION} \
+    --build-arg TF_PACKAGE=tensorflow-gpu \
+    --build-arg PY_VERSION=${PY_VERSION} \
     --target base_container \
-    -t hailinfufu/deepray-release:latest-py${PY_VERSION}-tf${TF_VERSION}-cu116-ubuntu20.04 ./
+    -t hailinfufu/deepray-release:latest-py${PY_VERSION}-tf${TF_VERSION}-cu${CUDA_VERSION}-ubuntu20.04 ./

tools/build_dev_container.sh

@@ -4,11 +4,13 @@ set -x -e
 
 PY_VERSION=${1:-"3.8"}
 TF_VERSION=${2:-"2.9.3"}
+CUDA_VERSION=${3:-"11.6.2"}
 
 docker build \
     -f tools/docker/dev_container.Dockerfile \
-    --build-arg TF_VERSION=2.9.3 \
+    --build-arg CUDA_VERSION=${CUDA_VERSION} \
+    --build-arg TF_VERSION=${TF_VERSION} \
     --build-arg TF_PACKAGE=tensorflow \
     --build-arg PY_VERSION=$PY_VERSION \
     --target dev_container \
-    -t hailinfufu/deepray-dev:latest-py${PY_VERSION}-tf${TF_VERSION}-cu116-ubuntu20.04 ./
+    -t hailinfufu/deepray-dev:latest-py${PY_VERSION}-tf${TF_VERSION}-cu${CUDA_VERSION}-ubuntu20.04 ./

tools/docker/base_container.Dockerfile

@@ -1,11 +1,9 @@
 #syntax=docker/dockerfile:1.1.5-experimental
-ARG IMAGE_TYPE
-
+ARG CUDA_VERSION=11.6.2
 # Currenly all of our dev images are GPU capable but at a cost of being quite large.
 # See https://github.com/tensorflow/build/pull/47
-ARG CUDA_DOCKER_VERSION=11.6.2-cudnn8-devel-ubuntu20.04
-FROM nvidia/cuda:${CUDA_DOCKER_VERSION} as base_container
-ARG TF_PACKAGE
+FROM nvidia/cuda:${CUDA_VERSION}-cudnn8-devel-ubuntu20.04 as base_container
+ARG TF_PACKAGE=tensorflow-gpu
 ARG TF_VERSION=2.9.3
 ARG PY_VERSION=3.8
 
@@ -41,6 +39,8 @@ RUN bash /install_deps/install_cmake.sh
 RUN bash /install_deps/install_openmpi.sh
 
 RUN pip install --default-timeout=1000 $TF_PACKAGE==$TF_VERSION
+# RUN mkdir -p /usr/local/lib/python${PY_VERSION}/dist-packages/tensorflow/include/third_party/gpus/cuda && \
+#     ln -s /usr/local/cuda/include /usr/local/lib/python${PY_VERSION}/dist-packages/tensorflow/include/third_party/gpus/cuda
 
 COPY requirements.txt /tmp/requirements.txt
 RUN pip install -r /install_deps/yapf.txt \

tools/docker/dev_container.Dockerfile

@@ -1,11 +1,11 @@
 #syntax=docker/dockerfile:1.1.5-experimental
-ARG IMAGE_TYPE
+ARG CUDA_VERSION=11.6.2
 ARG TF_VERSION=2.9.3
 ARG PY_VERSION=3.8
 
 # Currenly all of our dev images are GPU capable but at a cost of being quite large.
 # See https://github.com/tensorflow/build/pull/47
-ARG CUDA_DOCKER_VERSION=latest-py${PY_VERSION}-tf${TF_VERSION}-cu116-ubuntu20.04
+ARG CUDA_DOCKER_VERSION=latest-py${PY_VERSION}-tf${TF_VERSION}-cu${CUDA_VERSION}-ubuntu20.04
 FROM hailinfufu/deepray-release:${CUDA_DOCKER_VERSION} as dev_container
 
 COPY tools/install_deps /install_deps