Kandinsky6 audio VAE configuration.
Mirrors the checkpoint's audio_vae/config.json: { "_class_name": "MMAudioVAE", "mode": "44k", "sample_rate": 44100, "downsample_factor": 1024, "scaling_factor": 0.5302, "vocoder_config": {...BigVGAN-v2 hyperparams, ignored here -- redundant backward-compat copy; the pipeline loads a separate "vocoder" component instead, see fastvideo/models/audio/kandinsky6_audio_vae.py} }