config.json

{
  "architectures": [
    "Lfm2DSparkDraftModel"
  ],
  "model_type": "qwen3",
  "dtype": "bfloat16",
  "hidden_size": 2048,
  "num_hidden_layers": 5,
  "num_attention_heads": 32,
  "num_key_value_heads": 8,
  "head_dim": 64,
  "intermediate_size": 6144,
  "hidden_act": "silu",
  "rms_norm_eps": 1e-05,
  "vocab_size": 65536,
  "rope_theta": 1000000.0,
  "max_position_embeddings": 128000,
  "layer_types": [
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention"
  ],
  "block_size": 9,
  "dflash_config": {
    "mask_token_id": 64402,
    "target_layer_ids": [
      2,
      5,
      8,
      11,
      14
    ],
    "num_target_layers": 16
  },
  "markov_rank": 256,
  "rope_is_neox_style": false,
  "enable_confidence_head": true,
  "markov_head_type": "vanilla"
}