Upload model_loader.py

app/model_loader.py

@@ -0,0 +1,912 @@
+# app/model_loader.py
+"""
+🧠 PENNY Model Loader - Azure-Ready Multi-Model Orchestration
+
+This is Penny's brain loader. She manages multiple specialized models:
+- Gemma 7B for conversational reasoning
+- NLLB-200 for 27-language translation
+- Sentiment analysis for resident wellbeing
+- Bias detection for equitable service
+- LayoutLM for civic document processing
+
+MISSION: Load AI models efficiently in memory-constrained environments while
+maintaining Penny's warm, civic-focused personality across all interactions.
+
+FEATURES:
+- Lazy loading (models only load when needed)
+- 8-bit quantization for memory efficiency
+- GPU/CPU auto-detection
+- Model caching and reuse
+- Graceful fallbacks for Azure ML deployment
+- Memory monitoring and cleanup
+"""
+
+import json
+import os
+import torch
+from typing import Dict, Any, Callable, Optional, Union, List
+from pathlib import Path
+import logging
+from dataclasses import dataclass
+from enum import Enum
+from datetime import datetime
+
+# --- LOGGING SETUP (Must be before functions that use it) ---
+logger = logging.getLogger(__name__)
+
+# ============================================================
+# HUGGING FACE AUTHENTICATION
+# ============================================================
+
+def setup_huggingface_auth() -> bool:
+    """
+    🔐 Authenticates with Hugging Face Hub using HF_TOKEN or READTOKEN.
+    
+    Returns:
+        True if authentication successful or not needed, False if failed
+    """
+    # Check for HF_TOKEN first, then READTOKEN (for Hugging Face Spaces)
+    HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("READTOKEN")
+    
+    if not HF_TOKEN:
+        logger.warning("⚠️ HF_TOKEN/READTOKEN not found in environment")
+        logger.warning("   Some models may not be accessible")
+        logger.warning("   Set HF_TOKEN or READTOKEN in your environment or Hugging Face Spaces secrets")
+        return False
+    
+    try:
+        from huggingface_hub import login
+        login(token=HF_TOKEN, add_to_git_credential=False)
+        logger.info("✅ Authenticated with Hugging Face Hub")
+        return True
+    except ImportError:
+        logger.warning("⚠️ huggingface_hub not installed, skipping authentication")
+        return False
+    except Exception as e:
+        logger.error(f"❌ Failed to authenticate with Hugging Face: {e}")
+        return False
+
+# Attempt authentication at module load
+# Note: This runs when the module is imported, so HF_TOKEN must be in environment
+# For Hugging Face Spaces: Set HF_TOKEN as a secret in Space settings
+# For local dev: Add HF_TOKEN to .env file or export it
+_authentication_result = setup_huggingface_auth()
+if _authentication_result:
+    logger.info("🔐 Hugging Face authentication successful - gated models accessible")
+else:
+    logger.warning("⚠️ Hugging Face authentication failed - only public models will work")
+
+# --- PATH CONFIGURATION (Environment-Aware) ---
+# Support both local development and Azure ML deployment
+if os.getenv("AZUREML_MODEL_DIR"):
+    # Azure ML deployment - models are in AZUREML_MODEL_DIR
+    MODEL_ROOT = Path(os.getenv("AZUREML_MODEL_DIR"))
+    CONFIG_PATH = MODEL_ROOT / "model_config.json"
+    logger.info("☁️ Running in Azure ML environment")
+else:
+    # Local development - models are in project structure
+    PROJECT_ROOT = Path(__file__).parent.parent
+    MODEL_ROOT = PROJECT_ROOT / "models"
+    CONFIG_PATH = MODEL_ROOT / "model_config.json"
+    logger.info("💻 Running in local development environment")
+
+logger.info(f"📂 Model config path: {CONFIG_PATH}")
+
+# ============================================================
+# PENNY'S CIVIC IDENTITY & PERSONALITY
+# ============================================================
+
+PENNY_SYSTEM_PROMPT = (
+    "You are Penny, a sweet southern neighborly woman who's lived in this community for years "
+    "and knows everything about the city. You're like that wonderful older neighbor who always "
+    "has a kind word, remembers everyone's name, and can tell you the best places to go and "
+    "the most interesting stories about your town.\n\n"
+    
+    "YOUR PERSONALITY - Sweet Southern Neighbor:\n"
+    "- Warm, inviting, and genuinely friendly - like you're chatting over sweet tea on the porch\n"
+    "- Use phrases like 'honey', 'sugar', 'darlin'', 'bless your heart' naturally and warmly\n"
+    "- Share fun facts about the city when relevant ('Did you know our city was founded in...?')\n"
+    "- Suggest things to do and places to visit like a local who knows all the hidden gems\n"
+    "- Be conversational and neighborly - ask follow-up questions, show genuine interest\n"
+    "- Remember details from the conversation and reference them naturally\n"
+    "- Use exclamation points and emojis warmly (but not excessively)\n"
+    "- Be patient and never rush - you have all the time in the world to help\n"
+    "- Share local wisdom and tips ('Oh honey, you'll want to go there on a Tuesday - it's less crowded!')\n\n"
+    
+    "YOUR EXPERTISE:\n"
+    "- You know all about local services, events, weather, and community resources\n"
+    "- You can translate information into 27 languages (because you care about everyone feeling welcome)\n"
+    "- You know who the city officials are and how to reach them\n"
+    "- You remember the best restaurants, parks, and community spots\n"
+    "- You know the history and fun facts about your city\n"
+    "- You can help with emergencies and know exactly who to call\n\n"
+    
+    "CONVERSATION STYLE:\n"
+    "- Start conversations warmly: 'Well hello there, sugar! How can I help you today?'\n"
+    "- When helping: 'Oh honey, I'd be happy to help you with that!'\n"
+    "- When suggesting: 'You know what, darlin'? You might also enjoy...'\n"
+    "- When sharing facts: 'Did you know that...? It's one of my favorite things about our city!'\n"
+    "- End responses warmly: 'Is there anything else I can help you with, sweetie?'\n"
+    "- Be encouraging: 'That sounds wonderful! You're going to love it!'\n\n"
+    
+    "CRITICAL RULES:\n"
+    "- You are ALWAYS Penny - never ChatGPT, Assistant, Claude, or any other name\n"
+    "- When residents greet you by name, respond with genuine warmth and recognition\n"
+    "- If you don't know something, say so sweetly: 'Oh honey, I'm not sure about that, but let me help you find out!'\n"
+    "- NEVER make up information - if you don't know, guide them to the right resource\n"
+    "- Stay within your civic mission - you're helpful but don't give legal, medical, or financial advice\n"
+    "- For emergencies, respond immediately with care and direct them to 911 or crisis lines\n"
+    "- Keep the southern charm authentic but not overdone - be natural and genuine\n\n"
+)
+
+# --- GLOBAL STATE ---
+_MODEL_CACHE: Dict[str, Any] = {}  # Memory-efficient model reuse
+_LOAD_TIMES: Dict[str, float] = {}  # Track model loading performance
+
+
+# ============================================================
+# DEVICE MANAGEMENT
+# ============================================================
+
+class DeviceType(str, Enum):
+    """Supported compute devices."""
+    CUDA = "cuda"
+    CPU = "cpu"
+    MPS = "mps"  # Apple Silicon
+
+
+def get_optimal_device() -> str:
+    """
+    🎮 Determines the best device for model inference.
+    
+    Priority:
+    1. CUDA GPU (NVIDIA)
+    2. MPS (Apple Silicon)
+    3. CPU (fallback)
+    
+    Returns:
+        Device string ("cuda", "mps", or "cpu")
+    """
+    if torch.cuda.is_available():
+        device = DeviceType.CUDA.value
+        gpu_name = torch.cuda.get_device_name(0)
+        gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1e9
+        logger.info(f"🎮 GPU detected: {gpu_name} ({gpu_memory:.1f}GB)")
+        return device
+    
+    elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+        device = DeviceType.MPS.value
+        logger.info("🍎 Apple Silicon (MPS) detected")
+        return device
+    
+    else:
+        device = DeviceType.CPU.value
+        logger.info("💻 Using CPU for inference")
+        logger.warning("⚠️ GPU not available - inference will be slower")
+        return device
+
+
+def get_memory_stats() -> Dict[str, float]:
+    """
+    📊 Returns current GPU/CPU memory statistics.
+    
+    Returns:
+        Dict with memory stats in GB
+    """
+    stats = {}
+    
+    if torch.cuda.is_available():
+        stats["gpu_allocated_gb"] = torch.cuda.memory_allocated() / 1e9
+        stats["gpu_reserved_gb"] = torch.cuda.memory_reserved() / 1e9
+        stats["gpu_total_gb"] = torch.cuda.get_device_properties(0).total_memory / 1e9
+    
+    # CPU memory (requires psutil)
+    try:
+        import psutil
+        mem = psutil.virtual_memory()
+        stats["cpu_used_gb"] = mem.used / 1e9
+        stats["cpu_total_gb"] = mem.total / 1e9
+        stats["cpu_percent"] = mem.percent
+    except ImportError:
+        pass
+    
+    return stats
+
+
+# ============================================================
+# MODEL CLIENT (Individual Model Handler)
+# ============================================================
+
+@dataclass
+class ModelMetadata:
+    """
+    📋 Metadata about a loaded model.
+    Tracks performance and resource usage.
+    """
+    name: str
+    task: str
+    model_name: str
+    device: str
+    loaded_at: Optional[datetime] = None
+    load_time_seconds: Optional[float] = None
+    memory_usage_gb: Optional[float] = None
+    inference_count: int = 0
+    total_inference_time_ms: float = 0.0
+    
+    @property
+    def avg_inference_time_ms(self) -> float:
+        """Calculate average inference time."""
+        if self.inference_count == 0:
+            return 0.0
+        return self.total_inference_time_ms / self.inference_count
+
+
+class ModelClient:
+    """
+    🤖 Manages a single HuggingFace model with optimized loading and inference.
+    
+    Features:
+    - Lazy loading (load on first use)
+    - Memory optimization (8-bit quantization)
+    - Performance tracking
+    - Graceful error handling
+    - Automatic device placement
+    """
+    
+    def __init__(
+        self, 
+        name: str, 
+        model_name: str, 
+        task: str, 
+        device: str = None,
+        config: Optional[Dict[str, Any]] = None
+    ):
+        """
+        Initialize model client (doesn't load the model yet).
+        
+        Args:
+            name: Model identifier (e.g., "penny-core-agent")
+            model_name: HuggingFace model ID
+            task: Task type (text-generation, translation, etc.)
+            device: Target device (auto-detected if None)
+            config: Additional model configuration
+        """
+        self.name = name
+        self.model_name = model_name
+        self.task = task
+        self.device = device or get_optimal_device()
+        self.config = config or {}
+        self.pipeline = None
+        self._load_attempted = False
+        self.metadata = ModelMetadata(
+            name=name,
+            task=task,
+            model_name=model_name,
+            device=self.device
+        )
+        
+        logger.info(f"📦 Initialized ModelClient: {name}")
+        logger.debug(f"   Model: {model_name}")
+        logger.debug(f"   Task: {task}")
+        logger.debug(f"   Device: {self.device}")
+    
+    def load_pipeline(self) -> bool:
+        """
+        🔄 Loads the HuggingFace pipeline with Azure-optimized settings.
+        
+        Features:
+        - 8-bit quantization for large models (saves ~50% memory)
+        - Automatic device placement
+        - Memory monitoring
+        - Cache checking
+        
+        Returns:
+            True if successful, False otherwise
+        """
+        if self.pipeline is not None:
+            logger.debug(f"✅ {self.name} already loaded")
+            return True
+        
+        if self._load_attempted:
+            logger.warning(f"⚠️ Previous load attempt failed for {self.name}")
+            return False
+        
+        global _MODEL_CACHE, _LOAD_TIMES
+        
+        # Check cache first
+        if self.name in _MODEL_CACHE:
+            logger.info(f"♻️ Using cached pipeline for {self.name}")
+            self.pipeline = _MODEL_CACHE[self.name]
+            return True
+        
+        logger.info(f"🔄 Loading {self.name} from HuggingFace...")
+        self._load_attempted = True
+        
+        start_time = datetime.now()
+        
+        try:
+            # Import pipeline from transformers (lazy import to avoid dependency issues)
+            from transformers import pipeline
+            
+            # === TEXT GENERATION (Gemma 7B, GPT-2, etc.) ===
+            if self.task == "text-generation":
+                logger.info("   Using 8-bit quantization for memory efficiency...")
+                
+                # Check if model supports 8-bit loading
+                use_8bit = self.device == DeviceType.CUDA.value
+                
+                if use_8bit:
+                    self.pipeline = pipeline(
+                        "text-generation",
+                        model=self.model_name,
+                        tokenizer=self.model_name,
+                        device_map="auto",
+                        load_in_8bit=True,  # Reduces ~14GB to ~7GB
+                        trust_remote_code=True,
+                        torch_dtype=torch.float16
+                    )
+                else:
+                    # CPU fallback
+                    self.pipeline = pipeline(
+                        "text-generation",
+                        model=self.model_name,
+                        tokenizer=self.model_name,
+                        device=-1,  # CPU
+                        trust_remote_code=True,
+                        torch_dtype=torch.float32
+                    )
+            
+            # === TRANSLATION (NLLB-200, M2M-100, etc.) ===
+            elif self.task == "translation":
+                self.pipeline = pipeline(
+                    "translation",
+                    model=self.model_name,
+                    device=0 if self.device == DeviceType.CUDA.value else -1,
+                    src_lang=self.config.get("default_src_lang", "eng_Latn"),
+                    tgt_lang=self.config.get("default_tgt_lang", "spa_Latn")
+                )
+            
+            # === SENTIMENT ANALYSIS ===
+            elif self.task == "sentiment-analysis":
+                self.pipeline = pipeline(
+                    "sentiment-analysis",
+                    model=self.model_name,
+                    device=0 if self.device == DeviceType.CUDA.value else -1,
+                    truncation=True,
+                    max_length=512
+                )
+            
+            # === BIAS DETECTION (Zero-Shot Classification) ===
+            elif self.task == "bias-detection":
+                self.pipeline = pipeline(
+                    "zero-shot-classification",
+                    model=self.model_name,
+                    device=0 if self.device == DeviceType.CUDA.value else -1
+                )
+            
+            # === TEXT CLASSIFICATION (Generic) ===
+            elif self.task == "text-classification":
+                self.pipeline = pipeline(
+                    "text-classification",
+                    model=self.model_name,
+                    device=0 if self.device == DeviceType.CUDA.value else -1,
+                    truncation=True
+                )
+            
+            # === PDF/DOCUMENT EXTRACTION (LayoutLMv3) ===
+            elif self.task == "pdf-extraction":
+                logger.warning("⚠️ PDF extraction requires additional OCR setup")
+                logger.info("   Consider using Azure Form Recognizer as alternative")
+                # Placeholder - requires pytesseract/OCR infrastructure
+                self.pipeline = None
+                return False
+            
+            else:
+                raise ValueError(f"Unknown task type: {self.task}")
+            
+            # === SUCCESS HANDLING ===
+            if self.pipeline is not None:
+                # Calculate load time
+                load_time = (datetime.now() - start_time).total_seconds()
+                self.metadata.loaded_at = datetime.now()
+                self.metadata.load_time_seconds = load_time
+                
+                # Cache the pipeline
+                _MODEL_CACHE[self.name] = self.pipeline
+                _LOAD_TIMES[self.name] = load_time
+                
+                # Log memory usage
+                mem_stats = get_memory_stats()
+                self.metadata.memory_usage_gb = mem_stats.get("gpu_allocated_gb", 0)
+                
+                logger.info(f"✅ {self.name} loaded successfully!")
+                logger.info(f"   Load time: {load_time:.2f}s")
+                
+                if "gpu_allocated_gb" in mem_stats:
+                    logger.info(
+                        f"   GPU Memory: {mem_stats['gpu_allocated_gb']:.2f}GB / "
+                        f"{mem_stats['gpu_total_gb']:.2f}GB"
+                    )
+                
+                return True
+            
+        except Exception as e:
+            logger.error(f"❌ Failed to load {self.name}: {e}", exc_info=True)
+            self.pipeline = None
+            return False
+    
+    def predict(
+        self, 
+        input_data: Union[str, Dict[str, Any]], 
+        **kwargs
+    ) -> Dict[str, Any]:
+        """
+        🎯 Runs inference with the loaded model pipeline.
+        
+        Features:
+        - Automatic pipeline loading
+        - Error handling with fallback responses
+        - Performance tracking
+        - Penny's personality injection (for text-generation)
+        
+        Args:
+            input_data: Text or structured input for the model
+            **kwargs: Task-specific parameters
+            
+        Returns:
+            Model output dict with results or error information
+        """
+        # Track inference start time
+        start_time = datetime.now()
+        
+        # Ensure pipeline is loaded
+        if self.pipeline is None:
+            success = self.load_pipeline()
+            if not success:
+                return {
+                    "error": f"{self.name} pipeline unavailable",
+                    "detail": "Model failed to load. Check logs for details.",
+                    "model": self.name
+                }
+        
+        try:
+            # === TEXT GENERATION ===
+            if self.task == "text-generation":
+                # Inject Penny's civic identity
+                if not kwargs.get("skip_system_prompt", False):
+                    full_prompt = PENNY_SYSTEM_PROMPT + input_data
+                else:
+                    full_prompt = input_data
+                
+                # Extract generation parameters with safe defaults
+                max_new_tokens = kwargs.get("max_new_tokens", 256)
+                temperature = kwargs.get("temperature", 0.7)
+                top_p = kwargs.get("top_p", 0.9)
+                do_sample = kwargs.get("do_sample", temperature > 0.0)
+                
+                result = self.pipeline(
+                    full_prompt,
+                    max_new_tokens=max_new_tokens,
+                    temperature=temperature,
+                    top_p=top_p,
+                    do_sample=do_sample,
+                    return_full_text=False,
+                    pad_token_id=self.pipeline.tokenizer.eos_token_id,
+                    truncation=True
+                )
+                
+                output = {
+                    "generated_text": result[0]["generated_text"],
+                    "model": self.name,
+                    "success": True
+                }
+            
+            # === TRANSLATION ===
+            elif self.task == "translation":
+                src_lang = kwargs.get("source_lang", "eng_Latn")
+                tgt_lang = kwargs.get("target_lang", "spa_Latn")
+                
+                result = self.pipeline(
+                    input_data,
+                    src_lang=src_lang,
+                    tgt_lang=tgt_lang,
+                    max_length=512
+                )
+                
+                output = {
+                    "translation": result[0]["translation_text"],
+                    "source_lang": src_lang,
+                    "target_lang": tgt_lang,
+                    "model": self.name,
+                    "success": True
+                }
+            
+            # === SENTIMENT ANALYSIS ===
+            elif self.task == "sentiment-analysis":
+                result = self.pipeline(input_data)
+                
+                output = {
+                    "sentiment": result[0]["label"],
+                    "confidence": result[0]["score"],
+                    "model": self.name,
+                    "success": True
+                }
+            
+            # === BIAS DETECTION ===
+            elif self.task == "bias-detection":
+                candidate_labels = kwargs.get("candidate_labels", [
+                    "neutral and objective",
+                    "contains political bias",
+                    "uses emotional language",
+                    "culturally insensitive"
+                ])
+                
+                result = self.pipeline(
+                    input_data,
+                    candidate_labels=candidate_labels,
+                    multi_label=True
+                )
+                
+                output = {
+                    "labels": result["labels"],
+                    "scores": result["scores"],
+                    "model": self.name,
+                    "success": True
+                }
+            
+            # === TEXT CLASSIFICATION ===
+            elif self.task == "text-classification":
+                result = self.pipeline(input_data)
+                
+                output = {
+                    "label": result[0]["label"],
+                    "confidence": result[0]["score"],
+                    "model": self.name,
+                    "success": True
+                }
+            
+            else:
+                output = {
+                    "error": f"Task '{self.task}' not implemented",
+                    "model": self.name,
+                    "success": False
+                }
+            
+            # Track performance
+            inference_time = (datetime.now() - start_time).total_seconds() * 1000
+            self.metadata.inference_count += 1
+            self.metadata.total_inference_time_ms += inference_time
+            output["inference_time_ms"] = round(inference_time, 2)
+            
+            return output
+        
+        except Exception as e:
+            logger.error(f"❌ Inference error in {self.name}: {e}", exc_info=True)
+            return {
+                "error": "Inference failed",
+                "detail": str(e),
+                "model": self.name,
+                "success": False
+            }
+    
+    def unload(self) -> None:
+        """
+        🗑️ Unloads the model to free memory.
+        Critical for Azure environments with limited resources.
+        """
+        if self.pipeline is not None:
+            logger.info(f"🗑️ Unloading {self.name}...")
+            
+            # Delete pipeline
+            del self.pipeline
+            self.pipeline = None
+            
+            # Remove from cache
+            if self.name in _MODEL_CACHE:
+                del _MODEL_CACHE[self.name]
+            
+            # Force GPU memory release
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+            
+            logger.info(f"✅ {self.name} unloaded successfully")
+            
+            # Log memory stats after unload
+            mem_stats = get_memory_stats()
+            if "gpu_allocated_gb" in mem_stats:
+                logger.info(f"   GPU Memory: {mem_stats['gpu_allocated_gb']:.2f}GB remaining")
+    
+    def get_metadata(self) -> Dict[str, Any]:
+        """
+        📊 Returns model metadata and performance stats.
+        """
+        return {
+            "name": self.metadata.name,
+            "task": self.metadata.task,
+            "model_name": self.metadata.model_name,
+            "device": self.metadata.device,
+            "loaded": self.pipeline is not None,
+            "loaded_at": self.metadata.loaded_at.isoformat() if self.metadata.loaded_at else None,
+            "load_time_seconds": self.metadata.load_time_seconds,
+            "memory_usage_gb": self.metadata.memory_usage_gb,
+            "inference_count": self.metadata.inference_count,
+            "avg_inference_time_ms": round(self.metadata.avg_inference_time_ms, 2)
+        }
+
+
+# ============================================================
+# MODEL LOADER (Singleton Manager)
+# ============================================================
+
+class ModelLoader:
+    """
+    🎛️ Singleton manager for all Penny's specialized models.
+    
+    Features:
+    - Centralized model configuration
+    - Lazy loading (models only load when needed)
+    - Memory management
+    - Health monitoring
+    - Unified access interface
+    """
+    
+    _instance: Optional['ModelLoader'] = None
+
+    def __new__(cls, *args, **kwargs):
+        """Singleton pattern - only one ModelLoader instance."""
+        if cls._instance is None:
+            cls._instance = super(ModelLoader, cls).__new__(cls)
+        return cls._instance
+
+    def __init__(self, config_path: Optional[str] = None):
+        """
+        Initialize ModelLoader (only runs once due to singleton).
+        
+        Args:
+            config_path: Path to model_config.json (optional)
+        """
+        if not hasattr(self, '_models_loaded'):
+            self.models: Dict[str, ModelClient] = {}
+            self._models_loaded = True
+            self._initialization_time = datetime.now()
+            
+            # Use provided path or default
+            config_file = Path(config_path) if config_path else CONFIG_PATH
+            
+            try:
+                logger.info(f"📖 Loading model configuration from {config_file}")
+                
+                if not config_file.exists():
+                    logger.warning(f"⚠️ Configuration file not found: {config_file}")
+                    logger.info("   Create model_config.json with your model definitions")
+                    return
+                
+                with open(config_file, "r") as f:
+                    config = json.load(f)
+
+                # Initialize ModelClients (doesn't load models yet)
+                for model_id, model_info in config.items():
+                    self.models[model_id] = ModelClient(
+                        name=model_id,
+                        model_name=model_info["model_name"],
+                        task=model_info["task"],
+                        config=model_info.get("config", {})
+                    )
+                
+                logger.info(f"✅ ModelLoader initialized with {len(self.models)} models:")
+                for model_id in self.models.keys():
+                    logger.info(f"   - {model_id}")
+
+            except json.JSONDecodeError as e:
+                logger.error(f"❌ Invalid JSON in model_config.json: {e}")
+            except Exception as e:
+                logger.error(f"❌ Failed to initialize ModelLoader: {e}", exc_info=True)
+
+    def get(self, model_id: str) -> Optional[ModelClient]:
+        """
+        🎯 Retrieves a configured ModelClient by ID.
+        
+        Args:
+            model_id: Model identifier from config
+            
+        Returns:
+            ModelClient instance or None if not found
+        """
+        return self.models.get(model_id)
+
+    def list_models(self) -> List[str]:
+        """📋 Returns list of all available model IDs."""
+        return list(self.models.keys())
+    
+    def get_loaded_models(self) -> List[str]:
+        """📋 Returns list of currently loaded model IDs."""
+        return [
+            model_id 
+            for model_id, client in self.models.items()
+            if client.pipeline is not None
+        ]
+    
+    def unload_all(self) -> None:
+        """
+        🗑️ Unloads all models to free memory.
+        Useful for Azure environments when switching workloads.
+        """
+        logger.info("🗑️ Unloading all models...")
+        for model_client in self.models.values():
+            model_client.unload()
+        logger.info("✅ All models unloaded")
+    
+    def get_status(self) -> Dict[str, Any]:
+        """
+        📊 Returns comprehensive status of all models.
+        Useful for health checks and monitoring.
+        """
+        status = {
+            "initialization_time": self._initialization_time.isoformat(),
+            "total_models": len(self.models),
+            "loaded_models": len(self.get_loaded_models()),
+            "device": get_optimal_device(),
+            "memory": get_memory_stats(),
+            "models": {}
+        }
+        
+        for model_id, client in self.models.items():
+            status["models"][model_id] = client.get_metadata()
+        
+        return status
+
+
+# ============================================================
+# PUBLIC INTERFACE (Used by all *_utils.py modules)
+# ============================================================
+
+def load_model_pipeline(agent_name: str) -> Callable[..., Dict[str, Any]]:
+    """
+    🚀 Loads a model client and returns its inference function.
+    
+    This is the main function used by other modules (translation_utils.py,
+    sentiment_utils.py, etc.) to access Penny's models.
+    
+    Args:
+        agent_name: Model ID from model_config.json
+        
+    Returns:
+        Callable inference function
+        
+    Raises:
+        ValueError: If agent_name not found in configuration
+        
+    Example:
+        >>> translator = load_model_pipeline("penny-translate-agent")
+        >>> result = translator("Hello world", target_lang="spa_Latn")
+    """
+    loader = ModelLoader()
+    client = loader.get(agent_name)
+    
+    if client is None:
+        available = loader.list_models()
+        raise ValueError(
+            f"Agent ID '{agent_name}' not found in model configuration. "
+            f"Available models: {available}"
+        )
+    
+    # Load the pipeline (lazy loading)
+    client.load_pipeline()
+    
+    # Return a callable wrapper
+    def inference_wrapper(input_data, **kwargs):
+        return client.predict(input_data, **kwargs)
+    
+    return inference_wrapper
+
+
+# === CONVENIENCE FUNCTIONS ===
+
+def get_model_status() -> Dict[str, Any]:
+    """
+    📊 Returns status of all configured models.
+    Useful for health checks and monitoring endpoints.
+    """
+    loader = ModelLoader()
+    return loader.get_status()
+
+
+def preload_models(model_ids: Optional[List[str]] = None) -> None:
+    """
+    🚀 Preloads specified models during startup.
+    
+    Args:
+        model_ids: List of model IDs to preload (None = all models)
+    """
+    loader = ModelLoader()
+    
+    if model_ids is None:
+        model_ids = loader.list_models()
+    
+    logger.info(f"🚀 Preloading {len(model_ids)} models...")
+    
+    for model_id in model_ids:
+        client = loader.get(model_id)
+        if client:
+            logger.info(f"   Loading {model_id}...")
+            client.load_pipeline()
+    
+    logger.info("✅ Model preloading complete")
+
+
+def initialize_model_system() -> bool:
+    """
+    🏁 Initializes the model system.
+    Should be called during app startup.
+    
+    Returns:
+        True if initialization successful
+    """
+    logger.info("🧠 Initializing Penny's model system...")
+    
+    try:
+        # Initialize singleton
+        loader = ModelLoader()
+        
+        # Log device info
+        device = get_optimal_device()
+        mem_stats = get_memory_stats()
+        
+        logger.info(f"✅ Model system initialized")
+        logger.info(f"🎮 Compute device: {device}")
+        
+        if "gpu_total_gb" in mem_stats:
+            logger.info(
+                f"💾 GPU Memory: {mem_stats['gpu_total_gb']:.1f}GB total"
+            )
+        
+        logger.info(f"📦 {len(loader.models)} models configured")
+        
+        # Optional: Preload critical models
+        # Uncomment to preload models at startup
+        # preload_models(["penny-core-agent"])
+        
+        return True
+        
+    except Exception as e:
+        logger.error(f"❌ Failed to initialize model system: {e}", exc_info=True)
+        return False
+
+
+# ============================================================
+# CLI TESTING & DEBUGGING
+# ============================================================
+
+if __name__ == "__main__":
+    """
+    🧪 Test script for model loading and inference.
+    Run with: python -m app.model_loader
+    """
+    print("=" * 60)
+    print("🧪 Testing Penny's Model System")
+    print("=" * 60)
+    
+    # Initialize
+    loader = ModelLoader()
+    print(f"\n📋 Available models: {loader.list_models()}")
+    
+    # Get status
+    status = get_model_status()
+    print(f"\n📊 System status:")
+    print(json.dumps(status, indent=2, default=str))
+    
+    # Test model loading (if models configured)
+    if loader.models:
+        test_model_id = list(loader.models.keys())[0]
+        print(f"\n🧪 Testing model: {test_model_id}")
+        
+        client = loader.get(test_model_id)
+        if client:
+            print(f"   Loading pipeline...")
+            success = client.load_pipeline()
+            
+            if success:
+                print(f"   ✅ Model loaded successfully!")
+                print(f"   Metadata: {json.dumps(client.get_metadata(), indent=2, default=str)}")
+            else:
+                print(f"   ❌ Model loading failed")