Hugging Face's logo

IDEAS-Lab-Northwestern
/
lingbot-vla2-datagen-v1-stack-joint-2cam-yanZ

Robotics
Safetensors
lingbotvla
vla
lingbot-vla-2
maniguard
Model card Files
xet
 Community
yypeng666 commited on
Commit
9d5f813
·
verified ·
1 Parent(s): 3b22555

stack: config (local paths scrubbed)

Browse files
Files changed (1) hide show
  1. config.json +173 -0
config.json ADDED
@@ -0,0 +1,173 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_moe_implementation": "fused",
3
+ "action_dim": 55,
4
+ "action_fp32": false,
5
+ "action_head_dim": 128,
6
+ "action_num_attention_heads": 32,
7
+ "action_num_key_value_heads": 8,
8
+ "adanorm_time": true,
9
+ "align_params": {
10
+ "depth": {
11
+ "block_future_depth_to_action": false,
12
+ "detach_future_image_feats": true,
13
+ "dim_head": 32,
14
+ "dim_out": 1024,
15
+ "ff_mult": 1,
16
+ "input_size": 224,
17
+ "model_type": "MoRGBD",
18
+ "moge_path": "assets/pretrained/moge-2-vitb-normal/model.pt",
19
+ "morgbd_path": "assets/pretrained/lingbot-vla-v2-6b/depth/model.pt",
20
+ "num_backbone_tokens": 256,
21
+ "num_heads": 4,
22
+ "num_layers": 1,
23
+ "token_size": 16,
24
+ "use_future_depth": true
25
+ },
26
+ "depth_loss_weight": 0.004,
27
+ "future_depth_loss_weight": 0.004,
28
+ "llm": {
29
+ "dim_out": 2560,
30
+ "image_input_size": 224,
31
+ "image_token_size": 8
32
+ },
33
+ "mode": "query",
34
+ "num_task_tokens": 8,
35
+ "use_future_video": true,
36
+ "video": {
37
+ "attention_mode": "flex_block_causal",
38
+ "block_suffix_to_future_video": false,
39
+ "ckpt_path": "assets/pretrained/lingbot-vla-v2-6b/dino_video/teacher_step_10000.pth",
40
+ "cls_loss_type": "mse",
41
+ "cls_loss_weight": 0.2,
42
+ "cls_pool": "last",
43
+ "config_path": "assets/pretrained/lingbot-vla-v2-6b/dino_video/config.yaml",
44
+ "cosine_loss_weight": 0.2,
45
+ "detach_image_feats": true,
46
+ "dim_head": 32,
47
+ "dim_out": 1024,
48
+ "effective_fps": 1.0,
49
+ "ff_mult": 1,
50
+ "future_video_loss_weight": 0.004,
51
+ "input_size": 256,
52
+ "log_max_samples": 32,
53
+ "log_scale": 16,
54
+ "mse_loss_weight": 1.0,
55
+ "n_blocks": 1,
56
+ "num_backbone_tokens": 256,
57
+ "num_future_frames": 1,
58
+ "num_heads": 4,
59
+ "num_layers": 1,
60
+ "share_future_depth_query": true,
61
+ "use_cls_loss": false,
62
+ "use_cosine_loss": true,
63
+ "use_current_patch_loss": true,
64
+ "use_current_shared_task_proj": true,
65
+ "use_mse_loss": true,
66
+ "use_patch_loss": true,
67
+ "use_shared_future_task_proj": true,
68
+ "use_smooth_l1_loss": false,
69
+ "use_warmup_frame": true
70
+ },
71
+ "visual_dir": "outputs/images",
72
+ "visual_steps": 5000
73
+ },
74
+ "architectures": [
75
+ "LingbotVlaV2Policy"
76
+ ],
77
+ "attention_implementation": "flex_cached",
78
+ "bias_update_speed": 0.0,
79
+ "chunk_size": 50,
80
+ "depth_incremental_training": false,
81
+ "enable_expert_vision": false,
82
+ "expert_hidden_size": 768,
83
+ "expert_intermediate_size": 2752,
84
+ "expert_vision_path": null,
85
+ "expert_vision_type": null,
86
+ "final_norm_adanorm": false,
87
+ "freeze_vision_encoder": false,
88
+ "incremental_training": false,
89
+ "initializer_range": 0.02,
90
+ "loss_type": "L1_fm",
91
+ "max_action_dim": 55,
92
+ "max_state_dim": 55,
93
+ "model_type": "lingbotvla",
94
+ "moe_implementation": "fused",
95
+ "n_action_steps": 50,
96
+ "n_obs_steps": 1,
97
+ "norm_qkv": false,
98
+ "nosplit_gate_liner": false,
99
+ "num_steps": 10,
100
+ "post_training": true,
101
+ "precompute_grid_thw": true,
102
+ "proj_width": 768,
103
+ "qwen3vl_use_vision_boundaries": true,
104
+ "reinit_mismatched_weights": false,
105
+ "return_image_grid_thw": true,
106
+ "routed_scaling_factor": 4.0,
107
+ "router_activation": "sigmoid",
108
+ "router_z_loss_coeff": 0.0001,
109
+ "separate_time_proj": false,
110
+ "sequence_wise_loss_coeff": 0.001,
111
+ "sequence_wise_mode": "per_sequence",
112
+ "split_fused_experts_from_decoder_fsdp": false,
113
+ "split_gate_liner": false,
114
+ "token_moe_intermediate_size": 512,
115
+ "token_moe_layers": [
116
+ 0,
117
+ 1,
118
+ 2,
119
+ 3,
120
+ 4,
121
+ 5,
122
+ 6,
123
+ 7,
124
+ 8,
125
+ 9,
126
+ 10,
127
+ 11,
128
+ 12,
129
+ 13,
130
+ 14,
131
+ 15,
132
+ 16,
133
+ 17,
134
+ 18,
135
+ 19,
136
+ 20,
137
+ 21,
138
+ 22,
139
+ 23,
140
+ 24,
141
+ 25,
142
+ 26,
143
+ 27,
144
+ 28,
145
+ 29,
146
+ 30,
147
+ 31,
148
+ 32,
149
+ 33,
150
+ 34,
151
+ 35
152
+ ],
153
+ "token_num_experts": 32,
154
+ "token_shared_intermediate_size": 704,
155
+ "token_top_k": 4,
156
+ "tokenizer_max_length": 72,
157
+ "tokenizer_path": "assets/pretrained/Qwen3-VL-4B-Instruct",
158
+ "train_expert_only": false,
159
+ "train_state_proj": true,
160
+ "transformers_version": "4.57.3",
161
+ "use_cache": false,
162
+ "use_compile": true,
163
+ "use_lm_head": false,
164
+ "use_moe": true,
165
+ "use_qwen3_chat_template": true,
166
+ "use_qwen3_fixed_grid_cache": true,
167
+ "use_shared_expert_gate": false,
168
+ "vit_attn_implementation": "flash_attention_2",
169
+ "vlm_causal": true,
170
+ "vlm_family": "qwen3_vl",
171
+ "vlm_repo_id": null,
172
+ "vocab_size": 257152
173
+ }