Devre is a strong open-source UCI-compatible chess engine written in C++. While writing the engine, I got great help from chessprogramming wiki, talkchess forum, Stockfish discord, and some open-source engines: Ethereal, Vice, and Koivisto.
Devre uses a NNUE for evaluation. The Network architecture is 768 -> (1536x2) -> 1. The default net was trained with Leela data. The training code is written in C/CUDA and can be found at https://github.com/OmerFarukTutkun/CUDA-Trainer. The training resources and other useful information about NNUE can be found in Stockfish Discord. Thanks to the Stockfish and Leela teams for publishing their training data publicly.
To compile in Linux/Windows with a CPU that supports AVX512/AVX2/SSSE3:
make
make build=avx512
make build=avx2
make build=ssse3