infinifi/fal_app.py

import io
from pathlib import Path
from audiocraft.data.audio import audio_write
import fal
from fastapi import WebSocket
import torch

from prompts import PROMPTS

DATA_DIR = Path("/data/audio")


class InfinifiFalApp(fal.App, keep_alive=300):
    machine_type = "GPU-A6000"
    requirements = [
        "torch==2.1.0",
        "audiocraft==1.3.0",
        "torchaudio==2.1.0",
        "websockets==11.0.3",
    ]

    def setup(self):
        import torchaudio
        from audiocraft.models.musicgen import MusicGen

        self.model = MusicGen.get_pretrained("facebook/musicgen-large")
        self.model.set_generation_params(duration=60)

    @fal.endpoint("/generate")
    def run(self):
        wav = self.model.generate(PROMPTS)

        serialized = []
        for one_wav in wav:
            buf = io.BytesIO()
            torch.save(one_wav.cpu(), buf)
            serialized.append(buf.getvalue())

        return serialized

    @fal.endpoint("/ws")
    async def run_ws(self, ws: WebSocket):
        await ws.accept()

        wav = self.model.generate(PROMPTS)

        for i, one_wav in enumerate(wav):
            path = DATA_DIR.joinpath(f"{i}")
            audio_write(
                path,
                one_wav.cpu(),
                self.model.sample_rate,
                format="mp3",
                strategy="loudness",
                loudness_compressor=True,
            )
            with open(path, "rb") as f:
                data = f.read()
                await ws.send_bytes(data)

        await ws.close()