tool_trainer_m4_max.py

"""
tool_trainer_m4_max.py - Optimized training for M4 Max Apple Silicon + SmolLM3-3B

This script is specifically optimized for:
- M4 Max 40-core GPU Apple Silicon
- SmolLM3-3B (larger, more capable model)
- Large training dataset (100+ examples)
- Aggressive but stable hyperparameters for fast, high-quality training
"""

import json
import torch
import torch.backends.mps
from transformers import (
    AutoTokenizer, 
    AutoModelForCausalLM, 
    TrainingArguments,
    Trainer,
    DataCollatorForLanguageModeling
)
from peft import LoraConfig, get_peft_model, TaskType
from datasets import Dataset
import os
import time

def setup_mps_optimization():
    """Configure optimal settings for M4 Max."""
    print("🍎 Configuring M4 Max optimizations...")
    
    # Check MPS availability
    if torch.backends.mps.is_available():
        print("✅ MPS (Metal Performance Shaders) is available")
        print(f"🚀 Using all 40 GPU cores of M4 Max")
        device = torch.device("mps")
    else:
        print("⚠️ MPS not available, falling back to CPU")
        device = torch.device("cpu")
    
    # Optimize memory allocation
    os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"  # Aggressive memory usage
    os.environ["TOKENIZERS_PARALLELISM"] = "false"  # Avoid fork warnings
    
    return device

def load_training_data(file_path="tool_pairs_enhanced.jsonl"):
    """Load the comprehensive training dataset."""
    pairs = []
    with open(file_path, 'r') as f:
        for line in f:
            pairs.append(json.loads(line.strip()))
    return pairs

def format_for_sft(pairs, tokenizer):
    """Convert pairs to SFT format optimized for function calling."""
    formatted = []
    for pair in pairs:
        # Create training example: prompt + chosen response
        full_text = pair["prompt"] + pair["chosen"] + tokenizer.eos_token
        formatted.append({"text": full_text})
    return formatted

def tokenize_function(examples, tokenizer, max_length=512):
    """Tokenize with consistent padding for variable length sequences."""
    # Reduced max_length to handle variable sequences better
    tokenized = tokenizer(
        examples["text"],
        truncation=True,
        padding="max_length",  # Consistent padding
        max_length=max_length,
        return_tensors=None
    )
    
    # For causal LM, labels are the same as input_ids
    tokenized["labels"] = tokenized["input_ids"]
    return tokenized

def main():
    print("🚀 M4 Max Optimized Training: SmolLM3-3B Function Calling")
    print("=" * 70)
    
    # Setup M4 Max optimizations
    device = setup_mps_optimization()
    start_time = time.time()
    
    # 1. Load SmolLM3-3B (the real deal!)
    print("📥 Loading SmolLM3-3B model and tokenizer...")
    model_name = "HuggingFaceTB/SmolLM3-3B"  # Using the actual SmolLM3-3B!
    
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    
    # Ensure consistent tokenizer settings
    tokenizer.padding_side = "right"
    
    # Load model with MPS optimization
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype=torch.float32,  # Use float32 for MPS compatibility
        trust_remote_code=True,
        attn_implementation="eager"  # More stable for training
    )
    
    # Move to MPS if available
    if str(device) == "mps":
        model = model.to(device)
    
    print(f"✅ Loaded model: {model_name}")
    print(f"🔧 Model dtype: {model.dtype}")
    print(f"💾 Model size: ~{sum(p.numel() for p in model.parameters()) / 1e9:.1f}B parameters")
    print(f"🎯 Device: {device}")
    
    # 2. Setup LoRA with optimized config for larger model
    print("\n🔩 Setting up LoRA adapter (rank 16 for SmolLM3-3B)...")
    lora_config = LoraConfig(
        r=16,                   # Higher rank for 3B model (more capacity)
        lora_alpha=32,          # 2x rank
        target_modules=[        # Target more modules for better coverage
            "q_proj", "v_proj", "k_proj", "o_proj", 
            "gate_proj", "up_proj", "down_proj",
            "embed_tokens", "lm_head"  # Include embeddings for better learning
        ],
        lora_dropout=0.05,      # Lower dropout for stability
        bias="none",
        task_type=TaskType.CAUSAL_LM,
        modules_to_save=["embed_tokens", "lm_head"]  # Save these for better function calling
    )
    
    model = get_peft_model(model, lora_config)
    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
    total_params = sum(p.numel() for p in model.parameters())
    
    print(f"✅ LoRA adapter attached")
    print(f"🎯 Trainable parameters: {trainable_params:,} ({trainable_params/total_params*100:.2f}%)")
    
    # 3. Load comprehensive training data
    print("\n📊 Loading comprehensive training dataset...")
    pairs = load_training_data()
    formatted_pairs = format_for_sft(pairs, tokenizer)
    
    print(f"✅ Loaded {len(pairs)} training pairs")
    print(f"📈 Dataset is {len(pairs)/8:.1f}x larger than before!")
    
    # Create and tokenize dataset
    train_dataset = Dataset.from_list(formatted_pairs)
    tokenized_dataset = train_dataset.map(
        lambda x: tokenize_function(x, tokenizer),
        batched=True,
        remove_columns=train_dataset.column_names,
        num_proc=1  # Single process for MPS compatibility
    )
    
    print(f"📊 Tokenized dataset: {len(tokenized_dataset)} examples")
    
    # 4. Optimized training arguments for M4 Max
    print("\n⚙️ Configuring M4 Max optimized training...")
    training_args = TrainingArguments(
        output_dir="./smollm3_tool_adapter",
        num_train_epochs=5,                    # More epochs with larger dataset
        per_device_train_batch_size=4,         # Larger batch size for M4 Max
        gradient_accumulation_steps=2,         # Effective batch size = 8
        learning_rate=3e-4,                    # Higher LR for faster convergence
        weight_decay=0.01,                     # Regularization
        warmup_steps=50,                       # More warmup for stability
        logging_steps=5,
        save_steps=25,
        save_total_limit=3,
        remove_unused_columns=False,
        fp16=False,                           # Disable mixed precision for MPS compatibility
        dataloader_pin_memory=False,          # Disable for MPS
        report_to=None,
        logging_dir="./logs",
        gradient_checkpointing=True,          # Memory optimization
        optim="adamw_torch",                  # Optimized optimizer
        lr_scheduler_type="cosine",           # Better convergence
        save_strategy="steps",
        eval_strategy="no",
        load_best_model_at_end=False,
    )
    
    # 5. Data collator with proper padding
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer,
        mlm=False,
        pad_to_multiple_of=8,  # Efficient padding for performance
    )
    
    # 6. Initialize optimized trainer
    print("🏋️ Initializing M4 Max optimized trainer...")
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=tokenized_dataset,
        data_collator=data_collator,
        remove_unused_columns=False,
    )
    
    print("✅ Trainer ready for M4 Max acceleration")
    
    # 7. Start accelerated training
    print("\n🎯 Starting accelerated training on M4 Max...")
    print("⏱️  Expected time: ~3-5 minutes with 40 GPU cores")
    print("📊 Monitoring loss for quality improvement...")
    
    # Train with progress monitoring
    train_result = trainer.train()
    
    end_time = time.time()
    training_time = end_time - start_time
    
    print("\n🎉 M4 Max training completed!")
    print(f"📊 Final training loss: {train_result.training_loss:.4f}")
    print(f"⏱️ Total training time: {training_time:.1f} seconds")
    print(f"🚀 Training speed: {len(pairs) * 5 / training_time:.1f} examples/second")
    
    # 8. Save the optimized model
    print("\n💾 Saving optimized model adapter...")
    model.save_pretrained("./smollm3_tool_adapter")
    tokenizer.save_pretrained("./smollm3_tool_adapter")
    
    print("✅ Model saved to './smollm3_tool_adapter'")
    
    # 9. Enhanced functionality test
    print("\n🧪 Enhanced functionality test...")
    test_schemas = [
        {
            "schema": {
                "name": "get_stock_price",
                "description": "Get current stock price",
                "parameters": {
                    "type": "object",
                    "properties": {"ticker": {"type": "string"}},
                    "required": ["ticker"]
                }
            },
            "question": "What's Google stock price?",
            "expected_ticker": "GOOGL"
        },
        {
            "schema": {
                "name": "process_payment",
                "description": "Process a payment transaction",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "amount": {"type": "number"},
                        "currency": {"type": "string"},
                        "recipient": {"type": "string"}
                    },
                    "required": ["amount", "recipient"]
                }
            },
            "question": "Send $150 to Alice",
            "expected": "process_payment"
        }
    ]
    
    model.eval()
    for i, test in enumerate(test_schemas, 1):
        test_prompt = f"""<|im_start|>system
You are a helpful assistant that calls functions by responding with valid JSON when given a schema. Always respond with JSON function calls only, never prose.<|im_end|>

<schema>
{json.dumps(test['schema'], indent=2)}
</schema>

<|im_start|>user
{test['question']}<|im_end|>
<|im_start|>assistant
"""
        
        inputs = tokenizer(test_prompt, return_tensors="pt")
        if str(device) == "mps":
            inputs = {k: v.to(device) for k, v in inputs.items()}
        
        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=80,
                temperature=0.1,
                do_sample=True,
                pad_token_id=tokenizer.eos_token_id
            )
        
        response = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)
        print(f"🧪 Test {i}: {test['question']}")
        print(f"🤖 Response: {response.strip()}")
        
        # Try to parse JSON
        try:
            json_response = json.loads(response.strip())
            print(f"✅ Valid JSON: {json_response}")
        except:
            print(f"❌ Invalid JSON")
        print("-" * 50)
    
    print("\n🏆 M4 Max Optimized Training Complete!")
    print(f"📈 Loss reduction with {len(pairs)} examples should be significant")
    print(f"🎯 Ready for comprehensive testing with schema_tester.py")
    
    return model, tokenizer

if __name__ == "__main__":
    model, tokenizer = main()