app.py

from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
from transformers_stream_generator.main import NewGenerationMixin, StreamGenerationConfig
from threading import Thread
import gradio as gr
import torch


# lora_folder = ''
# model_folder = ''
#
# config = PeftConfig.from_pretrained(("Junity/Genshin-World-Model" if lora_folder == ''
#                                      else lora_folder),
#                                     trust_remote_code=True)
# model = AutoModelForCausalLM.from_pretrained(("baichuan-inc/Baichuan-13B-Base" if model_folder == ''
#                                               else model_folder), torch_dtype=torch.float32, trust_remote_code=True)
# model = PeftModel.from_pretrained(model,
#                                   ("Junity/Genshin-World-Model" if lora_folder == ''
#                                    else lora_folder)
#                                   , torch_dtype=torch.float32, trust_remote_code=True)
# tokenizer = AutoTokenizer.from_pretrained(("baichuan-inc/Baichuan-13B-Base" if model_folder == ''
#                                            else model_folder),
#                                           trust_remote_code=True)
# history = []
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# if device == "cuda":
#     model.cuda()
#     model = model.half()


def respond(role_name, character_name, msg, textbox, temp, rep, max_len, top_p, top_k):
    if textbox != '':
        textbox = (textbox
                   + "\n"
                   + role_name
                   + ("：" if role_name != '' else '')
                   + msg
                   + ('。\n' if msg[-1] not in ['。', '！', '？'] else ''))
        yield ["", textbox]
    else:
        textbox = (textbox
                   + role_name
                   + ("：" if role_name != '' else '')
                   + msg
                   + ('。' if msg[-1] not in ['。', '！', '？', ')', '}', '：', ':', '('] else '')
                   + ('\n' if msg[-1] in ['。', '！', '？', ')', '}'] else ''))
        yield ["", textbox]
    if character_name != '':
        textbox += ('\n' if textbox[-1] != '\n' else '') + character_name + '：'
    input_ids = tokenizer.encode(textbox)[-4096:]
    input_ids = torch.LongTensor([input_ids]).to(device)
    generation_config = model.generation_config
    stream_config = StreamGenerationConfig(**generation_config.to_dict(), do_stream=True)
    gen_kwargs = {}
    gen_kwargs.update(dict(
        input_ids=input_ids,
        temperature=1.5,
        top_p=0.7,
        top_k=50,
        repetition_penalty=1.0,
        max_new_tokens=256,
        do_sample=True,
    ))
    outputs = []
    print(input_ids)
    streamer = TextIteratorStreamer(tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True)
    gen_kwargs["streamer"] = streamer

    thread = Thread(target=model.generate, kwargs=gen_kwargs)
    thread.start()

    for new_text in streamer:
        textbox += new_text
        yield ["", textbox]


with gr.Blocks() as demo:
    gr.Markdown(
        """
        ## Genshin-World-Model
        - 模型地址 [https://huggingface.co/Junity/Genshin-World-Model](https://huggingface.co/Junity/Genshin-World-Model)
        - 此模型不支持要求对方回答什么，只支持续写。
        - 目前运行不了，因为没有钱租卡。
        """
    )
    with gr.Tab("创作") as chat:
        role_name = gr.Textbox(label="你将扮演的角色(可留空)")
        character_name = gr.Textbox(label="对方的角色(可留空)")
        msg = gr.Textbox(label="你说的话")
    with gr.Row():
        clear = gr.ClearButton()
        sub = gr.Button("Submit", variant="primary")
    with gr.Row():
        temp = gr.Slider(minimum=0, maximum=2.0, step=0.1, value=1.5, label="温度(调大则更随机)", interactive=True)
        rep = gr.Slider(minimum=0, maximum=2.0, step=0.1, value=1.0, label="对重复生成的惩罚", interactive=True)
        max_len = gr.Slider(minimum=4, maximum=512, step=4, value=256, label="对方回答的最大长度", interactive=True)
        top_p = gr.Slider(minimum=0, maximum=1.0, step=0.1, value=0.7, label="Top-p(调大则更随机)", interactive=True)
        top_k = gr.Slider(minimum=0, maximum=100, step=1, value=50, label="Top-k(调大则更随机)", interactive=True)
    textbox = gr.Textbox(interactive=True, label="全部文本(可修改)")
    clear.add([msg, role_name, textbox])
    sub.click(fn=respond,
              inputs=[role_name, character_name, msg, textbox, temp, rep, max_len, top_p, top_k],
              outputs=[msg, textbox])
    demo.queue().launch(server_port=6006, share=True)