[fix] added cuda 11.8 support for manylinux

.github/workflows/publish_manylinux_image.yml

@@ -37,16 +37,16 @@ jobs:
           tags: |
             vectorchai/scalellm_manylinux:cuda12.1
 
-      # - name: Build base for cuda 11.8
-      #   uses: docker/build-push-action@v5
-      #   with:
-      #     context: ./docker
-      #     file: ./docker/Dockerfile.manylinux
-      #     push: true
-      #     cache-from: type=local,src=$CI_CACHE_DIR/.buildx-cache
-      #     cache-to: type=local,dest=$CI_CACHE_DIR/.buildx-cache
-      #     build-args: |
-      #       CUDA_VERSION=11.8
-      #     tags: |
-      #       vectorchai/scalellm_manylinux:cuda11.8
+      - name: Build base for cuda 11.8
+        uses: docker/build-push-action@v5
+        with:
+          context: ./docker
+          file: ./docker/Dockerfile.manylinux_gcc11
+          push: true
+          cache-from: type=local,src=$CI_CACHE_DIR/.buildx-cache
+          cache-to: type=local,dest=$CI_CACHE_DIR/.buildx-cache
+          build-args: |
+            CUDA_VERSION=11.8
+          tags: |
+            vectorchai/scalellm_manylinux:cuda11.8
 

.github/workflows/release_wheel.yml

@@ -16,7 +16,7 @@ jobs:
       fail-fast: false
       matrix:
         python: ["3.9", "3.10", "3.11"]
-        cuda: ["12.1"]
+        cuda: ["11.8", "12.1"]
         torch: ["2.2", "2.3"]
     runs-on: [self-hosted, linux, release]
     steps:
@@ -27,7 +27,7 @@ jobs:
 
       - name: Build wheel
         run: |
-          docker pull vectorchai/scalellm_builder:cuda${{ matrix.cuda }}-ubuntu22.04
+          docker pull vectorchai/scalellm_manylinux:cuda${{ matrix.cuda }}
           docker run --rm -t \
               -v "$CI_CACHE_DIR":/ci_cache \
               -v "$GITHUB_WORKSPACE":/ScaleLLM \

docker/Dockerfile.base

@@ -23,7 +23,7 @@ RUN rm install_python.sh
 
 # Install cuda, cudnn and nccl
 ARG CUDA_VERSION=12.1
-RUN wget -q https://raw.githubusercontent.com/pytorch/builder/main/common/install_cuda.sh -O install_cuda.sh
+COPY ./common/install_cuda.sh install_cuda.sh
 RUN bash ./install_cuda.sh ${CUDA_VERSION} && rm install_cuda.sh
 ENV DESIRED_CUDA ${CUDA_VERSION}
 ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:$PATH

docker/Dockerfile.devel

@@ -9,25 +9,6 @@ ENV DEBIAN_FRONTEND noninteractive
 COPY ./common/install_base.sh install_base.sh
 RUN bash ./install_base.sh && rm install_base.sh
 
-# Install user
-COPY ./common/install_user.sh install_user.sh
-RUN bash ./install_user.sh && rm install_user.sh
-
-# Install multiple python versions
-COPY ./common/install_python.sh install_python.sh
-RUN bash ./install_python.sh "3.9.0"
-RUN bash ./install_python.sh "3.10.1"
-RUN bash ./install_python.sh "3.11.0"
-RUN bash ./install_python.sh "3.12.0"
-RUN rm install_python.sh
-
-# Install cuda, cudnn and nccl
-ARG CUDA_VERSION=12.1
-RUN wget -q https://raw.githubusercontent.com/pytorch/builder/main/common/install_cuda.sh -O install_cuda.sh
-RUN bash ./install_cuda.sh ${CUDA_VERSION} && rm install_cuda.sh
-ENV DESIRED_CUDA ${CUDA_VERSION}
-ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:$PATH
-
 # Install gcc
 ARG GCC_VERSION=12
 RUN apt-get update \
@@ -36,16 +17,23 @@ RUN apt-get update \
 COPY ./common/install_gcc.sh install_gcc.sh
 RUN bash ./install_gcc.sh && rm install_gcc.sh
 
-ARG CMAKE_VERSION=3.18.5
+ARG CMAKE_VERSION=3.29.3
 COPY ./common/install_cmake.sh install_cmake.sh
 RUN if [ -n "${CMAKE_VERSION}" ]; then bash ./install_cmake.sh; fi
 RUN rm install_cmake.sh
 
-ARG NINJA_VERSION=1.9.0
+ARG NINJA_VERSION=1.11.1
 COPY ./common/install_ninja.sh install_ninja.sh
 RUN if [ -n "${NINJA_VERSION}" ]; then bash ./install_ninja.sh; fi
 RUN rm install_ninja.sh
 
+# Install cuda, cudnn and nccl
+ARG CUDA_VERSION=12.1
+COPY ./common/install_cuda.sh install_cuda.sh
+RUN bash ./install_cuda.sh ${CUDA_VERSION} && rm install_cuda.sh
+ENV DESIRED_CUDA ${CUDA_VERSION}
+ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:$PATH
+
 # install rust
 ENV RUSTUP_HOME=/usr/local/rustup
 ENV CARGO_HOME=/usr/local/cargo

docker/Dockerfile.manylinux

@@ -21,7 +21,7 @@ COPY ./common/install_cuda.sh install_cuda.sh
 RUN bash ./install_cuda.sh ${CUDA_VERSION} && rm install_cuda.sh
 ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:$PATH
 
-# ARG CMAKE_VERSION=3.18.5
+# ARG CMAKE_VERSION=3.29.3
 # COPY ./common/install_cmake.sh install_cmake.sh
 # RUN if [ -n "${CMAKE_VERSION}" ]; then bash ./install_cmake.sh; fi
 # RUN rm install_cmake.sh

docker/Dockerfile.manylinux_gcc11

@@ -0,0 +1,56 @@
+FROM quay.io/pypa/manylinux_2_28_x86_64 as base
+
+LABEL maintainer="mi@vectorch.com"
+ENV DEBIAN_FRONTEND noninteractive
+
+ENV LC_ALL en_US.UTF-8
+ENV LANG en_US.UTF-8
+ENV LANGUAGE en_US.UTF-8
+
+# Install common dependencies
+COPY ./common/install_base.sh install_base.sh
+RUN bash ./install_base.sh && rm install_base.sh
+
+# Install user
+COPY ./common/install_user.sh install_user.sh
+RUN bash ./install_user.sh && rm install_user.sh
+
+# Install gcc-11
+RUN rm -rf /opt/rh/gcc-toolset-12
+RUN yum install -y gcc-toolset-11-toolchain
+ENV PATH=/opt/rh/gcc-toolset-11/root/usr/bin:$PATH
+ENV LD_LIBRARY_PATH=/opt/rh/gcc-toolset-11/root/usr/lib64:/opt/rh/gcc-toolset-11/root/usr/lib:$LD_LIBRARY_PATH
+RUN gcc --version; g++ --version
+
+
+# Install cuda, cudnn and nccl
+ARG CUDA_VERSION=11.8
+COPY ./common/install_cuda.sh install_cuda.sh
+RUN bash ./install_cuda.sh ${CUDA_VERSION} && rm install_cuda.sh
+ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:$PATH
+
+# ARG CMAKE_VERSION=3.29.3
+# COPY ./common/install_cmake.sh install_cmake.sh
+# RUN if [ -n "${CMAKE_VERSION}" ]; then bash ./install_cmake.sh; fi
+# RUN rm install_cmake.sh
+
+ARG NINJA_VERSION=1.11.1
+COPY ./common/install_ninja.sh install_ninja.sh
+RUN if [ -n "${NINJA_VERSION}" ]; then bash ./install_ninja.sh; fi
+RUN rm install_ninja.sh
+
+ARG CCACHE_VERSION=4.8.3
+COPY ./common/install_ccache.sh install_ccache.sh
+RUN if [ -n "${CCACHE_VERSION}" ]; then bash ./install_ccache.sh; fi
+RUN rm install_ccache.sh
+
+# install rust
+ENV RUSTUP_HOME=/usr/local/rustup
+ENV CARGO_HOME=/usr/local/cargo
+ENV PATH=/usr/local/cargo/bin:$PATH
+RUN curl https://sh.rustup.rs -sSf | sh -s -- -y
+# give everyone permission to use rust
+RUN chmod -R a+w ${RUSTUP_HOME} ${CARGO_HOME}
+RUN rustup --version; cargo --version; rustc --version
+
+CMD ["bash"]

docker/common/install_cmake.sh

@@ -19,13 +19,12 @@ case "$ID" in
     ;;
 esac
 
-# Turn 3.6.3 into v3.6
-path=$(echo "${CMAKE_VERSION}" | sed -e 's/\([0-9].[0-9]\+\).*/v\1/')
-file="cmake-${CMAKE_VERSION}-Linux-x86_64.tar.gz"
+path="v${CMAKE_VERSION}"
+file="cmake-${CMAKE_VERSION}-linux-x86_64.tar.gz"
 
 # Download and install specific CMake version in /usr/local
 pushd /tmp
-curl -Os --retry 3 "https://cmake.org/files/${path}/${file}"
-tar -C /usr/local --strip-components 1 --no-same-owner -zxf cmake-*.tar.gz
+wget -q "https://github.com/Kitware/CMake/releases/download/${path}/${file}"
+tar -C /usr/local --strip-components 1 --no-same-owner -zxf ${file}
 rm -f cmake-*.tar.gz
 popd

tools/run_in_docker.sh

@@ -62,19 +62,25 @@ RUN_OPTS+=("-v $(pwd):$(pwd)")
 RUN_OPTS+=("-v /tmp:/tmp")
 RUN_OPTS+=("-v ${HOME}:${HOME}")
 
-# carry over some environment variables
-RUN_OPTS+=("-e VCPKG_DEFAULT_BINARY_CACHE=${VCPKG_DEFAULT_BINARY_CACHE}")
-RUN_OPTS+=("-e CCACHE_DIR=${CCACHE_DIR}")
+# carry over cache settings
+if [[ -n "${VCPKG_DEFAULT_BINARY_CACHE}" ]]; then
+  RUN_OPTS+=("-v ${VCPKG_DEFAULT_BINARY_CACHE}:${VCPKG_DEFAULT_BINARY_CACHE}")
+  RUN_OPTS+=("-e VCPKG_DEFAULT_BINARY_CACHE=${VCPKG_DEFAULT_BINARY_CACHE}")
+fi
+
+if [[ -n "${CCACHE_DIR}" ]]; then
+  RUN_OPTS+=("-v ${CCACHE_DIR}:${CCACHE_DIR}")
+  RUN_OPTS+=("-e CCACHE_DIR=${CCACHE_DIR}")
+fi
 
 CMD="sh -c 'cd $(pwd); $@'"
 
 [[ "${CMD}" = "" ]] && usage
 [[ ! -x $(command -v docker) ]] && echo "ERROR: 'docker' command missing from PATH." && usage
 
-echo "== Pulling docker image: ${IMAGE}"
 if ! docker pull ${IMAGE} ; then
   echo "WARNING: Failed to docker pull image ${IMAGE}"
 fi
 
-echo "docker run ${RUN_OPTS[@]} ${IMAGE} bash -c \"$(get_switch_user_cmd) ${CMD}\""
+# echo "docker run ${RUN_OPTS[@]} ${IMAGE} bash -c \"$(get_switch_user_cmd) ${CMD}\""
 docker run ${RUN_OPTS[@]} ${IMAGE} bash -c "$(get_switch_user_cmd) ${CMD}"