Linux build instructions

[paboum]

I'm lazy so this is not a process documentation, but rather a workaround for project documentation lacking Linux build instructions.

git clone https://github.com/JonathanFly/bark #bark-infinity release zipfile is useless, only .bat files, depend on miniconda
cd bark/
cp old_setup_files/requirements-pip.txt . #docker fails to build without it, fails in step 8 or so
sed "s/python3-pip /python3-pip libpython3.10-dev gcc g++ ffmpeg //" Dockerfile #docker fails due to x86_64-linux-gnu-gcc missing while building fairseq, also ffmpeg is needed and no linux users would ever need ffmpeg-downloader for this
sed "s/rich-argparse/rich-argparse devtools/" Dockerfile # otherwise it won't run
sed "s/devtools/devtools ffprobe ffmpeg" Dockerfile # otherwise it won't generate
sed 's/"bark_webui.py"/"bark_webui.py", "--share"/' Dockerfile # for some reason opening http://127.0.0.1:7860 will just cause connection reset error but public URL will work
docker build -t bark-infinity:latest .

I don't recommend using this as a shell script or anything like that. It wastes a lot of system resources to build this way, but at least it builds and yields a workable end-user environment on Linux that-won't-break-that-much if the author decides to update anything. Just use it as a set of gotchas needed to build it.

To run it, nvidia-container-toolkit needs to be installed on your host so that docker can use gpu.

Note: it costed me over 2 hours of frequent docker image rebuilding to try and fail until getting my first audio generated on Linux.

[JonathanFly]

On linux I think conda/mamba is the way to go. I ran into windows specific problems with it which why the installer went back to pip but on my personal system I use conda in WSL. You can ping me in Discord for Bark same name questions, hang out often.