app.py

import torch
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer

# Dictionary of available models
MODELS = {
    "SmolLM2-135M-Instruct": "HuggingFaceTB/SmolLM2-135M-Instruct",
    "SmolLM2-360M-Instruct": "HuggingFaceTB/SmolLM2-360M-Instruct",
    "SmolLM2-1.7B-Instruct": "HuggingFaceTB/SmolLM2-1.7B-Instruct"
}

class ModelHandler:
    def __init__(self):
        self.current_model = None
        self.current_tokenizer = None
        self.device = "cpu" if torch.cuda.is_available() else "cpu"
    
    def load_model(self, model_name):
        try:
            checkpoint = MODELS[model_name]
            self.current_tokenizer = AutoTokenizer.from_pretrained(checkpoint)
            self.current_model = AutoModelForCausalLM.from_pretrained(
                checkpoint,
                torch_dtype=torch.bfloat16,
                device_map="auto"
            )
            return f"Successfully loaded {model_name}"
        except Exception as e:
            return f"Error loading model: {str(e)}"

model_handler = ModelHandler()

def generate_text(model_name, prompt, max_tokens, temperature, top_p):
    try:
        # Load model if it's different from the current one
        if model_handler.current_model is None or MODELS[model_name] != model_handler.current_model.name_or_path:
            load_status = model_handler.load_model(model_name)
            if "Error" in load_status:
                return load_status

        # Format input as chat message
        messages = [{"role": "user", "content": prompt}]
        input_text = model_handler.current_tokenizer.apply_chat_template(messages, tokenize=False)
        
        # Tokenize
        inputs = model_handler.current_tokenizer.encode(
            input_text, 
            return_tensors="pt"
        ).to(model_handler.device)
        
        # Generate
        outputs = model_handler.current_model.generate(
            inputs,
            max_new_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            do_sample=True
        )
        
        # Decode and return
        response = model_handler.current_tokenizer.decode(
            outputs[0], 
            skip_special_tokens=True
        )
        return response

    except Exception as e:
        return f"Error during generation: {str(e)}"

# Create Gradio interface
iface = gr.Interface(
    fn=generate_text,
    inputs=[
        gr.Dropdown(
            choices=list(MODELS.keys()),
            label="Select Model",
            value="SmolLM2-360M-Instruct"
        ),
        gr.Textbox(
            label="Enter your prompt",
            placeholder="What would you like to know?",
            lines=3
        ),
        gr.Slider(
            minimum=10,
            maximum=500,
            value=50,
            step=10,
            label="Maximum Tokens"
        ),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=0.2,
            step=0.1,
            label="Temperature"
        ),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=0.9,
            step=0.1,
            label="Top P"
        )
    ],
    outputs=gr.Textbox(label="Generated Response", lines=5),
    title="SmolLM2 Model Comparison",
    description="""
    Compare different sizes of SmolLM2 models:
    - SmolLM2-135M-Instruct: Smallest and fastest
    - SmolLM2-360M-Instruct: Balanced size and performance
    - SmolLM2-1.7B-Instruct: Largest and most capable
    """,
    examples=[
        ["SmolLM2-360M-Instruct", "What is the capital of France?", 50, 0.2, 0.9],
        ["SmolLM2-360M-Instruct", "Explain quantum computing in simple terms.", 200, 0.3, 0.9],
        ["SmolLM2-360M-Instruct", "Write a short poem about nature.", 100, 0.7, 0.9]
    ]
)

# Launch the application
if __name__ == "__main__":
    iface.launch(share=True)