update

2024-06-21 11:52:34 +08:00
parent 7e18293a54
commit 11ccadb0df
3 changed files with 50 additions and 23 deletions
--- a/finetune/Dockerfile
+++ b/finetune/Dockerfile
@@ -3,13 +3,8 @@ FROM nvidia/cuda:${CUDA_VERSION}-cudnn8-devel-ubuntu20.04
 ARG DEBIAN_FRONTEND="noninteractive"
 ENV DEBIAN_FRONTEND=${DEBIAN_FRONTEND}
 ENV MAMBA_ROOT_PREFIX=~/micromamba
-ARG CONDA_ENV_NAME="deepspeed"
-ENV CONDA_ENV_NAME=${CONDA_ENV_NAME}
-ARG PYTHON_VERSION=3.10
-ENV PYTHON_VERSION=${PYTHON_VERSION}
 ARG ROOT_PASSWD="root"
 ENV ROOT_PASSWD=${ROOT_PASSWD}
-ENV PATH /opt/conda/bin:/opt/conda/envs/${CONDA_ENV_NAME}/bin:$PATH
 WORKDIR /root
 SHELL ["/bin/bash", "-c"]
 # base tools
@@ -41,7 +36,6 @@ 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 
-echo "conda activate ${CONDA_ENV_NAME}" >> ~/.bashrc
 # 配置 .condarc 文件
 cat <<EOF > ~/.condarc
 channels:
@@ -76,6 +70,11 @@ 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/bin:/opt/conda/envs/${CONDA_ENV_NAME}/bin:$PATH
 ENV REF='main'
 ENV STAGE_DIR=/tmp
 ENV NV_PEER_MEM_VERSION=1.2
@@ -95,38 +94,47 @@ 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} pyyaml ipython -c conda-forge -y
+echo "conda activate ${CONDA_ENV_NAME}" >> ~/.bashrc
 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
-conda run -n ${CONDA_ENV_NAME} python ./configure.py --bootstrap
+python ./configure.py --bootstrap
 # 配置并构建 Ninja 测试，添加 pthread 链接选项
 # CXXFLAGS="-pthread" LDFLAGS="-pthread" ./configure.py --bootstrap --gtest-source-dir=$(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
-conda create -n ${CONDA_ENV_NAME} python=${PYTHON_VERSION} pyyaml ipython -c conda-forge -y
 python3 -m pip install --no-cache-dir --upgrade pip
-conda run -n ${CONDA_ENV_NAME} python -m pip install open_clip_torch nvidia-ml-py3 opencv-contrib-python 
+python -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 ..
-conda run -n ${CONDA_ENV_NAME} python -m pip install setuptools==${SETUPTOOLS_VERSION}
-conda run -n ${CONDA_ENV_NAME} python3 -m pip install --no-cache-dir ./transformers[deepspeed-testing]
+python -m pip install setuptools==${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)
-conda run -n ${CONDA_ENV_NAME} python3 -m pip uninstall -y torch torchvision torchaudio 
+python3 -m pip uninstall -y torch torchvision torchaudio 
 # # install pytorch create conda env aleay exists
-conda run -n ${CONDA_ENV_NAME} python3 -m pip install torch==${PYTORCH_VERSION}+${CUDA} torchvision==${TORCHVISION_VERSION}+${CUDA} torchaudio==${TORCHAUDIO_VERSION} --extra-index-url https://download.pytorch.org/whl/${CUDA}
-conda run -n ${CONDA_ENV_NAME} python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate
-conda run -n ${CONDA_ENV_NAME} python3 -m pip uninstall -y transformer-engine
-conda run -n ${CONDA_ENV_NAME} python3 -m pip uninstall -y torch-tensorrt
-conda run -n ${CONDA_ENV_NAME} python3 -m pip uninstall -y apex
+python3 -m pip install torch==${PYTORCH_VERSION}+${CUDA} torchvision==${TORCHVISION_VERSION}+${CUDA} torchaudio==${TORCHAUDIO_VERSION} --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
@@ -305,11 +313,11 @@ EOT
 # CUDA_ARCH_LIST="80;86;89;90"
 ARG DEEPSPEED_INSTALL_FLAGS="--allow_sudo --pip_sudo --no_clean"
 ENV DEEPSPEED_INSTALL_FLAGS=${DEEPSPEED_INSTALL_FLAGS}
-ARG CUDA_ARCH_LIST="80;86"
+ARG CUDA_ARCH_LIST="80;86;89;90"
 ENV CUDA_ARCH_LIST=${CUDA_ARCH_LIST}
 ARG DS_BUILD_SPARSE_ATTN=0
 ENV DS_BUILD_SPARSE_ATTN=${DS_BUILD_SPARSE_ATTN}
-ARG DS_BUILD_FUSED_ADAM=0
+ARG DS_BUILD_FUSED_ADAM=1
 ENV DS_BUILD_FUSED_ADAM=${DS_BUILD_FUSED_ADAM}
 ARG DS_BUILD_CPU_ADAM=0
 ENV DS_BUILD_CPU_ADAM=${DS_BUILD_CPU_ADAM}
@@ -341,9 +349,9 @@ cd ${STAGE_DIR}/DeepSpeed
 git checkout . 
 git checkout master 
 python setup.py bdist_wheel
-pip install dist/deepspeed*.whl --force-reinstall
+DS_BUILD_OPS=${DS_BUILD_OPS} pip install dist/deepspeed*.whl --force-reinstall
 # DS_BUILD_OPS=${DS_BUILD_OPS} pip install -r requirements/requirements.txt
-# DS_BUILD_OPS=0 DS_BUILD_SPARSE_ATTN=0 DS_BUILD_CPU_ADAM=0 DS_BUILD_FUSED_ADAM=0 pip install -U --no-cache-dir .
+# DS_BUILD_OPS=0 DS_BUILD_SPARSE_ATTN=0 DS_BUILD_CPU_ADAM=0 DS_BUILD_FUSED_ADAM=1 pip install -U --no-cache-dir .
 # ./install.sh ${DEEPSPEED_INSTALL_FLAGS} --hostfile /job/hostfile # ./install.sh --allow_sudo --pip_sudo --no_clean --hostfile /path/to/your/hostfile
 cd ..
 # rm -rf ${STAGE_DIR}/DeepSpeed
--- a/finetune/docker-compose_pytorch1.13.yml
+++ b/finetune/docker-compose_pytorch1.13.yml
@@ -13,7 +13,10 @@ services:
        DS_BUILD_OPS: 1
        DS_BUILD_SPARSE_ATTN: 0
        DS_BUILD_FUSED_ADAM: 1
-        DS_BUILD_CPU_ADAM: 0
+        DS_BUILD_CPU_ADAM: 1
+        USE_CUDA: 1
+        USE_ROCM: 0
+        USE_XPU: 0
        CUDA: cu117
        CUDA_ARCH_LIST: "80;86;89;90" # for RTX 4090, all : "80;86;89;90"
        SETUPTOOLS_VERSION: "69.5.1"
--- a/finetune/docker-compose_pytorch2.3.yml
+++ b/finetune/docker-compose_pytorch2.3.yml
@@ -1,5 +1,18 @@
 version: '3.8'

+# DeepSpeed支持多种C++/CUDA扩展（ops），这些ops旨在优化深度学习的训练和推理过程。以下是一些主要的DeepSpeed ops及其功能：
+
+# FusedAdam - 提供融合优化的Adam优化器，适用于GPU。
+# FusedLamb - 类似FusedAdam，针对LAMB优化器，适用于大规模分布式训练。
+# SparseAttention - 用于高效计算稀疏注意力机制。
+# Transformer - 提供Transformer模型的高效实现。
+# TransformerInference - 专门用于Transformer模型的推理优化。
+# CPUAdam - 针对CPU优化的Adam优化器。
+# CPULion - 针对CPU的Lion优化器。
+# Quantizer - 提供量化支持，以减少模型大小和提高推理速度。
+# RandomLTD - 用于随机层裁剪的优化器。
+# StochasticTransformer - 支持随机Transformer模型的训练和推理。
+
 services:
  ubuntu-finetune:
    build: 
@@ -13,7 +26,10 @@ services:
        DS_BUILD_OPS: 1
        DS_BUILD_SPARSE_ATTN: 0
        DS_BUILD_FUSED_ADAM: 1
-        DS_BUILD_CPU_ADAM: 0
+        DS_BUILD_CPU_ADAM: 1
+        USE_CUDA: 1
+        USE_ROCM: 0
+        USE_XPU: 0
        CUDA: cu121
        CUDA_ARCH_LIST: "80;86;89;90" # for RTX 4090, all : "80;86;89;90"
        SETUPTOOLS_VERSION: "69.5.1"