app.py

import subprocess
import sys
import os
import shutil
import site
import importlib

# Step 1: Delete wrong PyPI dia2 files
for path in site.getsitepackages():
    dia2_path = os.path.join(path, "dia2")
    if os.path.exists(dia2_path):
        print(f"Removing: {dia2_path}")
        shutil.rmtree(dia2_path)
    for entry in os.listdir(path):
        if entry.startswith("dia2") and (entry.endswith(".dist-info") or entry.endswith(".egg-info")):
            shutil.rmtree(os.path.join(path, entry))

# Step 2: Clone real repo
repo_dir = "/tmp/dia2_src"
if os.path.exists(repo_dir):
    shutil.rmtree(repo_dir)
subprocess.run(["git", "clone", "https://github.com/nari-labs/dia2.git", repo_dir], check=True)

# Step 3: Install deps from pyproject.toml (no-deps since torch already in requirements.txt)
subprocess.run([
    sys.executable, "-m", "pip", "install", "--quiet", "-e", repo_dir
], check=True)
# Step 4: Bust cache and prepend path
sys.path.insert(0, repo_dir)
importlib.invalidate_caches()

import gradio as gr
import tempfile
import numpy as np
import soundfile as sf
import scipy.io.wavfile as wavfile
from dia2 import Dia2, GenerationConfig, SamplingConfig

print("Loading Dia2-1B model...")
model = Dia2.from_repo("nari-labs/Dia2-1B", device="cuda", dtype="bfloat16")
print("Model loaded!")

def prepare_audio(audio_path, max_seconds=10):
    if audio_path is None:
        return None
    data, samplerate = sf.read(audio_path)
    if len(data.shape) > 1:
        data = data.mean(axis=1)
    max_samples = int(max_seconds * samplerate)
    if len(data) > max_samples:
        data = data[:max_samples]
    tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
    data_int16 = (data * 32767).astype(np.int16)
    wavfile.write(tmp.name, samplerate, data_int16)
    return tmp.name

def generate(text, speaker1_audio, speaker2_audio, cfg_scale, temperature, top_k):
    config = GenerationConfig(
        cfg_scale=cfg_scale,
        audio=SamplingConfig(temperature=temperature, top_k=int(top_k)),
        use_cuda_graph=True,
    )

    output_file = tempfile.NamedTemporaryFile(suffix=".wav", delete=False).name

    kwargs = dict(config=config, output_wav=output_file, verbose=True)

    if speaker1_audio:
        kwargs["prefix_speaker_1"] = prepare_audio(speaker1_audio)
    if speaker2_audio:
        kwargs["prefix_speaker_2"] = prepare_audio(speaker2_audio)

    model.generate(text, **kwargs)
    return output_file

demo = gr.Interface(
    fn=generate,
    inputs=[
        gr.Textbox(
            label="Script ([S1]/[S2] speaker tags)",
            placeholder="[S1] Hey, how are you?\n[S2] I'm doing great, thanks!",
            lines=5
        ),
        gr.Audio(label="Speaker 1 Voice Prompt (optional, max 10s)", type="filepath"),
        gr.Audio(label="Speaker 2 Voice Prompt (optional, max 10s)", type="filepath"),
        gr.Slider(1.0, 5.0, value=2.0, step=0.1, label="CFG Scale"),
        gr.Slider(0.1, 2.0, value=0.8, step=0.05, label="Temperature"),
        gr.Slider(10, 100, value=50, step=5, label="Top K"),
    ],
    outputs=gr.Audio(label="Generated Audio"),
    title="Dia2-2B TTS",
)

demo.launch()