app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel, PeftConfig
import torch
import os

def load_model(model_id, model_type="base"):
    try:
        if model_type == "base":
            tokenizer = AutoTokenizer.from_pretrained(model_id)
            model = AutoModelForCausalLM.from_pretrained(
                model_id,
                torch_dtype=torch.float16,
                device_map="auto"
            )
            return tokenizer, model
        else:  # finetuned model with PEFT
            # Load the base model first
            base_model_id = "satyanayak/gemma-3-base"
            tokenizer = AutoTokenizer.from_pretrained(base_model_id)
            base_model = AutoModelForCausalLM.from_pretrained(
                base_model_id,
                torch_dtype=torch.float16,
                device_map="auto"
            )
            
            # Load and merge the PEFT adapters
            model = PeftModel.from_pretrained(
                base_model,
                model_id,
                torch_dtype=torch.float16,
                device_map="auto"
            )
            return tokenizer, model
    except Exception as e:
        print(f"Error loading {model_type} model: {str(e)}")
        return None, None

# Load base model and tokenizer
base_model_id = "satyanayak/gemma-3-base"
base_tokenizer, base_model = load_model(base_model_id, "base")

# Load finetuned model and tokenizer
finetuned_model_id = "satyanayak/gemma-3-GRPO"
finetuned_tokenizer, finetuned_model = load_model(finetuned_model_id, "finetuned")

def generate_base_response(prompt, max_length=512):
    if base_model is None or base_tokenizer is None:
        return "Error: Base model failed to load. Please check if the model files are properly uploaded to Hugging Face."
    
    try:
        inputs = base_tokenizer(prompt, return_tensors="pt").to(base_model.device)
        outputs = base_model.generate(
            **inputs,
            max_length=max_length,
            num_return_sequences=1,
            temperature=0.7,
            do_sample=True,
            pad_token_id=base_tokenizer.eos_token_id
        )
        response = base_tokenizer.decode(outputs[0], skip_special_tokens=True)
        return response
    except Exception as e:
        return f"Error generating response with base model: {str(e)}"

def generate_finetuned_response(prompt, max_length=512):
    if finetuned_model is None or finetuned_tokenizer is None:
        return "Error: Finetuned model failed to load. Please check if the model files are properly uploaded to Hugging Face."
    
    try:
        inputs = finetuned_tokenizer(prompt, return_tensors="pt").to(finetuned_model.device)
        outputs = finetuned_model.generate(
            **inputs,
            max_length=max_length,
            num_return_sequences=1,
            temperature=0.7,
            do_sample=True,
            pad_token_id=finetuned_tokenizer.eos_token_id
        )
        response = finetuned_tokenizer.decode(outputs[0], skip_special_tokens=True)
        return response
    except Exception as e:
        return f"Error generating response with finetuned model: {str(e)}"

# Example prompts
examples = [
    ["What is the sqrt of 101"],
    ["How many r's are in strawberry?"],
    ["If Tom has 3 more apples than Jerry and Jerry has 5 apples, how many apples does Tom have?"]
]

# Create the Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# Gemma-3 Model Comparison Demo")
    gr.Markdown("Compare responses between the base model and the GRPO-finetuned model.")
    
    with gr.Row():
        # Base Model Column
        with gr.Column(scale=1):
            gr.Markdown("## Base Model (Gemma-3)")
            base_input = gr.Textbox(
                label="Enter your prompt",
                placeholder="Type your prompt here...",
                lines=5
            )
            base_generate_btn = gr.Button("Generate with Base Model")
            base_output = gr.Textbox(label="Base Model Output", lines=10)
            
            gr.Examples(
                examples=examples,
                inputs=base_input,
                outputs=base_output,
                fn=generate_base_response,
                cache_examples=True
            )
        
        # Finetuned Model Column
        with gr.Column(scale=1):
            gr.Markdown("## GRPO-Finetuned Model")
            finetuned_input = gr.Textbox(
                label="Enter your prompt",
                placeholder="Type your prompt here...",
                lines=5
            )
            finetuned_generate_btn = gr.Button("Generate with Finetuned Model")
            finetuned_output = gr.Textbox(label="Finetuned Model Output", lines=10)
            
            gr.Examples(
                examples=examples,
                inputs=finetuned_input,
                outputs=finetuned_output,
                fn=generate_finetuned_response,
                cache_examples=True
            )
    
    # Connect buttons to their respective functions
    base_generate_btn.click(
        fn=generate_base_response,
        inputs=base_input,
        outputs=base_output
    )
    
    finetuned_generate_btn.click(
        fn=generate_finetuned_response,
        inputs=finetuned_input,
        outputs=finetuned_output
    )

if __name__ == "__main__":
    demo.launch()