{
  "_debug_force_load_balance": false,
  "architectures": [
    "MotifForCausalLM"
  ],
  "attention_cls": "gdla",
  "attention_dropout": 0.0,
  "auto_map": {
    "AutoConfig": "configuration_motif.MotifConfig",
    "AutoModel": "modeling_motif.MotifForCausalLM",
    "AutoModelForCausalLM": "modeling_motif.MotifForCausalLM"
  },
  "diff_v2": true,
  "dtype": "bfloat16",
  "elementwise_attn_output_gate": true,
  "eos_token_id": 0,
  "experts_top_k": 8,
  "head_dim": 192,
  "headwise_attn_output_gate": false,
  "hidden_act": "poly_norm",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "interleave_moe_layer_step": 1,
  "intermediate_size": 12288,
  "k_ratio": 1,
  "kv_lora_rank": 512,
  "load_balance_coeff": 0.0001,
  "max_position_embeddings": 262144,
  "max_window_layers": 9,
  "mhc_enabled": true,
  "mhc_expansion_rate": 4,
  "mhc_identity_init": false,
  "mhc_sinkhorn_iters": 20,
  "model_type": "Motif",
  "moe_intermediate_size": 1280,
  "mscale": 1.0,
  "n_dense_first_layers": 2,
  "num_attention_heads": 80,
  "num_experts": 384,
  "num_hidden_layers": 53,
  "num_key_value_heads": 16,
  "num_noise_heads": 16,
  "num_shared_experts": 1,
  "output_router_logits": false,
  "q_lora_rank": 1024,
  "qk_rope_head_dim": 64,
  "rms_norm_eps": 1e-05,
  "rope_theta": 10000.0,
  "route_norm": true,
  "route_scale": 2.0,
  "router_aux_loss_coef": 0.0,
  "score_before_experts": false,
  "score_func": "sigmoid",
  "sliding_window": 128,
  "sliding_window_pattern": "interleave",
  "sliding_window_period": 4,
  "swa_rope_theta": 10000.0,
  "tie_word_embeddings": false,
  "transformers_version": "5.7.0",
  "use_cache": true,
  "use_sliding_window": true,
  "v_head_dim": 128,
  "vocab_size": 220160,
  "rope_factor": 64.0,
  "original_seq_len": 4096,
  "rope_scaling": {
    "original_max_position_embeddings": 4096,
    "factor": 64.0,
    "mscale": 1.0,
    "rope_type": "yarn",
    "rope_theta": 10000.0,
    "beta_fast": 32.0,
    "beta_slow": 1.0,
    "apply_yarn_scaling": false
  },
  "polynorm_output_scale": 0.5,
  "polynorm_output_scale_per_layer": {},
  "polynorm_bias_clamp": 0.5,
  "hidden_clamp": 1000000.0,
  "num_nextn_predict_layers": 1
}