llama.cpp

Support Cosyvoice2-0.5B By allowing Qwen2 architecture to have a optional bias tensor

#14711

Merged