config.yaml

run_name: glue_20251002_155411
seed: 6198
epoch: null
dry_run: false
model:
  d_model: 3584
  n_heads: 28
  n_kv_heads: 4
  qkv_bias: true
  clip_qkv: null
  n_layers: 28
  mlp_ratio: 4
  mlp_hidden_size: 37888
  activation_type: swiglu
  block_type: sequential
  block_group_size: 1
  rope: true
  rope_full_precision: true
  rope_theta: 1000000.0
  vision_backbone:
    image_model_type: openai
    image_default_input_size:
    - 336
    - 336
    image_patch_size: 14
    image_pos_patch_size: 14
    image_emb_dim: 1024
    image_num_heads: 16
    image_num_key_value_heads: 16
    image_num_layers: 23
    image_head_dim: 64
    image_mlp_dim: 4096
    image_mlp_activations: quick_gelu
    image_dropout_rate: 0.0
    image_num_pos: 577
    image_norm_eps: 1.0e-05
    attention_dropout: 0.0
    residual_dropout: 0.0
    initializer_range: 0.02
    fsdp_wrap: false
    resize_mode: default
  vit_load_path: /weka/oe-training-default/mm-olmo/pretrained_image_encoders/vit-l-14-336.pt
  llm_load_path: /weka/oe-training-default/mm-olmo/pretrained_llms/qwen2-7b.pt
  low_cpu_fsdp: true
  attention_type: sdpa
  float32_attention: true
  attention_dropout: 0.0
  attention_layer_norm: false
  residual_dropout: 0.1
  response_residual_dropout: 0.0
  embedding_dropout: 0.0
  layer_norm_type: rms
  layer_norm_with_affine: true
  layer_norm_eps: 1.0e-06
  attention_layer_norm_with_affine: true
  max_sequence_length: 4096
  max_position_embeddings: null
  include_bias: false
  bias_for_layer_norm: null
  scale_logits: false
  vocab_size: 152064
  embedding_size: 152064
  ff_out_size: null
  additional_vocab_size: 128
  new_embedding_init_range: 0.02
  weight_tying: false
  init_device: null
  init_fn: normal
  init_std: 0.02
  init_cutoff_factor: null
  norm_after: false
  precision: amp_bf16
  max_crops: 12
  crop_mode: overlap-and-resize-c2
  use_col_tokens: true
  prompt_type: uber_model
  system_prompt_kind: demo_or_style
  message_formatting: role
  always_start_with_space: true
  multi_annotation_weighting: root_subsegments
  default_inference_len: 65
  overlap_margins:
  - 4
  - 4
  pad_value: 0.0
  image_padding_embed: pad_and_partial_pad
  fix_image_padding: true
  vit_layers:
  - -2
  - -9
  image_pooling_h: 2
  image_pooling_w: 2
  image_pooling_2d: attention_meanq
  image_projector: mlp
  image_feature_dropout: 0.0
  initializer_range: 0.02
  normalize_input_embeds: false
  use_position_ids: true
  head_dim: null
  action_tokenizer:
    identifier: physical-intelligence/fast
    tokenizer_dir: null
  action_dim: 7
  horizon: 8
  tokenizer:
    identifier: Qwen/Qwen2-7B
    tokenizer_dir: null
  pad_tokenizer: true
  moe_num_experts: 8
  moe_top_k: 2
  moe_mlp_impl: sparse
  moe_log_expert_assignment: false
  moe_shared_expert: false
  moe_lbl_in_fp32: false
  moe_interleave: false
  moe_loss_weight: 0.1
  moe_zloss_weight: null
  moe_dropless: true
  moe_capacity_factor: 1.25
  action_head: l1_regression
  num_diffusion_steps: 1000
  num_diffusion_inference_steps: 30
  use_proprio: true
  action_head_dit_hidden_size: 1152
  action_head_dit_depth: 28
  action_head_dit_num_heads: 16
  llm_causal_attention: false
  action_use_left_eef: true
  action_use_mobile_base: false
allow_resume: false
ft_llm: true
ft_vit: false
ft_connector: false
ft_embedding: lm_head
lora: false
use_lora: true
lora_rank: 8
lora_llm: false
lora_vit: false
lora_connector: false
early_exit: false
train_exit_random_layer: false
optimizer:
  name: adamw
  learning_rate: 0.0001
  weight_decay: 0.01
  betas:
  - 0.9
  - 0.95
  eps: 1.0e-05
  connector_learning_rate: 0.0002
  vit_learning_rate: 6.0e-06
  llm_learning_rate: 5.0e-05
  connector_weight_decay: 0.0
  vit_weight_decay: 0.0
  llm_weight_decay: 0.0
  connector_betas:
  - 0.9
  - 0.95
  vit_betas:
  - 0.9
  - 0.95
  llm_betas:
  - 0.9
  - 0.95
  connector_eps: 1.0e-06
  vit_eps: 1.0e-06
  llm_eps: 1.0e-06
  metrics_log_interval: 20
scheduler:
  name: multimodal
  units: steps
  t_warmup: 100
  t_max: null
  alpha_f: 0.1
  connector_t_warmup: 200
  vit_t_warmup: 2000
  llm_t_warmup: 2000
  grad_clip_warmup_steps: null
  grad_clip_warmup_factor: null
  warmup_min_lr: 0.0
data:
  dataset: vla_dataset_realworld
  mixture: null
  root_size_mixture: null
  split: train
  seed: 95818
  shuffle_messages: false
  pad: to_max
  sequence_length: 1600
  shuffle: true
  for_inference: false
  multi_modal: torch
  num_workers: 0
  drop_last: true
  pin_memory: true
  prefetch_factor: null
  persistent_workers: false
  timeout: 0
  rlds_dataset_name: libero_4_task_suites_no_noops
  rlds_data_root_dir: /mnt/data/zhangjian/dataset/Simulation/datasets--openvla--modified_libero_rlds
  use_wrist_image: true
  use_proprio: true
  rlds_shuffle_buffer_size: 100000
  rlds_traj_threads: 8
  rlds_read_threads: 8
  lerobot_episode_index_start: null
  lerobot_episode_index_end: null
restore_dataloader: true
fast_forward_batches: null
evaluators:
- label: val
  data:
    dataset: vla_dataset_realworld
    mixture: null
    root_size_mixture: null
    split: validation
    seed: null
    shuffle_messages: false
    pad: to_max
    sequence_length: 1600
    shuffle: false
    for_inference: false
    multi_modal: torch
    num_workers: 0
    drop_last: true
    pin_memory: true
    prefetch_factor: null
    persistent_workers: true
    timeout: 0
    rlds_dataset_name: libero_4_task_suites_no_noops
    rlds_data_root_dir: /mnt/data/zhangjian/dataset/Simulation/datasets--openvla--modified_libero_rlds
    use_wrist_image: true
    use_proprio: true
    rlds_shuffle_buffer_size: 256000
    rlds_traj_threads: 8
    rlds_read_threads: 8
    lerobot_episode_index_start: 353
    lerobot_episode_index_end: 765
  device_eval_batch_size: null
  subset_num_batches: 64
  max_examples: null
  max_new_tokens: 448
  mm_evaluator: null
  save_dir: null
  save_to_checkpoint_dir: false
  eval_name: null
  skip_if_metrics_cached: true
eval_interval: 0
inf_eval_interval: -1
inf_evaluators: []
save_folder: /vast/users/xiaodan/workspace/minghao.guo/warehouse_a1/ckpt/
remote_save_folder: null
canceled_check_interval: 50
save_interval: 500
save_interval_unsharded: 500
save_interval_ephemeral: null
save_interval_action_head: 500
save_num_checkpoints_to_keep: 1
save_num_unsharded_checkpoints_to_keep: 1
save_num_action_head_checkpoints_to_keep: 2
save_overwrite: true
force_save_unsharded: false
no_pre_train_checkpoint: true
initial_model_checkpoint: /vast/users/xiaodan/zhangjian/molmo_data/Molmo-7B-D-0924
load_model_config: null
checkpoint_dir: /vast/users/xiaodan/zhangjian/molmo_data/Molmo-7B-D-0924
load_path: null
load_path_sharded_checkpointer: null
reset_optimizer_state: false
reset_trainer_state: false
save_dataloader_state: false
reset_dataloader_state: false
keep_lr_on_load: true
sharded_checkpointer: torch_legacy
max_duration: 500000
global_train_batch_size: 126
device_train_batch_size: 15
device_train_microbatch_size: 16
device_eval_batch_size: 4
eval_subset_num_batches: -1
eval_on_load: false
device_inf_eval_batch_size: 16
inf_eval_subset_num_batches: -1
device_train_grad_accum: 0
max_grad_norm: 1.0
multi_component_grad_norm: true
batch_divisor: global_batch
max_grad_norm_ratio: null
precision: amp_bf16
wandb:
  project: a1-realworld
  entity: henryeap
  group: null
  name: glue_20251002_155411
  tags:
  - watching
  log_artifacts: false
  rank_zero_only: true
  log_interval: 1
speed_monitor:
  window_size: 20
  gpu_flops_available: null
console_log_interval: 1
gen1_gc_interval: 1
compile: null
fsdp:
  use_orig_params: true
  sharding_strategy: FULL_SHARD
  wrapping_strategy: by_block_and_size
  precision: float
  hybrid_sharding_num_model_replicas: null
softmax_auxiliary_loss: true
softmax_auxiliary_loss_scale: 0.0001
time_limit: null
extra_steps_after_cancel: 10
python_profiling: false
torch_profiling: false
stop_at: 500000
stop_after: null
activation_checkpointing: whole_layer
fused_loss: null