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

[tempstudio]

Cosyvoice2-0.5B (https://github.com/FunAudioLLM/CosyVoice/blob/main/cosyvoice/vllm/cosyvoice2.py#L93) is a TTS model finetuned on top of the Qwen2-0.5B model, with an extra bias tensor on the decoder head.

This change allows this bias tensor to be loaded for better quality when running Cosyvoice2 in llama.cpp.

[hipudding]

@tempstudio Hello, may I ask how I should load and run inference with the cosyvoice2 model? Thanks.

[hipudding]

@tempstudio I was thinking about it — this approach only allows llama.cpp to be used as a library, correct? It can’t yet run inference for CosyVoice directly, as preprocessing and postprocessing would still need to be handled by external code.

[tempstudio]

@hipudding Yes you are right, this is a replacement for the LLM part of Cosyvoice. It doesn't cover the FLOW and HIFIGAN parts of Cosyvoice.

[hipudding]

@hipudding Yes you are right, this is a replacement for the LLM part of Cosyvoice. It doesn't cover the FLOW and HIFIGAN parts of Cosyvoice.

Thanks. Do you have plans to fully support CosyVoice?