This release is v0.10.0, codenamed "See You In The Afterlife", which is also the tenth track of our debut album. Since Sayuri’s passing in 2024, we wish to dedicate this version as a heartfelt tribute to her memory.
Changes in v0.10.0
- Add support for V5 networks, including the NestedBottleneck block and the RepLK policy head.
- Add benchmark mode, allowing you to tune optimal parameters on your own hardware.
- Add support for older network versions. The engine can now use UEC14 weights.
- Add support for automatic mixed precision.
- Add support for more analysis commands in GoGUI.
- Windows users can now compile the code using a .bat file.
- Set the default virtual loss to 1.
- Fixed bugs related to the territory scoring rule in training.
- Clean some code.
Benchmark
You can now tune the optimal threads and batch size on any computer. For example, to test whether (threads=12, batch size=6) or (threads=24, batch size=12) performs better, you can run:
./sayuri -w weights.txt --mode benchmark --benchmark-query tbg:12:6:10 --benchmark-query tbg:24:12:10
The program will output the estimated Elo score for each configuration.
Run since 2025-4-17
The latest weights adopt the Nested Bottleneck network architecture and fix issues related to Japanese rules in training. This set of weights outperforms all previous versions. In particular, sayuri-b12xc384nbt-s3592000-c4310000-w426561-swa.bin.txt is significantly stronger than the version used in UEC16, with a strength roughly equivalent to LZ226 (about 200 Elo stronger than ELF v2). This is currently the strongest set of weights we have trained.
- Download all weights here
- CGOS results: sayuri_b12s359_p1k