...It uses ONNX Runtime and other backends to accelerate inference, with notes on how different execution providers such as CUDA or DirectML affect operator support and numerical stability. Recent versions integrate with fish-speech via a dedicated fish-speech.cpp subproject using ggml.