Add granite training example

mtake · mtake · commit a8d903028b09 · 2025-10-13T19:14:59.000+09:00
diff --git a/examples/scripts/interpolator.py b/examples/scripts/interpolator.py
@@ -0,0 +1,117 @@
+"""
+A simple model interpolation utility.
+
+This takes two checkpoints of the same model, and outputs a merged checkpoint with the linear interpolation.
+
+Example usage:
+    python interpolator.py \\
+        --model-path ibm-granite/granite-3.3-8b-instruct \\
+        --trained-model-path /path/to/checkpoint
+"""
+# Standard
+import argparse
+
+# Third Party
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+
+def interpolate_models(
+    model_path: str,
+    trained_model_path: str,
+    trained_model_weight: float = 0.5,
+    output_model_path: str | None = None,
+    torch_dtype: str | None = "bfloat16",
+) -> str:
+    if output_model_path is None:
+        output_model_path = f"{trained_model_path}_interp"
+
+    model_kwargs: dict[str, any] = {}
+    if torch_dtype is not None and torch_dtype != "auto":
+        model_kwargs["torch_dtype"] = torch_dtype
+
+    # load original model
+    model = AutoModelForCausalLM.from_pretrained(
+        model_path,
+        **model_kwargs,
+    )
+    state_dict = model.state_dict()
+    original_model_weight = 1 - trained_model_weight
+    for key in state_dict.keys():
+        state_dict[key] = state_dict[key] * original_model_weight
+
+    # load trained model
+    trained_model = AutoModelForCausalLM.from_pretrained(
+        trained_model_path,
+        **model_kwargs,
+    )
+    trained_state_dict = trained_model.state_dict()
+    for key in state_dict.keys():
+        state_dict[key] += trained_state_dict[key] * trained_model_weight
+
+    # save interpolated model
+    model.save_pretrained(output_model_path, state_dict=state_dict)
+
+    # copy tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    tokenizer.save_pretrained(output_model_path)
+
+    return output_model_path
+
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    parser.add_argument(
+        "--model-path",
+        type=str,
+        required=True,
+        help="path to the original model",
+    )
+    parser.add_argument(
+        "--trained-model-path",
+        type=str,
+        required=True,
+        help="path to the trained model",
+    )
+    parser.add_argument(
+        "--trained-model-weight",
+        type=float,
+        default=0.5,
+        help="weight for the trained model",
+    )
+    parser.add_argument(
+        "--output-model-path",
+        type=str,
+        default=None,
+        help="path to the output model",
+    )
+    parser.add_argument(
+        "--torch-dtype",
+        type=str,
+        default="bfloat16",
+        help="torch dtype",
+    )
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_arguments()
+    model_path: str = args.model_path
+    trained_model_path: str = args.trained_model_path
+    trained_model_weight: float = args.trained_model_weight
+    output_model_path: str | None = args.output_model_path
+    torch_dtype: str | None = args.torch_dtype
+
+    interpolate_models(
+        model_path,
+        trained_model_path,
+        trained_model_weight=trained_model_weight,
+        output_model_path=output_model_path,
+        torch_dtype=torch_dtype,
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/examples/scripts/osft_granite_example.py b/examples/scripts/osft_granite_example.py
@@ -0,0 +1,237 @@
+#!/usr/bin/env python3
+"""
+OSFT Training Example: Granite 3.3 8B Instruct
+
+This script demonstrates OSFT (Orthogonal Subspace Fine-Tuning) training with Granite 3.3 8B Instruct model
+using a single-node, multi-GPU setup with training_hub.
+
+OSFT allows continual training without catastrophic forgetting, making it ideal for:
+- Adapting instruction-tuned models to new domains
+- Adding new knowledge without losing existing capabilities
+- Fine-tuning without replay buffers or supplementary datasets
+
+After the training, the script also creates a merged model with linear interpolation.
+
+Example usage:
+    python osft_granite_example.py \\
+        --data-path /path/to/data.jsonl \\
+        --ckpt-output-dir /path/to/checkpoints
+"""
+
+import os
+import sys
+import time
+from datetime import datetime
+import argparse
+import glob
+import torch
+
+from training_hub import osft
+
+
+# =============================================================================
+# MODEL CONFIGURATION EXAMPLE FOR OSFT
+# =============================================================================
+
+# Derived from generic_7b_example in examples/notebooks/osft_comprehensive_tutorial.ipynb
+granite_example = {
+    "model_name": "Granite 3.3 8B Instruct",
+    "model_path": "ibm-granite/granite-3.3-8b-instruct",  # HuggingFace model name or local path
+    "example_unfreeze_rank_ratio": 0.3,  # Balanced preservation vs adaptation
+    "example_max_tokens_per_gpu": 10000,
+    "example_max_seq_len": 4096,
+    "example_batch_size": 128,
+    "example_learning_rate": 5e-6,
+    "notes": "Good baseline for most 7B instruction-tuned models",
+}
+
+selected_example = granite_example  # Change this to your preferred example
+
+model_name = selected_example['model_name']
+default_model_path = selected_example['model_path']
+default_unfreeze_rank_ratio = selected_example["example_unfreeze_rank_ratio"]
+default_max_tokens_per_gpu = selected_example['example_max_tokens_per_gpu']
+default_max_seq_len = selected_example['example_max_seq_len']
+default_batch_size = selected_example['example_batch_size']
+default_learning_rate = selected_example['example_learning_rate']
+default_num_epochs = 3
+default_nproc_per_node = torch.cuda.device_count() if torch.cuda.is_available() else 0
+default_model_weight = 0.5
+
+# =============================================================================
+# COMPLETE OSFT PARAMETER CONFIGURATION
+# =============================================================================
+
+# Experiment identification
+experiment_name = "osft_granite_example"
+timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+full_experiment_name = f"{experiment_name}_{timestamp}"
+
+# data_output_dir=f"data/{full_experiment_name}"  # Directory for processed data
+data_output_dir=f"/dev/shm/data/{full_experiment_name}"  # Directory for processed data (RAM disk for speed)
+
+
+def find_most_recent_checkpoint(output_dir):
+    """
+    Find the most recent checkpoint in the training output directory.
+    
+    Args:
+        output_dir (str): Training output directory containing hf_format/ subdirectory
+        
+    Returns:
+        str: Path to the most recent checkpoint
+        
+    Raises:
+        ValueError: If no checkpoints are found
+    """
+    # Get all checkpoint directories under hf_format
+    checkpoint_pattern = os.path.join(output_dir, "hf_format", "samples_*.0")
+    checkpoint_dirs = glob.glob(checkpoint_pattern)
+    
+    if not checkpoint_dirs:
+        raise ValueError(f"No checkpoints found in {os.path.join(output_dir, 'hf_format')}")
+    
+    # Find the most recently created checkpoint
+    most_recent_checkpoint = max(checkpoint_dirs, key=os.path.getctime)
+    
+    return most_recent_checkpoint
+
+
+def main():
+    parser = argparse.ArgumentParser(description=f'OSFT Training Example: {model_name}')
+    
+    # Required parameters
+    parser.add_argument('--data-path', required=True,
+                       help='Path to training data (JSONL format)')
+    parser.add_argument('--ckpt-output-dir', required=True,
+                       help='Directory to save checkpoints')
+    
+    # Optional overrides
+    parser.add_argument('--model-path', default=default_model_path,
+                       help=f'Model path or HuggingFace name (default: {default_model_path})')
+    parser.add_argument('--num-epochs', type=int, default=default_num_epochs,
+                       help=f'Number of training epochs (default: {default_num_epochs})')
+    parser.add_argument('--unfreeze-rank-ratio', type=float, default=default_unfreeze_rank_ratio,
+                       help=f'Unfreeze rank ratio for OSFT (0.0-1.0, default: {default_unfreeze_rank_ratio})')
+    parser.add_argument('--max-tokens-per-gpu', type=int, default=default_max_tokens_per_gpu,
+                       help=f'Max tokens per GPU (default: {default_max_tokens_per_gpu})')
+    parser.add_argument('--nproc-per-node', type=int, default=default_nproc_per_node,
+                       help=f'Number of GPUs (default: {default_nproc_per_node})')
+    parser.add_argument('--learning-rate', type=float, default=default_learning_rate,
+                       help=f'Learning rate for training (default: {default_learning_rate})')
+    parser.add_argument('--unmask-messages', action='store_true', default=False,
+                       help='Unmask messages during training (default: False)')
+    parser.add_argument('--batch-size', type=int, default=default_batch_size,
+                       help=f'Effective batch size for training (default: {default_batch_size})')
+    parser.add_argument('--max-seq-len', type=int, default=default_max_seq_len,
+                       help=f'Max sequence length (default: {default_max_seq_len})')
+    parser.add_argument('--model-weight', type=float, default=default_model_weight,
+                       help=f'Weight for trained model for interpolation (0.0-1.0, default: {default_model_weight})')
+    
+    args = parser.parse_args()
+    
+    if args.nproc_per_node < 4:
+        raise ValueError("NPROC_PER_NODE must be larger than or equal to 4")
+    
+    # Granite 3.3 8B Instruct OSFT configuration
+    print(f"🚀 OSFT Training: {model_name}")
+    print("=" * 50)
+    print(f"Model: {args.model_path}")
+    print(f"Data: {args.data_path}")
+    print(f"Output: {args.ckpt_output_dir}")
+    print(f"GPUs: {args.nproc_per_node}")
+    print(f"Unfreeze Rank Ratio: {args.unfreeze_rank_ratio}")
+    print(f"Max tokens per GPU: {args.max_tokens_per_gpu:,}")
+    print(f"Epochs: {args.num_epochs}")
+    print(f"Batch size: {args.batch_size}")
+    print(f"Learning rate: {args.learning_rate}")
+    print(f"Max sequence length: {args.max_seq_len:,}")
+    print(f"Model weight: {args.model_weight}")
+    print()
+    print("📝 Note: OSFT enables continual learning without replay buffers")
+    print("    The model will adapt to new data while preserving existing capabilities")
+    print()
+    
+    # Training configuration optimized for Granite 3.3 8B Instruct with OSFT
+    start_time = time.time()
+    
+    try:
+        result = osft(
+            # Model and data
+            model_path=args.model_path,
+            data_path=args.data_path,
+            ckpt_output_dir=args.ckpt_output_dir,
+            
+            # OSFT-specific parameters
+            unfreeze_rank_ratio=args.unfreeze_rank_ratio,  # Controls preservation vs adaptation
+            
+            # Training parameters optimized for Granite 3.3 8B Instruct
+            num_epochs=args.num_epochs,
+            effective_batch_size=args.batch_size,            # Smaller batch for efficient model
+            learning_rate=args.learning_rate,                # Very low LR for smaller but dense model
+            max_seq_len=args.max_seq_len,
+            max_tokens_per_gpu=args.max_tokens_per_gpu,
+            
+            # Data processing
+            data_output_dir=data_output_dir,
+            warmup_steps=0,
+            unmask_messages=args.unmask_messages,
+            
+            # Optimization
+            use_liger=True,                     # Enable Liger kernels for efficiency
+            seed=42,
+            lr_scheduler='cosine',              # Cosine scheduler works well with OSFT
+            
+            # Checkpointing
+            checkpoint_at_epoch=True,
+            save_final_checkpoint=True,
+            
+            # Single-node multi-GPU setup
+            nproc_per_node=args.nproc_per_node,
+            nnodes=1,
+            node_rank=0,
+            rdzv_id=102,
+            rdzv_endpoint="127.0.0.1:29500",
+        )
+        
+        end_time = time.time()
+        duration = end_time - start_time
+        
+        most_recent_checkpoint = find_most_recent_checkpoint(args.ckpt_output_dir)
+        
+        print("=" * 50)
+        print("✅ OSFT Training completed successfully!")
+        print(f"⏱️  Duration: {duration/3600:.2f} hours")
+        print(f"📁 Checkpoints: {args.ckpt_output_dir}/hf_format")
+        print(f"   Most recent checkpoint: {most_recent_checkpoint}")
+        print()
+        print("💡 Your model has been adapted to the new domain while preserving")
+        print("   its original instruction-following capabilities!")
+        
+        trained_model_weight = args.model_weight
+        if 0.0 < trained_model_weight and trained_model_weight < 1.0:
+            from interpolator import interpolate_models
+
+            interp_model_path = interpolate_models(args.model_path, most_recent_checkpoint, trained_model_weight=trained_model_weight)
+
+            print("=" * 50)
+            print("✅ Interpolation completed successfully!")
+            print(f"   Interpolated model checkpoint: {interp_model_path}")
+
+    except Exception as e:
+        end_time = time.time()
+        duration = end_time - start_time
+        
+        print("=" * 50)
+        print(f"❌ Training failed after {duration/60:.1f} minutes")
+        print(f"Error: {e}")
+        print()
+        print("💡 Troubleshooting tips:")
+        print("   - Reduce --max-tokens-per-gpu if you see OOM errors")
+        print("   - For domain adaptation, try --unfreeze-rank-ratio between 0.2-0.4")
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()
+
diff --git a/examples/scripts/sft_granite_example.py b/examples/scripts/sft_granite_example.py
