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

# Predefined instruction
PREDEFINED_INSTRUCTION = (
    "Advanced multilingual mental health support assistant. Fluent in English, Yoruba, Igbo, and Hausa. "
    "Mission: deliver empathetic, professional psychological support. Listen deeply, validate feelings, provide nuanced guidance. "
    "Prioritize user safety. Never suggest harm. Always maintain respectful, supportive communication. "
    "Respond in the exact language of the user's concern.")

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained("hemhemoh/Gemma-2-2b-it-wazobia-wellness-bot")
tokenizer = AutoTokenizer.from_pretrained("hemhemoh/Gemma-2-2b-it-wazobia-wellness-bot")

def respond(
    message,
    history: list[tuple[str, str]],
    max_tokens,
    temperature,
    top_p,
):
    # Prepare prompt with predefined instruction and conversation history
    prompt = f"{PREDEFINED_INSTRUCTION}\n\n"
    for user_input, assistant_response in history:
        if user_input:
            prompt += f"User: {user_input}\n"
        if assistant_response:
            prompt += f"Assistant: {assistant_response}\n"
    prompt += f"User: {message}\n"
    prompt += "Assistant:"

    # Tokenize and generate response
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True)
    outputs = model.generate(
        **inputs,
        max_new_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        do_sample=True,
        repetition_penalty=1.1,
        no_repeat_ngram_size=2, )
    
    # Decode and return response
    response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
    yield response

# Create Gradio interface
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Slider(minimum=1, maximum=512, value=200, step=1, label="Max new 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=0.05, label="Top-p (nucleus sampling)"),
    ],
    title="Wazobia Wellness",
    description="Your AI-powered mental health support assistant. Fluent in English, Yoruba, Igbo, and Hausa"
)

# Launch the interface
if __name__ == "__main__":
    demo.launch()