app.py

import torch
import gradio as gr
import numpy as np
import soundfile as sf

from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_ID = "m-a-p/YuE-s1-0.5B"

# Load tokenizer (slow is REQUIRED)
tokenizer = AutoTokenizer.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    use_fast=False
)

# Load model (GPU REQUIRED)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    torch_dtype=torch.float16,
    device_map="auto"
)
model.eval()

# ----------------------------
# SIMPLE AUDIO TOKEN DECODER
# ----------------------------
# NOTE:
# YuE uses xcodec tokens.
# This is a *placeholder decoder*.
# Official decoder is required for best quality.

def fake_xcodec_decode(token_ids, sample_rate=44100):
    """
    This is a minimal placeholder decoder.
    For REAL quality, you must use YuE's official xcodec decoder.
    """
    duration = len(token_ids) // 50
    t = np.linspace(0, duration, duration * sample_rate)
    audio = 0.1 * np.sin(2 * np.pi * 440 * t)
    return audio.astype(np.float32), sample_rate


def generate_music(prompt, max_tokens=2048, temperature=1.0):
    if not prompt.strip():
        return None

    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        output = model.generate(
            **inputs,
            max_new_tokens=max_tokens,
            do_sample=True,
            temperature=temperature,
            eos_token_id=tokenizer.eos_token_id
        )

    token_ids = output[0].cpu().numpy()

    audio, sr = fake_xcodec_decode(token_ids)

    return (sr, audio)


# ----------------------------
# GRADIO UI
# ----------------------------

with gr.Blocks(title="YuE Music Generator") as demo:
    gr.Markdown(
        """
        # 🎵 YuE Song Generator
        Text → AI Music

        **Model:** m-a-p/YuE-s1-0.5B  
        ⚠️ Requires GPU
        """
    )

    prompt = gr.Textbox(
        label="Music Prompt",
        placeholder="A sad lo-fi song with piano and rain ambience",
        lines=3
    )

    max_tokens = gr.Slider(512, 4096, 2048, step=256, label="Max Tokens")
    temperature = gr.Slider(0.7, 1.5, 1.0, step=0.1, label="Creativity")

    btn = gr.Button("Generate Music")
    audio_out = gr.Audio(label="Generated Music", type="numpy")

    btn.click(
        generate_music,
        inputs=[prompt, max_tokens, temperature],
        outputs=audio_out
    )

demo.launch()