app_local.py

import llama_cpp
import os
import json
import subprocess
from llama_cpp import Llama
from llama_cpp_agent import LlamaCppAgent, MessagesFormatterType
from llama_cpp_agent.providers import LlamaCppPythonProvider
from llama_cpp_agent.chat_history import BasicChatHistory
from llama_cpp_agent.chat_history.messages import Roles
import gradio as gr

llm = None
llm_model = None

# Dropdown for Model Selection
model_dropdown = gr.Dropdown(
    [
        'qwen2-0_5b-instruct-q4_k_m.gguf',
        'qwen2_500m.gguf',
        'mistrallite.Q4_K_M.gguf',
    ],
    value="qwen2-0_5b-instruct-q4_k_m.gguf",
    label="Model"
)

def respond(
    message,
    history: list[tuple[str, str]],
        system_message,
    max_tokens,
    temperature,
    top_p,
    top_k,
    repeat_penalty,
	selected_model,  # This is now a parameter received from the interface
):
    chat_template = MessagesFormatterType.GEMMA_2

    global llm
    global llm_model
    
    # Update the model if it has changed
    if llm is None or llm_model != selected_model:
        llm = Llama(
            model_path=f"models/{selected_model}",
            flash_attn=True,
            n_gpu_layers=81,
            n_batch=1024,
            n_ctx=8192,
        )
        llm_model = selected_model

    provider = LlamaCppPythonProvider(llm)

    agent = LlamaCppAgent(
        provider,
        system_prompt=f"{system_message}",
        predefined_messages_formatter_type=chat_template,
        debug_output=True
    )
    
    settings = provider.get_provider_default_settings()
    settings.temperature = temperature
    settings.top_k = top_k
    settings.top_p = top_p
    settings.max_tokens = max_tokens
    settings.repeat_penalty = repeat_penalty
    settings.stream = True

    messages = BasicChatHistory()

    for msn in history:
        user = {
            'role': Roles.user,
            'content': msn[0]
        }
        assistant = {
            'role': Roles.assistant,
            'content': msn[1]
        }
        messages.add_message(user)
        messages.add_message(assistant)
    
    stream = agent.get_chat_response(
        message,
        llm_sampling_settings=settings,
        chat_history=messages,
        returns_streaming_generator=True,
        print_output=False
    )
    
    outputs = ""
    for output in stream:
        outputs += output
        yield outputs

description = """<p align="center">Defaults to Qwen 500M</p>
"""

# Create the Gradio interface
with gr.Blocks() as demo:  # Create a Gradio Blocks context
    
    # Model selection dropdown above the chat
    model_dropdown.render()

    # Main chat interface
    chat_interface = gr.ChatInterface(
        respond,
        additional_inputs=[
            gr.Textbox(value="You are a helpful assistant.", label="System message"),
            gr.Slider(minimum=1, maximum=4096, value=2048, step=1, label="Max tokens"),
            gr.Slider(minimum=0.1, maximum=4.0, value=0.7, 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",
            ),
            model_dropdown  # Pass the dropdown directly
        ],
        retry_btn="Retry",
        undo_btn="Undo",
        clear_btn="Clear",
        submit_btn="Send",
        title="Chat with Qwen 2 and friends using llama.cpp",
        description=description,
    )

demo.queue().launch()