"""Prepare MMAudio once on CPU, and mux sound without re-encoding video frames.""" from pathlib import Path import subprocess import threading _bundle = None _prepare_lock = threading.Lock() def prepare(cfg): global _bundle with _prepare_lock: if _bundle is not None: return _bundle import torch from mmaudio.model.networks import get_my_mmaudio from mmaudio.model.utils.features_utils import FeaturesUtils with torch.device("cpu"): net = get_my_mmaudio(cfg.model_name) net.load_weights(torch.load(cfg.model_path, map_location="cpu", weights_only=True)) features = FeaturesUtils(tod_vae_ckpt=cfg.vae_path, synchformer_ckpt=cfg.synchformer_ckpt, enable_conditions=True, mode=cfg.mode, bigvgan_vocoder_ckpt=cfg.bigvgan_16k_path, need_vae_encoder=False) # The GPU path already used BF16. Keep the same precision, without # retaining gradients or constructing/downloading models per clip. net = net.eval().requires_grad_(False).to(dtype=torch.bfloat16) features = features.eval().requires_grad_(False).to(dtype=torch.bfloat16) _bundle = (cfg, net, features) return _bundle def ready_models(): if _bundle is None: raise RuntimeError("MMAudio was not prepared at startup. Restart the Space and check its startup logs.") return _bundle def mux_audio(video_path, wav_path, output_path): result = subprocess.run( ["ffmpeg", "-nostdin", "-hide_banner", "-loglevel", "error", "-y", "-i", str(video_path), "-i", str(wav_path), "-map", "0:v:0", "-map", "1:a:0", "-c:v", "copy", "-c:a", "aac", "-b:a", "192k", "-af", "apad", "-shortest", "-movflags", "+faststart", str(output_path)], capture_output=True, ) if result.returncode or not Path(output_path).is_file(): Path(output_path).unlink(missing_ok=True) raise RuntimeError("Audio/video mux failed: " + result.stderr.decode("utf-8", "replace")[-1200:]) return str(output_path)