Merge pull request #272 from jakirkham/cleanup_old_cuda_bits

Cleanup and consolidate CUDA images

.github/workflows/ci.yaml

@@ -36,22 +36,28 @@ jobs:
             SHORT_DESCRIPTION: "conda-forge build image for Cent0S 7 on aarch64"
 
           - DOCKERIMAGE: linux-anvil-cuda
+            DOCKERFILE: linux-anvil-cuda
             DOCKERTAG: "11.8"
             CUDA_VER: "11.8.0"
+            DISTRO_ARCH: "amd64"
             DISTRO_NAME: "centos"
             DISTRO_VER: "7"
             SHORT_DESCRIPTION: "conda-forge build image for Cent0S 7 on x86_64 with CUDA"
 
           - DOCKERIMAGE: linux-anvil-ppc64le-cuda
+            DOCKERFILE: linux-anvil-cuda
             DOCKERTAG: "11.8"
             CUDA_VER: "11.8.0"
+            DISTRO_ARCH: "ppc64le"
             DISTRO_NAME: "ubi"
             DISTRO_VER: "8"
             SHORT_DESCRIPTION: "conda-forge build image for Cent0S 8 on ppc64le with CUDA"
 
           - DOCKERIMAGE: linux-anvil-aarch64-cuda
+            DOCKERFILE: linux-anvil-cuda
             DOCKERTAG: "11.8"
             CUDA_VER: "11.8.0"
+            DISTRO_ARCH: "arm64"
             DISTRO_NAME: "ubi"
             DISTRO_VER: "8"
             SHORT_DESCRIPTION: "conda-forge build image for Cent0S 8 on aarch64 with CUDA"

README.md

@@ -18,15 +18,15 @@ environment variables passed in to be able to build. In this case, you
 will want to use a command similar to the following:
 
 ```sh
-docker build --rm --build-arg DISTRO_NAME=centos --build-arg DISTRO_VER=6 --build-arg CUDA_VER=10.2 -f linux-anvil-cuda/Dockerfile .
+docker build --rm --build-arg DISTRO_NAME=centos --build-arg DISTRO_VER=6 --build-arg CUDA_VER=11.8.0 -f linux-anvil-cuda/Dockerfile .
 ```
 
 ## Environment variables
 
-* `$CUDA_VER`: This is the cuda & cudatoolkit version that will be used. The
-  value of this variable should be in major-minor for, e.g. `9.2` for versions
-  `9.x` and `10.x`. For versions `11.x` the variable should be in
-  major-minor-patch format, e.g. `11.2.0`.
+* `CUDA_VER`: This is the cuda & cudatoolkit version that will be used. The
+  value of this variable should be in major-minor-patch format, e.g. `11.8.0`.
+* `DISTRO_ARCH`: This is the Linux architecture that the image should use.
+  Should match the upstream Docker image, e.g. `amd64`.
 * `DISTRO_NAME`: This is the Linux distro image name that should be built with.
   Should match the upstream Docker image, e.g. `centos`.
 * `DISTRO_VER`: This is version of Linux distro (typical CentOS) that the image

linux-anvil-cuda/Dockerfile

@@ -1,63 +1,72 @@
 # Set environment variables during runtime.
 ARG CUDA_VER
+ARG DISTRO_ARCH
 ARG DISTRO_NAME
 ARG DISTRO_VER
-FROM --platform=linux/amd64 nvidia/cuda:${CUDA_VER}-devel-${DISTRO_NAME}${DISTRO_VER}
+FROM --platform=linux/${DISTRO_ARCH} nvidia/cuda:${CUDA_VER}-devel-${DISTRO_NAME}${DISTRO_VER}
 
 LABEL maintainer="conda-forge <[email protected]>"
 
+# Set `ARG`s during runtime.
 ARG CUDA_VER
+ARG DISTRO_ARCH
 ARG DISTRO_NAME
 ARG DISTRO_VER
 ENV CUDA_VER=${CUDA_VER} \
+    DISTRO_ARCH=${DISTRO_ARCH} \
     DISTRO_NAME=${DISTRO_NAME} \
     DISTRO_VER=${DISTRO_VER}
 
 # Set an encoding to make things work smoothly.
 ENV LANG en_US.UTF-8
+ENV LANGUAGE en_US.UTF-8
 
 # Set path to CUDA install (this is a symlink to /usr/local/cuda-${CUDA_VER})
 ENV CUDA_HOME /usr/local/cuda
 
-# we want to persist a path in ldconfig (to avoid having to always set LD_LIBRARY_PATH), but *after* the existing entries;
-# since entries in ld.so.conf.d have precedence before the preconfigured directories, we first add the latter to the former
-RUN ldconfig -v 2>/dev/null | grep -v ^$'\t' | cut -f1 -d":" >> /etc/ld.so.conf.d/cuda-$CUDA_VER.conf \
-    && if [ ${CUDA_VER} != "9.2" ]; then \
-        # the upstream images for 10.x all have libcuda.so under $CUDA_HOME/compat;
-        # add this to the ldconfig so it will be found correctly.
-        echo "$CUDA_HOME/compat" >> /etc/ld.so.conf.d/cuda-$CUDA_VER.conf ; \
-    else \
-        # For 9.2, the image nvidia/cuda:9.2-devel-centos6 contains neither
-        # $CUDA_HOME/compat, nor any (non-stub) libcuda.so. We fix this by
-        # adding cuda-compat-10.0 (which is not used for building, but to
-        # test if loading the respective library/package works). However,
-        # due to licensing reasons, these cannot be part of the conda-forge
-        # docker images, but are instead added for CI purposes in:
-        # github.com/conda-forge/conda-forge-ci-setup-feedstock/blob/master/recipe/run_conda_forge_build_setup_linux
-        # Here we only set the ldconfig accordingly.
-        echo "/usr/local/cuda-10.0/compat" >> /etc/ld.so.conf.d/cuda-$CUDA_VER.conf ; \
-    fi \
-    # don't forget to update settings by running ldconfig
-    && ldconfig
-
 # bust the docker cache so that we always rerun the installs below
 ADD https://loripsum.net/api /opt/docker/etc/gibberish
 
+# Add qemu in here so that we can use this image on regular linux hosts with qemu user installed
+ADD qemu-aarch64-static /usr/bin/qemu-aarch64-static
+ADD qemu-ppc64le-static /usr/bin/qemu-ppc64le-static
+
+# we want to persist a path in ldconfig (to avoid having to always set LD_LIBRARY_PATH), but *after* the existing entries;
+# since entries in ld.so.conf.d have precedence before the preconfigured directories, we first add the latter to the former
+# the upstream images all have libcuda.so under $CUDA_HOME/compat;
+# add this to the ldconfig so it will be found correctly.
+# don't forget to update settings by running ldconfig
+RUN ldconfig -v 2>/dev/null | grep -v ^$'\t' | cut -f1 -d":" >> /etc/ld.so.conf.d/cuda-$CUDA_VER.conf && \
+    echo "$CUDA_HOME/compat" >> /etc/ld.so.conf.d/cuda-$CUDA_VER.conf && \
+    ldconfig
+
 # Add the archived repo URL and fix RPM imports
 ADD centos7-repos /tmp/centos7-repos
 ADD scripts/fix_rpm /opt/docker/bin/fix_rpm
 RUN /opt/docker/bin/fix_rpm
 
-# Install basic requirements.
+# Add custom `yum_clean_all` script before using `yum`
 COPY scripts/yum_clean_all /opt/docker/bin/
+
+# Install basic requirements.
 RUN yum update -y --disablerepo=cuda && \
     yum install -y \
         bzip2 \
         sudo \
         tar \
-        which && \
+        which \
+    && \
     /opt/docker/bin/yum_clean_all
 
+# Fix locale in UBI 8 images
+# See https://github.com/CentOS/sig-cloud-instance-images/issues/154
+RUN if [ "${DISTRO_NAME}${DISTRO_VER}" = "ubi8" ]; then \
+        yum install -y \
+            glibc-langpack-en \
+        && \
+        /opt/docker/bin/yum_clean_all; \
+    fi
+
 # Remove preinclude system compilers
 RUN rpm -e --nodeps --verbose gcc gcc-c++
 
@@ -76,26 +85,14 @@ RUN source /opt/conda/etc/profile.d/conda.sh && \
     chgrp -R lucky /opt/conda && \
     chmod -R g=u /opt/conda
 
-# Symlink CUDA headers that were moved from $CUDA_HOME/include to /usr/include
-# in CUDA 10.1.
-RUN for HEADER_FILE in cublas_api.h cublas.h cublasLt.h cublas_v2.h cublasXt.h nvblas.h; do \
-    if [[ ! -f "${CUDA_HOME}/include/${HEADER_FILE}" ]]; \
-      then ln -s "/usr/include/${HEADER_FILE}" "${CUDA_HOME}/include/${HEADER_FILE}"; \
-    fi; \
-    done
-
-# Add qemu in here so that we can use this image on regular linux hosts with qemu user installed
-ADD qemu-aarch64-static /usr/bin/qemu-aarch64-static
-ADD qemu-ppc64le-static /usr/bin/qemu-ppc64le-static
-
 # Add a file for users to source to activate the `conda`
 # environment `base`. Also add a file that wraps that for
 # use with the `ENTRYPOINT`.
 COPY linux-anvil-cuda/entrypoint_source /opt/docker/bin/entrypoint_source
 COPY scripts/entrypoint /opt/docker/bin/entrypoint
 
 # Ensure that all containers start with tini and the user selected process.
-# Activate the `conda` environment `base` and the devtoolset compiler.
+# Activate the `conda` environment `base`.
 # Provide a default command (`bash`), which will start if the user doesn't specify one.
 ENTRYPOINT [ "/opt/conda/bin/tini", "--", "/opt/docker/bin/entrypoint" ]
 CMD [ "/bin/bash" ]