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")

# Globals for lazy loading
llama_pipeline = None
musicgen_model = None
musicgen_processor = None

# ---------------------------------------------------------------------
# Load Llama 3 Model with Zero GPU (Lazy Loading)
# ---------------------------------------------------------------------
@spaces.GPU(duration=300)  # Increased duration to 300 seconds
def load_llama_pipeline_zero_gpu(model_id: str, token: str):
    global llama_pipeline
    if llama_pipeline is None:
        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",  # Automatically handles GPU allocation
                trust_remote_code=True,
            )
            llama_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)
        except Exception as e:
            return str(e)
    return llama_pipeline

# ---------------------------------------------------------------------
# Generate Radio Script
# ---------------------------------------------------------------------
def generate_script(user_input: str, pipeline_llama):
    try:
        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_input}\nRefined script:"
        result = pipeline_llama(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 (Lazy Loading)
# ---------------------------------------------------------------------
@spaces.GPU(duration=300)
def load_musicgen_model():
    global musicgen_model, musicgen_processor
    if musicgen_model is None or musicgen_processor is None:
        try:
            musicgen_model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small")
            musicgen_processor = AutoProcessor.from_pretrained("facebook/musicgen-small")
        except Exception as e:
            return None, str(e)
    return musicgen_model, musicgen_processor

# ---------------------------------------------------------------------
# Generate Audio
# ---------------------------------------------------------------------
@spaces.GPU(duration=300)
def generate_audio(prompt: str, audio_length: int):
    global musicgen_model, musicgen_processor
    if musicgen_model is None or musicgen_processor is None:
        musicgen_model, musicgen_processor = load_musicgen_model()
        if isinstance(musicgen_model, str):
            return musicgen_model
    try:
        musicgen_model.to("cuda")  # Move the model to GPU
        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 radio_imaging_app(user_prompt, llama_model_id, audio_length):
    # Load Llama 3 Pipeline with Zero GPU
    pipeline_llama = load_llama_pipeline_zero_gpu(llama_model_id, hf_token)
    if isinstance(pipeline_llama, str):
        return pipeline_llama, None

    # Generate Script
    script = generate_script(user_prompt, pipeline_llama)

    # Generate Audio
    audio_data = generate_audio(script, audio_length)
    return script, audio_data


# ---------------------------------------------------------------------
# Interface
# ---------------------------------------------------------------------
with gr.Blocks(css="""
#app-title {
    text-align: center;
    font-size: 2rem;
    font-weight: bold;
    color: #4CAF50;
}
#subsection {
    margin: 20px 0;
    font-size: 1.2rem;
    color: #333;
    text-align: center;
}
""") as demo:
    gr.Markdown('<div id="app-title">🎧 AI Radio Imaging with Llama 3 + MusicGen (Zero GPU)</div>')

    with gr.Tab("Step 1: Generate Promo Script"):
        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-3-70B"
            )

        generate_script_button = gr.Button("Generate Script")
        script_output = gr.Textbox(label="Generated Promo Script", interactive=False)

        generate_script_button.click(
            fn=radio_imaging_app,
            inputs=[user_prompt, llama_model_id, gr.State(0)],
            outputs=[script_output, None],
        )

    with gr.Tab("Step 2: Generate Audio"):
        with gr.Row():
            audio_length = gr.Slider(
                label="Audio Length (tokens)",
                minimum=128,
                maximum=1024,
                step=64,
                value=512,
            )
            generate_audio_button = gr.Button("Generate Audio")
            audio_output = gr.Audio(label="Generated Audio", type="filepath")

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

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