app.py

import os
import io
import time
import logging
import threading
from flask import Flask, request, jsonify, send_file
from flask_cors import CORS
from PIL import Image
import torch
import torchvision.models as models
import torchvision.transforms as transforms
from ultralytics import YOLO

# Fix PyTorch 2.6+ weights_only loading issue
if not hasattr(torch, "_original_load"):
    torch._original_load = torch.load

def patched_torch_load(*args, **kwargs):
    kwargs["weights_only"] = False
    return torch._original_load(*args, **kwargs)

torch.load = patched_torch_load

# Configure logging
logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s [%(levelname)s] %(name)s - %(message)s"
)
logger = logging.getLogger("PlantClassifierAPI")

# ---------------------------------------------------------------------------
# Hugging Face repository configuration
# ---------------------------------------------------------------------------
# PRIMARY:  Set HF_MODEL_REPO to a dedicated HF *model* repo containing the
#           real binary .pth / .pt files (e.g. "yourname/plantcare-models").
#
# FALLBACK: If HF_MODEL_REPO is not set, the app will try to download the
#           weights from the Space's own repository using SPACE_ID (which HF
#           injects automatically into every running Space). This works because
#           HF stores the real LFS bytes even though the Docker build context
#           only receives the pointer files.
#
# Either way, set HF_TOKEN for private repos.
# ---------------------------------------------------------------------------
HF_MODEL_REPO  = os.environ.get("HF_MODEL_REPO", "")
# SPACE_ID is auto-injected by HF Spaces; fallback to the known Space ID
SPACE_ID       = os.environ.get("SPACE_ID", "Karim31003/classifier")
HF_CACHE_DIR   = os.environ.get("HF_CACHE_DIR", "/tmp/hf")
HF_TOKEN       = os.environ.get("HF_TOKEN", None)

# ---------------------------------------------------------------------------

# Initialize Flask App
app = Flask(__name__)
CORS(app)

# Limit request payload to 10MB
app.config['MAX_CONTENT_LENGTH'] = 10 * 1024 * 1024

# Configuration options
DEBUG_SAVE_IMAGES = os.environ.get("DEBUG_SAVE_IMAGES", "false").lower() == "true"
UPLOAD_FOLDER = os.path.join(os.path.dirname(os.path.abspath(__file__)), "uploads")
ALLOWED_EXTENSIONS = {'.jpg', '.jpeg', '.png'}
ALLOWED_MIME_TYPES = {'image/jpeg', 'image/png', 'application/octet-stream'}

# Global cache for loaded models
models_cache = {
    "yolo": None,
    "disease_models": {}
}

# Detailed load reports for debugging
model_load_reports = []

# Force CPU device
device = torch.device("cpu")
logger.info(f"Using hardware device: {device}")
models_lock = threading.RLock()
models_ready = threading.Event()
models_load_error = None

LFS_POINTER_PREFIX = "version https://git-lfs.github.com/spec/v1"

BASE_DIR = os.path.dirname(os.path.abspath(__file__))
MODEL_DIR = os.path.join(BASE_DIR, "model")

if not os.path.isdir(MODEL_DIR):
    logger.warning(f"Model directory not found at {MODEL_DIR}; falling back to {BASE_DIR}.")
    MODEL_DIR = BASE_DIR


# ---------------------------------------------------------------------------
# Git LFS / file diagnostics
# ---------------------------------------------------------------------------

def _log_file_diagnostics(path: str, label: str) -> None:
    """Log path, size, and the first 200 characters of a file for debugging."""
    if not os.path.exists(path):
        logger.warning(f"[DIAG] {label}: file does not exist at {path}")
        return
    size = os.path.getsize(path)
    try:
        with open(path, "r", encoding="utf-8", errors="ignore") as fh:
            head = fh.read(200)
    except Exception as exc:
        head = f"<could not read: {exc}>"
    logger.info(
        f"[DIAG] {label}\n"
        f"  path : {path}\n"
        f"  size : {size:,} bytes\n"
        f"  head : {repr(head)}"
    )


def _is_lfs_pointer(path: str) -> bool:
    """Return True if the file is a Git LFS pointer text file."""
    if not os.path.exists(path):
        return False
    try:
        with open(path, "r", encoding="utf-8", errors="ignore") as fh:
            header = fh.read(256)
        return header.startswith(LFS_POINTER_PREFIX)
    except Exception:
        return False


def validate_model_artifact(model_path: str, artifact_name: str) -> None:
    """
    Ensure the model artifact is a real binary checkpoint, not a Git LFS pointer.
    Logs full diagnostics before raising so the Space logs are informative.
    """
    _log_file_diagnostics(model_path, artifact_name)

    if not os.path.exists(model_path):
        raise FileNotFoundError(f"{artifact_name} not found at: {model_path}")

    if _is_lfs_pointer(model_path):
        raise RuntimeError(
            f"{artifact_name} at {model_path} is a Git LFS pointer, not the actual model file.\n"
            "Fix options:\n"
            "  1. Set HF_MODEL_REPO env var to your HF model repo ID so weights are downloaded automatically.\n"
            "  2. Run `git lfs pull` locally, verify files are binary, then push again.\n"
            "  3. Upload real weight files directly via the Hugging Face web UI."
        )


# ---------------------------------------------------------------------------
# Hugging Face model download helpers
# ---------------------------------------------------------------------------

def _hf_hub_available() -> bool:
    """Return True if huggingface_hub is importable."""
    try:
        import huggingface_hub  # noqa: F401
        return True
    except ImportError:
        logger.warning("huggingface_hub not installed — add it to requirements.txt.")
        return False


def _download_from_hf(local_filename: str) -> "str | None":
    """
    Download *local_filename* from the best available HF source, in priority order:

      1. HF_MODEL_REPO  (dedicated model repo, file at repo root)
         repo_type = "model"

      2. SPACE_ID  (the Space's own repo — LFS bytes are stored in HF's
         backend even though the Docker build only sees pointer files)
         repo_type = "space",  path inside repo = "model/<filename>"

    Returns the local cached path on success, None on failure.
    """
    if not _hf_hub_available():
        return None

    from huggingface_hub import hf_hub_download

    os.makedirs(HF_CACHE_DIR, exist_ok=True)

    # Build candidate list: (repo_id, repo_path, repo_type)
    candidates = []
    if HF_MODEL_REPO:
        candidates.append((HF_MODEL_REPO, local_filename, "model"))
    if SPACE_ID:
        candidates.append((SPACE_ID, f"model/{local_filename}", "space"))

    if not candidates:
        logger.error(
            f"Cannot auto-download '{local_filename}': neither HF_MODEL_REPO nor SPACE_ID is set.\n"
            "  • Set HF_MODEL_REPO to a dedicated model repo  (e.g. 'yourname/plantcare-models'), OR\n"
            "  • The SPACE_ID env var should be injected automatically by HF Spaces — "
            "if missing, add it manually in Space Settings → Variables."
        )
        return None

    for repo_id, repo_path, repo_type in candidates:
        try:
            logger.info(
                f"Downloading '{repo_path}' from {repo_type} repo '{repo_id}' → {HF_CACHE_DIR} ..."
            )
            local_path = hf_hub_download(
                repo_id=repo_id,
                filename=repo_path,
                repo_type=repo_type,
                cache_dir=HF_CACHE_DIR,
                token=HF_TOKEN,
                force_download=False,   # reuse cached copy on restarts
            )
            size = os.path.getsize(local_path)
            logger.info(f"Downloaded '{local_filename}' → {local_path} ({size:,} bytes)")
            return local_path
        except Exception as exc:
            logger.warning(
                f"Download attempt failed for '{repo_path}' from '{repo_id}' ({repo_type}): {exc}"
            )

    logger.error(f"All HF download attempts failed for '{local_filename}'.")
    return None


def resolve_model_path(local_filename: str) -> "str | None":
    """
    Return the best usable path for *local_filename*:
      1. Local file in MODEL_DIR — if it exists and is a real binary, use it directly.
      2. Otherwise attempt to download from HF (model repo or Space repo).
      3. Return None if nothing works.
    """
    local_path = os.path.join(MODEL_DIR, local_filename)

    # Fast path — local binary exists
    if os.path.exists(local_path) and not _is_lfs_pointer(local_path):
        logger.info(f"Using local model file: {local_path}")
        return local_path

    # Diagnostics before attempting download
    _log_file_diagnostics(local_path, local_filename)

    if _is_lfs_pointer(local_path):
        logger.warning(f"'{local_filename}' is a Git LFS pointer — downloading real weights from HF.")
    else:
        logger.warning(f"'{local_filename}' not found locally — attempting HF download.")

    downloaded = _download_from_hf(local_filename)
    if downloaded and os.path.exists(downloaded) and not _is_lfs_pointer(downloaded):
        return downloaded

    if downloaded:
        logger.error(f"HF download for '{local_filename}' produced another LFS pointer or empty file.")
    return None


def find_model_path(base_dir, *candidates):
    """Return the first existing, non-LFS model file path from candidate names."""
    for candidate in candidates:
        candidate_path = os.path.join(base_dir, candidate)
        # Prefer a resolved (possibly downloaded) path over a raw local path
        resolved = resolve_model_path(candidate)
        if resolved:
            if candidate != candidates[0]:
                logger.warning(f"Using fallback model file '{candidate}' instead of '{candidates[0]}'")
            return resolved
        # Fall back to bare existence check (covers edge-cases where MODEL_DIR != base_dir)
        if os.path.exists(candidate_path) and not _is_lfs_pointer(candidate_path):
            return candidate_path
    raise FileNotFoundError(
        f"No valid (non-LFS) model file found in {base_dir}. Tried: {', '.join(candidates)}"
    )


# ---------------------------------------------------------------------------
# Model architecture definitions
# ---------------------------------------------------------------------------

class MangoCNN(torch.nn.Module):
    def __init__(self, num_classes=8):
        super().__init__()
        self.conv1 = torch.nn.Conv2d(3, 32, kernel_size=3, padding=1)
        self.bn1 = torch.nn.BatchNorm2d(32)
        self.conv2 = torch.nn.Conv2d(32, 64, kernel_size=3, padding=1)
        self.bn2 = torch.nn.BatchNorm2d(64)
        self.conv3 = torch.nn.Conv2d(64, 128, kernel_size=3, padding=1)
        self.bn3 = torch.nn.BatchNorm2d(128)
        self.conv4 = torch.nn.Conv2d(128, 256, kernel_size=3, padding=1)
        self.bn4 = torch.nn.BatchNorm2d(256)
        self.pool = torch.nn.MaxPool2d(2, 2)
        self.fc1 = torch.nn.Linear(256 * 14 * 14, 512)
        self.fc2 = torch.nn.Linear(512, num_classes)

    def forward(self, x):
        x = self.pool(torch.nn.functional.relu(self.bn1(self.conv1(x))))
        x = self.pool(torch.nn.functional.relu(self.bn2(self.conv2(x))))
        x = self.pool(torch.nn.functional.relu(self.bn3(self.conv3(x))))
        x = self.pool(torch.nn.functional.relu(self.bn4(self.conv4(x))))
        x = x.view(x.size(0), -1)
        x = torch.nn.functional.relu(self.fc1(x))
        x = self.fc2(x)
        return x


def get_preprocess(plant_type):
    """Returns model-specific image preprocessing pipeline."""
    size = 380 if plant_type == "potato" else 224
    return transforms.Compose([
        transforms.Resize((size, size)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    ])


# ---------------------------------------------------------------------------
# Model loading
# ---------------------------------------------------------------------------

def load_disease_model(plant_type, base_dir):
    """Load and return a (model, classes) tuple for the given plant type."""
    plant_type = plant_type.lower()

    if plant_type == "apple":
        model_path = find_model_path(base_dir, "apple_efficientnet_best.pth", "apple_model.pth")
        model_type = "efficientnet_b0"
    elif plant_type == "tomato":
        model_path = find_model_path(base_dir, "tomato_efficientnet_best.pth", "tomato_model.pth")
        model_type = "efficientnet_b0"
    elif plant_type == "potato":
        model_path = find_model_path(base_dir, "potato_efficientnet_best.pth", "potato_model.pth")
        model_type = "efficientnet_b4"
    elif plant_type == "cucumber":
        model_path = find_model_path(base_dir, "best_cucumber_resnet18.pth", "best_model.pth")
        model_type = "resnet18"
    elif plant_type == "lemon":
        model_path = find_model_path(base_dir, "best_lemon_model_ema.pth", "lemon_model.pth")
        model_type = "efficientnet_b0_timm"
    elif plant_type == "mango":
        model_path = find_model_path(base_dir, "best_mango_model.pth", "mango_model.pth")
        model_type = "mango_cnn"
    elif plant_type in ["soybean", "soyabean"]:
        model_path = find_model_path(base_dir, "soybean_disease_model_efficientnetb2.pth", "soybean_model.pth")
        model_type = "efficientnet_b2"
    else:
        logger.warning(f"No disease classification mapping defined for plant type: {plant_type}")
        return None, None

    # validate_model_artifact is now only a final guard – resolve_model_path already handles
    # the LFS-pointer case before we get here, but keep it as a safety net.
    validate_model_artifact(model_path, f"{plant_type} disease model")

    logger.info(f"Loading {plant_type} disease model from {os.path.basename(model_path)} ...")
    checkpoint = torch.load(model_path, map_location="cpu")

    # Resolve class list
    if isinstance(checkpoint, dict) and "classes" in checkpoint:
        classes = checkpoint["classes"]
    elif isinstance(checkpoint, dict) and "class_names" in checkpoint:
        classes = checkpoint["class_names"]
    else:
        if plant_type == "cucumber":
            classes = [
                'Anthracnose', 'Bacterial Wilt', 'Belly Rot', 'Downy Mildew',
                'Fresh Cucumber', 'Fresh Leaf', 'Gummy Stem Blight', 'Pythium Fruit Rot'
            ]
        elif plant_type == "lemon":
            classes = [
                'Anthracnose', 'Bacterial Blight', 'Citrus Canker', 'Curl Virus',
                'Deficiency Leaf', 'Dry Leaf', 'Healthy Leaf', 'Sooty Mould', 'Spider Mites'
            ]
        elif plant_type == "mango":
            classes = [
                'Anthracnose', 'Bacterial Canker', 'Cutting Weevil', 'Die Back',
                'Gall Midge', 'Healthy', 'Powdery Mildew', 'Sooty Mould'
            ]
        elif plant_type in ["soybean", "soyabean"]:
            classes = [
                'Bacterial Blight', 'Cercospora Leaf Blight', 'Healthy',
                'Soybean Rust', 'Sudden Death Syndrome'
            ]
        else:
            classes = []

    # Instantiate architecture and load weights
    if model_type == "efficientnet_b0":
        model = models.efficientnet_b0(weights=None)
        in_features = model.classifier[1].in_features
        model.classifier[1] = torch.nn.Linear(in_features, len(classes))
        model.load_state_dict(checkpoint["model_state_dict"])

    elif model_type == "efficientnet_b2":
        model = models.efficientnet_b2(weights=None)
        in_features = model.classifier[1].in_features
        model.classifier[1] = torch.nn.Linear(in_features, len(classes))
        model.load_state_dict(checkpoint)

    elif model_type == "efficientnet_b4":
        import timm
        model = timm.create_model('efficientnet_b4', pretrained=False, num_classes=len(classes))
        model.load_state_dict(checkpoint["model_state_dict"])

    elif model_type == "efficientnet_b0_timm":
        import timm
        model = timm.create_model('efficientnet_b0', pretrained=False)
        model.classifier = torch.nn.Sequential(
            torch.nn.Dropout(p=0.2),
            torch.nn.Linear(1280, len(classes))
        )
        new_state_dict = {
            k[7:] if k.startswith("module.") else k: v
            for k, v in checkpoint.items()
            if k != "n_averaged"
        }
        model.load_state_dict(new_state_dict)

    elif model_type == "resnet18":
        model = models.resnet18(weights=None)
        model.fc = torch.nn.Sequential(
            torch.nn.Linear(512, 256),
            torch.nn.ReLU(),
            torch.nn.Dropout(0.5),
            torch.nn.Linear(256, len(classes))
        )
        model.load_state_dict(checkpoint)

    elif model_type == "mango_cnn":
        model = MangoCNN(num_classes=len(classes))
        model.load_state_dict(checkpoint)

    else:
        logger.error(f"Unknown model architecture type: {model_type}")
        return None, None

    model = model.to(device)
    model.eval()
    logger.info(f"Successfully loaded {plant_type} model ({len(classes)} classes).")
    return model, classes


# ---------------------------------------------------------------------------
# Startup model pre-loading
# ---------------------------------------------------------------------------

def preload_all_models():
    """Load and cache all detection and classification models at server startup."""
    global models_load_error, model_load_reports

    with models_lock:
        start_time = time.time()
        logger.info("=== Starting model pre-load ===")
        logger.info(f"HF_MODEL_REPO : '{HF_MODEL_REPO or '(not set)'}'")
        logger.info(f"SPACE_ID      : '{SPACE_ID or '(not set)'}'")
        logger.info(f"HF_CACHE_DIR  : '{HF_CACHE_DIR}'")
        model_load_reports = []

        try:
            model_files = os.listdir(MODEL_DIR)
        except Exception:
            model_files = []
        logger.info(f"MODEL_DIR: {MODEL_DIR}  |  files: {model_files}")

        # ------------------------------------------------------------------ #
        # 1. YOLO detector                                                    #
        # ------------------------------------------------------------------ #
        try:
            yolo_path = find_model_path(MODEL_DIR, "best.pt", "detector.pt")
            validate_model_artifact(yolo_path, "YOLO model")
            logger.info(f"Loading YOLO model from {yolo_path} ...")
            models_cache["yolo"] = YOLO(yolo_path)
            logger.info("YOLO model loaded successfully.")
        except Exception as exc:
            models_load_error = str(exc)
            logger.critical(f"YOLO model loading failed: {exc}", exc_info=True)
            models_ready.set()
            return

        # ------------------------------------------------------------------ #
        # 2. Plant disease CNN models                                         #
        # ------------------------------------------------------------------ #
        supported_plants = ["apple", "tomato", "potato", "cucumber", "lemon", "mango", "soybean"]
        for plant in supported_plants:
            try:
                model, classes = load_disease_model(plant, MODEL_DIR)
                if model is not None:
                    models_cache["disease_models"][plant] = (model, classes)
                    if plant == "soybean":
                        models_cache["disease_models"]["soyabean"] = (model, classes)
                    model_load_reports.append({
                        "plant": plant,
                        "status": "loaded",
                        "classes_count": len(classes)
                    })
            except Exception as exc:
                logger.error(f"Failed to load model for {plant}: {exc}", exc_info=True)
                model_load_reports.append({
                    "plant": plant,
                    "status": "error",
                    "error": str(exc)
                })

        elapsed = time.time() - start_time
        loaded = [r["plant"] for r in model_load_reports if r["status"] == "loaded"]
        failed = [r["plant"] for r in model_load_reports if r["status"] == "error"]
        logger.info(
            f"=== Model pre-load done in {elapsed:.1f}s | "
            f"loaded: {loaded} | failed: {failed} ==="
        )

        models_load_error = None
        models_ready.set()


def ensure_models_loaded():
    """Trigger model loading on demand if not already done."""
    global models_load_error

    if models_cache["yolo"] is not None:
        models_ready.set()
        return True

    with models_lock:
        if models_cache["yolo"] is not None:
            models_ready.set()
            return True
        try:
            preload_all_models()
        except Exception as exc:
            models_load_error = str(exc)
            logger.error(f"Model loading failed: {exc}", exc_info=True)
            return False

    return models_cache["yolo"] is not None and models_load_error is None


# ---------------------------------------------------------------------------
# Flask helpers
# ---------------------------------------------------------------------------

def allowed_file(filename, mime_type):
    ext = os.path.splitext(filename)[1].lower()
    valid = ext in ALLOWED_EXTENSIONS and mime_type in ALLOWED_MIME_TYPES
    if not valid:
        logger.warning(f"File rejected — name: '{filename}', ext: '{ext}', mime: '{mime_type}'")
    return valid


# ---------------------------------------------------------------------------
# Flask routes
# ---------------------------------------------------------------------------

@app.errorhandler(413)
def request_entity_too_large(error):
    return jsonify({"success": False, "message": "File size exceeds the 10 MB limit"}), 413


@app.route("/", methods=["GET"])
def index():
    return send_file("main.html")


@app.route("/health", methods=["GET"])
def health_check():
    loaded_disease_models = list(models_cache["disease_models"].keys())
    if "soyabean" in loaded_disease_models and "soybean" in loaded_disease_models:
        loaded_disease_models.remove("soyabean")

    return jsonify({
        "status": "healthy" if models_cache["yolo"] is not None and models_load_error is None else "starting",
        "device": str(device),
        "yolo_loaded": models_cache["yolo"] is not None,
        "loaded_disease_models": loaded_disease_models,
        "model_load_reports": model_load_reports,
        "debug_save_images": DEBUG_SAVE_IMAGES,
        "models_loading": models_cache["yolo"] is None,
        "load_error": models_load_error,
        "hf_model_repo": HF_MODEL_REPO or "(not configured)",
        "space_id": SPACE_ID or "(not detected)",
    }), 200


@app.route("/predict", methods=["POST"])
def predict():
    """Two-stage inference: YOLO plant detection → per-plant disease classification."""
    if not ensure_models_loaded():
        return jsonify({
            "success": False,
            "message": models_load_error or "Models are still loading. Please retry shortly."
        }), 503

    if 'image' not in request.files:
        return jsonify({"success": False, "message": "No file in form-data key 'image'"}), 400

    file = request.files['image']
    if file.filename == '':
        return jsonify({"success": False, "message": "No file selected for upload"}), 400

    mime_type = file.content_type
    logger.info(f"Received file: '{file.filename}', MIME: '{mime_type}'")
    if not allowed_file(file.filename, mime_type):
        return jsonify({
            "success": False,
            "message": "Invalid file format. Only JPG, JPEG, and PNG are allowed."
        }), 400

    try:
        img_bytes = file.read()

        if DEBUG_SAVE_IMAGES:
            os.makedirs(UPLOAD_FOLDER, exist_ok=True)
            debug_filename = f"img_{int(time.time() * 1000)}{os.path.splitext(file.filename)[1]}"
            debug_path = os.path.join(UPLOAD_FOLDER, debug_filename)
            with open(debug_path, "wb") as fh:
                fh.write(img_bytes)
            logger.info(f"Debug: saved image to {debug_path}")

        img = Image.open(io.BytesIO(img_bytes)).convert("RGB")

        yolo_model = models_cache["yolo"]
        if yolo_model is None:
            return jsonify({"success": False, "message": "YOLO model is not loaded"}), 500

        device_str = "cuda" if device.type == "cuda" else "cpu"
        results = yolo_model.predict(
            source=img, conf=0.25, iou=0.45,
            save=False, save_txt=False, verbose=False, device=device_str
        )
        result = results[0]

        if len(result.boxes) == 0:
            return jsonify({"success": False, "message": "No plant detected in the image"}), 200

        detections = []
        for box in result.boxes:
            class_id = int(box.cls[0])
            detection_confidence = float(box.conf[0])
            plant_type = yolo_model.names[class_id].lower()
            x1, y1, x2, y2 = map(int, box.xyxy[0].tolist())

            if plant_type in models_cache["disease_models"]:
                disease_model, classes = models_cache["disease_models"][plant_type]
                cropped_img = img.crop((x1, y1, x2, y2))
                preprocess = get_preprocess(plant_type)
                img_tensor = preprocess(cropped_img).unsqueeze(0).to(device)
                with torch.no_grad():
                    outputs = disease_model(img_tensor)
                    probs = torch.nn.functional.softmax(outputs[0], dim=0)
                    pred_idx = torch.argmax(probs).item()
                    disease_name = classes[pred_idx]
                    disease_confidence = float(probs[pred_idx].item())
            else:
                disease_name = "Unknown Disease (Unsupported Plant)"
                disease_confidence = 0.0
                logger.warning(f"No disease model cached for '{plant_type}'.")

            detections.append({
                "plant_type": plant_type,
                "detection_confidence": round(detection_confidence, 4),
                "box": [x1, y1, x2, y2],
                "disease": disease_name,
                "disease_confidence": round(disease_confidence, 4)
            })

        return jsonify({"success": True, "detections": detections}), 200

    except Exception as exc:
        logger.error(f"Inference pipeline failed: {exc}", exc_info=True)
        return jsonify({
            "success": False,
            "message": f"Server error during inference: {exc}"
        }), 500


# ---------------------------------------------------------------------------
# Entry point
# ---------------------------------------------------------------------------

if __name__ == "__main__":
    # Kick off model loading in a background thread so the HTTP server starts immediately.
    threading.Thread(target=preload_all_models, daemon=True).start()
    port = int(os.environ.get("PORT", 7860))
    app.run(host="0.0.0.0", port=port, debug=False)