leela-zero mit Homebrew, apt, Nix installieren

Projektlinks

• Homepage
• Repository

Befehle und Aliase

• leelaz

Projektgeschichte

The project appeared after the AlphaGo Zero paper as a public reimplementation effort. Its README explicitly framed the engine as a faithful open-source implementation of the AlphaGo Zero system, while also noting that the repository did not contain trained network weights.

A major part of the project was the distributed AutoGTP training loop: volunteers ran clients that generated self-play games, uploaded results, and helped train stronger networks. The official training site records 21,219,288 self-play games and 1,264,007 match games before the server shut down on 2021-02-15.

The release line evolved through practical engine work: OpenCL/GPU support, GTP integration, GUI-facing analysis commands, batching, Tensor Core support, Winograd optimizations, and AutoGTP features. Version 0.17 was released on 2019-04-04.

Adoptionsgeschichte

Leela Zero reached adoption well beyond a niche package: the GitHub repository shows thousands of stars and forks, the README documents Homebrew installation for macOS, and the official site linked to community successor projects after training ended.

Go players and engine users usually paired it with GTP-compatible GUIs rather than running the engine directly. The README names Lizzie, Sabaki, LeelaSabaki, and GoReviewPartner as related clients or review tools.

Wie es verwendet wird

Leela Zero is used as a Go engine through the Go Text Protocol. A user supplies a neural-network weights file, enables GTP with `--gtp`, and connects the engine to a GUI or analysis tool for playing, review, and study.

During the distributed-training phase, contributors ran AutoGTP to generate self-play or match games for the public training effort.

Warum Paket-Nerds sich dafür interessieren

For package people, Leela Zero is a snapshot of the post-AlphaGo open-source AI-engine wave: C++14, OpenCL, GPU tuning, neural-network weights as separate data artifacts, and package-manager builds that had to expose a command-line engine while leaving training data outside the formula.

It also links Go-engine packaging with AI reproducibility culture: the executable, weights, GUI clients, training server, and community forks are all separate moving parts.

Zeitleiste

• 2017: AlphaGo Zero paper prompted open-source reimplementation efforts including Leela Zero.
• 2018-12-08: GitHub releases page shows Leela Zero 0.9 + AutoGTP v8.
• 2019-04-04: Version 0.17 + AutoGTP v18 was released with batching, Tensor Core support, and search/analysis improvements.
• 2021-02-15: Official training server shut down after the project finished its distributed training run.

Related projects

• The README links Leela Zero to AlphaGo, AlphaGo Zero, AlphaZero, Lizzie, Sabaki, LeelaSabaki, GoReviewPartner, Leela Chess Zero, and lc0.
• The official training site points former contributors toward SAI and KataGo as successor-style Go AI projects.

Quellen

• https://github.com/leela-zero/leela-zero
• https://github.com/leela-zero/leela-zero/releases
• https://zero.sjeng.org/