import gradio as gr
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
import torch  # Import torch

# --- Model Loading (Do this only once, outside the function) ---

# Option 1: Pipeline (High-Level, Easier)
use_pipeline = True  # Set to False to use the manual method

if use_pipeline:
    pipe = pipeline("text-generation", model="kakaocorp/kanana-nano-2.1b-base", device="cpu") #Explicitly on CPU
else:
    # Option 2: Manual Tokenizer and Model (More Control)
    tokenizer = AutoTokenizer.from_pretrained("kakaocorp/kanana-nano-2.1b-base")
    model = AutoModelForCausalLM.from_pretrained("kakaocorp/kanana-nano-2.1b-base")
    #  No need to move to GPU. It will default to CPU.
    print("Model loaded on CPU")

# --- Generation Function ---

def generate_text(prompt, max_length=50, temperature=1.0, top_k=50, top_p=1.0, no_repeat_ngram_size=0, num_return_sequences=1):
    """Generates text based on the given prompt and parameters."""

    if use_pipeline:
        messages = [{"role": "user", "content": prompt}]  # Format for pipeline
        try:
            result = pipe(
                messages,
                max_length=max_length,
                temperature=temperature,
                top_k=top_k,
                top_p=top_p,
                no_repeat_ngram_size=no_repeat_ngram_size,
                num_return_sequences=num_return_sequences,
                return_full_text=False, # Important:  We only want generated text
                pad_token_id=pipe.tokenizer.eos_token_id  # Prevent warning, pipeline knows the EOS token
            )
            # Pipeline returns a list of dictionaries, each with 'generated_text'
            return "\n\n".join([res['generated_text'] for res in result])

        except Exception as e:
            return f"Error during generation: {e}"

    else:  # Manual method
        try:
            inputs = tokenizer(prompt, return_tensors="pt")
            # No need to move to GPU.  Inputs will default to CPU.
            outputs = model.generate(
                **inputs,
                max_length=max_length,
                temperature=temperature,
                top_k=top_k,
                top_p=top_p,
                no_repeat_ngram_size=no_repeat_ngram_size,
                num_return_sequences=num_return_sequences,
                pad_token_id=tokenizer.eos_token_id,  # Ensure padding is correct
                do_sample=True # Ensure sampling happens.
            )

            generated_texts = []
            for i in range(outputs.shape[0]):
                 generated_text = tokenizer.decode(outputs[i], skip_special_tokens=True)
                 generated_texts.append(generated_text)

            return "\n\n".join(generated_texts)
        except Exception as e:
            return f"Error during generation: {e}"



# --- Gradio Interface ---

with gr.Blocks() as demo:
    gr.Markdown("# Text Generation with kakaocorp/kanana-nano-2.1b-base")

    with gr.Row():
        with gr.Column():
            prompt_input = gr.Textbox(label="Prompt", placeholder="Enter your prompt here...")
            with gr.Accordion("Generation Parameters", open=False):
                max_length_slider = gr.Slider(label="Max Length", minimum=10, maximum=512, value=50, step=1)
                temperature_slider = gr.Slider(label="Temperature", minimum=0.1, maximum=2.0, value=1.0, step=0.1)
                top_k_slider = gr.Slider(label="Top K", minimum=0, maximum=100, value=50, step=1)
                top_p_slider = gr.Slider(label="Top P", minimum=0.0, maximum=1.0, value=1.0, step=0.05)
                no_repeat_ngram_size_slider = gr.Slider(label="No Repeat N-gram Size", minimum=0, maximum=10, value=0, step=1)  # Add the slider
                num_return_sequences_slider = gr.Slider(label="Number of Return Sequences", minimum=1, maximum=5, value=1, step=1)

            generate_button = gr.Button("Generate")

        with gr.Column():
            output_text = gr.Textbox(label="Generated Text", interactive=False) # Use interactive=False

    generate_button.click(
        generate_text,
        inputs=[
            prompt_input,
            max_length_slider,
            temperature_slider,
            top_k_slider,
            top_p_slider,
            no_repeat_ngram_size_slider,
            num_return_sequences_slider
        ],
        outputs=output_text,
    )

demo.launch() # Remove share=True for local testing, add it back for deployment