AlphaDiamond is the C++26 engine and training repository for Soo and Min. It owns the rules, encoders, MCTS, inference, self-play, replay, training, checkpoint and model release tools. The Qt/QML application lives in the separate AlphaDiamondDesktop repository and consumes an installed Runtime SDK.
Start with the documentation index, then read the
code guide and
build cycle. For training, use the
training panel guide and the current configurations
in configs/current/.
| Folder | Purpose |
|---|---|
src/ |
C++ implementations, executable entrypoints and target definitions |
include/ |
Engine, inference, training and orchestration interfaces |
benchmarks/ |
Optional diagnostics, probes and evaluation entrypoints |
cmake/ |
CMake package and build support |
configs/current/ |
Seven supported panel training presets |
configs/experiments/ |
Earlier and explicit CLI experiment configurations |
tools/ |
Build, model publication and research helpers |
training-panel/ |
Local training monitor and launcher |
data/ |
Canonical topology and runtime data |
models/ |
Shipped deployment models, weights and model index |
docs/ |
Current guides, contracts and historical documentation |
archive/ |
Preserved benchmark runs, Python runtime data and acceptance evidence |
build/<preset>/ |
Local build trees; executables are in bin/ |
artifacts/ |
Local runs, checkpoints, exports and optional downloads |
graphify-out/ |
Generated architecture graph |
The former native/ source wrapper is removed. Historical paths and measurements
are documented in the archive map. The former
TrainAlphaDiamond/ contained storage documentation and two manifests, rather
than another codebase or local bucket contents; those files now live in
docs/archive/storage/.
Use a C++26 compiler, CMake 3.21+ and Ninja. Linux development uses GCC 15,
sccache and mold. Debug uses -O0 -g1 -pipe, and LTO defaults to off. Make
selects workers from CPU and memory limits; override BUILD_JOBS when needed.
make build
# Add a compatible standalone LibTorch prefix for inference or training.
export CMAKE_PREFIX_PATH=/path/to/libtorch
make build PRESET=training-linux
make build PRESET=asanThe default Linux preset is core-linux. Portable Debug presets are core,
runtime and training; optimized final outputs use release or
training-release. Windows toolchain setup uses tools/cmake.sh.
Executables are written to build/<preset>/bin/, for example
build/training-linux/bin/alphadiamond-train.
Benchmarks and probes require -DDIAMOND_BUILD_BENCHMARKS=ON; ordinary training
and Runtime SDK builds omit them. Select model and training products with
DIAMOND_BUILD_MODEL and DIAMOND_BUILD_TRAINING. Direct builds should supply
--parallel "$(tools/build_resources.sh build)" and keep nested OpenMP/MKL threads
within that resource budget. Existing regression suites and GitHub CI have been
removed; the build guide explains validation.
Runtime consumers link the shared core, search, support, artifact, rating and model targets. The SDK excludes trainer, replay orchestration and release tools.
cmake --preset runtime-linux -DCMAKE_PREFIX_PATH=/path/to/cpu-libtorch
cmake --build --preset runtime-linux --parallel "$(tools/build_resources.sh build)"
cmake --install build/runtime-linux --component Runtime --prefix /path/to/sdkRebuild the complete runtime graph, including engine_sdk_provenance, before
installation when source content or HEAD changes. SDK manifests validate source
revision, configuration, compiler/Torch ABI and file hashes. Desktop must consume
a matching Debug or Release SDK. See
repository split and SDK upgrades.
alphadiamond-train owns CPU/CUDA training, transactional checkpoint-v3 and exact
resume. alphadiamond-checkpoint validates and exports checkpoints;
alphadiamond-release validates and stages format-v3 deployment artifacts.
The seven current presets are three Min expert local pilots, Min 512x12 from
scratch, and Soo bootstrap, production and six-hour runs. The larger Min uses
width 512, 12 residual blocks, all six turn orders, vacancy-prior weight 0.5 and a
three-hour iteration-boundary budget. A strict wall-clock limit also needs a
process/service cap because the native budget finishes the current iteration.
Earlier Min presets and the expert arena configuration are in
configs/experiments/.
Set run_seed for production self-play: game seeds derive from run, iteration
and job identities. Replay and optimizer seeds remain separate. Resume with the
run's saved configuration, preserving immutable metadata. The standalone
selfplay_benchmark --config also uses run_seed; --seed overrides it.
Reproducing an old benchmark may require its recorded self_play.seed as that
explicit override; equal seeds alone do not imply identical production games.
The training panel displays active CLI and Agent jobs,
configuration, output, loss and performance. To run it, execute
npm ci, npm run build and npm start inside training-panel/, then open
http://127.0.0.1:5178.
Model publishing keeps models/index.json, GitHub ubraket513/alphadiamond and
Hugging Face ubraket513/AlphaDiamond. Desktop's Git remote does not change model
URLs. For optional local bucket downloads, use artifacts/downloads/; the
storage guide explains native checkpoint
validation and the historical manifests.
Continue with product ownership, format-v3 model contracts and goal-aware training. Dated plans, measurements and original command transcripts are historical evidence; use the current build guide for supported commands.