app.py

import gradio as gr
import os
import torch
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    pipeline,
    AutoProcessor,
    MusicgenForConditionalGeneration,
)
from scipy.io.wavfile import write
import tempfile
from dotenv import load_dotenv
import spaces

load_dotenv()
hf_token = os.getenv("HF_TOKEN")

# ---------------------------------------------------------------------
# Load Llama 3 Pipeline with Zero GPU (Encapsulated)
# ---------------------------------------------------------------------
@spaces.GPU(duration=300)
def generate_script(user_prompt: str, model_id: str, token: str):
    try:
        tokenizer = AutoTokenizer.from_pretrained(model_id, use_auth_token=token)
        model = AutoModelForCausalLM.from_pretrained(
            model_id,
            use_auth_token=token,
            torch_dtype=torch.float16,
            device_map="auto",
            trust_remote_code=True,
        )
        llama_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)

        system_prompt = (
            "You are an expert radio imaging producer specializing in sound design and music. "
            "Take the user's concept and craft a concise, creative promo script with a strong focus on auditory elements and musical appeal."
        )

        combined_prompt = f"{system_prompt}\nUser concept: {user_prompt}\nRefined script:"
        result = llama_pipeline(combined_prompt, max_new_tokens=200, do_sample=True, temperature=0.9)
        return result[0]["generated_text"].split("Refined script:")[-1].strip()
    except Exception as e:
        return f"Error generating script: {e}"


# ---------------------------------------------------------------------
# Load MusicGen Model (Encapsulated)
# ---------------------------------------------------------------------
@spaces.GPU(duration=300)
def generate_audio(prompt: str, audio_length: int):
    try:
        musicgen_model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small")
        musicgen_processor = AutoProcessor.from_pretrained("facebook/musicgen-small")

        musicgen_model.to("cuda")
        inputs = musicgen_processor(text=[prompt], padding=True, return_tensors="pt")
        outputs = musicgen_model.generate(**inputs, max_new_tokens=audio_length)
        musicgen_model.to("cpu")

        sr = musicgen_model.config.audio_encoder.sampling_rate
        audio_data = outputs[0, 0].cpu().numpy()
        normalized_audio = (audio_data / max(abs(audio_data)) * 32767).astype("int16")

        output_path = f"{tempfile.gettempdir()}/generated_audio.wav"
        write(output_path, sr, normalized_audio)

        return output_path
    except Exception as e:
        return f"Error generating audio: {e}"


# ---------------------------------------------------------------------
# Gradio Interface Functions
# ---------------------------------------------------------------------
def interface_generate_script(user_prompt, llama_model_id):
    return generate_script(user_prompt, llama_model_id, hf_token)


def interface_generate_audio(script, audio_length):
    return generate_audio(script, audio_length)


# ---------------------------------------------------------------------
# Interface
# ---------------------------------------------------------------------
with gr.Blocks() as demo:
    # Header
    gr.Markdown("""
        # 🎙️ AI-Powered Radio Imaging Studio 🚀  
        ### Create stunning **radio promos** with **Llama 3** and **MusicGen**  
        🔥 **Zero GPU** integration for efficiency and ease!  
        ❤️ A huge thanks to the **Hugging Face community** for making this possible.
    """)

    # Script Generation Section
    gr.Markdown("## ✍️ Step 1: Generate Your Promo Script")
    with gr.Row():
        user_prompt = gr.Textbox(
            label="🎤 Enter Promo Idea",
            placeholder="E.g., A 15-second energetic jingle for a morning talk show.",
            lines=2,
            info="Describe your promo idea clearly to generate a creative script."
        )
        llama_model_id = gr.Textbox(
            label="🎛️ Llama 3 Model ID",
            value="meta-llama/Meta-Llama-3-8B-Instruct",
            info="Enter the Hugging Face model ID for Llama 3."
        )
    generate_script_button = gr.Button("Generate Script ✨")
    script_output = gr.Textbox(
        label="📜 Generated Promo Script",
        lines=4,
        interactive=False,
        info="Your generated promo script will appear here."
    )

    # Audio Generation Section
    gr.Markdown("## 🎧 Step 2: Generate Audio from Your Script")
    with gr.Row():
        audio_length = gr.Slider(
            label="🎵 Audio Length (tokens)",
            minimum=128,
            maximum=1024,
            step=64,
            value=512,
            info="Select the desired audio token length."
        )
        generate_audio_button = gr.Button("Generate Audio 🎶")
        audio_output = gr.Audio(
            label="🎶 Generated Audio File",
            type="filepath",
            interactive=False
        )

    # Footer
    gr.Markdown("""
        <br><hr>
        <p style="text-align: center; font-size: 0.9em;">
            Created with ❤️ by <a href="https://bilsimaging.com" target="_blank">bilsimaging.com</a>  
            Special thanks to the <strong>Hugging Face community</strong> for their incredible support and tools!
        </p>
    """, elem_id="footer")

    # Button Actions
    generate_script_button.click(
        fn=interface_generate_script,
        inputs=[user_prompt, llama_model_id],
        outputs=script_output,
    )

    generate_audio_button.click(
        fn=interface_generate_audio,
        inputs=[script_output, audio_length],
        outputs=audio_output,
    )

# ---------------------------------------------------------------------
# Launch App
# ---------------------------------------------------------------------
demo.launch(debug=True)