{
 "model": "vjepa2-ac-vitg",
 "source": "/home/overseer2/workdir/jepa.cpp/models/vjepa2_ac/vjepa2-ac-vitg.pt",
 "dtype": "float32",
 "npy_dtype_note": "all tensors float32 C-order except frames_u8 (uint8) and top5_idx (int64)",
 "framework": {
  "torch": "2.13.0+cpu",
  "transformers": "5.16.1",
  "python": "3.12.13",
  "threads": 16,
  "cpu": "x86_64"
 },
 "created": "2026-09-01T21:43:50+0300",
 "source_url": "https://dl.fbaipublicfiles.com/vjepa2/vjepa2-ac-vitg.pt",
 "code": "https://github.com/facebookresearch/vjepa2 (hubconf vjepa2_ac_vit_giant, src/models/ac_predictor.py, notebooks/utils/mpc_utils.py + world_model_wrapper.py), local clone at /home/overseer2/workdir/jepa.cpp/tmp/vjepa2-src",
 "checkpoint": {
  "top_level_keys": [
   "encoder",
   "predictor",
   "opt",
   "scaler",
   "target_encoder",
   "epoch",
   "loss",
   "batch_size",
   "world_size",
   "lr"
  ],
  "encoder_key_used": "encoder",
  "key_cleanup": "strip 'module.' and 'backbone.'",
  "epoch": 315,
  "encoder_equals_target_encoder": true
 },
 "trajectory": {
  "file": "notebooks/franka_example_traj.npz",
  "frames": 2,
  "frame_size_hw": [
   256,
   256
  ],
  "states": [
   [
    0.5804390907287598,
    -0.0020212342496961355,
    0.24755726754665375,
    -3.066615581512451,
    0.030529797077178955,
    -1.9127511978149414,
    0.9969713687896729
   ],
   [
    0.6728084683418274,
    0.028965970501303673,
    0.33183684945106506,
    -3.0514373779296875,
    0.017709508538246155,
    -1.9124950170516968,
    0.9969713687896729
   ]
  ]
 },
 "hparams": {
  "embed_dim": 1408,
  "n_layer": 40,
  "n_head": 22,
  "ffn_dim": 6144,
  "patch_size": 16,
  "tubelet_size": 2,
  "img_size": 256,
  "n_frames": 64,
  "ln_eps": 1e-06,
  "act": "gelu_erf",
  "cls_token": false,
  "pos_type": "rope3d",
  "rope_theta": 10000.0,
  "tokens_per_frame": 256,
  "pred": {
   "kind": "ac",
   "embed_dim": 1024,
   "n_layer": 24,
   "n_head": 16,
   "head_dim": 64,
   "ffn_dim": 4096,
   "ln_eps": 1e-06,
   "action_dim": 7,
   "state_dim": 7,
   "out_dim": 1408,
   "grid_size": 16,
   "n_cond_tokens": 2,
   "cond_order": "action,state",
   "frame_causal": true,
   "use_extrinsics": false,
   "attn_mask_shape": [
    8256,
    8256
   ],
   "rope_dims_per_axis": [
    20,
    20,
    20
   ],
   "rope_grid_size": 16,
   "rope_freq_layout": "tiled"
  }
 },
 "preprocessing": {
  "description": "app/vjepa_droid/transforms.py::make_transforms(random_horizontal_flip=False, random_resize_scale=(1,1), random_resize_aspect_ratio=(1,1), crop_size=256): random_resized_crop degenerates to a centre crop of the largest square followed by torch F.interpolate(bilinear, align_corners=False, no antialias) to 256x256, then (x - 255*mean) / (255*std) on the CTHW float tensor. On the 256x256 Franka renders both the crop and the resize are the identity.",
  "resize": {
   "shortest_edge": 256,
   "resample": "bilinear",
   "antialias": false,
   "on_dtype": "float32",
   "reference_order": "centre-crop to the largest square, THEN resize to 256"
  },
  "center_crop": 256,
  "rescale": 0.00392156862745098,
  "mean": [
   0.485,
   0.456,
   0.406
  ],
  "std": [
   0.229,
   0.224,
   0.225
  ]
 },
 "world_model": {
  "encode": "one observation frame -> a 2-frame clip (the frame repeated along T, tubelet 2) -> 256 tokens after the encoder's final LayerNorm; then a NON-AFFINE LayerNorm over the feature dim (torch F.layer_norm default eps 1e-5). world_model_wrapper.py:42-52.",
  "predict": "predictor(context, actions, states) returns T*tpf rows; the world model keeps the LAST tpf (the next frame) and normalises them again before feeding them back. world_model_wrapper.py:56-64.",
  "state": "the pose of frame t; the pose after a step is compute_new_pose(pose, action) = (xyz + d_xyz, euler_xyz(R(d) @ R(pose)), clip(gripper + d_gripper, 0, 1)). mpc_utils.py:166-190.",
  "energy": "l1(a, b) = mean |a - b| over the flattened [tpf * D] final state and goal. mpc_utils.py:17-18 and the loss_fn of energy_landscape_example.ipynb.",
  "candidates": "4 fixed action sequences of 2 steps: the trajectory's ground-truth first action poses_to_diff(states[0], states[1]) and that action shifted by +0.05 along x, y and z, each repeated over the horizon."
 },
 "outputs": {
  "frames_u8": "the uint8 observation frame, repeated to the 2-frame clip, THWC",
  "input": "NCTHW float32 (the 2-frame clip of one observation frame)",
  "last_hidden_state": "encoder(input)[0]: [256, 1408] tokens after the final LN, h-major then w",
  "context": "F.layer_norm(last_hidden_state) -- the predictor's input",
  "goal": "context of the LAST trajectory frame",
  "action/state": "[7] the action driving the step and the pose at the frame",
  "pred_next": "predictor(context, action, state)[-256:] -- the next frame's latents, un-normalised",
  "pred_seq": "predictor over T=2 frames, ALL T*tpf rows (row block t predicts frame t+1)",
  "rollout": "[K, H, 256, 1408] normalised latents per candidate per step",
  "rollout_energy": "[K] l1(rollout[:, -1], goal)",
  "actions/states_seq": "[K, H, 7] the candidate actions and the poses compute_new_pose produced",
  "next_state_ref": "[K, 7] compute_new_pose(state0, actions[:, 0]) -- the jepa_ac_next_state check"
 },
 "timing_s": {
  "model_load": 8.863,
  "forward_total": 21.136,
  "forward_mean": 2.642,
  "wall_total": 21.193
 },
 "samples": [
  {
   "name": "frame0",
   "media": "franka_example_traj.npz[0]",
   "timing_s": {
    "preprocess_s": 0.0,
    "forward_s": 0.7271
   },
   "tensors": {
    "frames_u8": {
     "file": "frame0.frames_u8.npy",
     "shape": [
      2,
      256,
      256,
      3
     ],
     "dtype": "uint8",
     "layout": "THWC uint8 (the observation frame, repeated to the 2-frame clip)"
    },
    "input": {
     "file": "frame0.input.npy",
     "shape": [
      1,
      3,
      2,
      256,
      256
     ],
     "dtype": "float32",
     "layout": "NCTHW (T=2, the frame repeated)"
    },
    "last_hidden_state": {
     "file": "frame0.last_hidden_state.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D] h-major then w"
    },
    "pooled_mean": {
     "file": "frame0.pooled_mean.npy",
     "shape": [
      1408
     ],
     "dtype": "float32",
     "layout": "[D]"
    },
    "context": {
     "file": "frame0.context.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D] after the non-affine LayerNorm"
    }
   },
   "frame_index": 0,
   "frames": 2
  },
  {
   "name": "frame1",
   "media": "franka_example_traj.npz[1]",
   "timing_s": {
    "preprocess_s": 0.0,
    "forward_s": 0.6842
   },
   "tensors": {
    "frames_u8": {
     "file": "frame1.frames_u8.npy",
     "shape": [
      2,
      256,
      256,
      3
     ],
     "dtype": "uint8",
     "layout": "THWC uint8 (the observation frame, repeated to the 2-frame clip)"
    },
    "input": {
     "file": "frame1.input.npy",
     "shape": [
      1,
      3,
      2,
      256,
      256
     ],
     "dtype": "float32",
     "layout": "NCTHW (T=2, the frame repeated)"
    },
    "last_hidden_state": {
     "file": "frame1.last_hidden_state.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D] h-major then w"
    },
    "pooled_mean": {
     "file": "frame1.pooled_mean.npy",
     "shape": [
      1408
     ],
     "dtype": "float32",
     "layout": "[D]"
    },
    "context": {
     "file": "frame1.context.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D] after the non-affine LayerNorm"
    }
   },
   "frame_index": 1,
   "frames": 2
  },
  {
   "name": "goalframe",
   "media": "franka_example_traj.npz[1]",
   "timing_s": {
    "preprocess_s": 0.0,
    "forward_s": 0.6708
   },
   "tensors": {
    "frames_u8": {
     "file": "goalframe.frames_u8.npy",
     "shape": [
      2,
      256,
      256,
      3
     ],
     "dtype": "uint8",
     "layout": "THWC uint8 (the observation frame, repeated to the 2-frame clip)"
    },
    "input": {
     "file": "goalframe.input.npy",
     "shape": [
      1,
      3,
      2,
      256,
      256
     ],
     "dtype": "float32",
     "layout": "NCTHW (T=2, the frame repeated)"
    },
    "last_hidden_state": {
     "file": "goalframe.last_hidden_state.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D] h-major then w"
    },
    "pooled_mean": {
     "file": "goalframe.pooled_mean.npy",
     "shape": [
      1408
     ],
     "dtype": "float32",
     "layout": "[D]"
    },
    "context": {
     "file": "goalframe.context.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D] after the non-affine LayerNorm"
    }
   },
   "frame_index": 1,
   "frames": 2
  },
  {
   "name": "step1",
   "media": "franka_example_traj.npz[0]",
   "timing_s": {
    "preprocess_s": 0.0,
    "forward_s": 0.2757
   },
   "tensors": {
    "context": {
     "file": "step1.context.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D]"
    },
    "action": {
     "file": "step1.action.npy",
     "shape": [
      7
     ],
     "dtype": "float32",
     "layout": "[7]"
    },
    "state": {
     "file": "step1.state.npy",
     "shape": [
      7
     ],
     "dtype": "float32",
     "layout": "[7]"
    },
    "pred_next": {
     "file": "step1.pred_next.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D] predictor output, T=1"
    }
   },
   "frames": 1
  },
  {
   "name": "step2",
   "media": "franka_example_traj.npz[0]",
   "timing_s": {
    "preprocess_s": 0.0,
    "forward_s": 0.4742
   },
   "tensors": {
    "context_seq": {
     "file": "step2.context_seq.npy",
     "shape": [
      512,
      1408
     ],
     "dtype": "float32",
     "layout": "[2*256, D]"
    },
    "action_seq": {
     "file": "step2.action_seq.npy",
     "shape": [
      2,
      7
     ],
     "dtype": "float32",
     "layout": "[2, 7]"
    },
    "state_seq": {
     "file": "step2.state_seq.npy",
     "shape": [
      2,
      7
     ],
     "dtype": "float32",
     "layout": "[2, 7]"
    },
    "pred_seq": {
     "file": "step2.pred_seq.npy",
     "shape": [
      512,
      1408
     ],
     "dtype": "float32",
     "layout": "[2*256, D] all rows; block t predicts frame t+1"
    }
   },
   "frames": 2
  },
  {
   "name": "rollout",
   "media": "franka_example_traj.npz",
   "timing_s": {
    "preprocess_s": 0.0,
    "forward_s": 2.2028
   },
   "tensors": {
    "context": {
     "file": "rollout.context.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D] shared seed"
    },
    "goal": {
     "file": "rollout.goal.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D]"
    },
    "state0": {
     "file": "rollout.state0.npy",
     "shape": [
      7
     ],
     "dtype": "float32",
     "layout": "[7]"
    },
    "actions": {
     "file": "rollout.actions.npy",
     "shape": [
      4,
      2,
      7
     ],
     "dtype": "float32",
     "layout": "[4, 2, 7]"
    },
    "states_seq": {
     "file": "rollout.states_seq.npy",
     "shape": [
      4,
      2,
      7
     ],
     "dtype": "float32",
     "layout": "[4, 2, 7]"
    },
    "rollout": {
     "file": "rollout.rollout.npy",
     "shape": [
      4,
      2,
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[4, 2, 256, D]"
    },
    "rollout_energy": {
     "file": "rollout.rollout_energy.npy",
     "shape": [
      4
     ],
     "dtype": "float32",
     "layout": "[4]"
    },
    "next_state_ref": {
     "file": "rollout.next_state_ref.npy",
     "shape": [
      4,
      7
     ],
     "dtype": "float32",
     "layout": "[4, 7]"
    }
   },
   "candidates": 4,
   "horizon": 2
  },
  {
   "name": "rollout_seed2",
   "media": "franka_example_traj.npz",
   "timing_s": {
    "preprocess_s": 0.0,
    "forward_s": 3.5616
   },
   "tensors": {
    "context": {
     "file": "rollout_seed2.context.npy",
     "shape": [
      512,
      1408
     ],
     "dtype": "float32",
     "layout": "[2*256, D] two OBSERVED frames"
    },
    "goal": {
     "file": "rollout_seed2.goal.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D]"
    },
    "seed_actions": {
     "file": "rollout_seed2.seed_actions.npy",
     "shape": [
      1,
      7
     ],
     "dtype": "float32",
     "layout": "[n_seed-1, 7] the action between the observed frames"
    },
    "seed_states": {
     "file": "rollout_seed2.seed_states.npy",
     "shape": [
      2,
      7
     ],
     "dtype": "float32",
     "layout": "[2, 7]"
    },
    "actions": {
     "file": "rollout_seed2.actions.npy",
     "shape": [
      4,
      2,
      7
     ],
     "dtype": "float32",
     "layout": "[4, 2, 7]"
    },
    "states_seq": {
     "file": "rollout_seed2.states_seq.npy",
     "shape": [
      4,
      2,
      7
     ],
     "dtype": "float32",
     "layout": "[4, 2, 7]"
    },
    "rollout": {
     "file": "rollout_seed2.rollout.npy",
     "shape": [
      4,
      2,
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[4, 2, 256, D]"
    },
    "rollout_energy": {
     "file": "rollout_seed2.rollout_energy.npy",
     "shape": [
      4
     ],
     "dtype": "float32",
     "layout": "[4]"
    }
   },
   "candidates": 4,
   "horizon": 2,
   "n_seed": 2
  },
  {
   "name": "cem",
   "media": "franka_example_traj.npz",
   "timing_s": {
    "preprocess_s": 0.0,
    "forward_s": 12.5391
   },
   "tensors": {
    "context": {
     "file": "cem.context.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D]"
    },
    "goal": {
     "file": "cem.goal.npy",
     "shape": [
      256,
      1408
     ],
     "dtype": "float32",
     "layout": "[256, D]"
    },
    "state0": {
     "file": "cem.state0.npy",
     "shape": [
      7
     ],
     "dtype": "float32",
     "layout": "[7]"
    },
    "cem_noise": {
     "file": "cem.cem_noise.npy",
     "shape": [
      3,
      2,
      8,
      4
     ],
     "dtype": "float32",
     "layout": "[3, 2, 8, 4] every torch.randn draw, in Meta's order"
    },
    "cem_action": {
     "file": "cem.cem_action.npy",
     "shape": [
      2,
      7
     ],
     "dtype": "float32",
     "layout": "[2, 7] the plan cem() returned"
    }
   },
   "cem": {
    "samples": 8,
    "topk": 4,
    "cem_steps": 3,
    "rollout": 2,
    "maxnorm": 0.05,
    "momentum_mean": 0.15,
    "momentum_std": 0.75,
    "momentum_mean_gripper": 0.15,
    "momentum_std_gripper": 0.15
   }
  }
 ]
}