app.py

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

# Load model + tokenizer
model_name = "Samurai719214/gptneo-mythology-storyteller"
device = "cuda" if torch.cuda.is_available() else "cpu"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name).to(device)

# Story generation with history
def generate_full_story_chunks(excerpt, history_state):
    if not excerpt or not excerpt.strip():
        history_state.append(("❌", "⚠️ Enter a story excerpt."))
        yield history_state, gr.update(visible=False), gr.update(interactive=True)
        return

    inputs = tokenizer(excerpt, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}

    output_ids = model.generate(
        inputs["input_ids"],
        attention_mask=inputs["attention_mask"],
        max_new_tokens=400,
        do_sample=True,
        temperature=0.1,
        top_p=0.9,
        no_repeat_ngram_size=2,
    )

    generated_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)

    # Append user input
    history_state.append(("You", excerpt))

    # Stream response in chunks
    response = ""
    for i in range(0, len(generated_text), 200):
        response += generated_text[i:i+200]
        if len(history_state) > 0 and history_state[-1][0] == "AI":
            history_state[-1] = ("AI", response)
        else:
            history_state.append(("AI", response))
        yield history_state, gr.update(visible=False), gr.update(interactive=True)

# Clear conversation
def clear_history():
    return [], gr.update(interactive=False)

# Enable/disable generate button
def toggle_generate_button(text):
    return gr.update(interactive=bool(text.strip()))

# Build UI
with gr.Blocks() as demo:
    gr.Markdown("## 🏺 Mythology Storyteller")
    gr.Markdown("Enter a phrase from a chapter of your choice (include Parv, key event, and section for better results).")

    with gr.Row():
        with gr.Column():
            user_input = gr.Textbox(
                label="Incomplete story excerpt",
                placeholder="Enter an excerpt from the Mahabharata here...",
                lines=4,
            )
            summary_input = gr.Textbox(
                label="Chapter summary (optional)",
                placeholder="Enter summary if available...",
                lines=2,
            )
            generate_btn = gr.Button("✨ Generate Story", interactive=False)

        with gr.Column():
            output_text = gr.Chatbot(
                label="Generated Story",
                height=400,
                placeholder="⚔️ Legends are being written..." 
            )
            spinner = gr.Markdown("", visible=False)  # spinner placeholder
            clear_btn = gr.Button("🗑️ Clear Conversation", interactive=False)

    gr.Markdown("---")
    gr.Markdown("🔌 Use via API (see Hugging Face Inference docs).")

    # Toggle generate button when input changes
    user_input.change(
        fn=toggle_generate_button,
        inputs=user_input,
        outputs=generate_btn,
    )

    # Show spinner when generating
    def show_spinner():
        return gr.update(value="⏳ Generating story...", visible=True)

    def hide_spinner():
        return gr.update(visible=False)

    generate_btn.click(
        fn=show_spinner,
        inputs=None,
        outputs=spinner,
    ).then(
        fn=generate_full_story_chunks,
        inputs=[user_input, output_text],
        outputs=[output_text, spinner, clear_btn],
    ).then(
        fn=hide_spinner,
        inputs=None,
        outputs=spinner,
    )

    # Clear history
    clear_btn.click(
        fn=clear_history,
        inputs=None,
        outputs=[output_text, clear_btn],
    )

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