From 68eb3d3442bf0bf117837db8fb550604c5409b7c Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Sun, 16 Aug 2026 23:53:26 +0800 Subject: [PATCH] docs: add recommended GGML EP config comment (sync into allinone3) --- OpenUtau.Core/Analysis/GameGgmlBackend.cs | 20 ++++++++++++++++++++ 1 file changed, 20 insertions(+) diff --git a/OpenUtau.Core/Analysis/GameGgmlBackend.cs b/OpenUtau.Core/Analysis/GameGgmlBackend.cs index 22d1d0eae..419dab54e 100644 --- a/OpenUtau.Core/Analysis/GameGgmlBackend.cs +++ b/OpenUtau.Core/Analysis/GameGgmlBackend.cs @@ -26,6 +26,26 @@ namespace OpenUtau.Core.Analysis; /// One instance owns exactly one subprocess; the /// subclass disposes it after the whole transcription. /// +/// +/// Recommended GGML configuration (7-channel 60 s benchmark, nsteps=8, +/// frame-level metrics vs torch-CUDA fp32 no-cache baseline — all EPS ≥ 0.97 RPA): +/// +/// | platform | GPU | weights | EP offered by oudep pkg | +/// |---------------|---------------|---------|--------------------------| +/// | Windows | NVIDIA | F32 | CUDA (fallback Vulkan) | +/// | Windows | Intel/AMD/etc | F32 | Vulkan | +/// | Windows | integrated | F32/Q8 | CPU (+ DBCache): 0.44x wall | +/// | Linux | NVIDIA | F32 | CUDA (fallback Vulkan) | +/// | Linux | Nouveau/AMD | F32 | Vulkan | +/// | macOS | Apple Silicon | F32 | Metal (only EP) | +/// | macOS | Intel | F32 | Metal (cross-compiled) | +/// +/// Rules of thumb (no user config required — CLI picks backend from GGUF/EP): +/// * GPU present → F32 weights; CPU-only → CPU EP with DBCache on by default +/// * on GPU, Vulkan is our smallest-VRAM (≈+0.3 GiB) and CUDA the fastest +/// * Q8 saves VRAM/RAM (~3.4x smaller weights) at near-lossless quality but may +/// flip a boundary note on Vulkan — prefer the -full package if you need +/// bit-consistent output; both are equally valid in practice. public class GameGgmlBackend : IGameBackend { private const string PackageId = "game"; // Single oudep package contains both the CLI binary and the GGUF weights.