config.json

{
  "speculators_model_type": "mtp",
  "architectures": [
    "MTPSpeculator"
  ],
  "num_speculative_steps": 3,
  "num_nextn_predict_layers": 1,
  "mtp_loss_step_weights": [
    0.51,
    0.31,
    0.18
  ],
  "hidden_size": 2048,
  "intermediate_size": 5120,
  "num_attention_heads": 16,
  "num_key_value_heads": 2,
  "vocab_size": 151936,
  "max_position_embeddings": 262144,
  "rms_norm_eps": 1e-06,
  "transformer_config": {
    "model_type": "qwen3_next",
    "vocab_size": 151936,
    "hidden_size": 2048,
    "intermediate_size": 5120,
    "num_hidden_layers": 1,
    "num_attention_heads": 16,
    "num_key_value_heads": 2,
    "max_position_embeddings": 262144,
    "rms_norm_eps": 1e-06,
    "rope_theta": 10000000,
    "tie_word_embeddings": false,
    "attention_bias": true
  },
  "speculators_config": {
    "algorithm": "mtp",
    "proposal_methods": [
      {
        "proposal_type": "greedy",
        "speculative_tokens": 3,
        "verifier_accept_k": 1,
        "accept_tolerance": 0.0
      }
    ],
    "default_proposal_method": "greedy",
    "verifier": {
      "name_or_path": "Qwen/Qwen3-Next-80B-A3B-Instruct",
      "architectures": ["Qwen3MoeForCausalLM"]
    }
  }
}