app.py

# app.py
import asyncio, logging
import gradio as gr

from config import settings
from rabbit_base import RabbitBase
from listener import RabbitListenerBase
from rabbit_repo import RabbitRepo
from oa_server import OpenAIServers
from vllm_backend import VLLMChatBackend, StubImagesBackend   # ✅ use your backend module

# ---- vLLM imports ----
from vllm.engine.async_llm_engine import AsyncLLMEngine
from vllm.engine.arg_utils import AsyncEngineArgs

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s [%(levelname)s] %(name)s: %(message)s"
)
log = logging.getLogger("app")

try:
    import spaces
    @spaces.GPU(duration=60)
    def gpu_entrypoint() -> str: 
        return "gpu: ready"
except Exception:
    def gpu_entrypoint() -> str: 
        return "gpu: not available (CPU only)"

# ----------------- vLLM globals -----------------
vllm_engine: AsyncLLMEngine | None = None

async def init_vllm():
    """Initialize vLLM engine with a Hugging Face model."""
    global vllm_engine
    if vllm_engine is not None:
        return vllm_engine

    model_id = getattr(settings, "LlmHFModelID", "Qwen/Qwen2.5-7B-Instruct")
    log.info(f"Loading vLLM model: {model_id}")

    args = AsyncEngineArgs(
        model=model_id,
        trust_remote_code=True,
        max_model_len=getattr(settings, "LlmOpenAICtxSize", 32768),
    )
    vllm_engine = AsyncLLMEngine.from_engine_args(args)
    return vllm_engine

# ----------------- RabbitMQ wiring -----------------
publisher = RabbitRepo(external_source="openai.mq.server")
resolver = (lambda name: "direct" if name.startswith("oa.") else settings.RABBIT_EXCHANGE_TYPE)
base = RabbitBase(exchange_type_resolver=resolver)

servers = OpenAIServers(
    publisher,
    chat_backend=VLLMChatBackend(),    # ✅ now from llm_backend.py
    images_backend=StubImagesBackend()
)

handlers = {
    "oaChatCreate": servers.handle_chat_create,
    "oaImagesGenerate": servers.handle_images_generate,
}

DECLS = [
    {"ExchangeName": "oa.chat.create", "FuncName": "oaChatCreate",
     "MessageTimeout": 600_000, "RoutingKeys": [settings.RABBIT_ROUTING_KEY]},
    {"ExchangeName": "oa.images.generate", "FuncName": "oaImagesGenerate",
     "MessageTimeout": 600_000, "RoutingKeys": [settings.RABBIT_ROUTING_KEY]},
]

listener = RabbitListenerBase(base, instance_name=settings.RABBIT_INSTANCE_NAME, handlers=handlers)

# ----------------- Startup init -----------------
async def _startup_init():
    try:
        await init_vllm()             # load vLLM model
        await base.connect()          # connect to RabbitMQ
        await listener.start(DECLS)   # start queue listeners
        return "OpenAI MQ + vLLM: ready"
    except Exception as e:
        log.exception("Startup init failed")
        return f"ERROR: {e}"

async def ping(): 
    return "ok"

# ----------------- Gradio UI -----------------
with gr.Blocks(title="OpenAI over RabbitMQ (local vLLM)", theme=gr.themes.Soft()) as demo:
    gr.Markdown("## OpenAI-compatible over RabbitMQ — using vLLM locally inside Space")
    with gr.Tabs():
        with gr.Tab("Service"):
            btn = gr.Button("Ping")
            out = gr.Textbox(label="Ping result")
            btn.click(ping, inputs=None, outputs=out)
            init_status = gr.Textbox(label="Startup status", interactive=False)
            demo.load(fn=_startup_init, inputs=None, outputs=init_status)

        with gr.Tab("@spaces.GPU Probe"):
            gpu_btn = gr.Button("GPU Ready Probe", variant="primary")
            gpu_out = gr.Textbox(label="GPU Probe Result", interactive=False)
            gpu_btn.click(gpu_entrypoint, inputs=None, outputs=gpu_out)

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True, debug=True)