gptmed 0.3.3__py3-none-any.whl → 0.3.4__py3-none-any.whl

gptmed/api.py

@@ -39,6 +39,8 @@ from gptmed.configs.train_config import TrainingConfig
 from gptmed.training.dataset import create_dataloaders
 from gptmed.training.trainer import Trainer
 from gptmed.inference.generator import TextGenerator
+from gptmed.services.device_manager import DeviceManager
+from gptmed.services.training_service import TrainingService
 
 
 def create_config(output_path: str = 'training_config.yaml') -> None:
@@ -58,7 +60,11 @@ def create_config(output_path: str = 'training_config.yaml') -> None:
     create_default_config_file(output_path)
 
 
-def train_from_config(config_path: str, verbose: bool = True) -> Dict[str, Any]:
+def train_from_config(
+    config_path: str, 
+    verbose: bool = True, 
+    device: Optional[str] = None
+) -> Dict[str, Any]:
     """
     Train a GPT model using a YAML configuration file.
     
@@ -68,6 +74,8 @@ def train_from_config(config_path: str, verbose: bool = True) -> Dict[str, Any]:
     Args:
         config_path: Path to YAML configuration file
         verbose: Whether to print training progress (default: True)
+        device: Device to use ('cuda', 'cpu', or 'auto'). If None, uses config value.
+                'auto' will select best available device.
     
     Returns:
         Dictionary with training results:
@@ -82,13 +90,16 @@ def train_from_config(config_path: str, verbose: bool = True) -> Dict[str, Any]:
         >>> gptmed.create_config('config.yaml')
         >>> # ... edit config.yaml ...
         >>> 
-        >>> # Train the model
-        >>> results = gptmed.train_from_config('config.yaml')
+        >>> # Train the model on CPU
+        >>> results = gptmed.train_from_config('config.yaml', device='cpu')
         >>> print(f"Best model: {results['best_checkpoint']}")
+        >>> 
+        >>> # Train with auto device selection
+        >>> results = gptmed.train_from_config('config.yaml', device='auto')
     
     Raises:
         FileNotFoundError: If config file or data files don't exist
-        ValueError: If configuration is invalid
+        ValueError: If configuration is invalid or device is invalid
     """
     if verbose:
         print("=" * 60)
@@ -111,47 +122,42 @@ def train_from_config(config_path: str, verbose: bool = True) -> Dict[str, Any]:
     # Convert to arguments
     args = config_to_args(config)
     
-    # Import here to avoid circular imports
-    import random
-    import numpy as np
+    # Override device if provided as parameter
+    if device is not None:
+        # Validate and normalize device
+        device = DeviceManager.validate_device(device)
+        if verbose:
+            print(f"\n⚙️  Device override: {device} (from parameter)")
+        args['device'] = device
     
-    # Set random seed
-    def set_seed(seed: int):
-        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
+    # Create DeviceManager with the selected device
+    device_manager = DeviceManager(
+        preferred_device=args['device'],
+        allow_fallback=True
+    )
+    
+    # Print device information
+    device_manager.print_device_info(verbose=verbose)
+    
+    # Create TrainingService with DeviceManager
+    training_service = TrainingService(
+        device_manager=device_manager,
+        verbose=verbose
+    )
     
+    # Set random seed
     if verbose:
         print(f"\n🎲 Setting random seed: {args['seed']}")
-    set_seed(args['seed'])
+    training_service.set_seed(args['seed'])
     
-    # Check device
-    device = args['device']
-    if device == 'cuda' and not torch.cuda.is_available():
-        if verbose:
-            print("⚠️  CUDA not available, using CPU")
-        device = 'cpu'
+    # Get actual device to use
+    actual_device = device_manager.get_device()
     
     # Load model config
     if verbose:
         print(f"\n🧠 Creating model: {args['model_size']}")
     
-    if args['model_size'] == 'tiny':
-        model_config = get_tiny_config()
-    elif args['model_size'] == 'small':
-        model_config = get_small_config()
-    elif args['model_size'] == 'medium':
-        model_config = get_medium_config()
-    else:
-        raise ValueError(f"Unknown model size: {args['model_size']}")
-    
-    # Create model
-    model = GPTTransformer(model_config)
+    model = training_service.create_model(args['model_size'])
     total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
     
     if verbose:
@@ -159,7 +165,7 @@ def train_from_config(config_path: str, verbose: bool = True) -> Dict[str, Any]:
         print(f"  Parameters: {total_params:,}")
         print(f"  Memory: ~{total_params * 4 / 1024 / 1024:.2f} MB")
     
-    # Load data
+    # Load data using TrainingService
     if verbose:
         print(f"\n📊 Loading data...")
         print(f"  Train: {args['train_data']}")
@@ -176,7 +182,7 @@ def train_from_config(config_path: str, verbose: bool = True) -> Dict[str, Any]:
         print(f"  Train batches: {len(train_loader)}")
         print(f"  Val batches: {len(val_loader)}")
     
-    # Create training config
+    # Create training config with actual device
     train_config = TrainingConfig(
         batch_size=args['batch_size'],
         learning_rate=args['learning_rate'],
@@ -195,7 +201,7 @@ def train_from_config(config_path: str, verbose: bool = True) -> Dict[str, Any]:
         val_data_path=args['val_data'],
         checkpoint_dir=args['checkpoint_dir'],
         log_dir=args['log_dir'],
-        device=device,
+        device=actual_device,  # Use actual device from DeviceManager
         seed=args['seed'],
     )
     
@@ -213,70 +219,17 @@ def train_from_config(config_path: str, verbose: bool = True) -> Dict[str, Any]:
         weight_decay=args['weight_decay'],
     )
     
-    # Create trainer
-    if verbose:
-        print(f"\n🎯 Initializing trainer...")
-    
-    trainer = Trainer(
+    # Execute training using TrainingService
+    results = training_service.execute_training(
         model=model,
         train_loader=train_loader,
         val_loader=val_loader,
         optimizer=optimizer,
-        config=train_config,
-        device=device,
+        train_config=train_config,
+        device=actual_device,
+        model_config_dict=model.config.to_dict()
     )
     
-    # Resume if requested
-    if args['resume_from'] is not None:
-        if verbose:
-            print(f"\n📥 Resuming from checkpoint: {args['resume_from']}")
-        trainer.resume_from_checkpoint(Path(args['resume_from']))
-    
-    # Start training
-    if verbose:
-        print(f"\n{'='*60}")
-        print("🚀 Starting Training!")
-        print(f"{'='*60}\n")
-    
-    try:
-        trainer.train()
-    except KeyboardInterrupt:
-        if 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.to_dict(),
-            train_config=train_config.to_dict(),
-        )
-        if 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 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
 
 

gptmed/configs/config_loader.py

@@ -76,6 +76,15 @@ def validate_config(config: Dict[str, Any]) -> None:
         raise ValueError("batch_size must be positive")
     if config['training']['learning_rate'] <= 0:
         raise ValueError("learning_rate must be positive")
+    
+    # Validate device
+    valid_devices = ['cuda', 'cpu', 'auto']
+    device_value = config.get('device', {}).get('device', 'cuda').lower()
+    if device_value not in valid_devices:
+        raise ValueError(
+            f"Invalid device: {device_value}. "
+            f"Must be one of {valid_devices}"
+        )
 
 
 def config_to_args(config: Dict[str, Any]) -> Dict[str, Any]:
@@ -169,7 +178,7 @@ def create_default_config_file(output_path: str = 'training_config.yaml') -> Non
             'log_interval': 10
         },
         'device': {
-            'device': 'cuda',
+            'device': 'cuda',  # Options: 'cuda', 'cpu', or 'auto'
             'seed': 42
         },
         'advanced': {

{gptmed-0.3.3.dist-info → gptmed-0.3.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gptmed
-Version: 0.3.3
+Version: 0.3.4
 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>
@@ -61,27 +61,6 @@ A lightweight GPT-based language model framework for training custom question-an
 - 📦 **Lightweight**: Small model size suitable for edge deployment
 - 🛠️ **Complete Toolkit**: Includes tokenizer training, model training, and inference utilities
 
-## Table of Contents
-
-- [Features](#features)
-- [Installation](#installation)
-- [Quick Start](#quick-start)
-- [Package Structure](#package-structure)
-  - [Core Modules](#core-modules)
-  - [Model Components](#model-components)
-  - [Training Components](#training-components)
-  - [Inference Components](#inference-components)
-  - [Data Processing](#data-processing)
-  - [Utilities](#utilities)
-- [Model Architecture](#model-architecture)
-- [Configuration](#configuration)
-- [Documentation](#documentation)
-- [Performance](#performance)
-- [Examples](#examples)
-- [Contributing](#contributing)
-- [License](#license)
-- [Support](#support)
-
 ## Installation
 
 ### From PyPI (Recommended)
@@ -208,134 +187,27 @@ config = TrainingConfig(
 )
 ```
 
-## Package Structure
-
-### Core Modules
-
-The `gptmed` package contains the following main modules:
-
-```
-gptmed/
-├── model/                  # Model architecture and configurations
-├── inference/              # Text generation and sampling
-├── training/               # Training loops and datasets
-├── tokenizer/              # Tokenizer training and data processing
-├── data/                   # Data parsers and formatters
-├── configs/                # Training configurations
-└── utils/                  # Utilities (checkpoints, logging)
-```
-
-### Model Components
-
-**`gptmed.model.architecture`** - GPT Transformer Implementation
-
-- `GPTTransformer` - Main model class
-- `TransformerBlock` - Individual transformer layers
-- `MultiHeadAttention` - Attention mechanism
-- `FeedForward` - Feed-forward networks
-- `RoPEPositionalEncoding` - Rotary position embeddings
-
-**`gptmed.model.configs`** - Model Configurations
-
-- `get_tiny_config()` - ~2M parameters (testing)
-- `get_small_config()` - ~10M parameters (recommended)
-- `get_medium_config()` - ~50M parameters (high quality)
-- `ModelConfig` - Custom configuration class
-
-### Training Components
-
-**`gptmed.training`** - Training Pipeline
-
-- `train.py` - Main training script (CLI: `gptmed-train`)
-- `Trainer` - Training loop with checkpointing
-- `TokenizedDataset` - PyTorch dataset for tokenized data
-- `create_dataloaders()` - DataLoader creation utilities
-
-**`gptmed.configs`** - Training Configurations
-
-- `TrainingConfig` - Training hyperparameters
-- `get_default_config()` - Default training settings
-- `get_quick_test_config()` - Fast testing configuration
-
-### Inference Components
-
-**`gptmed.inference`** - Text Generation
-
-- `TextGenerator` - Main generation class
-- `generator.py` - CLI command (CLI: `gptmed-generate`)
-- `sampling.py` - Sampling strategies (top-k, top-p, temperature)
-- `decoding_utils.py` - Decoding utilities
-- `GenerationConfig` - Generation parameters
-
-### Data Processing
-
-**`gptmed.tokenizer`** - Tokenizer Training & Data Processing
-
-- `train_tokenizer.py` - Train SentencePiece tokenizer
-- `tokenize_data.py` - Convert text to token sequences
-- SentencePiece BPE tokenizer support
-
-**`gptmed.data.parsers`** - Data Parsing & Formatting
-
-- `MedQuADParser` - XML Q&A parser (example)
-- `CausalTextFormatter` - Format Q&A pairs for training
-- `FormatConfig` - Formatting configuration
-
-### Utilities
-
-**`gptmed.utils`** - Helper Functions
-
-- `checkpoints.py` - Model checkpoint management
-- `logging.py` - Training metrics logging
-
----
-
-## Detailed Project Structure
+## Project Structure
 
 ```
 gptmed/
 ├── model/
-│   ├── architecture/
-│   │   ├── gpt.py              # GPT transformer model
-│   │   ├── attention.py        # Multi-head attention
-│   │   ├── feedforward.py      # Feed-forward networks
-│   │   └── embeddings.py       # Token + positional embeddings
-│   └── configs/
-│       └── model_config.py     # Model size configurations
+│   ├── architecture/      # GPT transformer implementation
+│   └── configs/           # Model configurations
 ├── inference/
-│   ├── generator.py            # Text generation (CLI command)
-│   ├── sampling.py             # Sampling strategies
-│   ├── decoding_utils.py       # Decoding utilities
-│   └── generation_config.py    # Generation parameters
+│   ├── generator.py       # Text generation
+│   └── sampling.py        # Sampling strategies
 ├── training/
-│   ├── train.py                # Main training script (CLI command)
-│   ├── trainer.py              # Training loop
-│   ├── dataset.py              # PyTorch dataset
-│   └── utils.py                # Training utilities
+│   ├── train.py          # Training script
+│   ├── trainer.py        # Training loop
+│   └── dataset.py        # Data loading
 ├── tokenizer/
-│   ├── train_tokenizer.py      # Train SentencePiece tokenizer
-│   └── tokenize_data.py        # Tokenize text data
-├── data/
-│   └── parsers/
-│       ├── medquad_parser.py   # Example XML parser
-│       └── text_formatter.py   # Q&A text formatter
+│   └── train_tokenizer.py # SentencePiece tokenizer
 ├── configs/
-│   └── train_config.py         # Training configurations
+│   └── train_config.py   # Training configurations
 └── utils/
-    ├── checkpoints.py          # Model checkpointing
-    └── logging.py              # Training logging
-```
-
-### Command-Line Interface
-
-The package provides two main CLI commands:
-
-```bash
-# Train a model
-gptmed-train --model-size small --num-epochs 10 --batch-size 16
-
-# Generate text
-gptmed-generate --prompt "Your question?" --max-length 100
+    ├── checkpoints.py    # Model checkpointing
+    └── logging.py        # Training logging
 ```
 
 ## Requirements

{gptmed-0.3.3.dist-info → gptmed-0.3.4.dist-info}/RECORD

@@ -1,7 +1,7 @@
 gptmed/__init__.py,sha256=mwzeW2Qc6j1z5f6HOvZ_BNOnFSncWEK2KEkdqq91yYY,1676
-gptmed/api.py,sha256=gUWooWsXDaGb1r22YnzS3w-sU-n-b4gB4-gh0fMsT4A,11109
+gptmed/api.py,sha256=k9a_1F2h__xgKnH2l0FaJqAqu-iTYt5tu_VfVO0UhrA,9806
 gptmed/configs/__init__.py,sha256=yRa-zgPQ-OCzu8fvCrfWMG-CjF3dru3PZzknzm0oUaQ,23
-gptmed/configs/config_loader.py,sha256=ZWdH63XOOu0T8seWBiJFZtzlyFmzHzKmMxon6ZgZHlg,6000
+gptmed/configs/config_loader.py,sha256=3GQ1iCNpdJ5yALWXA3SPPHRkaUO-117vdArEL6u7sK8,6354
 gptmed/configs/train_config.py,sha256=KqfNBh9hdTTd_6gEAlrClU8sVFSlVDmZJOrf3cPwFe8,4657
 gptmed/configs/training_config.yaml,sha256=EEZZa3kcsZr3g-_fKDPYZt4_NTpmS-3NvJrTYSWNc8g,2874
 gptmed/data/__init__.py,sha256=iAHeakB5pBAd7MkmarPPY0UKS9bTaO_winLZ23Y2O90,54
@@ -33,9 +33,9 @@ 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-0.3.3.dist-info/licenses/LICENSE,sha256=v2spsd7N1pKFFh2G8wGP_45iwe5S0DYiJzG4im8Rupc,1066
-gptmed-0.3.3.dist-info/METADATA,sha256=0ohKwsi3802GMhVUIx2n76i4QHhY0dkzdG4a_g1p_Hw,13605
-gptmed-0.3.3.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-gptmed-0.3.3.dist-info/entry_points.txt,sha256=ATqOzTtPVdUiFX5ZSeo3n9JkUCqocUxEXTgy1CfNRZE,110
-gptmed-0.3.3.dist-info/top_level.txt,sha256=mhyEq3rG33t21ziJz5w3TPgx0RjPf4zXMNUx2JTiNmE,7
-gptmed-0.3.3.dist-info/RECORD,,
+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,,