config.json

{
  "architectures": [
    "EncoderTuningModel"
  ],
  "audio_encoder_name": "models/flamingo_qwen/flamingo_qwen2_moe.py",
  "audio_encoder_params": {
    "adapter_dropout": 0.1,
    "adapter_type": "adapter",
    "cls_task_ids": "None",
    "fusion_mode": "moe_swiglu",
    "num_tasks": 12,
    "qformer_d_model": 1024,
    "qformer_drop_path": 0.05,
    "qformer_dropout": 0.1,
    "qformer_heads": 16,
    "qformer_layers": 6,
    "qformer_num_global_queries": 32,
    "qformer_num_local_queries": 96,
    "qformer_pool_stride": 4,
    "task_cond": false
  },
  "decoder_type": "/private/models/SmolLM2-135M",
  "dtype": "float32",
  "lora_alpha": 32,
  "lora_dropout": 0.1,
  "lora_r": 8,
  "lora_target_modules": "all-linear",
  "model_type": "encoder_tuning_model",
  "train_encoder_stage1": false,
  "train_encoder_stage2": false,
  "transformers_version": "5.0.0"
}