config.json

{
  "action_dim": 7,
  "action_model_type": "DiT-L",
  "consolidate_type": "tome",
  "data_root_dir": "/training-results/haodong/Real/memoryvla_rlds",
  "dataloader_type": "stream",
  "ema_decay": 0.999,
  "fusion_type": "gate",
  "future_action_window_size": 15,
  "group_size": 16,
  "hf_token": "HF_TOKEN",
  "image_aug": false,
  "image_aug_mode": "spatial",
  "is_resume": false,
  "load_all_data_for_training": true,
  "mem_length": 256,
  "per_token_size": 256,
  "pretrained_checkpoint": "/training-results/haodong/MemoryVLA/code/pretrained/CogACT-Large/checkpoints/CogACT-Large.pt",
  "repeated_diffusion_steps": 4,
  "resume_epoch": 0,
  "resume_step": 0,
  "retain_optimizer_checkpoints": 2,
  "retrieval_layers": 2,
  "run_id": "piper-color-sorting-frozen16-ema-real21-20260823-053817",
  "run_id_note": null,
  "run_root_dir": "/training-results/haodong/MemoryVLA/runs/piper-color-sorting",
  "save_interval": 2000,
  "save_on_terminate": true,
  "seed": 42,
  "trackers": [
    "jsonl",
    "wandb"
  ],
  "update_fused": false,
  "use_ema": true,
  "use_timestep_pe": true,
  "vla": {
    "base_vlm": "prism-dinosiglip-224px+7b",
    "data_mix": "custom_finetuning",
    "enable_gradient_checkpointing": true,
    "enable_mixed_precision_training": true,
    "epochs": 100,
    "expected_world_size": 16,
    "freeze_llm_backbone": true,
    "freeze_vision_backbone": true,
    "global_batch_size": 32,
    "learning_rate": 2e-05,
    "lr_scheduler_type": "linear-warmup+cosine-decay",
    "max_grad_norm": 1.0,
    "max_steps": 20000,
    "per_device_batch_size": 2,
    "reduce_in_full_precision": true,
    "shuffle_buffer_size": 64,
    "train_strategy": "fsdp-full-shard",
    "type": "prism-dinosiglip-224px+oxe+diffusion",
    "unfreeze_last_llm_layer": false,
    "vla_id": "prism-dinosiglip-224px+oxe+diffusion",
    "warmup_ratio": 0.03,
    "weight_decay": 0.0
  },
  "wandb_entity": "spikingtransformer",
  "wandb_project": "Memory World Model"
}