Some test GGUFs
#2
by dignome - opened
For anyone that wants to test this:
A new NVFP4 GGUF and a few CPU optimized GGUFs that are capable of realtime/near-realtime (RTF).
TQ2_1 CPU (uses default decoder best quality)
TQ2_1 CPU w8 (uses student w8 decoder)
TQ2_1 CPU w4 (uses student w4 decoder)
The NVFP4 GGUF is usable on RTX 50, 40 or possibly earlier CUDA platforms and has been tested on RTX 4060 Ti (uses compat but still faster than previously published GGUFs).
Use this branch on my fork of audio.cpp:
https://github.com/dignome/audio.cpp-custom/tree/kitten-tts2-nvfp4-tq2-student
Weights will be uploaded here shortly.
The GGUF and code are experimental and subject to change at any time until merged.
Feel free to leave comments here and I will consider making a merge request soon on it.