token-channel/ref/dspark-b15/config.json

{
  "architectures": [
    "Qwen3DSparkModel"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "auto_map": {
    "AutoModel": "dspark.DSparkDraftModel"
  },
  "block_size": 15,
  "block_size_convention": "Without-anchor (DeepSpec/DSpark): block_size=15 trains labels anchor+1..anchor+15 = 15 real drafts. Matches jetspec-qwen3-8b draft depth (DFlash with-anchor bs16 = 15 drafts). DSpark stays without-anchor (follow DeepSeek, per Lanxiang).",
  "bos_token_id": 151643,
  "ce_loss_alpha": 0.1,
  "confidence_head_alpha": 1.0,
  "confidence_head_with_markov": true,
  "dflash_config": {
    "causal_head": false,
    "mask_token_id": 151669,
    "target_layer_ids": [
      1,
      9,
      17,
      25,
      33
    ]
  },
  "dtype": "bfloat16",
  "enable_confidence_head": true,
  "eos_token_id": 151645,
  "head_dim": 128,
  "hidden_act": "silu",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "intermediate_size": 12288,
  "l1_loss_alpha": 0.9,
  "layer_types": [
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention"
  ],
  "loss_decay_gamma": 7.0,
  "loss_decay_gamma_note": "7.0 matches our DFlash block_size=16 convention (run_qwen3_8b_dflash*.sh), not DeepSpec's 4.0 which pairs with their block_size=7. Gamma scales with block size.",
  "markov_head_type": "vanilla",
  "markov_rank": 256,
  "mask_token_id": 151669,
  "max_position_embeddings": 40960,
  "max_window_layers": 5,
  "model_type": "qwen3",
  "num_anchors": 512,
  "num_attention_heads": 32,
  "num_hidden_layers": 5,
  "num_key_value_heads": 8,
  "num_target_layers": 36,
  "pad_token_id": null,
  "rms_norm_eps": 1e-06,
  "rope_parameters": {
    "rope_theta": 1000000,
    "rope_type": "default"
  },
  "sliding_window": null,
  "target_layer_ids": [
    1,
    9,
    17,
    25,
    33
  ],
  "tie_word_embeddings": false,
  "transformers_version": "5.9.0",
  "use_cache": true,
  "use_sliding_window": false,
  "vocab_size": 151936
}