[Feature] Support ChatGLM3-6B-Base

[LZHgrla]

Due to the injection attack prevention of ChatGLM3 tokenizer, the default prompt_template is currently used to fine-tune ChatGLM3-6B-Base.

TODO (in other PRs, future)

• ChatGLM3 prompt_template fine-tuning
• Support ChatGLM3-6B (a chat model)

Quick Start

1. Fine-tune

xtuner train ${CONFIG}

2. Convert

xtuner convert pth_to_hf ${CONFIG} ${PTH} ${SAVE_PATH}

3. Chat

xtuner chat THUDM/chatglm3-6b-base --adapter ${ADAPTER_PATH} --prompt-template default --system-template ${SYSTEM_TEMPLATE}