app.py

import gradio as gr
import torch
from transformers import AutoTokenizer
from peft import AutoPeftModelForCausalLM

# Load the fine-tuned model and tokenizer
model_name = "richardcsuwandi/llama2-javanese"
model = AutoPeftModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype=torch.bfloat16)

# Merge adapter with base
model = model.merge_and_unload()
model.eval()
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token_id = 0
tokenizer.padding_side = "left"

def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    # Format the input text
    input_text = f"<s>[INST] <<SYS>> {system_message} <</SYS>> {message} [/INST]"
    
    # Tokenize the input text
    inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

    # Generate response
    output_sequences = model.generate(
        input_ids=inputs['input_ids'],
        max_length=max_tokens + inputs['input_ids'].shape[1],  # Adjust for input length
        repetition_penalty=1.2,
        temperature=temperature,
        top_p=top_p
    )
    
    # Decode the generated response
    generated_text = tokenizer.decode(output_sequences[0], skip_special_tokens=True)
    
    return generated_text

demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="Sampeyan minangka chatbot umum sing tansah mangsuli nganggo basa Jawa.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=0.95,
            step=0.05,
            label="Top-p (nucleus sampling)",
        ),
    ],
)

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