config.json

{
  "architectures": [
    "FlowMatchingWithBigVGan"
  ],
  "dtype": "float32",
  "model_config": {
    "attention_dropout": 0.0,
    "cfg_dropout": 0.2,
    "cfg_strength": 0.7,
    "dt": 0.1,
    "embedding_dim": 768,
    "hidden_size": 512,
    "intermediate_size": 1024,
    "max_position_embeddings": null,
    "mean": -5.8843,
    "model_type": "",
    "num_attention_heads": 2,
    "num_encoder_layers": 2,
    "num_hidden_layers": 4,
    "num_mel_bins": 80,
    "rope_parameters": {
      "rope_theta": 10000.0,
      "rope_type": "default"
    },
    "rope_theta": 10000.0,
    "std": 2.2615,
    "vocab_size": 8192
  },
  "model_type": "flow_matching_with_bigvgan",
  "transformers_version": "5.8.0.dev0",
  "vocoder_config": {
    "architectures": [
      "Qwen2_5OmniToken2WavBigVGANModel"
    ],
    "dtype": "float32",
    "mel_dim": 80,
    "model_type": "qwen2_5_omni_bigvgan",
    "resblock_dilation_sizes": [
      [
        1,
        3,
        5
      ],
      [
        1,
        3,
        5
      ],
      [
        1,
        3,
        5
      ]
    ],
    "resblock_kernel_sizes": [
      3,
      7,
      11
    ],
    "upsample_initial_channel": 512,
    "upsample_kernel_sizes": [
      10,
      9,
      8,
      4,
      4
    ],
    "upsample_rates": [
      5,
      4,
      4,
      2,
      2
    ]
  }
}