app.py

import gradio as gr
from huggingface_hub import hf_hub_download
from llama_cpp import Llama
from llama_cpp_agent import LlamaCppAgent, MessagesFormatterType
from llama_cpp_agent.chat_history import BasicChatHistory
from llama_cpp_agent.chat_history.messages import Roles

# ⬇️ دانلود مدل
hf_hub_download(
    repo_id="mradermacher/Arsh-llm-GGUF",
    filename="Arsh-llm.Q4_K_M.gguf",
    local_dir="./models"
)

# 🧠 لود مدل (فقط یکبار در ابتدای اجرا)
llm = Llama(
    model_path="./models/Arsh-llm.Q4_K_M.gguf",
    n_batch=512,
    n_ctx=8192,
    verbose=False
)

provider = LlamaCppPythonProvider(llm)

agent = LlamaCppAgent(
    provider,
    system_prompt="You are Arsh, a helpful assistant.",
    predefined_messages_formatter_type=MessagesFormatterType.CHATML,
    debug_output=False
)

# 💬 تابع پاسخ‌دهنده
def respond(message, chat_history):
    messages = BasicChatHistory()

    for user_msg, assistant_msg in chat_history:
        messages.add_message({
            'role': Roles.user,
            'content': user_msg
        })
        messages.add_message({
            'role': Roles.assistant,
            'content': assistant_msg
        })

    stream = agent.get_chat_response(
        message,
        chat_history=messages,
        returns_streaming_generator=True,
        print_output=False
    )

    response = ""
    for token in stream:
        response += token
        yield response

# 🎛️ ساخت رابط Gradio
with gr.Blocks() as demo:
    gr.Markdown("# Arsh-LLM Q4_K_M Chat Demo")
    gr.ChatInterface(
        respond,
        additional_inputs=[
            gr.Textbox("You are Arsh, a helpful assistant.", label="System Message", interactive=True),
            gr.Slider(minimum=1, maximum=4096, value=2048, step=1, label="Max Tokens"),
            gr.Slider(minimum=0.1, maximum=4.0, value=0.6, step=0.1, label="Temperature"),
            gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p"),
            gr.Slider(minimum=0, maximum=100, value=40, step=1, label="Top-k"),
            gr.Slider(minimum=0.0, maximum=2.0, value=1.1, step=0.1, label="Repetition Penalty"),
        ],
        chatbot=gr.Chatbot(
            label="Chat with Arsh-LLM",
            bubble_full_width=False,
            show_copy_button=True
        )
    )

# 🚀 اجرای برنامه
if __name__ == "__main__":
    demo.launch()