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  # Assumes Hugging Face Spaces library supports `@spaces.GPU`

# Load environment variables (e.g., Hugging Face token)
load_dotenv()
hf_token = os.getenv("HF_TOKEN")


# ---------------------------------------------------------------------
# Load Llama 3 Pipeline with Zero GPU (Encapsulated)
# ---------------------------------------------------------------------
@spaces.GPU(duration=300)  # Adjust GPU allocation duration
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 a top-tier radio imaging producer using Llama 3. "
            "Take the user's concept and craft a short, creative promo script."
        )
        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")  # Return the 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")

        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_wav:
            write(temp_wav.name, sr, normalized_audio)
            return temp_wav.name
    except Exception as e:
        return f"Error generating audio: {e}"


# ---------------------------------------------------------------------
# Gradio Interface
# ---------------------------------------------------------------------
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:
    gr.Markdown("# 🎧 AI Radio Imaging with Llama 3 + MusicGen (Zero GPU)")

    with gr.Row():
        user_prompt = gr.Textbox(
            label="Enter your promo idea",
            placeholder="E.g., A 15-second hype jingle for a morning talk show.",
        )
        llama_model_id = gr.Textbox(
            label="Llama 3 Model ID", value="meta-llama/Meta-Llama-3-8B-Instruct"
        )
        audio_length = gr.Slider(label="Audio Length (tokens)", minimum=128, maximum=1024, step=64, value=512)

    generate_script_button = gr.Button("Generate Script")
    script_output = gr.Textbox(label="Generated Script")
    generate_audio_button = gr.Button("Generate Audio")
    audio_output = gr.Audio(label="Generated Audio", type="filepath")

    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)