Porting/chaiNNer

Installation on Debian Bookworm

• Install at least one of PyTorch, ONNX, or ncnn
• Install Electron
• Pin Debian's python3 so that Electron will use Debian's PyTorch package (See Bedrock pinning instructions)
• sudo brl apply
• Download and extract a portable zip Linux x86_64 binary of chaiNNer.
• Rename chainner executable in chaiNNer directory to chainner.bak (this is an Electron executable; we will replace it with the ppc64le Electron that we installed).
• Rename portable file in chaiNNer directory to portable.bak (skip this step if you want chaiNNer configuration to be stored in chaiNNer's directory instead of $HOME)
• In chaiNNer's directory, create the following shell script named chainner:

#!/usr/bin/env bash

cd "$(dirname "$0")"
electron ./resources/app "$@" --no-sandbox

• Give the chainner script executable permissions, and (if you like) add the chaiNNer directory to your PATH.
• Run chaiNNer once, and quickly exit.
• Run the following commands to take care of some remaining dependencies:

echo true > ~/.config/chaiNNer/settings/use-system-python
ln -s -T /bedrock/cross/bin/ffprobe ~/.config/chaiNNer/ffmpeg/ffprobe
ln -s -T /bedrock/cross/bin/ffmpeg ~/.config/chaiNNer/ffmpeg/ffmpeg
sudo apt install python3-requests python3-wcmatch libre2-dev
pip install --user --break-system-packages sanic Sanic-Cors ffmpeg-python google-re2

• chaiNNer is installed!

Finished

• --no-sandbox Electron flag causes ChaiNNer to error
• Fix argument parsing to allow passthrough of extra args
• Fix backend script path detection with external Electron
• Add --devtools argument
• Support NCNN CPU inference
• Use load_param_mem with ncnn CPU inference

Merged, Awaiting Stable Tag

(None at the moment.)

In Progress

• Make ncnn memory budget configurable

See Also

• Electron
• ncnn
• ONNX
• PyTorch