mach9243

/

s2-pro-gguf

like 10

Thanks for releasing the quantized S2 Pro models and the original s2.cpp implementation.

I forked the project and added CUDA support for both the transformer and audio codec:

https://github.com/ryuryu0x08/s2.cpp

On my RTX 5060 Ti with the Q8 models, CUDA reduced the total RTF from about 1.11 to 0.76 compared with Vulkan, making generation faster than real time.

Thanks again for the original work.