llama.cpp

1bd57a3b - convert: add dequant function for compressed_tensor (kimi-k2-thinking)