VJEPA2ACConfig
ModelConfigVJEPA2ACConfig(image_size: int = 256, patch_size: int = 16, tubelet_size: int = 2, num_frames: int = 64, in_channels: int = 3, encoder_dim: int = 1408, encoder_depth: int = 40, encoder_heads: int = 22, mlp_ratio: float = 48.0 / 11.0, predictor_mlp_ratio: float = 4.0, layer_norm_eps: float = 1e-06, qkv_bias: bool = True, init_std: float = 0.02, use_silu: bool = False, wide_silu: bool = True, use_rope: bool = True, rope_base: float = 10000.0, predictor_dim: int = 1024, predictor_depth: int = 24, predictor_heads: int = 16, action_dim: int = 7, state_dim: int = 7, extrinsics_dim: int = 6, use_extrinsics: bool = False, frame_causal: bool = True, normalize_targets: bool = True)Frozen architecture configuration for V-JEPA 2-AC.
Inputs use Lucid's video convention (B, T, C, H, W). The encoder
is applied independently to each frame, producing T action-model
steps — one per frame, which is the rate the released checkpoint was
post-trained at.
num_frames records the released clip length and sizes
token_grid; it does not constrain a call. One action-model
step is one frame — the released loop repeats each frame to fill the
encoder's tubelet — so a rollout is as long as the actions handed to
it and any frame count of two or more is valid.
normalize_targets layer-normalises the latents on both sides of
the transition loss, which the released post-training does.
Used by 3
Constructors
1__init__
→None__init__(image_size: int = 256, patch_size: int = 16, tubelet_size: int = 2, num_frames: int = 64, in_channels: int = 3, encoder_dim: int = 1408, encoder_depth: int = 40, encoder_heads: int = 22, mlp_ratio: float = 48.0 / 11.0, predictor_mlp_ratio: float = 4.0, layer_norm_eps: float = 1e-06, qkv_bias: bool = True, init_std: float = 0.02, use_silu: bool = False, wide_silu: bool = True, use_rope: bool = True, rope_base: float = 10000.0, predictor_dim: int = 1024, predictor_depth: int = 24, predictor_heads: int = 16, action_dim: int = 7, state_dim: int = 7, extrinsics_dim: int = 6, use_extrinsics: bool = False, frame_causal: bool = True, normalize_targets: bool = True)Properties
2Configured step, row and column counts.
One step is one frame, so the temporal entry is the clip length rather than a tubelet count.
Spatial tokens emitted for one encoded frame.