config.json

{
  "attention_bias": false,
  "head_dim": 128,
  "hidden_act": "silu",
  "hidden_size": 2048,
  "intermediate_size": 8192,
  "max_position_embeddings": 262144,
  "model_type": "laguna",
  "num_attention_heads": 64,
  "num_hidden_layers": 5,
  "num_key_value_heads": 8,
  "rms_norm_eps": 1e-06,
  "rope_theta": 500000.0,
  "sliding_window": 512,
  "vocab_size": 100352,
  "layer_types": [
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "sliding_attention"
  ],
  "gating": "per-head",
  "architectures": [
    "DFlashLagunaForCausalLM"
  ],
  "draft_vocab_size": 100352,
  "torch_dtype": "bfloat16",
  "eagle_aux_hidden_state_layer_ids": [
    2,
    14,
    26,
    34,
    40
  ],
  "dflash_config": {
    "block_size": 16,
    "mask_token_id": 12,
    "num_target_layers": 40,
    "target_layer_ids": [
      1,
      13,
      25,
      33,
      39
    ],
    "causal": true
  },
  "num_experts": 0
}