import gradio as gr
from transformers import GPT2Tokenizer, GPT2LMHeadModel, GPT2Config
import torch
from langchain.memory import ConversationBufferMemory

# Move model to device (GPU if available)
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

# Load the tokenizer (you can use the pre-trained tokenizer for GPT-2 family)
tokenizer = GPT2Tokenizer.from_pretrained("distilgpt2")

# Manually create a configuration for the model (since we don't have config.json)
config = GPT2Config.from_pretrained("distilgpt2")

# Initialize the model using the manually created configuration
model = GPT2LMHeadModel(config)

# Load the weights from the pytorch_model.bin file
model_path = "./pytorch_model_100.bin"  # Path to local model file
state_dict = torch.load(model_path, map_location=device)  # Load the state_dict
model.load_state_dict(state_dict)  # Load the state dict into the model

# Move model to the device (GPU or CPU)
model.to(device)

# Set up conversational memory using LangChain's ConversationBufferMemory
memory = ConversationBufferMemory()

# Define the chatbot function with memory and additional parameters
def chat_with_distilgpt2(input_text, temperature, top_p, top_k):
    # Retrieve conversation history
    conversation_history = memory.load_memory_variables({})['history']
    
    # Combine the (possibly summarized) history with the current user input
    no_memory_input = f"Question: {input_text}\nAnswer:"
    
    # Tokenize the input and convert to tensor
    input_ids = tokenizer.encode(no_memory_input, return_tensors="pt").to(device)
    
    # Generate the response using the model with adjusted parameters
    outputs = model.generate(
        input_ids,
        max_length=input_ids.shape[1] + 50,  # Limit total length
        max_new_tokens=15,
        num_return_sequences=1,
        no_repeat_ngram_size=3,
        repetition_penalty=1.2,
        early_stopping=True,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id,
        temperature=temperature,  # Add temperature from slider
        top_p=top_p,              # Add top_p from slider
        top_k=top_k               # Add top_k from slider
    )
    
    # Decode the model output
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # Update the memory with the user input and model response
    memory.save_context({"input": input_text}, {"output": response})
    
    return response

# Set up the Gradio interface with additional sliders
interface = gr.Interface(
    fn=chat_with_distilgpt2,
    inputs=[
        gr.Textbox(label="Chat with DistilGPT-2"),  # User input text
        gr.Slider(0.1, 1.0, step=0.1, value=1.0, label="Temperature"),  # Slider for temperature
        gr.Slider(0.0, 1.0, step=0.1, value=1.0, label="Top-p"),       # Slider for top-p
        gr.Slider(1, 100, step=1, value=50, label="Top-k")             # Slider for top-k
    ],
    outputs=gr.Textbox(label="DistilGPT-2's Response"),  # Model response
    title="DistilGPT-2 Chatbot with Memory and Adjustable Parameters",
    description="This is a simple chatbot powered by the DistilGPT-2 model with conversational memory, using LangChain. You can adjust temperature, top-p, and top-k using the sliders.",
)

# Launch the Gradio app
interface.launch()

How can this be modified to give the results for both a baseline DistilGPT2 and the fine tuned version for each input prompt?