config.json

{
  "_moe_implementation": "fused",
  "action_dim": 55,
  "action_fp32": false,
  "action_head_dim": 128,
  "action_num_attention_heads": 32,
  "action_num_key_value_heads": 8,
  "adanorm_time": true,
  "align_params": {},
  "architectures": [
    "LingbotVlaV2Policy"
  ],
  "attention_implementation": "flex_cached",
  "bias_update_speed": 0.0,
  "chunk_size": 50,
  "depth_incremental_training": false,
  "enable_expert_vision": false,
  "expert_hidden_size": 768,
  "expert_intermediate_size": 2752,
  "expert_vision_path": null,
  "expert_vision_type": null,
  "final_norm_adanorm": false,
  "freeze_vision_encoder": false,
  "incremental_training": false,
  "initializer_range": 0.02,
  "loss_type": "fm",
  "max_action_dim": 55,
  "max_state_dim": 55,
  "model_type": "lingbotvla",
  "moe_implementation": "fused",
  "n_action_steps": 50,
  "n_obs_steps": 1,
  "norm_qkv": false,
  "nosplit_gate_liner": false,
  "num_steps": 10,
  "post_training": false,
  "precompute_grid_thw": true,
  "proj_width": 768,
  "qwen3vl_use_vision_boundaries": true,
  "reinit_mismatched_weights": false,
  "return_image_grid_thw": true,
  "routed_scaling_factor": 4.0,
  "router_activation": "sigmoid",
  "router_z_loss_coeff": 0.0001,
  "separate_time_proj": false,
  "sequence_wise_loss_coeff": 0.001,
  "sequence_wise_mode": "per_sequence",
  "split_fused_experts_from_decoder_fsdp": false,
  "split_gate_liner": false,
  "token_moe_intermediate_size": 512,
  "token_moe_layers": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15,
    16,
    17,
    18,
    19,
    20,
    21,
    22,
    23,
    24,
    25,
    26,
    27,
    28,
    29,
    30,
    31,
    32,
    33,
    34,
    35
  ],
  "token_num_experts": 32,
  "token_shared_intermediate_size": 704,
  "token_top_k": 4,
  "tokenizer_max_length": 72,
  "tokenizer_path": "/home/k_humanoid_3/lerobot_lab/models/Qwen3-VL-4B-Instruct",
  "train_expert_only": true,
  "train_state_proj": true,
  "transformers_version": "4.57.3",
  "use_cache": false,
  "use_compile": false,
  "use_lm_head": false,
  "use_moe": true,
  "use_qwen3_chat_template": true,
  "use_qwen3_fixed_grid_cache": true,
  "use_shared_expert_gate": false,
  "vit_attn_implementation": "flash_attention_2",
  "vlm_causal": true,
  "vlm_family": "qwen3_vl",
  "vlm_repo_id": null,
  "vocab_size": 257152
}