I failed to get it working with the default script, something always went wrong, so I tweaked it to prefer the platform-independent .whl files over other archives, and that finally allowed me to build a flatpak :)
See the Makefile for local test builds.
For torch version 2.1.0 I had no issues running the flatpak, but users reported a strange error that indicated a version incompatibility between torch and torchvision. The versions matched, but I did notice that upon startup, torch output an error message about missing some libjpg library which apparently would've been caused by a bad compilation.
To figure out why my version worked, I created a fresh venv and installed torch using the pip command from https://pytorch.org/get-started/locally/: pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cpu Rather than using the index site myself. Pip grabbed a different file than the one on the index site, torchvision remained the same. The new file worked in the venv, so I incorporated that into the flatpak.
this package would be required to provide support for the --notify command line option, but in the gui we don't need to support that, and since it only causes strange problems, it shall be forcefully ignored.
Simple lama inpainting introduces many optional nvidia dependencies. Be sure to nuke them from auto-generated pypi dependencies.