app.py

import os
import json
import gradio as gr
import torch
import pandas as pd
from transformers import (
    TrainingArguments,
    Trainer,
    AutoModelForCausalLM,
    AutoTokenizer,
    DataCollatorForLanguageModeling
)
from datasets import Dataset
from peft import (
    prepare_model_for_kbit_training,
    LoraConfig,
    get_peft_model
)

# Constants
MODEL_NAME = "deepseek-ai/DeepSeek-R1"
OUTPUT_DIR = "finetuned_models"
LOGS_DIR = "training_logs"

def save_uploaded_file(file_obj):
    """Save uploaded file and return its path"""
    try:
        os.makedirs('uploads', exist_ok=True)
        import tempfile
        
        # Create a temporary file
        temp_file = tempfile.NamedTemporaryFile(delete=False, suffix='.csv', dir='uploads')
        
        # Write the content
        if isinstance(file_obj, (bytes, bytearray)):
            temp_file.write(file_obj)
        else:
            content = file_obj.read()
            if isinstance(content, str):
                temp_file.write(content.encode('utf-8'))
            else:
                temp_file.write(content)
        
        temp_file.close()
        return temp_file.name
        
    except Exception as e:
        raise Exception(f"Error saving file: {str(e)}")

def prepare_training_data(df):
    """Convert DataFrame into Q&A format"""
    formatted_data = []
    try:
        for _, row in df.iterrows():
            # Clean and validate the data
            chunk_id = str(row['chunk_id']).strip()
            text = str(row['text']).strip()
            
            if chunk_id and text:  # Only include non-empty pairs
                # Format each conversation in the required structure
                formatted_text = f"User: {chunk_id}\nAssistant: {text}"
                formatted_data.append({"text": formatted_text})
        
        if not formatted_data:
            raise ValueError("No valid training pairs found in the data")
            
        return formatted_data
    except Exception as e:
        raise Exception(f"Error preparing training data: {str(e)}")

def prepare_training_components(
    data_path,
    learning_rate,
    num_epochs,
    batch_size,
    model_name=MODEL_NAME
):
    """Prepare model, tokenizer, and training arguments"""
    print(f"Loading data from: {data_path}")  # Debug logging
    """Prepare model, tokenizer, and training arguments"""
    
    # Create output directory with timestamp
    import time
    timestamp = time.strftime("%Y%m%d_%H%M%S")
    specific_output_dir = os.path.join(OUTPUT_DIR, f"run_{timestamp}")
    os.makedirs(specific_output_dir, exist_ok=True)
    os.makedirs(LOGS_DIR, exist_ok=True)

    # Load data and convert to Q&A format
    try:
        df = pd.read_csv(data_path, encoding='utf-8')
        print(f"Loaded CSV with {len(df)} rows")  # Debug logging
        formatted_data = prepare_training_data(df)
        print(f"Prepared {len(formatted_data)} training examples")  # Debug logging
    except Exception as e:
        print(f"Error loading CSV: {str(e)}")  # Debug logging
        raise

    # Load tokenizer and model
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        trust_remote_code=True,
        torch_dtype=torch.float16,
        load_in_8bit=True
    )

    # LoRA Configuration
    lora_config = LoraConfig(
        r=16,
        lora_alpha=32,
        target_modules=[
            "q_proj", "k_proj", "v_proj", "o_proj",
            "gate_proj", "up_proj", "down_proj"
        ],
        lora_dropout=0.05,
        bias="none",
        task_type="CAUSAL_LM"
    )

    # Prepare model
    model = prepare_model_for_kbit_training(model)
    model = get_peft_model(model, lora_config)

    # Training Arguments
    training_args = TrainingArguments(
        output_dir=specific_output_dir,
        num_train_epochs=num_epochs,
        per_device_train_batch_size=batch_size,
        learning_rate=learning_rate,
        fp16=True,
        gradient_accumulation_steps=8,
        gradient_checkpointing=True,
        logging_dir=os.path.join(LOGS_DIR, f"run_{timestamp}"),
        logging_steps=10,
        save_strategy="epoch",
        evaluation_strategy="epoch",
        save_total_limit=2,
    )

    # Convert to datasets format
    dataset = Dataset.from_dict({
        'text': [item['text'] for item in formatted_data]
    })

    # Create data collator
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer,
        mlm=False
    )

    return {
        'model': model,
        'tokenizer': tokenizer,
        'training_args': training_args,
        'dataset': dataset,
        'data_collator': data_collator,
        'output_dir': specific_output_dir
    }

def train_model(
    file,
    learning_rate=2e-4,
    num_epochs=3,
    batch_size=4,
    progress=gr.Progress()
):
    """Training function for Gradio interface"""
    if file is None:
        return "Please upload a file first."
    
    try:
        # File validation
        progress(0.1, desc="Validating file...")
        file_path = save_uploaded_file(file)
        
        # Prepare components
        progress(0.2, desc="Preparing training components...")
        components = prepare_training_components(
            file_path,
            learning_rate,
            num_epochs,
            batch_size
        )
        
        # Initialize trainer
        progress(0.4, desc="Initializing trainer...")
        trainer = Trainer(
            model=components['model'],
            args=components['training_args'],
            train_dataset=components['dataset'],
            data_collator=components['data_collator'],
        )
        
        # Train
        progress(0.5, desc="Training model...")
        trainer.train()
        
        # Save model and tokenizer
        progress(0.9, desc="Saving model...")
        trainer.save_model()
        components['tokenizer'].save_pretrained(components['output_dir'])
        
        progress(1.0, desc="Training complete!")
        return f"Training completed! Model saved in {components['output_dir']}"
        
    except Exception as e:
        error_msg = f"Error during training: {str(e)}"
        print(error_msg)  # Log the error
        return error_msg
    """Training function for Gradio interface"""
    try:
        # Save uploaded file
        file_path = save_uploaded_file(file)
        
        # Prepare components
        progress(0.2, desc="Preparing training components...")
        components = prepare_training_components(
            file_path,
            learning_rate,
            num_epochs,
            batch_size
        )
        
        # Initialize trainer
        progress(0.4, desc="Initializing trainer...")
        trainer = Trainer(
            model=components['model'],
            args=components['training_args'],
            train_dataset=components['dataset'],
            data_collator=components['data_collator'],
        )
        
        # Train
        progress(0.5, desc="Training model...")
        trainer.train()
        
        # Save model and tokenizer
        progress(0.9, desc="Saving model...")
        trainer.save_model()
        components['tokenizer'].save_pretrained(components['output_dir'])
        
        progress(1.0, desc="Training complete!")
        return f"Training completed! Model saved in {components['output_dir']}"
    
    except Exception as e:
        return f"Error during training: {str(e)}"

def create_interface():
    """Create Gradio interface with proper file upload handling"""
    with gr.Blocks(analytics_enabled=False) as demo:
        gr.Markdown("# DeepSeek-R1 Model Finetuning Interface")
        
        with gr.Row():
            with gr.Column():
                file_input = gr.File(
                    label="Upload Training Data (CSV)",
                    type="binary",
                    file_types=[".csv"],
                    file_count="single"
                )
                
                learning_rate = gr.Slider(
                    minimum=1e-5,
                    maximum=1e-3,
                    value=2e-4,
                    label="Learning Rate"
                )
                
                num_epochs = gr.Slider(
                    minimum=1,
                    maximum=10,
                    value=3,
                    step=1,
                    label="Number of Epochs"
                )
                
                batch_size = gr.Slider(
                    minimum=1,
                    maximum=8,
                    value=4,
                    step=1,
                    label="Batch Size"
                )
                
                train_button = gr.Button("Start Training")
                
            with gr.Column():
                output = gr.Textbox(label="Training Status")
                
        train_button.click(
            fn=train_model,
            inputs=[file_input, learning_rate, num_epochs, batch_size],
            outputs=output
        )
        
        gr.Markdown("""
        ## Instructions
        1. Upload your training data in CSV format with columns:
           - chunk_id (questions)
           - text (answers)
        2. Adjust training parameters if needed
        3. Click 'Start Training'
        4. Wait for training to complete
        """)
    
    return demo

if __name__ == "__main__":
    # Create necessary directories
    os.makedirs(OUTPUT_DIR, exist_ok=True)
    os.makedirs(LOGS_DIR, exist_ok=True)
    
    # Launch Gradio interface
    demo = create_interface()
    demo.launch(
        share=True, 
        server_port=7860, 
        server_name="0.0.0.0"
    )