Fixing docker build for DeepSpeed, introducing build for conda environment

.gitignore

@@ -172,5 +172,6 @@ cython_debug/
 
 .idea/*
 
-# Deepspeed build
+# Docker build
-deepspeed/build
+docker/conda/build
+docker/deepspeed/build

DOCKER_README.md

@@ -2,9 +2,11 @@
 The Docker image currently works on Windows and Linux, optionally supporting NVIDIA GPUs.
 
 ## General Remarks
-- The resulting Docker image is 22 GB in size. Building might require even more disk space temporarily.
+- The resulting Docker image is 21 GB in size. Building might require even more disk space temporarily.
-- Build time depends on your hardware and internet connection. Expect at least 10min to be normal.
+    - Another 15 GB is required for building DeepSpeed
-- The Docker build:
+- Build time depends on your hardware and internet connection. Expect at least 20-30min to be normal for a full build.
+  - This includes building the conda environment as well as DeepSpeed, which is the basis for the alltalk Docker image.
+- The Docker build for alltalk:
   - Downloads XTTS as default TTS engine
   - Enables RVC by default
   - Downloads all supported RVC models
@@ -55,18 +57,22 @@ There are various arguments to customize the build and start of the docker image
   - Example: `docker-build.sh --tts_model piper`
 - `--tag` allows to choose the docker tag. Defaults to `latest`.
   - Example: `docker-build.sh --tag mytag`
+- `--clean` allows remove existing dependency build like conda environment or DeepSpeed.
+    - Example: `docker-build.sh --clean`
 
 ### Arguments for `docker-start.sh`
 - `--config` lets you choose a config JSON file which can subset of `confignew.json`. This allows you to change only 
   few values and leave the rest as defined in the default `confignew.json` file.
-  - Example: `docker-build.sh --config /my/config/file.json` with content `{"branding": "My Brand "}` will just change
+  - Example: `docker-start.sh --config /my/config/file.json` with content `{"branding": "My Brand "}` will just change
     the branding in `confignew.json`.
 - `--voices` lets you add voices for the TTS engine in WAV format. You have to specify the folder containing all
   voice files.
-  - Example: `docker-build.sh --voices /my/voices/dir`
+  - Example: `docker-start.sh --voices /my/voices/dir`
 - `--rvc_voices` similar to voices, this option lets you pick the folder containing the RVC models.
-  - Example: `docker-build.sh --rvc_vices /my/rvc/voices/dir`
+  - Example: `docker-start.sh --rvc_vices /my/rvc/voices/dir`
 - `--no_ui` allows you to not expose port 7852 for the gradio interface. Note that you still have to set `launch_gradio`
   to `false` via JSON file passed to `--config`.
+- `--tag` allows to choose the docker tag of the image to run. Defaults to `latest`.
+    - Example: `docker-start.sh --tag mytag`
 - Since the above commands only address the most important options, you might pass additional arbitrary docker commands
     to the `docker-start.sh`.

Dockerfile

@@ -4,23 +4,21 @@ FROM continuumio/miniconda3:24.7.1-0
 ARG TTS_MODEL="xtts"
 ENV TTS_MODEL=$TTS_MODEL
 
-ARG CUDA_VERSION="12.1.1"
+ARG ALLTALK_DIR=/opt/alltalk
-ENV CUDA_VERSION=$CUDA_VERSION
-
-ARG PYTHON_VERSION=3.11.9
-ENV PYTHON_VERSION=$PYTHON_VERSION
-
-ARG PYTORCH_VERSION=2.2.1
-ENV PYTORCH_VERSION=$PYTORCH_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 GRADIO_SERVER_NAME="0.0.0.0"
 ENV NVIDIA_VISIBLE_DEVICES=all
+ENV CONDA_AUTO_UPDATE_CONDA="false"
 
+ENV GRADIO_SERVER_NAME="0.0.0.0"
+
+##############################################################################
+# Installation/Basic Utilities
+##############################################################################
 RUN <<EOR
     apt-get update
     apt-get upgrade -y
@@ -34,59 +32,47 @@ RUN <<EOR
     apt-get clean && rm -rf /var/lib/apt/lists/*
 EOR
 
-WORKDIR /alltalk
+WORKDIR ${ALLTALK_DIR}
 
+##############################################################################
 # Create a conda environment and install dependencies:
-ARG INSTALL_ENV_DIR=/alltalk/alltalk_environment/env
+##############################################################################
-ENV CONDA_AUTO_UPDATE_CONDA="false"
+COPY docker/conda/build/environment-*.yml environment.yml
 RUN <<EOR
-    CUDA_SHORT_VERSION=${CUDA_VERSION%.*}
+    RESULT=$( { conda env create -f environment.yml ; } 2>&1 )
-
-    conda create -y -n "alltalk" -c conda-forge python=${PYTHON_VERSION}
-    conda activate alltalk
-    RESULT=$( { conda install -y \
-      gcc_linux-64 \
-      gxx_linux-64 \
-      pytorch=${PYTORCH_VERSION} \
-      pytorch-cuda=${CUDA_SHORT_VERSION} \
-      torchvision \
-      torchaudio \
-      libaio \
-      nvidia/label/cuda-${CUDA_SHORT_VERSION}.0::cuda-toolkit \
-      faiss-gpu=1.9.0 \
-      conda-forge::ffmpeg=7.1.0 \
-      conda-forge::portaudio=19.7.0 \
-      -c pytorch \
-      -c anaconda \
-      -c nvidia ; } 2>&1 )
 
     if echo $RESULT | grep -izq error ; then
-      echo "Failed to install conda dependencies 2: $RESULT"
+      echo "Failed to install conda dependencies: $RESULT"
       exit 1
     fi
+
     conda clean -a && pip cache purge
 EOR
 
+##############################################################################
 # Install python dependencies (cannot use --no-deps because requirements are not complete)
+##############################################################################
 COPY system/config system/config
 COPY system/requirements/requirements_standalone.txt system/requirements/requirements_standalone.txt
 COPY system/requirements/requirements_parler.txt system/requirements/requirements_parler.txt
-ENV PIP_CACHE_DIR=/alltalk/pip_cache
+ENV PIP_CACHE_DIR=${ALLTALK_DIR}/pip_cache
 RUN <<EOR
     conda activate alltalk
 
-    mkdir /alltalk/pip_cache
+    mkdir ${ALLTALK_DIR}k/pip_cache
-    pip install --no-cache-dir --cache-dir=/alltalk/pip_cache -r system/requirements/requirements_standalone.txt
+    pip install --no-cache-dir --cache-dir=${ALLTALK_DIR}/pip_cache -r system/requirements/requirements_standalone.txt
-    pip install --no-cache-dir --cache-dir=/alltalk/pip_cache --upgrade gradio==4.32.2
+    pip install --no-cache-dir --cache-dir=${ALLTALK_DIR}/pip_cache --upgrade gradio==4.32.2
     # Parler:
-    pip install --no-cache-dir --cache-dir=/alltalk/pip_cache -r system/requirements/requirements_parler.txt
+    pip install --no-cache-dir --cache-dir=${ALLTALK_DIR}/pip_cache -r system/requirements/requirements_parler.txt
 
     conda clean --all --force-pkgs-dirs -y && pip cache purge
 EOR
 
-# Deepspeed:
+##############################################################################
+# Install DeepSpeed
+##############################################################################
 RUN mkdir -p /tmp/deepseped
-COPY deepspeed/build/*.whl /tmp/deepspeed/
+COPY docker/deepspeed/build/*.whl /tmp/deepspeed/
 RUN <<EOR
     DEEPSPEED_WHEEL=$(realpath /tmp/deepspeed/*.whl)
     conda activate alltalk
@@ -103,11 +89,9 @@ RUN <<EOR
     conda clean --all --force-pkgs-dirs -y && pip cache purge
 EOR
 
-### Deepspeed requires cutlass:
+##############################################################################
-###RUN git clone --depth 1 --branch "v3.5.1" https://github.com/NVIDIA/cutlass /alltalk/cutlass
-###ENV CUTLASS_PATH=/alltalk/cutlass
-
 # Writing scripts to start alltalk:
+##############################################################################
 RUN <<EOR
     cat << EOF > start_alltalk.sh
 #!/usr/bin/env bash
@@ -141,7 +125,9 @@ EOR
 
 COPY . .
 
-# Create script to execute firstrun.py:
+##############################################################################
+# Create script to execute firstrun.py and run it:
+##############################################################################
 RUN echo $'#!/usr/bin/env bash \n\
 source ~/.bashrc \n\
 conda activate alltalk \n\
@@ -150,18 +136,24 @@ python ./system/config/firstrun.py $@' > ./start_firstrun.sh
 RUN chmod +x start_firstrun.sh
 RUN ./start_firstrun.sh --tts_model $TTS_MODEL
 
-RUN mkdir -p /alltalk/outputs
+RUN mkdir -p ${ALLTALK_DIR}/outputs
 RUN mkdir -p /root/.triton/autotune
 
-# Enabling deepspeed for all models:
+##############################################################################
+# Enable deepspeed for all models:
+##############################################################################
 RUN find . -name model_settings.json -exec sed -i -e 's/"deepspeed_enabled": false/"deepspeed_enabled": true/g' {} \;
 
-# Downloading all RVC models:
+##############################################################################
+# Download all RVC models:
+##############################################################################
 RUN <<EOR
   jq -r '.[]' system/tts_engines/rvc_files.json > /tmp/rvc_files.txt
   xargs -n 1 curl --create-dirs --output-dir models/rvc_base -LO < /tmp/rvc_files.txt
   rm -f /tmp/rvc_files.txt
 EOR
 
-## Start alltalk:
+##############################################################################
+# Start alltalk:
+##############################################################################
 ENTRYPOINT ["sh", "-c", "./start_alltalk.sh"]

docker-build.sh

@@ -1,10 +1,13 @@
 #!/usr/bin/env bash
 
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
+cd $SCRIPT_DIR
+
+. ${SCRIPT_DIR=}/docker/variables.sh
+
 TTS_MODEL=xtts
-CUDA_VERSION=12.1.1
-PYTHON_VERSION=3.11.9
-PYTORCH_VERSION=2.2.1
 DOCKER_TAG=latest
+CLEAN=false
 
 # Parse arguments
 while [ "$#" -gt 0 ]; do
@@ -17,10 +20,6 @@ while [ "$#" -gt 0 ]; do
       PYTHON_VERSION="$2"
       shift
       ;;
-    --pytorch-version)
-      PYTORCH_VERSION="$2"
-      shift
-      ;;
     --tts_model)
       TTS_MODEL="$2"
       shift
@@ -29,6 +28,9 @@ while [ "$#" -gt 0 ]; do
       DOCKER_TAG="$2"
       shift
       ;;
+    --clean)
+      CLEAN=true
+      ;;
     *)
       printf '%s\n' "Invalid argument ($1)"
       exit 1
@@ -37,29 +39,27 @@ while [ "$#" -gt 0 ]; do
   shift
 done
 
-echo "$PYTHON_VERSION -> ${PYTHON_VERSION%.*}"
+if [ "$CLEAN" = true ]; then
+  rm -rf ${SCRIPT_DIR=}/docker/conda/build
+  rm -rf ${SCRIPT_DIR=}/docker/deepspeed/build
+fi
 
-SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
+$SCRIPT_DIR/docker/conda/build-conda-env.sh \
-cd $SCRIPT_DIR
-
-PYTHON_MAJOR_MINOR=${PYTHON_VERSION%.*}
-$SCRIPT_DIR/deepspeed/build-deepspeed.sh \
   --cuda-version ${CUDA_VERSION} \
-  --python-version ${PYTHON_MAJOR_MINOR} \
+  --python-version ${PYTHON_VERSION}
-  --pytorch-version ${PYTORCH_VERSION}
+
+$SCRIPT_DIR/docker/deepspeed/build-deepspeed.sh \
+  --python-version ${PYTHON_VERSION}
 
 echo "Starting docker build process using TTS model '${TTS_MODEL}' and docker tag '${DOCKER_TAG}'"
-echo "Building for CUDA $CUDA_VERSION using python ${PYTHON_VERSION} with PyTorch ${PYTORCH_VERSION}"
+echo "Building for CUDA $CUDA_VERSION using python ${PYTHON_VERSION}"
-
 
 docker buildx \
   build \
   --progress=plain \
   --build-arg TTS_MODEL=$TTS_MODEL \
-  --build-arg CUDA_VERSION=$CUDA_VERSION \
+  --build-arg ALLTALK_DIR=$ALLTALK_DIR \
-  --build-arg PYTHON_VERSION=$PYTHON_VERSION \
-  --build-arg PYTORCH_VERSION=$PYTORCH_VERSION \
   -t alltalk_beta:${DOCKER_TAG} \
   .
 
-echo "Docker build process finished"
+echo "Docker build process finished. Use docker-start.sh to start the container."

docker-start.sh

@@ -1,7 +1,10 @@
 #!/usr/bin/env bash
 
-ALLTALK_DIR="/alltalk"
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
+. ${SCRIPT_DIR=}/docker/variables.sh
+
 WITH_UI=true
+DOCKER_TAG=latest
 declare -a ADDITIONAL_ARGS=()
 
 # Parse arguments
@@ -22,6 +25,10 @@ while [ "$#" -gt 0 ]; do
     --no_ui)
       WITH_UI=false
       ;;
+    --tag)
+      DOCKER_TAG="$2"
+      shift
+      ;;
     *)
       # Allow to pass arbitrary arguments to docker as well to be flexible:
       ADDITIONAL_ARGS+=( $1 )
@@ -58,4 +65,4 @@ docker run \
   --name alltalk \
  "${DOCKER_ARGS[@]}" \
  "${ADDITIONAL_ARGS[@]}" \
-  alltalk_beta:latest &> /dev/stdout
+  alltalk_beta:${DOCKER_TAG} &> /dev/stdout

docker/conda/Dockerfile

@@ -0,0 +1,77 @@
+FROM continuumio/miniconda3:24.7.1-0 AS build
+
+ARG CUDA_VERSION=12.4.1
+ENV CUDA_VERSION=$CUDA_VERSION
+
+ARG PYTHON_VERSION=3.11.11
+ENV PYTHON_VERSION=$PYTHON_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"
+
+##############################################################################
+# 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:
+##############################################################################
+RUN <<EOR
+    CUDA_SHORT_VERSION=${CUDA_VERSION%.*}
+
+    conda create -y -n "alltalk" -c conda-forge python=${PYTHON_VERSION}
+    conda activate alltalk
+    RESULT=$( { conda install -y \
+      gcc_linux-64 \
+      gxx_linux-64 \
+      pytorch \
+      pytorch-cuda=${CUDA_SHORT_VERSION} \
+      torchvision \
+      torchaudio \
+      libaio \
+      nvidia/label/cuda-${CUDA_SHORT_VERSION}.0::cuda-toolkit \
+      faiss-gpu=1.9.0 \
+      conda-forge::ffmpeg=7.1.0 \
+      conda-forge::portaudio=19.7.0 \
+      -c pytorch \
+      -c anaconda \
+      -c nvidia ; } 2>&1 )
+
+    if echo $RESULT | grep -izq error ; then
+      echo "Failed to install conda dependencies: $RESULT"
+      exit 1
+    fi
+
+    conda clean -a && pip cache purge
+EOR
+
+##############################################################################
+# Export conda environment:
+##############################################################################
+RUN <<EOR
+  conda activate alltalk
+  conda env export > /environment-cu-${CUDA_VERSION}-cp-${PYTHON_VERSION}.yml
+EOR
+
+##############################################################################
+# Make environment file available to docker build output:
+##############################################################################
+FROM scratch
+COPY --from=build /environment-*.yml /

docker/conda/build-conda-env.sh

@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
+cd $SCRIPT_DIR
+
+. ${SCRIPT_DIR=}/../variables.sh
+
+if [ -f $SCRIPT_DIR/build/environment-cu-${CUDA_VERSION}-cp-${PYTHON_VERSION}.yml ]; then
+   echo "Environment file already exists - skipping..."
+   exit 0
+fi
+
+# Parse arguments
+while [ "$#" -gt 0 ]; do
+  case "$1" in
+    --cuda-version)
+      CUDA_VERSION="$2"
+      shift
+      ;;
+    --python-version)
+      PYTHON_VERSION="$2"
+      shift
+      ;;
+    *)
+      # Allow to pass arbitrary arguments to docker as well to be flexible:
+      echo "Unknown argument '$1'"
+      exit 1
+      ;;
+  esac
+  shift
+done
+
+echo "Building conda environment using python ${PYTHON_VERSION} with CUDA ${CUDA_VERSION}"
+
+rm -rf build # make sure to properly clean up
+mkdir -p build
+
+docker buildx \
+  build \
+  --build-arg CUDA_VERSION=$CUDA_VERSION \
+  --build-arg PYTHON_VERSION=$PYTHON_VERSION \
+  --output=$SCRIPT_DIR/build \
+  .

docker/deepspeed/Dockerfile

@@ -1,16 +1,15 @@
-ARG CUDA_VERSION=12.1.1
+FROM continuumio/miniconda3:24.7.1-0
-FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
 
-ARG PYTHON_VERSION=3.11
+ARG DEEPSPEED_VERSION=0.16.2
-ENV PYTHON_VERSION=$PYTHON_VERSION
-
-ARG PYTORCH_VERSION=2.2.1
-ENV PYTORCH_VERSION=$PYTORCH_VERSION
-
-ARG DEEPSPEED_VERSION=0.16.1
 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:
@@ -21,62 +20,49 @@ RUN mkdir -p ${STAGE_DIR}
 ##############################################################################
 # Installation/Basic Utilities
 ##############################################################################
-SHELL ["/bin/bash", "-l", "-c"]
-ENV SHELL=/bin/bash
-
 RUN <<EOR
-  apt-get update
-  apt-get install --no-install-recommends -y \
-    software-properties-common \
-    ca-certificates
-
-  add-apt-repository ppa:git-core/ppa -y # for latest git
-  add-apt-repository ppa:deadsnakes/ppa -y # for python
-
   apt-get update
   apt-get upgrade -y
   apt-get install --no-install-recommends -y \
-    build-essential \
+      espeak-ng \
-    autotools-dev \
-    pdsh \
-    cmake \
-    g++ \
-    gcc \
       curl \
       wget \
-    vim \
+      jq \
-    unzip \
+      vim
-    llvm-dev \
+
-    git \
+   apt-get clean && rm -rf /var/lib/apt/lists/*
-    python${PYTHON_VERSION}-dev \
-    libcupti-dev \
-    libaio-dev
 EOR
 
 ##############################################################################
-# Python & pip
+# Create a conda environment and install dependencies:
 ##############################################################################
+COPY build/environment-*.yml environment.yml
 RUN <<EOR
-  # Correct symlinks to use the proper python version:
+    RESULT=$( { conda env create -f environment.yml ; } 2>&1 )
-  PYTHON_MAJOR_VERSION=${PYTHON_VERSION%%.*}
-  rm -f /usr/bin/python${PYTHON_MAJOR_VERSION}
-  ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python${PYTHON_MAJOR_VERSION}
-  ln -s /usr/bin/python3 /usr/bin/python
 
-  curl -O https://bootstrap.pypa.io/pip/3.7/get-pip.py
+    if echo $RESULT | grep -izq error ; then
-  python get-pip.py
+      echo "Failed to install conda dependencies: $RESULT"
-  rm get-pip.py
+      exit 1
-  pip install --upgrade pip
+    fi
+
+    conda clean -a && pip cache purge
 EOR
 
-# Minimal dependencies needed to build deepspeed:
+##############################################################################
-RUN pip install \
+# Install python dependencies
+##############################################################################
+ENV PIP_CACHE_DIR=/tmp/pip_cache
+RUN <<EOR
+  conda activate alltalk
+
+  pip install \
       deepspeed-kernels \
-    scikit-learn \
+      scikit-learn
-    torch==${PYTORCH_VERSION}
+EOR
+
 
 ##############################################################################
-# DeepSpeed
+# Clone DeepSpeed sources
 ##############################################################################
 RUN <<EOR
   git clone https://github.com/microsoft/DeepSpeed.git ${STAGE_DIR}/DeepSpeed
@@ -86,12 +72,15 @@ RUN <<EOR
 EOR
 
 ##############################################################################
-# DeepSpeed build file
+# Create DeepSpeed build file
 ##############################################################################
 RUN <<EOR
   cat << EOF > build_deepspeed.sh
 #!/usr/bin/env bash
+source ~/.bashrc
 mkdir -p /deepspeed
+
+conda activate alltalk
 cd ${STAGE_DIR}/DeepSpeed
 DS_BUILD_OPS=1 python setup.py build_ext -j8 bdist_wheel
 mv ${STAGE_DIR}/DeepSpeed/dist/*.whl /deepspeed/
@@ -100,4 +89,7 @@ EOR
 
 RUN chmod +x /build_deepspeed.sh
 
+##############################################################################
+# Build DeepSpeed:
+##############################################################################
 ENTRYPOINT ["/build_deepspeed.sh"]

docker/deepspeed/build-deepspeed.sh

@@ -1,28 +1,17 @@
 #!/usr/bin/env bash
 
-CUDA_VERSION=12.1.1
-PYTHON_VERSION=3.11
-PYTORCH_VERSION=2.2.1
-DEEPSPEED_VERSION=0.16.1
-
 SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
 cd $SCRIPT_DIR
 
+. ${SCRIPT_DIR=}/../variables.sh
+
 # Parse arguments
 while [ "$#" -gt 0 ]; do
   case "$1" in
-    --cuda-version)
-      CUDA_VERSION="$2"
-      shift
-      ;;
     --python-version)
       PYTHON_VERSION="$2"
       shift
       ;;
-    --pytorch-version)
-      PYTORCH_VERSION="$2"
-      shift
-      ;;
     --deepspeed-version)
       DEEPSPEED_VERSION="$2"
       shift
@@ -36,24 +25,30 @@ while [ "$#" -gt 0 ]; do
   shift
 done
 
-PYTHON_VERSION_NO_DOT=${PYTHON_VERSION//./}
+PYTHON_SHORT_VERSION=${PYTHON_VERSION%.*}
+PYTHON_VERSION_NO_DOT=${PYTHON_SHORT_VERSION//./}
 if [[ -n $(find build -name "deepspeed-${DEEPSPEED_VERSION}*-cp${PYTHON_VERSION_NO_DOT}-cp${PYTHON_VERSION_NO_DOT}-*.whl") ]]
 then
    echo "DeepSpeed was already built - skipping..."
    exit 0
 fi
 
-echo "Building DeepSpeed $DEEPSPEED_VERSION for CUDA $CUDA_VERSION using python ${PYTHON_VERSION} with PyTorch ${PYTORCH_VERSION}"
+CONDA_ENV=${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
 
-rm -rf build # make sure to properly clean up - we only want 1 wheel at the time
+echo "Building DeepSpeed $DEEPSPEED_VERSION using python ${PYTHON_VERSION}"
+
+rm -rf build # make sure to properly clean up
 mkdir -p build
+cp ${CONDA_ENV} ${SCRIPT_DIR=}/build
+
 docker buildx \
   build \
-  --build-arg CUDA_VERSION=$CUDA_VERSION \
-  --build-arg PYTHON_VERSION=$PYTHON_VERSION \
-  --build-arg PYTORCH_VERSION=$PYTORCH_VERSION \
   --build-arg DEEPSPEED_VERSION=$DEEPSPEED_VERSION \
-  -t deepspeed:cu-$CUDA_VERSION-ds-$DEEPSPEED_VERSION \
+  -t deepspeed:$DEEPSPEED_VERSION \
   .
 
 docker run \
@@ -62,4 +57,4 @@ docker run \
   --gpus=all \
   --name deepspeed \
   -v $SCRIPT_DIR/build:/deepspeed \
-  deepspeed:cu-$CUDA_VERSION-ds-$DEEPSPEED_VERSION
+  deepspeed:$DEEPSPEED_VERSION

docker/variables.sh

@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+
+export CUDA_VERSION=12.4.1
+export PYTHON_VERSION=3.11.11
+export DEEPSPEED_VERSION=0.16.2
+export ALLTALK_DIR=/opt/alltalk