CoolFace
Modelpublic

spatialtemporal-ai/a1_model

sourceHugging Faceupdated 9mo agoView on Hugging Face
0likes8downloads
config.yaml323 linesDownload Raw Back to root
1run_name: glue_20251002_1554112seed: 61983epoch: null4dry_run: false5model:6  d_model: 35847  n_heads: 288  n_kv_heads: 49  qkv_bias: true10  clip_qkv: null11  n_layers: 2812  mlp_ratio: 413  mlp_hidden_size: 3788814  activation_type: swiglu15  block_type: sequential16  block_group_size: 117  rope: true18  rope_full_precision: true19  rope_theta: 1000000.020  vision_backbone:21    image_model_type: openai22    image_default_input_size:23    - 33624    - 33625    image_patch_size: 1426    image_pos_patch_size: 1427    image_emb_dim: 102428    image_num_heads: 1629    image_num_key_value_heads: 1630    image_num_layers: 2331    image_head_dim: 6432    image_mlp_dim: 409633    image_mlp_activations: quick_gelu34    image_dropout_rate: 0.035    image_num_pos: 57736    image_norm_eps: 1.0e-0537    attention_dropout: 0.038    residual_dropout: 0.039    initializer_range: 0.0240    fsdp_wrap: false41    resize_mode: default42  vit_load_path: /weka/oe-training-default/mm-olmo/pretrained_image_encoders/vit-l-14-336.pt43  llm_load_path: /weka/oe-training-default/mm-olmo/pretrained_llms/qwen2-7b.pt44  low_cpu_fsdp: true45  attention_type: sdpa46  float32_attention: true47  attention_dropout: 0.048  attention_layer_norm: false49  residual_dropout: 0.150  response_residual_dropout: 0.051  embedding_dropout: 0.052  layer_norm_type: rms53  layer_norm_with_affine: true54  layer_norm_eps: 1.0e-0655  attention_layer_norm_with_affine: true56  max_sequence_length: 409657  max_position_embeddings: null58  include_bias: false59  bias_for_layer_norm: null60  scale_logits: false61  vocab_size: 15206462  embedding_size: 15206463  ff_out_size: null64  additional_vocab_size: 12865  new_embedding_init_range: 0.0266  weight_tying: false67  init_device: null68  init_fn: normal69  init_std: 0.0270  init_cutoff_factor: null71  norm_after: false72  precision: amp_bf1673  max_crops: 1274  crop_mode: overlap-and-resize-c275  use_col_tokens: true76  prompt_type: uber_model77  system_prompt_kind: demo_or_style78  message_formatting: role79  always_start_with_space: true80  multi_annotation_weighting: root_subsegments81  default_inference_len: 6582  overlap_margins:83  - 484  - 485  pad_value: 0.086  image_padding_embed: pad_and_partial_pad87  fix_image_padding: true88  vit_layers:89  - -290  - -991  image_pooling_h: 292  image_pooling_w: 293  image_pooling_2d: attention_meanq94  image_projector: mlp95  image_feature_dropout: 0.096  initializer_range: 0.0297  normalize_input_embeds: false98  use_position_ids: true99  head_dim: null100  action_tokenizer:101    identifier: physical-intelligence/fast102    tokenizer_dir: null103  action_dim: 7104  horizon: 8105  tokenizer:106    identifier: Qwen/Qwen2-7B107    tokenizer_dir: null108  pad_tokenizer: true109  moe_num_experts: 8110  moe_top_k: 2111  moe_mlp_impl: sparse112  moe_log_expert_assignment: false113  moe_shared_expert: false114  moe_lbl_in_fp32: false115  moe_interleave: false116  moe_loss_weight: 0.1117  moe_zloss_weight: null118  moe_dropless: true119  moe_capacity_factor: 1.25120  action_head: l1_regression121  num_diffusion_steps: 1000122  num_diffusion_inference_steps: 30123  use_proprio: true124  action_head_dit_hidden_size: 1152125  action_head_dit_depth: 28126  action_head_dit_num_heads: 16127  llm_causal_attention: false128  action_use_left_eef: true129  action_use_mobile_base: false130allow_resume: false131ft_llm: true132ft_vit: false133ft_connector: false134ft_embedding: lm_head135lora: false136use_lora: true137lora_rank: 8138lora_llm: false139lora_vit: false140lora_connector: false141early_exit: false142train_exit_random_layer: false143optimizer:144  name: adamw145  learning_rate: 0.0001146  weight_decay: 0.01147  betas:148  - 0.9149  - 0.95150  eps: 1.0e-05151  connector_learning_rate: 0.0002152  vit_learning_rate: 6.0e-06153  llm_learning_rate: 5.0e-05154  connector_weight_decay: 0.0155  vit_weight_decay: 0.0156  llm_weight_decay: 0.0157  connector_betas:158  - 0.9159  - 0.95160  vit_betas:161  - 0.9162  - 0.95163  llm_betas:164  - 0.9165  - 0.95166  connector_eps: 1.0e-06167  vit_eps: 1.0e-06168  llm_eps: 1.0e-06169  metrics_log_interval: 20170scheduler:171  name: multimodal172  units: steps173  t_warmup: 100174  t_max: null175  alpha_f: 0.1176  connector_t_warmup: 200177  vit_t_warmup: 2000178  llm_t_warmup: 2000179  grad_clip_warmup_steps: null180  grad_clip_warmup_factor: null181  warmup_min_lr: 0.0182data:183  dataset: vla_dataset_realworld184  mixture: null185  root_size_mixture: null186  split: train187  seed: 95818188  shuffle_messages: false189  pad: to_max190  sequence_length: 1600191  shuffle: true192  for_inference: false193  multi_modal: torch194  num_workers: 0195  drop_last: true196  pin_memory: true197  prefetch_factor: null198  persistent_workers: false199  timeout: 0200  rlds_dataset_name: libero_4_task_suites_no_noops201  rlds_data_root_dir: /mnt/data/zhangjian/dataset/Simulation/datasets--openvla--modified_libero_rlds202  use_wrist_image: true203  use_proprio: true204  rlds_shuffle_buffer_size: 100000205  rlds_traj_threads: 8206  rlds_read_threads: 8207  lerobot_episode_index_start: null208  lerobot_episode_index_end: null209restore_dataloader: true210fast_forward_batches: null211evaluators:212- label: val213  data:214    dataset: vla_dataset_realworld215    mixture: null216    root_size_mixture: null217    split: validation218    seed: null219    shuffle_messages: false220    pad: to_max221    sequence_length: 1600222    shuffle: false223    for_inference: false224    multi_modal: torch225    num_workers: 0226    drop_last: true227    pin_memory: true228    prefetch_factor: null229    persistent_workers: true230    timeout: 0231    rlds_dataset_name: libero_4_task_suites_no_noops232    rlds_data_root_dir: /mnt/data/zhangjian/dataset/Simulation/datasets--openvla--modified_libero_rlds233    use_wrist_image: true234    use_proprio: true235    rlds_shuffle_buffer_size: 256000236    rlds_traj_threads: 8237    rlds_read_threads: 8238    lerobot_episode_index_start: 353239    lerobot_episode_index_end: 765240  device_eval_batch_size: null241  subset_num_batches: 64242  max_examples: null243  max_new_tokens: 448244  mm_evaluator: null245  save_dir: null246  save_to_checkpoint_dir: false247  eval_name: null248  skip_if_metrics_cached: true249eval_interval: 0250inf_eval_interval: -1251inf_evaluators: []252save_folder: /vast/users/xiaodan/workspace/minghao.guo/warehouse_a1/ckpt/253remote_save_folder: null254canceled_check_interval: 50255save_interval: 500256save_interval_unsharded: 500257save_interval_ephemeral: null258save_interval_action_head: 500259save_num_checkpoints_to_keep: 1260save_num_unsharded_checkpoints_to_keep: 1261save_num_action_head_checkpoints_to_keep: 2262save_overwrite: true263force_save_unsharded: false264no_pre_train_checkpoint: true265initial_model_checkpoint: /vast/users/xiaodan/zhangjian/molmo_data/Molmo-7B-D-0924266load_model_config: null267checkpoint_dir: /vast/users/xiaodan/zhangjian/molmo_data/Molmo-7B-D-0924268load_path: null269load_path_sharded_checkpointer: null270reset_optimizer_state: false271reset_trainer_state: false272save_dataloader_state: false273reset_dataloader_state: false274keep_lr_on_load: true275sharded_checkpointer: torch_legacy276max_duration: 500000277global_train_batch_size: 126278device_train_batch_size: 15279device_train_microbatch_size: 16280device_eval_batch_size: 4281eval_subset_num_batches: -1282eval_on_load: false283device_inf_eval_batch_size: 16284inf_eval_subset_num_batches: -1285device_train_grad_accum: 0286max_grad_norm: 1.0287multi_component_grad_norm: true288batch_divisor: global_batch289max_grad_norm_ratio: null290precision: amp_bf16291wandb:292  project: a1-realworld293  entity: henryeap294  group: null295  name: glue_20251002_155411296  tags:297  - watching298  log_artifacts: false299  rank_zero_only: true300  log_interval: 1301speed_monitor:302  window_size: 20303  gpu_flops_available: null304console_log_interval: 1305gen1_gc_interval: 1306compile: null307fsdp:308  use_orig_params: true309  sharding_strategy: FULL_SHARD310  wrapping_strategy: by_block_and_size311  precision: float312  hybrid_sharding_num_model_replicas: null313softmax_auxiliary_loss: true314softmax_auxiliary_loss_scale: 0.0001315time_limit: null316extra_steps_after_cancel: 10317python_profiling: false318torch_profiling: false319stop_at: 500000320stop_after: null321activation_checkpointing: whole_layer322fused_loss: null323