app_fastapi.py

"""

FastAPI Backend mit optimiertem Modell-Loading für HF Spaces

Support für Quantization und Memory-Limited Environments

"""

from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
import torch
from transformers import (
    AutoModelForCausalLM, 
    AutoTokenizer, 
    BitsAndBytesConfig,
    pipeline
)
import logging
import time
from pathlib import Path
import os

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

app = FastAPI(title="Zephyr-7B API - Optimized")

# CORS
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Model Config - OPTIMIERT FÜR HF SPACES
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
USE_QUANTIZATION = True  # 8-bit loading

# Wähle Modell basierend auf verfügbarem Memory
def select_model():
    """Wählt das beste Modell für verfügbares Memory"""
    try:
        # GPU Memory check
        if torch.cuda.is_available():
            gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1e9
            logger.info(f"GPU Memory: {gpu_memory:.1f}GB")
            
            # Wähle Modell basierend auf Memory
            if gpu_memory >= 20:  # Genug für fp16
                return "HuggingFaceH4/zephyr-7b-beta"
            elif gpu_memory >= 10:  # AWQ 4-bit
                return "TheBloke/zephyr-7B-beta-AWQ"
            else:  # GGUF 4-bit (kompressester)
                return "TheBloke/zephyr-7B-beta-GGUF"
        else:
            # CPU - nutze kleineres Modell
            return "Qwen/Qwen2.5-1.5B-Instruct" # "HuggingFaceH4/zephyr-7b-alpha"
    except Exception as e:
        logger.warning(f"Could not detect GPU memory: {e}, using safe default")
        return "Qwen/Qwen2.5-1.5B-Instruct"# "TheBloke/zephyr-7B-beta-AWQ"

MODEL_NAME = os.getenv("MODEL_NAME", select_model())
logger.info(f"Using model: {MODEL_NAME}")

# Initialize Model mit Quantization
logger.info(f"Loading model {MODEL_NAME} on {DEVICE}...")

def load_model_optimized():
    """Lädt Modell mit optimaler Quantization für HF Spaces"""
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
    
    # Config für 8-bit Loading (spart 50% Memory!)
    if USE_QUANTIZATION and DEVICE == "cuda":
        try:
            bnb_config = BitsAndBytesConfig(
                load_in_8bit=True,
                bnb_8bit_compute_dtype=torch.float16,
                bnb_8bit_use_double_quant=True,
            )
            model = AutoModelForCausalLM.from_pretrained(
                MODEL_NAME,
                quantization_config=bnb_config,
                device_map="auto",
            )
            logger.info("✓ Model loaded with 8-bit quantization")
        except Exception as e:
            logger.warning(f"8-bit quantization failed: {e}, trying default")
            model = AutoModelForCausalLM.from_pretrained(
                MODEL_NAME,
                device_map="auto" if DEVICE == "cuda" else None,
                torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
            )
    else:
        # Standard Loading für CPU oder non-quantized
        model = AutoModelForCausalLM.from_pretrained(
            MODEL_NAME,
            device_map="auto" if DEVICE == "cuda" else None,
            torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
        )
    
    return tokenizer, model

try:
    tokenizer, model = load_model_optimized()
    pipe = pipeline(
        "text-generation",
        model=model,
        tokenizer=tokenizer,
        device=0 if DEVICE == "cuda" else -1,
    )
    logger.info("✓ Pipeline initialized successfully")
except Exception as e:
    logger.error(f"✗ Failed to load model: {e}")
    raise

# Request Model
class GenerateRequest(BaseModel):
    prompt: str
    system_prompt: str = None
    max_tokens: int = 512
    temperature: float = 0.7
    top_p: float = 0.9

@app.post("/api/generate")
async def generate(request: GenerateRequest):
    """Generate text - optimized for HF Spaces"""
    try:
        start = time.time()
        
        # Format prompt
        if request.system_prompt:
            messages = f"<|system|>\n{request.system_prompt}\n<|user|>\n{request.prompt}\n<|assistant|>\n"
        else:
            messages = f"<|user|>\n{request.prompt}\n<|assistant|>\n"
        
        # Generate
        outputs = pipe(
            messages,
            max_new_tokens=request.max_tokens,
            temperature=request.temperature,
            top_p=request.top_p,
            do_sample=True,
            return_full_text=False,
        )
        
        response_text = outputs[0]["generated_text"].strip()
        elapsed = time.time() - start
        
        return {
            "response": response_text,
            "tokens": len(tokenizer.encode(response_text)),
            "time_seconds": round(elapsed, 2),
            "model": MODEL_NAME,
        }
    except Exception as e:
        logger.error(f"Generation error: {e}")
        raise HTTPException(status_code=500, detail=str(e))

@app.get("/api/health")
async def health():
    """Health check"""
    return {
        "status": "ok",
        "model": MODEL_NAME,
        "device": DEVICE,
        "quantization": USE_QUANTIZATION,
    }

@app.get("/api/info")
async def info():
    """Model info"""
    gpu_memory = None
    if torch.cuda.is_available():
        gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1e9
    
    return {
        "model": MODEL_NAME,
        "device": DEVICE,
        "gpu_memory_gb": gpu_memory,
        "quantization_enabled": USE_QUANTIZATION,
    }

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)