Skip to content

Latest commit

 

History

History
84 lines (65 loc) · 4.99 KB

README_Windows_Dockerfile_tldr.md

File metadata and controls

84 lines (65 loc) · 4.99 KB
Raw

Note The purpose of this document is to capture the thought process and steps taken to build the Windows Dockerfile. This was a new experience for me. I stepped into many rabbit holes and wanted to capture some of the lessons learned for reference. spoiler alert - A big AHA moment was understanding the power in Linux-based Docker containers vs the agony of attempting to build a native Windows Docker container. AND installing the software goop required to gain access to the GPU is a R-E-A-L S-L-O-G. The amount of wasted hours - lifetimes of hours - by all of us who go through it must be staggering.

Windows Dockerfile

The Windows Dockerfile, dockerfile.wsl, runs the Obsidian Transcriber Service (OTS) using GPUs when available. The OTS is dependent on faster-whisper which is a layer on top of the ctranslate2 Python package to provide audio to text translation. The WSL2 compatibility layer is used so that Linux-based Docker containers (like those from NVidia and others) can run directly on the WSL2 layer.

The biggest challenge in building the OTS docker image is understanding what cuda components to install and where to install them. Looking at NVidia's cuda layers:

  NVidia software components

The CUDA toolkit and CUDA driver need to be available for ctranslate2 to access the GPUs.

Keeping that in mind, let's look at the contents of the Docker image.

Building the Windows Dockerfile

The boxes in this image are intended to give a good idea on how the Windows Dockerfile is put together:

  Windows Docker container

### Start with a Good Base - the ctranslate2 Docker Image

ctranslate2 dockerfile

A Yoga lesson I learned awhile back was to start with a Good Base. This project relies on ctranslate2. The ctranslate2 github includes a dockerfile that will be the base docker image used for the OTS dockerfile.

The ctranslate2 image relies on NVidia's base image. The NVidia base image includes :

  • Ubuntu 20.04 LTS is used as the container base image (AKA "Container OS User Space" in the NVidia image). This also includes Python 3.8
  • The CUDA Toolkit (see NVidia image) which include the cudnn and cuBLAS libraries used by ctranslate2 for GPU processing. The output of this image includes availability of the software installed by the NVidia image as well as ctranslate2 .

On to the OTS dockerfile.

OTS dockerfile

The OTS dockerfile:

  • Updates Python 3.8 to Python 3.12.

Note I had done development with Python 3.12. Errors occurred running the code using Python3.8 that did not occur using Python3.12.

  • Installs ffmpeg which is used by yt_dlp when creating the audio file to be used for transcription.
  • Installs the Obsidian Transcriber code.

Installing Python 3.12

There are two ways to do this:

  1. Use the deadsnakes PPA to install Python 3.12. Two resources that were helpful in understanding the process are:

  2. Build Python 3.12 from source for Ubuntu 20.04. A helpful resource is:

I ended up using the deadsnakes PPA to install Python 3.12. This was the easiest and created a smaller image.

Installing from Deadsnakes PPA Installation

The steps I used to build Python 3.12 from the deadsnakes PPA are:

RUN apt-get update && \
    apt-get install -y -qqq software-properties-common && \
    ppa:deadsnakes && \
    apt-get install -y -qqq python3.12
Installing from Source
# Update system and install required dependencies
RUN apt-get update -y && apt-get upgrade -y && \
    apt-get install -y build-essential libssl-dev zlib1g-dev libbz2-dev \
    libreadline-dev libsqlite3-dev wget curl llvm libncurses5-dev libncursesw5-dev \
    xz-utils tk-dev libffi-dev liblzma-dev python-openssl git && \
    rm -rf /var/lib/apt/lists/*

# Build Python 3.12
RUN wget https://www.python.org/ftp/python/3.12.0/Python-3.12.0.tgz && \
    tar -xf Python-3.12.0.tgz && \
    cd Python-3.12.0 && \
    ./configure --enable-optimizations && \
    make -j 8 && \
    make altinstall && \
    python3.12 --version && \
    cd .. && \
    rm -rf Python-3.12.0 Python-3.12.0.tgz

CUDA Driver

The final step to enable GPU access is installing the NVidia CUDA drivers on the Windows PC. The code and directions can be found on the NVidia's page, CUDA on WSL. The drivers must be installed before ctranslate2 can access the GPU.