app.py

from flask import Flask, request, jsonify, send_file
from flask_cors import CORS
from diffusers import DiffusionPipeline, LCMScheduler
import torch
import os
import json
import secrets
from io import BytesIO
import gc
from datetime import datetime
import traceback

app = Flask(__name__)
CORS(app)

# Configuration
BASE = "/home/sd"
WL_PATH = f"{BASE}/whitelist.txt"
USAGE_PATH = f"{BASE}/usage.json"
LIMITS_PATH = f"{BASE}/limits.json"
DEFAULT_LIMIT = 500

# Use a fast, reliable model: LCM version for speed + quality
# Alternatives: "segmind/SSD-1B" (smaller) or "stabilityai/sdxl-turbo" (fastest)
MODEL_ID = "Lykon/dreamshaper-8-lcm"

# Global pipeline with lazy loading
pipe = None

def init_pipeline():
    """Initialize the pipeline with optimizations"""
    global pipe
    
    if pipe is not None:
        return pipe
    
    print(f"Loading model: {MODEL_ID}")
    
    # Use half precision for speed and memory efficiency
    torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
    
    try:
        # Load pipeline with optimizations
        pipe = DiffusionPipeline.from_pretrained(
            MODEL_ID,
            torch_dtype=torch_dtype,
            variant="fp16" if torch_dtype == torch.float16 else None,
            use_safetensors=True,
            safety_checker=None,  # Disable for speed (optional)
            requires_safety_checker=False
        )
        
        # Move to GPU if available
        device = "cuda" if torch.cuda.is_available() else "cpu"
        pipe = pipe.to(device)
        
        # Enable optimizations
        if device == "cuda":
            pipe.enable_attention_slicing()  # Reduce memory usage
            if torch_dtype == torch.float16:
                pipe.enable_model_cpu_offload()  # Offload to CPU when not in use
        
        print(f"Model loaded successfully on {device}")
        return pipe
        
    except Exception as e:
        print(f"Error loading model: {e}")
        # Fallback to a simpler model
        try:
            pipe = DiffusionPipeline.from_pretrained(
                "SimianLuo/LCM_Dreamshaper_v7",
                torch_dtype=torch_dtype
            ).to("cuda" if torch.cuda.is_available() else "cpu")
            print("Loaded fallback model")
            return pipe
        except:
            raise Exception("Failed to load any model")

# Initialize storage
os.makedirs(BASE, exist_ok=True)
for path in [WL_PATH, USAGE_PATH, LIMITS_PATH]:
    if not os.path.exists(path):
        if path.endswith(".json"):
            with open(path, "w") as f:
                json.dump({}, f)
        else:
            with open(path, "w") as f:
                f.write("")

# Helper functions
def get_whitelist():
    try:
        with open(WL_PATH, "r") as f:
            return set(line.strip() for line in f if line.strip())
    except:
        return set()

def load_json(path):
    try:
        with open(path, "r") as f:
            return json.load(f)
    except:
        return {}

def save_json(path, data):
    with open(path, "w") as f:
        json.dump(data, f, indent=2)

def validate_api_key(key):
    """Validate API key and check rate limits"""
    if key not in get_whitelist():
        return False, "Unauthorized"
    
    limits = load_json(LIMITS_PATH)
    usage = load_json(USAGE_PATH)
    
    limit = limits.get(key, DEFAULT_LIMIT)
    if limit == "unlimited":
        return True, "OK"
    
    month = datetime.now().strftime("%Y-%m")
    used = usage.get(key, {}).get(month, 0)
    
    if used >= limit:
        return False, "Monthly limit reached"
    
    return True, "OK"

# Routes
@app.route("/", methods=["GET"])
def health():
    return jsonify({
        "status": "online",
        "model": MODEL_ID,
        "device": "cuda" if torch.cuda.is_available() else "cpu"
    }), 200

@app.route("/generate-key", methods=["POST"])
def generate_key():
    try:
        data = request.get_json() or {}
        unlimited = data.get("unlimited", False)
        limit = data.get("limit", DEFAULT_LIMIT)
        
        key = "sk-" + secrets.token_hex(16)
        
        # Add to whitelist
        with open(WL_PATH, "a") as f:
            f.write(key + "\n")
        
        # Set limits
        limits = load_json(LIMITS_PATH)
        limits[key] = "unlimited" if unlimited else int(limit)
        save_json(LIMITS_PATH, limits)
        
        # Initialize usage
        usage = load_json(USAGE_PATH)
        if key not in usage:
            usage[key] = {}
        save_json(USAGE_PATH, usage)
        
        return jsonify({
            "key": key,
            "limit": limits[key],
            "message": "Key generated successfully"
        })
        
    except Exception as e:
        return jsonify({"error": str(e)}), 500

@app.route("/api/generate", methods=["POST"])
def generate():
    try:
        # Validate API key
        key = request.headers.get("x-api-key", "")
        valid, message = validate_api_key(key)
        if not valid:
            return jsonify({"error": message}), 401 if message == "Unauthorized" else 429
        
        # Parse request
        data = request.get_json() or {}
        prompt = data.get("prompt", "").strip()
        
        if not prompt:
            return jsonify({"error": "Prompt is required"}), 400
        
        # Set generation parameters with safe defaults
        steps = min(max(int(data.get("steps", 4)), 1), 20)  # LCM models work with 4-8 steps
        guidance = float(data.get("guidance", 1.2))  # LCM uses low guidance
        width = min(max(int(data.get("width", 512)), 256), 1024)
        height = min(max(int(data.get("height", 512)), 256), 1024)
        
        # Ensure pipeline is loaded
        if pipe is None:
            init_pipeline()
        
        # Generate image
        print(f"Generating: {prompt[:50]}... (steps: {steps}, guidance: {guidance})")
        
        with torch.inference_mode():
            image = pipe(
                prompt=prompt,
                num_inference_steps=steps,
                guidance_scale=guidance,
                width=width,
                height=height,
                output_type="pil"
            ).images[0]
        
        # Update usage
        usage = load_json(USAGE_PATH)
        month = datetime.now().strftime("%Y-%m")
        usage.setdefault(key, {})
        usage[key][month] = usage[key].get(month, 0) + 1
        save_json(USAGE_PATH, usage)
        
        # Return image
        buf = BytesIO()
        image.save(buf, format="PNG", optimize=True)
        buf.seek(0)
        
        return send_file(buf, mimetype="image/png")
        
    except torch.cuda.OutOfMemoryError:
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        return jsonify({"error": "GPU out of memory. Try smaller image size."}), 507
        
    except Exception as e:
        error_details = traceback.format_exc()
        print(f"Generation error: {error_details}")
        return jsonify({
            "error": "Generation failed",
            "details": str(e)
        }), 500

@app.route("/api/status", methods=["GET"])
def status():
    """Check API key status and usage"""
    key = request.headers.get("x-api-key", "")
    if key not in get_whitelist():
        return jsonify({"error": "Invalid API key"}), 401
    
    limits = load_json(LIMITS_PATH)
    usage = load_json(USAGE_PATH)
    
    month = datetime.now().strftime("%Y-%m")
    used = usage.get(key, {}).get(month, 0)
    limit = limits.get(key, DEFAULT_LIMIT)
    
    return jsonify({
        "key": key[:8] + "..." + key[-4:] if len(key) > 12 else key,
        "usage": used,
        "limit": limit,
        "remaining": "unlimited" if limit == "unlimited" else max(0, limit - used),
        "month": month
    })

if __name__ == "__main__":
    # Initialize pipeline on startup
    print("Initializing pipeline...")
    init_pipeline()
    print("API starting on port 7860...")
    app.run(host="0.0.0.0", port=7860, debug=False)