model.json

{
  "name": "SongGen",
  "local_smoke_threshold_gb": 20.0,
  "package_dir": "songgen",
  "entry_point": [
    "songgen.modeling_songgen_mixed.SongGenMixedForConditionalGeneration",
    "songgen.modeling_songgen_dual_track.SongGenDualTrackForConditionalGeneration",
    "songgen.processing_songgen.SongGenProcessor"
  ],
  "checkpoint": {
    "size_mb": 14114,
    "note": "loaded via from_pretrained/hf_hub_download at runtime, not committed to Space repo. Both mode checkpoints held in memory simultaneously (dropdown selects mode at inference time): LiuZH-19/SongGen_mixed_pro (~5457MB) + LiuZH-19/SongGen_interleaving_A_V (~5422MB), each of which internally constructs its own full m-a-p/MERT-v1-330M model via AutoModel.from_pretrained(trust_remote_code=True) (~1262MB each, loaded twice total) for reference-voice conditioning. Plus xcodec ZhenYe234/xcodec xcodec_hubert_general_audio_v2.pth + config_hubert_general.yaml (~710MB, shared single copy) and htdemucs separation weights (~150MB) auto-downloaded by SongGenProcessor (also loaded once per mode).",
    "repos": {
      "mixed_pro": "LiuZH-19/SongGen_mixed_pro",
      "dual_track": "LiuZH-19/SongGen_interleaving_A_V",
      "xcodec": {"repo": "ZhenYe234/xcodec", "filename": "xcodec_hubert_general_audio_v2.pth", "config_filename": "config_hubert_general.yaml"},
      "mert": "m-a-p/MERT-v1-330M"
    }
  }
}