app.py

import json
import subprocess
import time
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
from huggingface_hub import hf_hub_download

# Global variables to store the model and agent
llm = None
llm_model = None
agent = None

# Download the new model
hf_hub_download(
    repo_id="hugging-quants/Llama-3.2-1B-Instruct-Q4_K_M-GGUF",
    filename="llama-3.2-1b-instruct-q4_k_m.gguf",
    local_dir="./models"
)

def get_messages_formatter_type(model_name):
    return MessagesFormatterType.LLAMA_3

def load_model(model_path):
    global llm
    global llm_model
    if llm is None or llm_model != model_path:
        llm = Llama(
            model_path=model_path,
            n_gpu_layers=0,  # Adjust based on your GPU
            n_batch=32398,     # Adjust based on your RAM
            n_ctx=512,      # Adjust based on your RAM and desired context length
        )
        llm_model = model_path
    return llm

def load_agent(llm, system_message, chat_template):
    global agent
    if agent is None:
        provider = LlamaCppPythonProvider(llm)
        agent = LlamaCppAgent(
            provider,
            system_prompt=system_message,
            predefined_messages_formatter_type=chat_template,
            debug_output=True
        )
    return agent

def respond(
    message,
    history: list[tuple[str, str]],
    model,
    system_message,
    max_tokens,
    temperature,
    top_p,
    top_k,
    repeat_penalty,
):
    global llm
    global agent

    chat_template = get_messages_formatter_type(model)
    llm = load_model(f"models/{model}")
    agent = load_agent(llm, system_message, chat_template)

    settings = agent.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)
    
    start_time = time.time()
    token_count = 0

    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
        token_count += len(output.split()) 
        yield outputs

    end_time = time.time()
    latency = end_time - start_time
    speed = token_count / (end_time - start_time)
    print(f"Latency: {latency} seconds")
    print(f"Speed: {speed} tokens/second")

description = """<p><center>
<a href="https://huggingface.co./hugging-quants/Llama-3.2-1B-Instruct-Q4_K_M-GGUF" target="_blank">[Meta Llama 3.2 (1B)]</a>

Meta Llama 3.2 (1B) is a multilingual large language model (LLM) optimized for conversational dialogue use cases, including agentic retrieval and summarization tasks. It outperforms many open-source and closed chat models on industry benchmarks, and is intended for commercial and research use in multiple languages.

</center></p>
"""

demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Dropdown([
                "llama-3.2-1b-instruct-q4_k_m.gguf"
            ],
            value="llama-3.2-1b-instruct-q4_k_m.gguf",
            label="Model"
        ),
        gr.TextArea(value="""You are Meta Llama 3.2 (1B), an advanced AI assistant created by Meta.""", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=1024, step=1, label="Max tokens"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(
            minimum=0.1,
            maximum=2.0,
            value=0.9,
            step=0.05,
            label="Top-p",
        ),
        gr.Slider(
            minimum=0,
            maximum=100,
            value=1,
            step=1,
            label="Top-k",
        ),
        gr.Slider(
            minimum=0.0,
            maximum=2.0,
            value=1.1,
            step=0.1,
            label="Repetition penalty",
        ),
    ],
    theme=gr.themes.Soft(primary_hue="violet", secondary_hue="violet", neutral_hue="gray",font=[gr.themes.GoogleFont("Exo"), "ui-sans-serif", "system-ui", "sans-serif"]).set(
        body_background_fill_dark="#16141c",
        block_background_fill_dark="#16141c",
        block_border_width="1px",
        block_title_background_fill_dark="#1e1c26",
        input_background_fill_dark="#292733",
        button_secondary_background_fill_dark="#24212b",
        border_color_accent_dark="#343140",
        border_color_primary_dark="#343140",
        background_fill_secondary_dark="#16141c",
        color_accent_soft_dark="transparent",
        code_background_fill_dark="#292733",
    ),
    title="Meta Llama 3.2 (1B)",
    description=description,
    chatbot=gr.Chatbot(
        scale=1, 
        likeable=True,
        show_copy_button=True
    ),
    examples=[
        ["Hello! Can you introduce yourself?"],
        ["What's the capital of France?"],
        ["Can you explain the concept of photosynthesis?"],
        ["Write a short story about a robot learning to paint."],
        ["Explain the difference between machine learning and deep learning."],
        ["Summarize the key points of climate change and its global impact."],
        ["Explain quantum computing to a 10-year-old."],
        ["Design a step-by-Step Meal Plan for Weight Loss and Muscle Gain."],
    ],
    cache_examples=False,
    autofocus = False,
    concurrency_limit = None
)

if __name__ == "__main__":
    demo.launch()