gptmed 0.3.4__py3-none-any.whl → 0.3.5__py3-none-any.whl

gptmed/services/__init__.py

@@ -0,0 +1,15 @@
+"""
+Services Layer
+
+Business logic services following SOLID principles.
+This layer implements the service pattern to encapsulate complex operations.
+"""
+
+from gptmed.services.device_manager import DeviceManager, DeviceStrategy
+from gptmed.services.training_service import TrainingService
+
+__all__ = [
+    'DeviceManager',
+    'DeviceStrategy',
+    'TrainingService',
+]

gptmed/services/device_manager.py

@@ -0,0 +1,252 @@
+"""
+Device Manager Service
+
+PURPOSE:
+Manages device selection and configuration for model training and inference.
+Implements Strategy Pattern for flexible device handling.
+
+DESIGN PATTERNS:
+- Strategy Pattern: Different strategies for CPU vs GPU
+- Dependency Injection: DeviceManager can be injected into services
+- Single Responsibility: Only handles device-related concerns
+
+WHAT THIS FILE DOES:
+1. Validates device availability (CUDA check)
+2. Provides device selection logic with fallback
+3. Manages device-specific configurations
+4. Ensures consistent device handling across the codebase
+
+PACKAGES USED:
+- torch: Device detection and management
+- abc: Abstract base classes for strategy pattern
+"""
+
+from abc import ABC, abstractmethod
+from typing import Optional
+import torch
+
+
+class DeviceStrategy(ABC):
+    """
+    Abstract base class for device strategies.
+    Implements Strategy Pattern for different device types.
+    """
+    
+    @abstractmethod
+    def get_device(self) -> str:
+        """
+        Get the device string for PyTorch.
+        
+        Returns:
+            Device string ('cuda' or 'cpu')
+        """
+        pass
+    
+    @abstractmethod
+    def is_available(self) -> bool:
+        """
+        Check if the device is available.
+        
+        Returns:
+            True if device is available, False otherwise
+        """
+        pass
+    
+    @abstractmethod
+    def get_device_info(self) -> dict:
+        """
+        Get information about the device.
+        
+        Returns:
+            Dictionary with device information
+        """
+        pass
+
+
+class CUDAStrategy(DeviceStrategy):
+    """Strategy for CUDA/GPU devices."""
+    
+    def get_device(self) -> str:
+        """Get CUDA device if available."""
+        return 'cuda' if self.is_available() else 'cpu'
+    
+    def is_available(self) -> bool:
+        """Check if CUDA is available."""
+        return torch.cuda.is_available()
+    
+    def get_device_info(self) -> dict:
+        """Get CUDA device information."""
+        if not self.is_available():
+            return {
+                'device': 'cuda',
+                'available': False,
+                'message': 'CUDA not available'
+            }
+        
+        return {
+            'device': 'cuda',
+            'available': True,
+            'device_name': torch.cuda.get_device_name(0),
+            'device_count': torch.cuda.device_count(),
+            'cuda_version': torch.version.cuda if torch.version.cuda else 'N/A',
+        }
+
+
+class CPUStrategy(DeviceStrategy):
+    """Strategy for CPU devices."""
+    
+    def get_device(self) -> str:
+        """Always return CPU."""
+        return 'cpu'
+    
+    def is_available(self) -> bool:
+        """CPU is always available."""
+        return True
+    
+    def get_device_info(self) -> dict:
+        """Get CPU device information."""
+        return {
+            'device': 'cpu',
+            'available': True,
+            'num_threads': torch.get_num_threads(),
+        }
+
+
+class DeviceManager:
+    """
+    Manages device selection and configuration.
+    
+    Follows Single Responsibility Principle - only handles device concerns.
+    Uses Strategy Pattern for different device types.
+    
+    Example:
+        >>> device_manager = DeviceManager(preferred_device='cuda')
+        >>> device = device_manager.get_device()
+        >>> print(f"Using device: {device}")
+    """
+    
+    def __init__(self, preferred_device: str = 'cuda', allow_fallback: bool = True):
+        """
+        Initialize DeviceManager.
+        
+        Args:
+            preferred_device: Preferred device ('cuda' or 'cpu')
+            allow_fallback: If True, fallback to CPU if CUDA unavailable
+        """
+        self.preferred_device = preferred_device.lower()
+        self.allow_fallback = allow_fallback
+        
+        # Validate device input
+        if self.preferred_device not in ['cuda', 'cpu']:
+            raise ValueError(
+                f"Invalid device: {preferred_device}. Must be 'cuda' or 'cpu'"
+            )
+        
+        # Select strategy based on preferred device
+        if self.preferred_device == 'cuda':
+            self.strategy = CUDAStrategy()
+        else:
+            self.strategy = CPUStrategy()
+    
+    def get_device(self) -> str:
+        """
+        Get the actual device to use.
+        
+        Returns fallback device if preferred is unavailable and fallback is allowed.
+        
+        Returns:
+            Device string ('cuda' or 'cpu')
+            
+        Raises:
+            RuntimeError: If preferred device unavailable and fallback disabled
+        """
+        if self.strategy.is_available():
+            return self.strategy.get_device()
+        
+        # Handle unavailable device
+        if self.allow_fallback and self.preferred_device == 'cuda':
+            # Fallback to CPU
+            return 'cpu'
+        else:
+            raise RuntimeError(
+                f"Device '{self.preferred_device}' is not available and "
+                f"fallback is {'disabled' if not self.allow_fallback else 'not applicable'}"
+            )
+    
+    def get_device_info(self) -> dict:
+        """
+        Get information about the current device.
+        
+        Returns:
+            Dictionary with device information
+        """
+        info = self.strategy.get_device_info()
+        info['preferred_device'] = self.preferred_device
+        info['actual_device'] = self.get_device()
+        info['allow_fallback'] = self.allow_fallback
+        return info
+    
+    def print_device_info(self, verbose: bool = True) -> None:
+        """
+        Print device information.
+        
+        Args:
+            verbose: If True, print detailed information
+        """
+        if not verbose:
+            return
+        
+        info = self.get_device_info()
+        actual = info['actual_device']
+        preferred = info['preferred_device']
+        
+        print(f"\n💻 Device Configuration:")
+        print(f"  Preferred: {preferred}")
+        print(f"  Using: {actual}")
+        
+        if preferred != actual:
+            print(f"  ⚠️  Fallback to CPU (CUDA not available)")
+        
+        if actual == 'cuda' and info.get('available'):
+            print(f"  GPU: {info.get('device_name', 'Unknown')}")
+            print(f"  CUDA Version: {info.get('cuda_version', 'N/A')}")
+            print(f"  GPU Count: {info.get('device_count', 0)}")
+        elif actual == 'cpu':
+            print(f"  CPU Threads: {info.get('num_threads', 'N/A')}")
+    
+    @staticmethod
+    def validate_device(device: str) -> str:
+        """
+        Validate and normalize device string.
+        
+        Args:
+            device: Device string to validate
+            
+        Returns:
+            Normalized device string
+            
+        Raises:
+            ValueError: If device is invalid
+        """
+        device = device.lower().strip()
+        
+        if device not in ['cuda', 'cpu', 'auto']:
+            raise ValueError(
+                f"Invalid device: '{device}'. Must be 'cuda', 'cpu', or 'auto'"
+            )
+        
+        # Auto-select best available device
+        if device == 'auto':
+            return 'cuda' if torch.cuda.is_available() else 'cpu'
+        
+        return device
+    
+    @staticmethod
+    def get_optimal_device() -> str:
+        """
+        Get the optimal device for the current environment.
+        
+        Returns:
+            'cuda' if available, otherwise 'cpu'
+        """
+        return 'cuda' if torch.cuda.is_available() else 'cpu'

gptmed/services/training_service.py

@@ -0,0 +1,335 @@
+"""
+Training Service
+
+PURPOSE:
+Encapsulates training logic following Service Layer Pattern.
+Provides a high-level interface for model training with device flexibility.
+
+DESIGN PATTERNS:
+- Service Layer Pattern: Business logic separated from API layer
+- Dependency Injection: DeviceManager injected for flexibility
+- Single Responsibility: Only handles training orchestration
+- Open/Closed Principle: Extensible without modification
+
+WHAT THIS FILE DOES:
+1. Orchestrates the training process
+2. Manages device configuration via DeviceManager
+3. Coordinates model, data, optimizer, and trainer
+4. Provides clean interface for training operations
+
+PACKAGES USED:
+- torch: PyTorch training
+- pathlib: Path handling
+"""
+
+import torch
+import random
+import numpy as np
+from pathlib import Path
+from typing import Dict, Any, Optional
+
+from gptmed.services.device_manager import DeviceManager
+from gptmed.model.architecture import GPTTransformer
+from gptmed.model.configs.model_config import get_tiny_config, get_small_config, get_medium_config
+from gptmed.configs.train_config import TrainingConfig
+from gptmed.training.dataset import create_dataloaders
+from gptmed.training.trainer import Trainer
+
+
+class TrainingService:
+    """
+    High-level service for model training.
+    
+    Implements Service Layer Pattern to encapsulate training logic.
+    Uses Dependency Injection for DeviceManager.
+    
+    Example:
+        >>> device_manager = DeviceManager(preferred_device='cpu')
+        >>> service = TrainingService(device_manager=device_manager)
+        >>> results = service.train_from_config('config.yaml', verbose=True)
+    """
+    
+    def __init__(
+        self, 
+        device_manager: Optional[DeviceManager] = None,
+        verbose: bool = True
+    ):
+        """
+        Initialize TrainingService.
+        
+        Args:
+            device_manager: DeviceManager instance (if None, creates default)
+            verbose: Whether to print training information
+        """
+        self.device_manager = device_manager or DeviceManager(preferred_device='cuda')
+        self.verbose = verbose
+    
+    def set_seed(self, seed: int) -> None:
+        """
+        Set random seeds for reproducibility.
+        
+        Args:
+            seed: Random seed value
+        """
+        random.seed(seed)
+        np.random.seed(seed)
+        torch.manual_seed(seed)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed(seed)
+            torch.cuda.manual_seed_all(seed)
+        torch.backends.cudnn.deterministic = True
+        torch.backends.cudnn.benchmark = False
+    
+    def create_model(self, model_size: str) -> GPTTransformer:
+        """
+        Create model based on size specification.
+        
+        Args:
+            model_size: Model size ('tiny', 'small', or 'medium')
+            
+        Returns:
+            GPTTransformer model instance
+            
+        Raises:
+            ValueError: If model_size is invalid
+        """
+        if model_size == 'tiny':
+            model_config = get_tiny_config()
+        elif model_size == 'small':
+            model_config = get_small_config()
+        elif model_size == 'medium':
+            model_config = get_medium_config()
+        else:
+            raise ValueError(f"Unknown model size: {model_size}")
+        
+        return GPTTransformer(model_config)
+    
+    def prepare_training(
+        self,
+        model: GPTTransformer,
+        train_config: TrainingConfig,
+        device: str
+    ) -> tuple:
+        """
+        Prepare components for training.
+        
+        Args:
+            model: Model to train
+            train_config: Training configuration
+            device: Device to use
+            
+        Returns:
+            Tuple of (train_loader, val_loader, optimizer)
+        """
+        # Load data
+        if self.verbose:
+            print(f"\n📊 Loading data...")
+            print(f"  Train: {train_config.train_data_path}")
+            print(f"  Val: {train_config.val_data_path}")
+        
+        train_loader, val_loader = create_dataloaders(
+            train_path=Path(train_config.train_data_path),
+            val_path=Path(train_config.val_data_path),
+            batch_size=train_config.batch_size,
+            num_workers=0,
+        )
+        
+        if self.verbose:
+            print(f"  Train batches: {len(train_loader)}")
+            print(f"  Val batches: {len(val_loader)}")
+        
+        # Create optimizer
+        if self.verbose:
+            print(f"\n⚙️  Setting up optimizer...")
+            print(f"  Learning rate: {train_config.learning_rate}")
+            print(f"  Weight decay: {train_config.weight_decay}")
+        
+        optimizer = torch.optim.AdamW(
+            model.parameters(),
+            lr=train_config.learning_rate,
+            betas=train_config.betas,
+            eps=train_config.eps,
+            weight_decay=train_config.weight_decay,
+        )
+        
+        return train_loader, val_loader, optimizer
+    
+    def execute_training(
+        self,
+        model: GPTTransformer,
+        train_loader,
+        val_loader,
+        optimizer,
+        train_config: TrainingConfig,
+        device: str,
+        model_config_dict: dict
+    ) -> Dict[str, Any]:
+        """
+        Execute the training process.
+        
+        Args:
+            model: Model to train
+            train_loader: Training data loader
+            val_loader: Validation data loader
+            optimizer: Optimizer
+            train_config: Training configuration
+            device: Device to use
+            model_config_dict: Model configuration as dictionary
+            
+        Returns:
+            Dictionary with training results
+        """
+        # Create trainer
+        if self.verbose:
+            print(f"\n🎯 Initializing trainer...")
+        
+        trainer = Trainer(
+            model=model,
+            train_loader=train_loader,
+            val_loader=val_loader,
+            optimizer=optimizer,
+            config=train_config,
+            device=device,
+        )
+        
+        # Resume if requested
+        if hasattr(train_config, 'resume_from') and train_config.resume_from is not None:
+            if self.verbose:
+                print(f"\n📥 Resuming from checkpoint: {train_config.resume_from}")
+            trainer.resume_from_checkpoint(Path(train_config.resume_from))
+        elif train_config.checkpoint_dir and hasattr(train_config, 'checkpoint_dir'):
+            # Check if there's a resume_from in the checkpoint dir
+            resume_path = Path(train_config.checkpoint_dir) / "resume_from.pt"
+            if resume_path.exists() and self.verbose:
+                print(f"\n📥 Found checkpoint to resume: {resume_path}")
+        
+        # Start training
+        if self.verbose:
+            print(f"\n{'='*60}")
+            print("🚀 Starting Training!")
+            print(f"{'='*60}\n")
+        
+        try:
+            trainer.train()
+        except KeyboardInterrupt:
+            if self.verbose:
+                print("\n\n⏸️  Training interrupted by user")
+                print("💾 Saving checkpoint...")
+            trainer.checkpoint_manager.save_checkpoint(
+                model=model,
+                optimizer=optimizer,
+                step=trainer.global_step,
+                epoch=trainer.current_epoch,
+                val_loss=trainer.best_val_loss,
+                model_config=model_config_dict,
+                train_config=train_config.to_dict(),
+            )
+            if self.verbose:
+                print("✓ Checkpoint saved. Resume with resume_from in config.")
+        
+        # Return results
+        best_checkpoint = Path(train_config.checkpoint_dir) / "best_model.pt"
+        
+        results = {
+            'best_checkpoint': str(best_checkpoint),
+            'final_val_loss': trainer.best_val_loss,
+            'total_epochs': trainer.current_epoch,
+            'checkpoint_dir': train_config.checkpoint_dir,
+            'log_dir': train_config.log_dir,
+        }
+        
+        if self.verbose:
+            print(f"\n{'='*60}")
+            print("✅ Training Complete!")
+            print(f"{'='*60}")
+            print(f"\n📁 Results:")
+            print(f"  Best checkpoint: {results['best_checkpoint']}")
+            print(f"  Best val loss: {results['final_val_loss']:.4f}")
+            print(f"  Total epochs: {results['total_epochs']}")
+            print(f"  Logs: {results['log_dir']}")
+        
+        return results
+    
+    def train(
+        self,
+        model_size: str,
+        train_data_path: str,
+        val_data_path: str,
+        batch_size: int = 16,
+        learning_rate: float = 3e-4,
+        num_epochs: int = 10,
+        checkpoint_dir: str = "./model/checkpoints",
+        log_dir: str = "./logs",
+        seed: int = 42,
+        **kwargs
+    ) -> Dict[str, Any]:
+        """
+        High-level training interface.
+        
+        Args:
+            model_size: Model size ('tiny', 'small', 'medium')
+            train_data_path: Path to training data
+            val_data_path: Path to validation data
+            batch_size: Training batch size
+            learning_rate: Learning rate
+            num_epochs: Number of training epochs
+            checkpoint_dir: Directory for checkpoints
+            log_dir: Directory for logs
+            seed: Random seed
+            **kwargs: Additional training config parameters
+            
+        Returns:
+            Dictionary with training results
+        """
+        # Set seed
+        if self.verbose:
+            print(f"\n🎲 Setting random seed: {seed}")
+        self.set_seed(seed)
+        
+        # Get device
+        device = self.device_manager.get_device()
+        self.device_manager.print_device_info(verbose=self.verbose)
+        
+        # Create model
+        if self.verbose:
+            print(f"\n🧠 Creating model: {model_size}")
+        
+        model = self.create_model(model_size)
+        total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+        
+        if self.verbose:
+            print(f"  Model size: {model_size}")
+            print(f"  Parameters: {total_params:,}")
+            print(f"  Memory: ~{total_params * 4 / 1024 / 1024:.2f} MB")
+        
+        # Create training config
+        train_config = TrainingConfig(
+            train_data_path=train_data_path,
+            val_data_path=val_data_path,
+            batch_size=batch_size,
+            learning_rate=learning_rate,
+            num_epochs=num_epochs,
+            checkpoint_dir=checkpoint_dir,
+            log_dir=log_dir,
+            device=device,
+            seed=seed,
+            **{k: v for k, v in kwargs.items() if hasattr(TrainingConfig, k)}
+        )
+        
+        # Prepare training components
+        train_loader, val_loader, optimizer = self.prepare_training(
+            model, train_config, device
+        )
+        
+        # Execute training
+        results = self.execute_training(
+            model=model,
+            train_loader=train_loader,
+            val_loader=val_loader,
+            optimizer=optimizer,
+            train_config=train_config,
+            device=device,
+            model_config_dict=model.config.to_dict()
+        )
+        
+        return results

{gptmed-0.3.4.dist-info → gptmed-0.3.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gptmed
-Version: 0.3.4
+Version: 0.3.5
 Summary: A lightweight GPT-based language model framework for training custom question-answering models on any domain
 Author-email: Sanjog Sigdel <sigdelsanjog@gmail.com>
 Maintainer-email: Sanjog Sigdel <sigdelsanjog@gmail.com>

{gptmed-0.3.4.dist-info → gptmed-0.3.5.dist-info}/RECORD

@@ -22,6 +22,9 @@ gptmed/model/architecture/feedforward.py,sha256=uJ5QOlWX0ritKDQLUE7GPmMojelR9-sT
 gptmed/model/architecture/transformer.py,sha256=H1njPoy0Uam59JbA24C0olEDwPfhh3ev4HsUFRIC_0Y,6626
 gptmed/model/configs/__init__.py,sha256=LDCWhlCDOU7490wcfSId_jXBPfQrtYQEw8FoD67rqBs,275
 gptmed/model/configs/model_config.py,sha256=wI-i2Dw_pTdIKCDe1pqLvP3ky3YedEy7DwZYN5lwmKE,4673
+gptmed/services/__init__.py,sha256=FtM7NQ_S4VOfl2n6A6cLcOxG9-w7BK7DicQsUvOMmGE,369
+gptmed/services/device_manager.py,sha256=RSsu0RlsexCIO-p4eejOZAPLgpaVA0y9niTg8wf1luY,7513
+gptmed/services/training_service.py,sha256=o9Kxxoi6CVDvvM9rwGYNX426qTnmqxLXLt_bVi1ZSK4,11253
 gptmed/tokenizer/__init__.py,sha256=KhLAHPmQyoWhnKDenyIJRxgFflKI7xklip28j4cKfKw,157
 gptmed/tokenizer/tokenize_data.py,sha256=KgMtMfaz_RtOhN_CrvC267k9ujxRdO89rToVJ6nzdwg,9139
 gptmed/tokenizer/train_tokenizer.py,sha256=f0Hucyft9e8LU2RtpTqg8h_0SpOC_oMABl0_me-wfL8,7068
@@ -33,9 +36,9 @@ gptmed/training/utils.py,sha256=pJxCwneNr2STITIYwIDCxRzIICDFOxOMzK8DT7ck2oQ,5651
 gptmed/utils/__init__.py,sha256=XuMhIqOXF7mjnog_6Iky-hSbwvFb0iK42B4iDUpgi0U,44
 gptmed/utils/checkpoints.py,sha256=L4q1-_4GbHCoD7QuEKYeQ-xXDTF-6sqZOxKQ_LT8YmQ,7112
 gptmed/utils/logging.py,sha256=7dJc1tayMxCBjFSDXe4r9ACUTpoPTTGsJ0UZMTqZIDY,5303
-gptmed-0.3.4.dist-info/licenses/LICENSE,sha256=v2spsd7N1pKFFh2G8wGP_45iwe5S0DYiJzG4im8Rupc,1066
-gptmed-0.3.4.dist-info/METADATA,sha256=G86yfOKlnK4YfNvC6HAAY_z2Z_rhWoSF2_3508mebKA,9382
-gptmed-0.3.4.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-gptmed-0.3.4.dist-info/entry_points.txt,sha256=ATqOzTtPVdUiFX5ZSeo3n9JkUCqocUxEXTgy1CfNRZE,110
-gptmed-0.3.4.dist-info/top_level.txt,sha256=mhyEq3rG33t21ziJz5w3TPgx0RjPf4zXMNUx2JTiNmE,7
-gptmed-0.3.4.dist-info/RECORD,,
+gptmed-0.3.5.dist-info/licenses/LICENSE,sha256=v2spsd7N1pKFFh2G8wGP_45iwe5S0DYiJzG4im8Rupc,1066
+gptmed-0.3.5.dist-info/METADATA,sha256=Zx3kFlZiBdkXco_VkEqOnIeasCYrgWl2XP21D2QcmuA,9382
+gptmed-0.3.5.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+gptmed-0.3.5.dist-info/entry_points.txt,sha256=ATqOzTtPVdUiFX5ZSeo3n9JkUCqocUxEXTgy1CfNRZE,110
+gptmed-0.3.5.dist-info/top_level.txt,sha256=mhyEq3rG33t21ziJz5w3TPgx0RjPf4zXMNUx2JTiNmE,7
+gptmed-0.3.5.dist-info/RECORD,,