app.py

import os
import requests
import subprocess
import json
import time
import spaces
import gradio as gr
from typing import List, Optional, Tuple, Dict

DEFAULT_SYSTEM = "You are a helpful assistant."
HF_MODEL_ID = "bartowski/Llama-3.2-1B-Instruct-GGUF"
HF_FILENAME = "Llama-3.2-1B-Instruct-Q4_K_M.gguf"

###############################################

API_PATH_HEALTH = "/health"
API_PATH_COMPLETIONS = "/chat/completions"
LLAMA_CPP_SERVER_BASE = "http://127.0.0.1:8080"
LLAMA_CPP_SERVER_START_TIMEOUT = 50  # seconds

if not os.path.exists('model.gguf'):
    url = f"https://huggingface.co/{HF_MODEL_ID}/resolve/main/{HF_FILENAME}"
    subprocess.check_call(["curl", "-o", "model.gguf", "-L", url])

if not os.path.exists("llama-server"):
  # FIXME: currently, we can't build inside gradio container because nvcc is missing
  subprocess.check_call("curl -o llama-server -L https://ngxson-llamacpp-builder.hf.space/llama-server", shell=True)
  subprocess.check_call("chmod +x llama-server", shell=True)

###############################################

class Role:
    SYSTEM = "system"
    USER = "user"
    ASSISTANT = "assistant"

History = List[Tuple[str, str]]
Messages = List[Dict[str, str]]

def clear_session() -> History:
    return "", []

def modify_system_session(system: str) -> str:
    if system is None or len(system) == 0:
        system = DEFAULT_SYSTEM
    return system, system, []

def history_to_messages(history: History, system: str) -> Messages:
    messages = [{"role": Role.SYSTEM, "content": system}]
    for h in history:
        messages.append({"role": Role.USER, "content": h[0]})
        messages.append({"role": Role.ASSISTANT, "content": h[1]})
    return messages


def messages_to_history(messages: Messages) -> Tuple[str, History]:
    assert messages[0]["role"] == Role.SYSTEM
    system = messages[0]["content"]
    history = []
    for q, r in zip(messages[1::2], messages[2::2]):
        history.append([q["content"], r["content"]])
    return system, history

def wait_until_llamacpp_ready():
    time.sleep(5)
    gr.Info("starting llama.cpp server...")
    trials = 0
    while True:
        try:
            response = requests.get(LLAMA_CPP_SERVER_BASE + API_PATH_HEALTH)
            if response.status_code == 200:
                print("Status 200 received. Exiting loop.")
                break
            else:
                print(f"Received status {response.status_code}. Retrying...")
        except requests.exceptions.RequestException as e:
            print(f"Request failed: {e}")
        trials += 1
        if trials > LLAMA_CPP_SERVER_START_TIMEOUT:
            raise TimeoutError("llama.cpp server did not start in time")
        time.sleep(1)  # Wait for 1 second before retrying
    gr.Info("llama.cpp server is ready.")
    print("llama.cpp server is ready.")


@spaces.GPU
def model_chat(query: Optional[str], history: Optional[History], system: str
) -> Tuple[str, str, History]:
    if query is None:
        query = ""
    if history is None:
        history = []

    # start llama.cpp server
    proc = subprocess.Popen(["./llama-server"], env=dict(
        os.environ,
        LLAMA_HOST="0.0.0.0",
        LLAMA_PORT="8080",
        LLAMA_ARG_CTX_SIZE=str(1024 * 32),
        LLAMA_ARG_FLASH_ATTN="1",
        LLAMA_ARG_MODEL="model.gguf",
        LLAMA_ARG_N_PARALLEL="1",
        LLAMA_ARG_N_GPU_LAYERS="9999",
        LLAMA_ARG_NO_MMAP="1",
    ))

    exception = None
    try:
        wait_until_llamacpp_ready()

        messages = history_to_messages(history, system)
        messages.append({"role": Role.USER, "content": query})

        # adapted from https://gist.github.com/ggorlen/7c944d73e27980544e29aa6de1f2ac54
        url = LLAMA_CPP_SERVER_BASE + API_PATH_COMPLETIONS
        headers = {
            # "Authorization": f"Bearer {api_key}",
            "Content-Type": "application/json"
        }
        data = {
            "temperature": 0.8,
            "messages": messages,
            "stream": True
        }
        response = requests.post(url, headers=headers, json=data, stream=True)
        response.raise_for_status()

        curr_text = ""
        for line in response.iter_lines():
            line = line.decode("utf-8")

            if line.startswith("data: ") and not line.endswith("[DONE]"):
                data = json.loads(line[len("data: "):])
                chunk = data["choices"][0]["delta"].get("content", "")
                # print(chunk, end="", flush=True)
                curr_text += chunk
                system, history = messages_to_history(messages + [{"role": Role.ASSISTANT, "content": curr_text}])
                yield "", history, system
    except Exception as e:
        print(e)
        exception = e
    finally:
        # clean up
        proc.kill()
        if exception is not None:
            # re-raise the exception if needed
            raise exception


with gr.Blocks() as demo:
    gr.Markdown(f"""<center><font size=6>{HF_MODEL_ID}</center>""")

    with gr.Row():
        with gr.Column(scale=3):
            system_input = gr.Textbox(value=DEFAULT_SYSTEM, lines=1, label="System")
        with gr.Column(scale=1):
            modify_system = gr.Button("🛠️ Set system prompt and clear history", scale=2)
        system_state = gr.Textbox(value=DEFAULT_SYSTEM, visible=False)
    chatbot = gr.Chatbot(label=HF_MODEL_ID)
    textbox = gr.Textbox(lines=2, label="Input")

    with gr.Row():
        clear_history = gr.Button("🧹 Clear history")
        sumbit = gr.Button("🚀 Send")

    sumbit.click(model_chat,
                 inputs=[textbox, chatbot, system_state],
                 outputs=[textbox, chatbot, system_input],
                 concurrency_limit = 5)
    clear_history.click(fn=clear_session,
                        inputs=[],
                        outputs=[textbox, chatbot])
    modify_system.click(fn=modify_system_session,
                        inputs=[system_input],
                        outputs=[system_state, system_input, chatbot])

demo.queue(api_open=False)
demo.launch(max_threads=5)