Refactoring docker images, updating CUDA to 12.6 and pytorch 2.6

This commit is contained in:
ekoc committed 2025-02-12 22:01:34 +01:00
1 parent 73d9d9450b
commit 284c100860
11 files changed
+123 -143

No files matched your search

@@ -1,6 +1,6 @@
FROM continuumio/miniconda3:24.7.1-0 AS build
ARG CUDA_VERSION=12.4.1
ARG CUDA_VERSION=12.6.0
ENV CUDA_VERSION=$CUDA_VERSION
ARG PYTHON_VERSION=3.11.11
@@ -14,6 +14,8 @@ ENV CUDA_DOCKER_ARCH=all
ENV NVIDIA_VISIBLE_DEVICES=all
ENV CONDA_AUTO_UPDATE_CONDA="false"
ENV ALLTALK_DIR=/opt/alltalk
##############################################################################
# Installation/Basic Utilities
##############################################################################
@@ -35,24 +37,24 @@ EOR
##############################################################################
RUN <<EOR
CUDA_SHORT_VERSION=${CUDA_VERSION%.*}
conda create -y -n "alltalk" -c conda-forge python=${PYTHON_VERSION}
conda activate alltalk
RESULT=$( { conda install -y \
conda install -y \
gcc_linux-64 \
gxx_linux-64 \
pytorch \
pytorch-cuda=${CUDA_SHORT_VERSION} \
torchvision \
torchaudio \
libaio \
nvidia/label/cuda-${CUDA_SHORT_VERSION}.0::cuda-toolkit \
faiss-gpu=1.9.0 \
cuda=${CUDA_VERSION} cuda-tools=${CUDA_VERSION} cuda-toolkit=${CUDA_VERSION} cuda-version=${CUDA_SHORT_VERSION} \
cuda-command-line-tools=${CUDA_VERSION} cuda-compiler=${CUDA_VERSION} cuda-runtime=${CUDA_VERSION} \
cuda-libraries=${CUDA_VERSION} cuda-libraries-dev=${CUDA_VERSION} cuda-libraries-static=${CUDA_VERSION} \
cuda-cudart=${CUDA_SHORT_VERSION} cuda-cudart_linux-64=${CUDA_SHORT_VERSION} \
cudnn=9.3 \
py-cpuinfo=9 \
conda-forge::ffmpeg=7.1.0 \
conda-forge::portaudio=19.7.0 \
-c pytorch \
-c anaconda \
-c nvidia ; } 2>&1 )
-c nvidia \
-c anaconda
if [ $? -ne 0 ] ; then
echo "Failed to install conda dependencies: $RESULT"
@@ -62,16 +64,17 @@ RUN <<EOR
conda clean -a && pip cache purge
EOR
##############################################################################
# Export conda environment:
##############################################################################
RUN <<EOR
conda activate alltalk
conda env export > /environment-cu-${CUDA_VERSION}-cp-${PYTHON_VERSION}.yml
EOR
CUDA_SHORT_VERSION=${CUDA_VERSION%.*}
CUDA_SHORT_VERSION_NO_DOT=${CUDA_SHORT_VERSION//./}
##############################################################################
# Make environment file available to docker build output:
##############################################################################
FROM scratch
COPY --from=build /environment-*.yml /
conda activate alltalk
mkdir -p ${ALLTALK_DIR}/pip_cache
pip install --no-cache-dir --cache-dir=${ALLTALK_DIR}/pip_cache \
torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu${CUDA_SHORT_VERSION_NO_DOT}
if [ $? -ne 0 ] ; then
echo "Failed to install pip dependencies: $RESULT"
exit 1
fi
EOR
@@ -5,10 +5,7 @@ cd $SCRIPT_DIR
. ${SCRIPT_DIR=}/../variables.sh
if [ -f $SCRIPT_DIR/build/environment-cu-${CUDA_VERSION}-cp-${PYTHON_VERSION}.yml ]; then
echo "Environment file already exists - skipping..."
exit 0
fi
DOCKER_TAG=latest
# Parse arguments
while [ "$#" -gt 0 ]; do
@@ -21,8 +18,17 @@ while [ "$#" -gt 0 ]; do
PYTHON_VERSION="$2"
shift
;;
--tag)
DOCKER_TAG="$2"
shift
;;
--github-repository)
if [ -n "${GITHUB_REPOSITORY}" ] && ! [[ $GITHUB_REPOSITORY =~ ^--.* ]]; then
GITHUB_REPOSITORY="$2"
shift
fi
;;
*)
# Allow to pass arbitrary arguments to docker as well to be flexible:
echo "Unknown argument '$1'"
exit 1
;;
@@ -32,12 +38,10 @@ done
echo "Building conda environment using python ${PYTHON_VERSION} with CUDA ${CUDA_VERSION}"
rm -rf build # make sure to properly clean up
mkdir -p build
docker buildx \
build \
--progress=plain \
--build-arg CUDA_VERSION=$CUDA_VERSION \
--build-arg PYTHON_VERSION=$PYTHON_VERSION \
--output=$SCRIPT_DIR/build \
-t ${GITHUB_REPOSITORY}alltalk_environment:${DOCKER_TAG} \
.
+13 -43
View File
@@ -1,53 +1,15 @@
FROM continuumio/miniconda3:24.7.1-0
ARG GITHUB_REPOSITORY
FROM ${GITHUB_REPOSITORY}alltalk_environment:latest
ARG DEEPSPEED_VERSION=0.16.2
ARG DEEPSPEED_VERSION=0.16.3
ENV DEEPSPEED_VERSION=$DEEPSPEED_VERSION
SHELL ["/bin/bash", "-l", "-c"]
ENV SHELL=/bin/bash
ENV HOST=0.0.0.0
ENV DEBIAN_FRONTEND=noninteractive
ENV CUDA_DOCKER_ARCH=all
ENV NVIDIA_VISIBLE_DEVICES=all
ENV CONDA_AUTO_UPDATE_CONDA="false"
##############################################################################
# Directories:
##############################################################################
ENV STAGE_DIR=/tmp
RUN mkdir -p ${STAGE_DIR}
##############################################################################
# Installation/Basic Utilities
##############################################################################
RUN <<EOR
apt-get update
apt-get upgrade -y
apt-get install --no-install-recommends -y \
espeak-ng \
curl \
wget \
jq \
vim
apt-get clean && rm -rf /var/lib/apt/lists/*
EOR
##############################################################################
# Create a conda environment and install dependencies:
##############################################################################
COPY build/environment-*.yml environment.yml
RUN <<EOR
RESULT=$( { conda env create -f environment.yml ; } 2>&1 )
if [ $? -ne 0 ] ; then
echo "Failed to install conda dependencies: $RESULT"
exit 1
fi
conda clean -a && pip cache purge
EOR
##############################################################################
# Install python dependencies
##############################################################################
@@ -56,8 +18,7 @@ RUN <<EOR
conda activate alltalk
pip install \
deepspeed-kernels \
scikit-learn
deepspeed-kernels
EOR
@@ -69,6 +30,15 @@ RUN <<EOR
cd ${STAGE_DIR}/DeepSpeed
git checkout .
git checkout "tags/v${DEEPSPEED_VERSION}" -b "v${DEEPSPEED_VERSION}"
# Patching some files due to bug https://github.com/microsoft/DeepSpeed/issues/6709
START_LINE=`awk '/#ifndef BF16_AVAILABLE/{ print NR; exit }' csrc/transformer/inference/csrc/gelu.cu`
TO_LINE=$((START_LINE + 2))
sed -i "${START_LINE},${TO_LINE}d" ./csrc/transformer/inference/csrc/gelu.cu
START_LINE=`awk '/#ifndef BF16_AVAILABLE/{ print NR; exit }' csrc/transformer/inference/csrc/transform.cu`
TO_LINE=$((START_LINE + 2))
sed -i "${START_LINE},${TO_LINE}d" ./csrc/transformer/inference/csrc/transform.cu
EOR
##############################################################################
+15 -10
View File
@@ -5,6 +5,8 @@ cd $SCRIPT_DIR
. ${SCRIPT_DIR=}/../variables.sh
DOCKER_TAG=latest
# Parse arguments
while [ "$#" -gt 0 ]; do
case "$1" in
@@ -16,8 +18,17 @@ while [ "$#" -gt 0 ]; do
DEEPSPEED_VERSION="$2"
shift
;;
--tag)
DOCKER_TAG="$2"
shift
;;
--github-repository)
if [ -n "${GITHUB_REPOSITORY}" ] && ! [[ $GITHUB_REPOSITORY =~ ^--.* ]]; then
GITHUB_REPOSITORY="$2"
shift
fi
;;
*)
# Allow to pass arbitrary arguments to docker as well to be flexible:
echo "Unknown argument '$1'"
exit 1
;;
@@ -33,22 +44,16 @@ then
exit 0
fi
CONDA_ENV=${SCRIPT_DIR=}/../conda/build/environment-cu-${CUDA_VERSION}-cp-${PYTHON_VERSION}.yml
if [ ! -f ${CONDA_ENV} ]; then
echo "No conda environment found. Please run 'build-conda-env.sh' first."
exit 1
fi
echo "Building DeepSpeed $DEEPSPEED_VERSION using python ${PYTHON_VERSION}"
rm -rf build # make sure to properly clean up
mkdir -p build
cp ${CONDA_ENV} ${SCRIPT_DIR=}/build
docker buildx \
build \
--build-arg GITHUB_REPOSITORY=$GITHUB_REPOSITORY \
--build-arg DEEPSPEED_VERSION=$DEEPSPEED_VERSION \
-t deepspeed:$DEEPSPEED_VERSION \
-t ${GITHUB_REPOSITORY}alltalk_deepspeed:${DOCKER_TAG} \
.
docker run \
@@ -57,4 +62,4 @@ docker run \
--gpus=all \
--name deepspeed \
-v $SCRIPT_DIR/build:/deepspeed \
deepspeed:$DEEPSPEED_VERSION
${GITHUB_REPOSITORY}alltalk_deepspeed:${DOCKER_TAG}
+3 -2
View File
@@ -1,6 +1,7 @@
#!/usr/bin/env bash
export CUDA_VERSION=12.4.1
export CUDA_VERSION=12.6.0
export PYTHON_VERSION=3.11.11
export DEEPSPEED_VERSION=0.16.2
export ALLTALK_DIR=/opt/alltalk
export ALLTALK_DIR=/opt/alltalk
export GITHUB_REPOSITORY=