lingbotvla_cli.yaml

data:
  cameras:
  - camera_top
  - camera_wrist_left
  - camera_wrist_right
  chat_template: default
  data_name: pot14
  data_type: conversation
  dataloader_type: native
  datasets_type: vla
  drop_last: true
  img_size: 224
  joints:
  - '{''arm.position'': 6}'
  - '{''effector.position'': 1}'
  max_seq_len: 2048
  norm_stats_file: assets/norm_stats/pot14_right_arm.json
  norm_type: meanstd
  num_workers: 4
  pin_memory: true
  prefetch_factor: 2
  robot_config_root: configs/robot_configs
  text_keys: messages
  train_path: /root/autodl-tmp/data/processed/pot14_right_arm
  train_size: 10000000
  video_backend: torchcodec
model:
  adanorm_time: true
  attn_implementation: flash_attention_2
  basic_modules: []
  config_path: /root/autodl-tmp/models/pretrained/lingbot-vla-4b
  decoders: {}
  encoders: {}
  final_norm_adanorm: false
  force_use_huggingface: false
  model_path: /root/autodl-tmp/models/pretrained/lingbot-vla-4b
  moge_path: null
  morgbd_path: null
  nosplit_gate_liner: false
  post_training: true
  separate_time_proj: false
  split_gate_liner: false
  tokenizer_path: /root/autodl-tmp/models/pretrained/Qwen2.5-VL-3B-Instruct
  use_lm_head: false
  vlm_repo_id: null
  vocab_size: 0
train:
  action_dim: null
  activation_gpu_limit: 0.0
  adapt_to_pi_aloha: false
  align_params: {}
  bsz_warmup_init_mbtoken: 200
  bsz_warmup_ratio: 0
  chunk_size: 50
  ckpt_manager: dcp
  context_parallel_size: 1
  data_parallel_mode: fsdp2
  data_parallel_replicate_size: 1
  data_parallel_shard_size: 1
  decayed_max_grad_norm: 1.0
  dyn_bsz: true
  dyn_bsz_buffer_size: 200
  dyn_bsz_margin: 0
  empty_cache_steps: 500
  enable_activation_offload: false
  enable_expert_vision: false
  enable_forward_prefetch: true
  enable_fp32: true
  enable_fsdp_offload: false
  enable_full_determinism: false
  enable_full_shard: false
  enable_gradient_checkpointing: true
  enable_mixed_precision: true
  enable_profiling: false
  enable_reentrant: false
  enable_resume: false
  expert_parallel_size: 1
  expert_vision_path: null
  expert_vision_type: null
  freeze_vision_encoder: false
  freeze_vit: false
  global_batch_size: 32
  gradient_accumulation_steps: 16
  init_device: cuda
  load_checkpoint_path: null
  loss_type: fm
  lr: 5.0e-05
  lr_decay_ratio: 1.0
  lr_decay_style: constant
  lr_min: 1.0e-07
  lr_start: 0.0
  lr_warmup_ratio: 0
  max_action_dim: 75
  max_checkpoints_to_keep: 1
  max_grad_norm: 1.0
  max_state_dim: 75
  max_steps: 1024
  micro_batch_size: 2
  module_fsdp_enable: true
  norm_qkv: false
  num_steps: 10
  num_train_epochs: 999999999
  optimizer: adamw
  output_dir: output/pot14/
  pipeline_parallel_size: 1
  profile_end_step: 2
  profile_profile_memory: true
  profile_record_shapes: true
  profile_start_step: 1
  profile_trace_dir: ./trace
  profile_with_stack: true
  resize_imgs_with_padding:
  - 224
  - 224
  resume_dataloader_state: true
  rmpad: false
  rmpad_with_pos_ids: false
  save_epochs: 0
  save_hf_weights: true
  save_optimizer: false
  save_steps: 0
  seed: 42
  stable_train_steps: 100000
  tensor_parallel_size: 1
  tokenizer_max_length: 72
  train_expert_only: false
  train_state_proj: true
  ulysses_parallel_size: 1
  use_cache: true
  use_compile: true
  use_delta_joint_actions_aloha: false
  use_wandb: false
  vit_lr: 1.0e-06
  vlm_causal: false
  wandb_name: null
  wandb_project: LingBotVLA
  weight_decay: 0