import gradio as gr
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM , BitsAndBytesConfig
import gc
import torch

def clear_memory():
    gc.collect()
    torch.cuda.empty_cache()


model_name = "GIGAParviz/Firooze_test"


bnb_config = BitsAndBytesConfig(llm_int8_threshold=6.0)
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=bnb_config)
model = model.to("cpu")
model.gradient_checkpointing_enable()

pipe = pipeline(task="text-generation", model=model, tokenizer=tokenizer, max_length=128)

def generate_response(prompt):
    clear_memory() 
    instruction = f"### Instruction:\n{prompt}\n\n### Response:\n"
    result = pipe(instruction)

    return result[0]['generated_text'][len(instruction):]

with gr.Blocks() as demo:
    gr.Markdown("<h1 style='text-align: center;'>🔮 Persian LLM made by A.M.Parviz</h1>")

    prompt_input = gr.Textbox(label="Enter Prompt", placeholder="Type your prompt here...", lines=2)
    
    generate_button = gr.Button("Generate Response")
    
    response_output = gr.Textbox(label="Generated Response", lines=5)
    
    generate_button.click(fn=generate_response, inputs=prompt_input, outputs=response_output)
    
    clear_button = gr.ClearButton([prompt_input, response_output])

demo.launch()