app.py

import gradio as gr
from gpt4all import GPT4All
from huggingface_hub import hf_hub_download
import subprocess
import asyncio

title = "Apollo-7B-GGUF Run On CPU"
 
description = """
🔎 [Apollo-7B](https://huggingface.co/FreedomIntelligence/Apollo-7B) [GGUF format model](https://huggingface.co/FreedomIntelligence/Apollo-7B-GGUF) , 8-bit quantization balanced quality gguf version, running on CPU. Using [GitHub - llama.cpp](https://github.com/ggerganov/llama.cpp) [GitHub - gpt4all](https://github.com/nomic-ai/gpt4all). 

🔨 Running on CPU-Basic free hardware. Suggest duplicating this space to run without a queue. 

Mistral does not support system prompt symbol (such as ```<<SYS>>```) now, input your system prompt in the first message if you need. Learn more: [Guardrailing Mistral 7B](https://docs.mistral.ai/usage/guardrailing). 
"""

"""
[Model From TheBloke/Mistral-6B-Instruct-v0.1-GGUF](https://huggingface.co/FreedomIntelligence/Apollo-6B-GGUF)
[Mistral-instruct-v0.1 System prompt](https://docs.mistral.ai/usage/guardrailing)
"""

model_path = "models"
model_name = "Apollo-6B-q8_0.gguf"
hf_hub_download(repo_id="FreedomIntelligence/Apollo-6B-GGUF", filename=model_name, local_dir=model_path, local_dir_use_symlinks=False)

print("Start the model init process")
model = model = GPT4All(model_name, model_path, allow_download = False, device="cpu")
print("Finish the model init process")

model.config["promptTemplate"] = "{0}"
model.config["systemPrompt"] = "You are a multiligual AI doctor, your name is Apollo."
model._is_chat_session_activated = False

max_new_tokens = 2048

# def generater(message, history, temperature, top_p, top_k):
#     prompt = "<s>"
#     for user_message, assistant_message in history:
#         prompt += model.config["promptTemplate"].format(user_message)
#         prompt += assistant_message + "</s>"
#     prompt += model.config["promptTemplate"].format(message)
#     outputs = []    
#     for token in model.generate(prompt=prompt, temp=temperature, top_k = top_k, top_p = top_p, max_tokens = max_new_tokens, streaming=True):
#         outputs.append(token)
#         yield "".join(outputs)
async def generater(message, history, temperature, top_p, top_k):
    # 构建prompt
    prompt = ""
    for user_message, assistant_message in history:
        prompt += model.config["promptTemplate"].format(user_message)
        prompt += assistant_message
    prompt += model.config["promptTemplate"].format(message)

    # Debug: 打印最终的prompt以验证其正确性
    print(f"Final prompt: {prompt}")

    cmd = [
        "./main",
        "-m", model_path+"/"+model_name,
        "--prompt", prompt
    ]

    # 使用subprocess.Popen调用./main并流式读取输出
    process = subprocess.Popen(
        cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True
    )

    # 初始占位符输出
    yield "Generating response..."
    # 异步等待并处理输出
    try:
        while True:
            line = process.stdout.readline()
            if not line:
                break  # 如果没有更多的输出，结束循环
            print(f"Generated line: {line.strip()}")  # Debug: 打印生成的每行
            yield line
    except Exception as e:
        print(f"Error during generation: {e}")
        yield "Sorry, an error occurred while generating the response."


def vote(data: gr.LikeData):
    if data.liked:
        return
    else:
        return

chatbot = gr.Chatbot(avatar_images=('resourse/user-icon.png', 'resourse/chatbot-icon.png'),bubble_full_width = False)

additional_inputs=[
    gr.Slider(
        label="temperature",
        value=0.5,
        minimum=0.0,
        maximum=2.0,
        step=0.05,
        interactive=True,
        info="Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic.",
    ),
    gr.Slider(
        label="top_p",
        value=1.0,
        minimum=0.0,
        maximum=1.0,
        step=0.01,
        interactive=True,
        info="0.1 means only the tokens comprising the top 10% probability mass are considered. Suggest set to 1 and use temperature. 1 means 100% and will disable it",
    ),
    gr.Slider(
        label="top_k",
        value=40,
        minimum=0,
        maximum=1000,
        step=1,
        interactive=True,
        info="limits candidate tokens to a fixed number after sorting by probability. Setting it higher than the vocabulary size deactivates this limit.",
    )
]

iface = gr.ChatInterface(
    fn = generater,
    title=title,
    description = description,
    chatbot=chatbot,
    additional_inputs=additional_inputs,
    examples=[
        ["枸杞有什么疗效"],
        ["I've taken several courses of antibiotics for recurring infections, and now they seem less effective. Am I developing antibiotic resistance?"],
    ]
)

with gr.Blocks(css="resourse/style/custom.css") as demo:
    chatbot.like(vote, None, None)
    iface.render()

if __name__ == "__main__":
    demo.queue(max_size=3).launch()