import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load the model and tokenizer
model_name = "migueldeguzmandev/GPT2XL_RLLMv19-4"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Set the pad token ID to the EOS token ID
model.config.pad_token_id = model.config.eos_token_id

# Define the inference function
def generate_response(input_text, temperature):
    # Tokenize the input text
    inputs = tokenizer(input_text, return_tensors="pt")
    input_ids = inputs["input_ids"]
    attention_mask = inputs["attention_mask"]

    # Generate the model's response
    output = model.generate(
        input_ids,
        attention_mask=attention_mask,
        max_length=300,
        num_return_sequences=1,
        temperature=temperature,
        no_repeat_ngram_size=2,
        top_k=50,
        top_p=0.95,
        do_sample=True,  # Set do_sample to True when using temperature
    )

    # Decode the generated response
    response = tokenizer.decode(output[0], skip_special_tokens=True)
    return response

# Create the Gradio interface
interface = gr.Interface(
    fn=generate_response,
    inputs=[
        gr.Textbox(label="User Input"),
        gr.Slider(minimum=0.000000000000000000000000001, maximum=1.0, value=0.7, step=0.1, label="Temperature"),
    ],
    outputs=gr.Textbox(label="Model Response"),
    title="TestOnlyRLLMv19Layer4",
    description=(
        """
        RLLMv19 is a spin-off experiment focusing on improving GPT2XL's robustness to jailbreaks. The 4th layer of RLLMv19 is compared to the 4th layer of RLLMv3. Why RLLMv3? This <a href='https://huggingface.co/spaces/migueldeguzmandev/RLLMv3.2-10'>prototype</a> demonstrated a capability to resist jailbreak attacks up to 67.8%, which you can read more about (<a href='https://www.lesswrong.com/posts/vZ5fM6FtriyyKbwi9/betterdan-ai-machiavelli-and-oppo-jailbreaks-vs-sota-models'>here</a>).
        """
    ),
)

# Launch the interface without the share option
interface.launch()