# NOTE: Building this image require's docker version >= 23.0.
#
# For reference:
# - https://docs.docker.com/build/dockerfile/frontend/#stable-channel
ARG CUDA_VERSION=12.1.0
FROM nvidia/cuda:${CUDA_VERSION}-cudnn8-devel-ubuntu22.04
ARG DEBIAN_FRONTEND="noninteractive"
ENV DEBIAN_FRONTEND=${DEBIAN_FRONTEND}
ENV MAMBA_ROOT_PREFIX=~/micromamba
ARG ROOT_PASSWD="root"
ENV ROOT_PASSWD=${ROOT_PASSWD}
WORKDIR /root
SHELL ["/bin/bash", "-c"]
# base tools
RUN <<EOT
#!/bin/bash
apt-get update
apt-get install -y wget curl htop jq vim bash libaio-dev build-essential openssh-server python3 python3-pip bzip2
apt-get install -y --no-install-recommends software-properties-common build-essential autotools-dev nfs-common pdsh cmake g++ gcc curl wget vim tmux emacs less unzip htop iftop iotop ca-certificates openssh-client openssh-server rsync iputils-ping net-tools sudo llvm-dev re2c
add-apt-repository ppa:git-core/ppa -y
apt-get install -y git libnuma-dev wget
# install latest cmake
wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc | sudo apt-key add -
sudo apt-add-repository "deb https://apt.kitware.com/ubuntu/ $(lsb_release -cs) main"
sudo apt-get update
sudo apt-get install -y cmake
# Configure SSH for password and public key authentication
sed -i 's/#PermitRootLogin prohibit-password/PermitRootLogin yes/' /etc/ssh/sshd_config
sed -i 's/#PasswordAuthentication yes/PasswordAuthentication yes/' /etc/ssh/sshd_config
sed -i 's/PubkeyAuthentication no/PubkeyAuthentication yes/' /etc/ssh/sshd_config
sed -i 's/^#Port 22/Port 22/' /etc/ssh/sshd_config
sed -i 's/^Port [0-9]*/Port 22/' /etc/ssh/sshd_config
mkdir /var/run/sshd
echo "root:${ROOT_PASSWD}" | chpasswd
mkdir -p ~/.pip
# install miniconda
wget -qO- https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O /tmp/miniconda.sh
bash /tmp/miniconda.sh -b -p /opt/conda 
rm /tmp/miniconda.sh 
conda init bash 
ln -s /opt/conda/etc/profile.d/conda.sh /etc/profile.d/conda.sh
echo ". /opt/conda/etc/profile.d/conda.sh" >> ~/.bashrc 
# 配置 .condarc 文件
cat <<EOF > ~/.condarc
channels:
  - conda-forge
  - bioconda
  - pytorch
  - pytorch-nightly
  - nvidia
  - defaults
show_channel_urls: true
EOF
# 安装 micromamba
# echo 1 | bash <(curl -s https://cdn.jsdelivr.net/gh/hotwa/MicroMamba_Installer@main/install.sh)
# micromamba shell init -s bash -p ~/micromamba
# cat <<'EOF' >> ~/.bashrc
# source ~/micromamba/etc/profile.d/micromamba.sh
# EOF
# # 配置 .mambarc 文件
# cat <<EOF > ~/.mambarc
# channels:
#   - conda-forge
#   - bioconda
#   - pytorch
#   - pytorch-nightly
#   - nvidia
#   - defaults
# show_channel_urls: true
# EOF
EOT

# reference: https://github.com/huggingface/transformers/blob/main/docker/transformers-pytorch-deepspeed-latest-gpu/Dockerfile
# PyTorch
ARG CONDA_ENV_NAME="deepspeed"
ENV CONDA_ENV_NAME=${CONDA_ENV_NAME}
ARG PYTHON_VERSION=3.10
ENV PYTHON_VERSION=${PYTHON_VERSION}
ENV PATH=/opt/conda/envs/${CONDA_ENV_NAME}/bin:/usr/bin:/opt/conda/bin:$PATH
ENV DEEPSPEED_PYTHON="/opt/conda/envs/${CONDA_ENV_NAME}/bin/python3"
ENV REF='main'
ENV STAGE_DIR=/tmp
ENV NV_PEER_MEM_VERSION=1.2
ENV NV_PEER_MEM_TAG=${NV_PEER_MEM_VERSION}-0
ENV OPENMPI_BASEVERSION=4.1
ENV OPENMPI_VERSION=${OPENMPI_BASEVERSION}.6
ARG CUDA='cu121'
ENV CUDA=${CUDA}
ARG PYTORCH_VERSION=2.3.1
ENV PYTORCH_VERSION=${PYTORCH_VERSION}
ARG TORCHVISION_VERSION=0.18.1
ENV TORCHVISION_VERSION=${TORCHVISION_VERSION}
ARG TORCHAUDIO_VERSION=2.3.1
ENV TORCHAUDIO_VERSION=${TORCHAUDIO_VERSION}
ARG PYTORCH_CUDA_VERSION=12.1
ENV PYTORCH_CUDA_VERSION=${PYTORCH_CUDA_VERSION}
ENV MLNX_OFED_VERSION=4.9-7.1.0.0
ARG SETUPTOOLS_VERSION=69.5.1
ENV SETUPTOOLS_VERSION=${SETUPTOOLS_VERSION}
ARG USE_CUDA=1
ENV USE_CUDA=${USE_CUDA}
ARG USE_ROCM=0
ENV USE_ROCM=${USE_ROCM}
ARG USE_XPU=0
ENV USE_XPU=${USE_XPU}
ARG _GLIBCXX_USE_CXX11_ABI=1
ENV _GLIBCXX_USE_CXX11_ABI=${_GLIBCXX_USE_CXX11_ABI}
RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda create -n ${CONDA_ENV_NAME} python=${PYTHON_VERSION} -c conda-forge -y
echo "conda activate ${CONDA_ENV_NAME}" >> ~/.bashrc
which python > ~/python_path.txt
conda activate ${CONDA_ENV_NAME}
# 克隆 ninja 源码并编译
git clone https://github.com/ninja-build/ninja.git ${STAGE_DIR}/ninja
cd ${STAGE_DIR}/ninja
# 克隆 GoogleTest 源码
git clone https://github.com/google/googletest.git
python ./configure.py --bootstrap
# 配置并构建 Ninja 测试，添加 pthread 链接选项
# CXXFLAGS="-pthread" LDFLAGS="-pthread" ./configure.py --bootstrap --gtest-source-dir=$(pwd)/googletest
conda run -n ${CONDA_ENV_NAME} bash -c "CXXFLAGS='-pthread' LDFLAGS='-pthread' python ./configure.py --bootstrap --gtest-source-dir=$(pwd)/googletest"
./ninja all
# 运行 Ninja 单元测试
./ninja_test
python3 -m pip install --no-cache-dir --upgrade pip
python3 -m pip install open_clip_torch nvidia-ml-py3 opencv-contrib-python 
conda clean -afy 
git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF && cd ..
python -m pip install setuptools==${SETUPTOOLS_VERSION}
python3 -m pip install --no-cache-dir ./transformers[deepspeed-testing]
# # (PyTorch must be installed before pre-compiling any DeepSpeed c++/cuda ops.)
# # (https://www.deepspeed.ai/tutorials/advanced-install/#pre-install-deepspeed-ops)
python3 -m pip uninstall -y torch torchvision torchaudio 
# # install pytorch create conda env aleay exists
python3 -m pip install torch==${PYTORCH_VERSION}+${CUDA} torchvision==${TORCHVISION_VERSION}+${CUDA} torchaudio==${TORCHAUDIO_VERSION} xformers --extra-index-url https://download.pytorch.org/whl/${CUDA}
python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate
python3 -m pip uninstall -y transformer-engine
python3 -m pip uninstall -y torch-tensorrt
python3 -m pip uninstall -y apex
EOT

# install apex
RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
git clone https://github.com/NVIDIA/apex ${STAGE_DIR}/apex
cd apex
# if pip >= 23.1 (ref: https://pip.pypa.io/en/stable/news/#v23-1) which supports multiple `--config-settings` with the same key... 
MAX_JOBS=1 python3 -m pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" ./
python -c "import apex.amp; print('Apex is installed and the amp module is available.')"
cd ..
rm -rf ${STAGE_DIR}/apex
EOT

RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
# Pre-build **latest** DeepSpeed, so it would be ready for testing (otherwise, the 1st deepspeed test will timeout)
python3 -m pip uninstall -y deepspeed
# This has to be run (again) inside the GPU VMs running the tests.
# The installation works here, but some tests fail, if we do not pre-build deepspeed again in the VMs running the tests.
# TODO: Find out why test fail. install deepspeed
# DS_BUILD_CPU_ADAM=${DS_BUILD_CPU_ADAM} DS_BUILD_FUSED_ADAM={DS_BUILD_FUSED_ADAM} python3 -m pip install "deepspeed<=0.14.0" --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check 2>&1
# from https://github.com/huggingface/transformers/blob/main/docker/transformers-pytorch-deepspeed-latest-gpu/Dockerfile install deepspeed fail 
# reference deepspeed install from https://github.com/microsoft/DeepSpeed/blob/master/docker/Dockerfile
# install deepspeed prepare
# install Mellanox OFED
mkdir -p ${STAGE_DIR}
wget -q -O - http://www.mellanox.com/downloads/ofed/MLNX_OFED-${MLNX_OFED_VERSION}/MLNX_OFED_LINUX-${MLNX_OFED_VERSION}-ubuntu20.04-x86_64.tgz | tar xzf -
cd MLNX_OFED_LINUX-${MLNX_OFED_VERSION}-ubuntu20.04-x86_64 
./mlnxofedinstall --user-space-only --without-fw-update --all -q 
cd ${STAGE_DIR} 
rm -rf ${STAGE_DIR}/MLNX_OFED_LINUX-${MLNX_OFED_VERSION}-ubuntu20.04-x86_64*
cd ..
# install nv_peer_mem
rm -rf ${STAGE_DIR}
mkdir -p ${STAGE_DIR}
git clone https://github.com/Mellanox/nv_peer_memory.git --branch ${NV_PEER_MEM_TAG} ${STAGE_DIR}/nv_peer_memory
cd ${STAGE_DIR}/nv_peer_memory
./build_module.sh
cd ${STAGE_DIR}
tar xzf ${STAGE_DIR}/nvidia-peer-memory_${NV_PEER_MEM_VERSION}.orig.tar.gz
cd ${STAGE_DIR}/nvidia-peer-memory-${NV_PEER_MEM_VERSION}
apt-get update
apt-get install -y dkms
dpkg-buildpackage -us -uc
dpkg -i ${STAGE_DIR}/nvidia-peer-memory_${NV_PEER_MEM_TAG}_all.deb
EOT

# install mpi
ENV PATH=/usr/local/mpi/bin:${PATH} 
ENV LD_LIBRARY_PATH=/usr/local/lib:/usr/local/mpi/lib:/usr/local/mpi/lib64:${LD_LIBRARY_PATH}
RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
# OPENMPI
rm -rf ${STAGE_DIR}
mkdir -p ${STAGE_DIR}
cd ${STAGE_DIR}
wget -q -O - https://download.open-mpi.org/release/open-mpi/v${OPENMPI_BASEVERSION}/openmpi-${OPENMPI_VERSION}.tar.gz | tar xzf - 
cd openmpi-${OPENMPI_VERSION} 
./configure --prefix=/usr/local/openmpi-${OPENMPI_VERSION} 
make -j"$(nproc)" install 
ln -s /usr/local/openmpi-${OPENMPI_VERSION} /usr/local/mpi 
# Sanity check:
test -f /usr/local/mpi/bin/mpic++ 
cd ${STAGE_DIR} 
rm -r ${STAGE_DIR}/openmpi-${OPENMPI_VERSION}
# Create a wrapper for OpenMPI to allow running as root by default
mv /usr/local/mpi/bin/mpirun /usr/local/mpi/bin/mpirun.real
echo '#!/bin/bash' > /usr/local/mpi/bin/mpirun
echo 'mpirun.real --allow-run-as-root --prefix /usr/local/mpi "$@"' >> /usr/local/mpi/bin/mpirun
chmod a+x /usr/local/mpi/bin/mpirun
EOT

# SSH daemon port inside container cannot conflict with host OS port
ENV SSH_PORT=2222
RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
cat /etc/ssh/sshd_config > ${STAGE_DIR}/sshd_config && \
sed "0,/^Port 22/s//Port ${SSH_PORT}/" ${STAGE_DIR}/sshd_config > /etc/ssh/sshd_config
EOT

# 29.78 Usage: install.sh [options...]
# 29.78 
# 29.78 By default will install deepspeed and all third party dependencies across all machines listed in
# 29.78 hostfile (hostfile: /job/hostfile). If no hostfile exists, will only install locally
# 29.78 
# 29.78 [optional]
# 29.78     -l, --local_only        Install only on local machine
# 29.78     -s, --pip_sudo          Run pip install with sudo (default: no sudo)
# 29.78     -r, --allow_sudo        Allow script to be run by root (probably don't want this, instead use --pip_sudo)
# 29.78     -n, --no_clean          Do not clean prior build state, by default prior build files are removed before building wheels
# 29.78     -m, --pip_mirror        Use the specified pip mirror (default: the default pip mirror)
# 29.78     -H, --hostfile          Path to MPI-style hostfile (default: /job/hostfile)
# 29.78     -e, --examples          Checkout deepspeed example submodule (no install)
# 29.78     -v, --verbose           Verbose logging
# 29.78     -h, --help              This help text

RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
useradd --create-home --uid 1000 --shell /bin/bash deepspeed
usermod -aG sudo deepspeed
echo "deepspeed ALL=(ALL) NOPASSWD: ALL" >> /etc/sudoers
EOT

# install cutlass https://github.com/NVIDIA/cutlass
# H100: architecture is Hopper (cutlass need add : cmake .. -DCUTLASS_NVCC_ARCHS="90a" )
# A100: architecture is Ampere 
# V100: architecture is Volta 
# T4: architecture is Turing 
# ENV CUDACXX=${CUDA_INSTALL_PATH}/bin/nvcc
# 70：适用于 NVIDIA Volta 架构（如 Tesla V100）。
# 75：适用于 NVIDIA Turing 架构（如 Tesla T4）。
# 80：适用于 NVIDIA Ampere 架构（如 A100）。
# 90a：适用于 NVIDIA Hopper 架构（如 H100）。
# 89:GeForce RTX 4090 
ARG DCUTLASS_NVCC_ARCHS="80;89;90a"
ENV DCUTLASS_NVCC_ARCHS=${DCUTLASS_NVCC_ARCHS}
RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
git clone https://github.com/NVIDIA/cutlass /opt/cutlass
cd /opt/cutlass
git checkout . 
git checkout master
mkdir build
cd build
cmake .. -DCUTLASS_NVCC_ARCHS=${DCUTLASS_NVCC_ARCHS} -DCUTLASS_ENABLE_TESTS=OFF -DCUTLASS_UNITY_BUILD_ENABLED=ON            # compiles for NVIDIA Hopper GPU architecture, like H100
make -j"$(nproc)" install
cd ..
# make test_unit -j"$(nproc)"
# make test_unit_gemm_warp -j"$(nproc)"
EOT

# Some Packages from https://github.com/microsoft/DeepSpeed/blob/master/docker/Dockerfile
# RUN <<EOT
# source /opt/conda/etc/profile.d/conda.sh
# conda activate ${CONDA_ENV_NAME}
# apt-get update 
# apt-get install -y --no-install-recommends libsndfile-dev libcupti-dev libjpeg-dev libpng-dev screen libaio-dev
# python -m pip install pipdeptree \
# psutil \
# yappi \
# cffi \
# ipdb \
# pandas \
# matplotlib \
# py3nvml \
# pyarrow \
# graphviz \
# astor \
# boto3 \
# tqdm \
# sentencepiece \
# msgpack \
# requests \
# pandas \
# sphinx \
# sphinx_rtd_theme \
# scipy \
# numpy \
# scikit-learn \
# nvidia-ml-py3 \
# mpi4py
# EOT

# install deepspeed step 1
RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
/opt/conda/envs/${CONDA_ENV_NAME}/bin/python -m pip install setuptools==${SETUPTOOLS_VERSION}
# install oneapi for deepspeed
git clone https://github.com/oneapi-src/oneCCL.git ${STAGE_DIR}/oneCCL
cd ${STAGE_DIR}/oneCCL
git checkout . 
git checkout master
mkdir build
cd build 
cmake .. -DCMAKE_INSTALL_PREFIX=/usr/local
make -j"$(nproc)" install
EOT

# install deepspeed step 2
ARG CUDA_ARCH_LIST="80;86;89;90"
ENV CUDA_ARCH_LIST=${CUDA_ARCH_LIST}
RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
git clone https://github.com/microsoft/DeepSpeed-Kernels.git ${STAGE_DIR}/DeepSpeed-Kernels
cd ${STAGE_DIR}/DeepSpeed-Kernels
# CUDA_ARCH_LIST=${CUDA_ARCH_LIST} python setup.py bdist_wheel
# pip install dist/deepspeed_kernels-*.whl
CUDA_ARCH_LIST=${CUDA_ARCH_LIST} python -m pip install -v .
EOT

ARG DEEPSPEED_VERSION="v0.14.3"
ENV DEEPSPEED_VERSION=${DEEPSPEED_VERSION}
ARG DEEPSPEED_INSTALL_FLAGS="--allow_sudo --pip_sudo --verbose"
ENV DEEPSPEED_INSTALL_FLAGS=${DEEPSPEED_INSTALL_FLAGS}
ARG DS_BUILD_SPARSE_ATTN=0
ENV DS_BUILD_SPARSE_ATTN=${DS_BUILD_SPARSE_ATTN}
ARG DS_BUILD_FUSED_ADAM=1
ENV DS_BUILD_FUSED_ADAM=${DS_BUILD_FUSED_ADAM}
ARG DS_BUILD_CPU_ADAM=1
ENV DS_BUILD_CPU_ADAM=${DS_BUILD_CPU_ADAM}
ARG DS_BUILD_OPS=1
ENV DS_BUILD_OPS=${DS_BUILD_OPS}
ARG HOSTFILE_CONTENT=""
ENV HOSTFILE_CONTENT=${HOSTFILE_CONTENT}
ENV CUTLASS_PATH='/opt/cutlass'
ENV CUDA_HOME='/usr/local/cuda'
ENV LD_LIBRARY_PATH=${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}
ENV PATH=${CUDA_HOME}/bin:${PATH}
# install deepspeed step 3
RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
git clone https://github.com/microsoft/DeepSpeed.git ${STAGE_DIR}/DeepSpeed
cd ${STAGE_DIR}/DeepSpeed
git checkout ${DEEPSPEED_VERSION}
sed 's/pip install/python -m pip install/' install.sh > install_modified.sh
chmod +x ./install_modified.sh
# 检查 HOSTFILE_CONTENT 并写入文件
if [ -n "${HOSTFILE_CONTENT}" ]; then
    echo "${HOSTFILE_CONTENT}" > /tmp/hostfile
    INSTALL_CMD="./install_modified.sh ${DEEPSPEED_INSTALL_FLAGS} --hostfile /tmp/hostfile"
else
    INSTALL_CMD="./install_modified.sh ${DEEPSPEED_INSTALL_FLAGS}"
fi
eval $INSTALL_CMD
# compile deepspeed ops
cat <<'EOF' >> ~/.bashrc
source ~/micromamba/etc/profile.d/micromamba.sh
echo "alias mamba=micromamba" >> ~/.bashrc
echo "alias mba=mamba" >> ~/.bashrc
EOF
# 配置 .mambarc 文件
cat <<EOF > ~/compile_deepspeed_ops.py
import deepspeed

def compile_ops():
    builders = [
        deepspeed.ops.op_builder.AsyncIOBuilder,
        deepspeed.ops.op_builder.FusedAdamBuilder,
        deepspeed.ops.op_builder.CPUAdamBuilder,
        deepspeed.ops.op_builder.CPUAdagradBuilder,
        deepspeed.ops.op_builder.CPULionBuilder,
        deepspeed.ops.op_builder.EvoformerAttnBuilder,
        deepspeed.ops.op_builder.FPQuantizerBuilder,
        deepspeed.ops.op_builder.FusedLambBuilder,
        deepspeed.ops.op_builder.FusedLionBuilder,
        deepspeed.ops.op_builder.QuantizerBuilder,
        deepspeed.ops.op_builder.RaggedOpsBuilder,
        deepspeed.ops.op_builder.RandomLTDBuilder,
        deepspeed.ops.op_builder.SparseAttnBuilder,
        deepspeed.ops.op_builder.SpatialInferenceBuilder,
        deepspeed.ops.op_builder.TransformerBuilder,
        deepspeed.ops.op_builder.StochasticTransformerBuilder,
    ]
    
    for builder in builders:
        print(f"Compiling {builder.__name__}")
        builder().load()

if __name__ == "__main__":
    compile_ops()
EOF
python compile_deepspeed_ops.py
ds_report
# clean up
# rm -f deepspeed/git_version_info_installed.py
# rm -rf dist build deepspeed.egg-info
# python setup.py bdist_wheel
# DS_BUILD_OPS=${DS_BUILD_OPS} pip install -v dist/deepspeed*.whl
# DS_BUILD_OPS=${DS_BUILD_OPS} pip install -v -r requirements/requirements.txt
# pip install numpy==1.22.4 # ImportError: cannot import name 'BUFSIZE' from 'numpy' (/opt/conda/envs/deepspeed/lib/python3.10/site-packages/numpy/__init__.py) wait for fix in numpy=2.0.0
EOT

# install transformers and flash-attn
RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
# install transformers
git clone https://github.com/huggingface/transformers ${STAGE_DIR}/transformers
cd ${STAGE_DIR}/transformers
python3 ./setup.py develop
python3 -m pip install -U --no-cache-dir "pydantic<2"
# install flash-attn
# pip install packaging -i https://pypi.org/simple/ --trusted-host pypi.org
pip install flash-attn --no-build-isolation -i https://pypi.org/simple/ --trusted-host pypi.org
EOT

# other packages
ENV TORCH_CUDA_ARCH_LIST="80;86;89;90"
RUN <<EOT
#!/bin/bash
source /opt/conda/etc/profile.d/conda.sh
conda activate ${CONDA_ENV_NAME}
pip3 install optimum
pip3 install peft tiktoken \
    tqdm matplotlib seaborn numpy pandas scikit-learn diffusers \
    huggingface_hub spacy blobfile pycocotools \
    open_clip_torch \
    zstandard -i https://pypi.org/simple/ --trusted-host pypi.org
EOT

CMD ["/usr/sbin/sshd", "-D"]