feat: run on Modal

- introduce stub to run Dia on Modal
- update README

Author: stukennedy

README.md

@@ -87,6 +87,24 @@ sf.write("simple.mp3", output, 44100)
 
 A pypi package and a working CLI tool will be available soon.
 
+### Run on Modal
+
+We are providing a modal stub for easy deployment.
+To do this, you need to have a modal account.
+You can sign up [here](https://modal.com/signup) and follow the onboarding instructions.
+
+Then, install the modal CLI.
+
+```bash
+pip install modal
+```
+
+Now, you can run the following command.
+
+```bash
+uv run dia_modal_stub.py --script "[S1] Hello, how are you? [S2] I'm fine, thank you." --out output.wav
+```
+
 ## 💻 Hardware and Inference Speed
 
 Dia has been tested on only GPUs (pytorch 2.0+, CUDA 12.6). CPU support is to be added soon.
@@ -133,7 +151,6 @@ Join our [Discord Server](https://discord.gg/pgdB5YRe) for discussions.
 - "Nari" is a pure Korean word for lily.
 - We thank Jason Y. for providing help with data filtering.
 
-
 ## ⭐ Star History
 
 <a href="https://www.star-history.com/#nari-labs/dia&Date">

dia_modal_stub.py

@@ -0,0 +1,66 @@
+import io
+import pathlib
+import modal
+
+# ---------- Modal App and Image ----------
+
+stub = modal.App("dia-tts-app")
+
+# Base CUDA‑ready Debian image + Python 3.10
+# We install torch with GPU wheels, Dia itself and a couple of audio helpers.
+image = (
+    modal.Image.debian_slim(python_version="3.10")
+    # SoundFile requires libsndfile at the OS level.
+    .apt_install("git", "ffmpeg", "libsndfile1").pip_install(
+        # CUDA enabled torch/torchaudio wheels
+        "torch==2.6.0",  # Modal already has the CUDA runtime so the PyPI wheel is fine
+        "torchaudio==2.6.0",
+        "soundfile",
+        # Install Dia straight from GitHub ― keeps you on the latest commit
+        "git+https://github.com/nari-labs/dia.git",
+    )
+)
+
+# ---------- Remote function ----------
+
+
+@stub.function(
+    image=image,
+    gpu="A10G",  # ~10 GB VRAM, plenty for Dia 1.6 B
+    timeout=60 * 15,  # 15‑minute max runtime (first run downloads ~4 GB weights)
+)
+def tts(text: str) -> bytes:
+    """Generate speech from *text* and return WAV bytes."""
+    from dia.model import Dia
+    import soundfile as sf
+
+    # Load the pretrained model (cached inside the container after the first call)
+    model = Dia.from_pretrained("nari-labs/Dia-1.6B")
+
+    # Dia uses 44.1 kHz, 16‑bit WAV output
+    wav = model.generate(text)
+
+    # Serialise the NumPy array returned by Dia to an in‑memory WAV file
+    buf = io.BytesIO()
+    sf.write(buf, wav, 44100, format="WAV", subtype="PCM_16")
+    buf.seek(0)
+    return buf.read()
+
+
+# ---------- Local entry‑point ----------
+
+
+@stub.local_entrypoint()
+def main(
+    script: str = "[S1] Dia running on Modal! [S2] Sounds good, doesn’t it?",
+    out: str = "output.wav",
+):
+    """CLI entry‑point executed on your MacBook via `modal run`.
+
+    Examples
+    --------
+    $ modal run dia_modal_stub.py --script "[S1] Hello world" --out hello.wav
+    """
+    audio = tts.remote(script)  # Remote GPU call ➜ audio bytes back to your laptop
+    pathlib.Path(out).write_bytes(audio)
+    print(f"\N{MUSICAL NOTE}  Saved {out} ({len(audio) // 1024} KB)")