app.py

import os
import logging
import time
from fastapi import FastAPI, HTTPException
from fastapi.responses import JSONResponse
from pydantic import BaseModel
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from llama_cpp import Llama
import asyncio
import uvicorn

# Set up logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

# Set up cache directory
CACHE_DIR = "/app/.cache/huggingface/hub"
os.environ["TRANSFORMERS_CACHE"] = CACHE_DIR
os.environ["HF_HOME"] = CACHE_DIR

# Create the FastAPI app
app = FastAPI(
    title="MGZON Smart Assistant",
    description="دمج نموذج T5 المدرب مع Mistral-7B (GGUF) داخل Space"
)

# Initialize model variables
t5_tokenizer = None
t5_model = None
mistral = None
t5_loaded = False
mistral_loaded = False

# Root endpoint
@app.get("/")
async def root():
    logger.info(f"Root endpoint called at {time.time()}")
    return JSONResponse(
        content={"message": "MGZON Smart Assistant is running"},
        headers={"Cache-Control": "no-cache", "Connection": "keep-alive"}
    )

# Health check endpoint
@app.get("/health")
async def health_check():
    logger.info(f"Health check endpoint called at {time.time()}")
    return JSONResponse(
        content={"status": "healthy" if t5_loaded else "loading"},
        headers={"Cache-Control": "no-cache", "Connection": "keep-alive"}
    )

# Async function to load T5 model
async def load_t5_model():
    global t5_tokenizer, t5_model, t5_loaded
    start_time = time.time()
    logger.info(f"Starting T5 model loading at {start_time}")
    try:
        T5_MODEL_PATH = os.path.join(CACHE_DIR, "models--MGZON--mgzon-flan-t5-base/snapshots")
        logger.info(f"Loading tokenizer for MGZON/mgzon-flan-t5-base from {T5_MODEL_PATH}")
        t5_tokenizer = AutoTokenizer.from_pretrained(
            T5_MODEL_PATH,
            local_files_only=True,
            torch_dtype="float16"  # Reduce memory usage
        )
        logger.info(f"Successfully loaded tokenizer for MGZON/mgzon-flan-t5-base in {time.time() - start_time} seconds")
        logger.info(f"Loading model for MGZON/mgzon-flan-t5-base from {T5_MODEL_PATH}")
        t5_model = AutoModelForSeq2SeqLM.from_pretrained(
            T5_MODEL_PATH,
            local_files_only=True,
            torch_dtype="float16"  # Reduce memory usage
        )
        logger.info(f"Successfully loaded model for MGZON/mgzon-flan-t5-base in {time.time() - start_time} seconds")
        t5_loaded = True
    except Exception as e:
        logger.error(f"Failed to load T5 model: {str(e)}", exc_info=True)
        t5_loaded = False
        raise RuntimeError(f"Failed to load T5 model: {str(e)}")
    finally:
        end_time = time.time()
        logger.info(f"T5 model loading completed in {end_time - start_time} seconds")

# Async function to load Mistral model
async def load_mistral_model():
    global mistral, mistral_loaded
    start_time = time.time()
    logger.info(f"Starting Mistral model loading at {start_time}")
    try:
        gguf_path = os.path.abspath("models/mistral-7b-instruct-v0.1.Q2_K.gguf")
        if not os.path.exists(gguf_path):
            logger.error(f"Mistral GGUF file not found at {gguf_path}")
            raise RuntimeError(f"Mistral GGUF file not found at {gguf_path}")
        logger.info(f"Loading Mistral model from {gguf_path}")
        mistral = Llama(
            model_path=gguf_path,
            n_ctx=512,
            n_threads=1,
            n_batch=128,
            verbose=True
        )
        logger.info(f"Successfully loaded Mistral model from {gguf_path} in {time.time() - start_time} seconds")
        mistral_loaded = True
    except Exception as e:
        logger.error(f"Failed to load Mistral model: {str(e)}", exc_info=True)
        mistral_loaded = False
        raise RuntimeError(f"Failed to load Mistral model: {str(e)}")
    finally:
        end_time = time.time()
        logger.info(f"Mistral model loading completed in {end_time - start_time} seconds")

# Run T5 model loading in the background
@app.on_event("startup")
async def startup_event():
    logger.info(f"Startup event triggered at {time.time()}")
    asyncio.create_task(load_t5_model())  # Load only T5 at startup

# Define request schema
class AskRequest(BaseModel):
    question: str
    max_new_tokens: int = 150

# Endpoint: /ask
@app.post("/ask")
async def ask(req: AskRequest):
    logger.info(f"Received ask request: {req.question} at {time.time()}")
    if not t5_loaded:
        logger.error("T5 model not loaded yet")
        raise HTTPException(status_code=503, detail="T5 model is still loading, please try again later")
    
    q = req.question.strip()
    if not q:
        logger.error("Empty question received")
        raise HTTPException(status_code=400, detail="Empty question")

    try:
        if any(tok in q.lower() for tok in ["mgzon", "flan", "t5"]):
            # Use T5 model
            logger.info("Using MGZON-FLAN-T5 model")
            inputs = t5_tokenizer(q, return_tensors="pt", truncation=True, max_length=256)
            out_ids = t5_model.generate(**inputs, max_length=req.max_new_tokens)
            answer = t5_tokenizer.decode(out_ids[0], skip_special_tokens=True)
            model_name = "MGZON-FLAN-T5"
        else:
            # Load Mistral model if not loaded
            if not mistral_loaded:
                logger.info("Mistral model not loaded, loading now...")
                await load_mistral_model()
                if not mistral_loaded:
                    raise HTTPException(status_code=503, detail="Failed to load Mistral model")
            # Use Mistral model
            logger.info("Using Mistral-7B-GGUF model")
            out = mistral(prompt=q, max_tokens=req.max_new_tokens, temperature=0.7)
            answer = out["choices"][0]["text"].strip()
            model_name = "Mistral-7B-GGUF"
        logger.info(f"Response generated by {model_name}: {answer}")
        return {"model": model_name, "response": answer}
    except Exception as e:
        logger.error(f"Error processing request: {str(e)}", exc_info=True)
        raise HTTPException(status_code=500, detail=f"خطأ أثناء معالجة الطلب: {str(e)}")

# Run the app
if __name__ == "__main__":
    uvicorn.run(
        app,
        host="0.0.0.0",
        port=8080,
        log_level="info",
        workers=1,
        timeout_keep_alive=15,
        limit_concurrency=5,
        limit_max_requests=50
    )