burrowdweller/rook-chess-v1-wgsl

rook-chess-v1 (WebGPU)

The rook-chess-v1 chess engine with a faster brain: the same network, run on hand-written WebGPU code, so it searches more moves in the same time.

Same weights and same search as rook-chess-v1 — only the evaluation step changes. The original runs the network through ONNX Runtime; this version runs it on small hand-written graphics shaders. In a 20-pair equal-time match it scored 0.825 against rook-chess-v1 (27 wins, 12 draws, 1 loss). Needs a browser with WebGPU; without one, or if the graphics device is lost mid-game, it switches once to ONNX Runtime and keeps playing.

What's in this repo

This is a browser chess engine package (chess-gpt-package-v1), not a Transformers checkpoint.

A compatible runner loads:

• browser/manifest.json — file list and hashes
• browser/entry.js — search and move picker
• the ONNX net named in that manifest

Use

Point a chess-gpt-compatible runner at burrowdweller/rook-chess-v1-wgsl. It calls loadPackage → newGame → chooseMove({ history, legalMoves }) and must return one of those legal moves.

The net is small on purpose. Strength comes from searching at move time.

Source: junisbuilding/chessdb