Add GGML backend for GAME transcription

[KakaruHayate]

Summary

• add a pluggable GAME inference backend abstraction while preserving the existing ONNX implementation
• add a native GGML backend using the long-lived game_ggml_cli serve protocol
• support independent GGML-only installation through a game-ggml-medium .oudep package
• add an ONNX/GGML backend selector in Preferences with automatic fallback to an installed backend
• load the active backend's own config.json, so GGML does not require the ONNX GAME package
• handle process cancellation, stderr draining, Unix executable permissions, and paths containing spaces

Origin & Credits

This GGML backend grew out of KCKT0112/GAME-ggml
(native C++/ggml GAME inference — CPU/Metal/CUDA/Vulkan) and its web variant
KCKT0112/web-game (Web Generative
Adaptive MIDI Extractor). The model, operator set, and weight layout ported
there form the basis of the game.cpp
backend; the GAME PyTorch original remains
the parity reference.

Motivation

The GGML implementation runs GAME without ONNX Runtime model sessions or a Python runtime. The native backend supports Vulkan/CUDA on Windows and Linux, plus Metal on macOS, and keeps one model process alive for all chunks in a transcription request.

The backend is implemented in a separate project: KakaruHayate/game.cpp. It uses ggml and packages the native executable, runtime libraries, medium GGUF model, and GAME config as one OpenUtau dependency package.

Official packages (v0.1.1)

Ready-to-install .oudep files are published to the
KakaruHayate/game.cpp Releases:

platform	full (F32)	q8 (Q8_0)
Windows x64	game_ggml-windows-x64-vulkan.oudep + -cuda.oudep	same names, -q8 suffix
Linux x64	...-linux-x64-vulkan.oudep + ...-cuda.oudep	same
Linux x64	...-linux-x64-cpu.oudep	same
macOS	game_ggml-macos-{arm64,x64}-metal.oudep	same

Download the archive for your platform, unzip it in OpenUtau's Dependencies
folder at Dependencies/game-ggml-medium/ (both the CLI and the GGUF live in
one package), restart OpenUtau, and select the GGML backend in Preferences.

Recommended GGML configuration

Full measurements in docs/benchmark-7channel.md
(60 s audio, no slicing, nsteps=8, frame-level metrics vs the torch-CUDA fp32
no-cache baseline — every EP ≥ 0.97 RPA):

platform	GPU	weights	EP	notes
Windows	NVIDIA	F32	CUDA (fallback Vulkan)	6–8 s wall
Windows	Intel/AMD/etc	F32	Vulkan	4–5 s, smallest VRAM (+0.3 GiB)
Windows	integrated	F32/Q8	CPU + DBCache	48.5→27.1 s (−44 %), quality-neutral
Linux	NVIDIA	F32	CUDA (fallback Vulkan)	as Windows CUDA
Linux	Nouveau/AMD	F32	Vulkan	as Windows Vulkan
macOS	Apple Silicon	F32	Metal (only EP)	CI-built
macOS	Intel	F32	Metal (cross-compiled)	CI-built

Rules of thumb:

• GPU present → F32 weights; CPU-only → CPU EP with DBCache on by default (−44 % wall, quality-neutral).
• On GPU, Vulkan = smallest VRAM, CUDA = fastest wall.
• Q8 saves ~3.4× VRAM/RAM (near-lossless quality) but may flip a boundary note on Vulkan — use the -full package if you need bit-consistent output; both are valid in practice.
• No user escalation needed: CLI picks backend from GGUF/EP and cache defaults by EP (GPU off, CPU auto).

Design

• IGameBackend isolates model lifecycle, inference, cancellation, and configuration.
• GameOnnxBackend is a behavior-preserving extraction of the existing four-model ONNX pipeline and retains native batching.
• GameGgmlBackend starts game_ggml_cli serve <model.gguf> lazily, sends framed float32 waveform requests through stdin, and reads one-line JSON note responses from stdout.
• GameBackendFactory resolves the user preference and falls back to an installed backend (ONNX first for compatibility).
• GGML installation is considered valid only when the CLI, GGUF model, and adjacent config.json are all present.

Verification

On current openutau/OpenUtau:master:

• GameGgmlBackendTest: 3 passed
• dotnet test OpenUtau.Test: 209 passed
• dotnet build OpenUtau.sln: 0 warnings, 0 errors
• the Windows x64 Vulkan artifact from the linked run was launched with serve and returned {"type":"ready"}
• the CUDA artifacts are built with CUDA Toolkit 12.6.3 (release architecture 7.5 with forward-compatible PTX) and verified for Toolkit discovery, compilation, linking, packaging, and PE import-table integrity; the GitHub-hosted runner has no NVIDIA driver, so the CUDA startup smoke test runs on a machine with an NVIDIA GPU
• Windows and Linux artifact layouts and oudep.yaml manifests were checked against OpenUtau PackageManager expectations
• CI (game.cpp @ main + v0.1.0) is green: prepare-model(F32+Q8) → 7 platform builds → 14 oudep packages → release

Notes

• ONNX remains the default and preferred fallback when both packages are installed.
• GGML currently processes one chunk per request; OpenUtau disables batch dispatch for this backend.
• The .oudep package id is game-ggml-medium, independent from the existing ONNX game package.