mstar.model.vjepa2.config

V-JEPA 2 configuration dataclasses.

Flat config matching HuggingFace VJEPA2Config plus an optional action-conditioned predictor block. A single VJepa2Config.from_hf_config(dict) handles every open V-JEPA 2 checkpoint (vitl/h/g at 256, vitg at 384, and the AC variant).

Classes

VJepa2ACPredictorConfig([img_size, ...])	Configuration for the action-conditioned predictor (upstream V-JEPA 2-AC).
VJepa2Config([patch_size, crop_size, ...])	Top-level V-JEPA 2 config (shared across the HF and upstream ports).

class mstar.model.vjepa2.config.VJepa2ACPredictorConfig(img_size=(256, 256), patch_size=16, num_frames=64, tubelet_size=2, embed_dim=1408, predictor_embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4.0, qkv_bias=True, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.0, layer_norm_eps=1e-06, is_frame_causal=True, use_rope=True, action_embed_dim=7, use_extrinsics=False)

Bases: object

Configuration for the action-conditioned predictor (upstream V-JEPA 2-AC).

Mirrors vjepa2/src/models/ac_predictor.py defaults.

Parameters:

• img_size (tuple[int, int])

• patch_size (int)

• num_frames (int)

• tubelet_size (int)

• embed_dim (int)

• predictor_embed_dim (int)

• depth (int)

• num_heads (int)

• mlp_ratio (float)

• qkv_bias (bool)

• drop_rate (float)

• attn_drop_rate (float)

• drop_path_rate (float)

• layer_norm_eps (float)

• is_frame_causal (bool)

• use_rope (bool)

• action_embed_dim (int)

• use_extrinsics (bool)

action_embed_dim: int = 7

attn_drop_rate: float = 0.0

depth: int = 24

drop_path_rate: float = 0.0

drop_rate: float = 0.0

embed_dim: int = 1408

img_size: tuple[int, int] = (256, 256)

is_frame_causal: bool = True

layer_norm_eps: float = 1e-06

mlp_ratio: float = 4.0

num_frames: int = 64

num_heads: int = 16

patch_size: int = 16

predictor_embed_dim: int = 1024

qkv_bias: bool = True

tubelet_size: int = 2

use_extrinsics: bool = False

use_rope: bool = True

class mstar.model.vjepa2.config.VJepa2Config(patch_size=16, crop_size=256, frames_per_clip=64, tubelet_size=2, hidden_size=1024, in_chans=3, num_attention_heads=16, num_hidden_layers=24, drop_path_rate=0.0, mlp_ratio=4.0, layer_norm_eps=1e-06, qkv_bias=True, attention_probs_dropout_prob=0.0, hidden_act='gelu', pred_hidden_size=384, pred_num_attention_heads=12, pred_num_hidden_layers=12, pred_num_mask_tokens=10, pred_zero_init_mask_tokens=True, pred_mlp_ratio=4.0, predictor_kind='masked', ac_predictor=None, max_rollout_horizon=16, rollout_num_output_frames=2, rollout_frames_per_second=4, rollout_anticipation_seconds=1.0, mpc_cost_fn='l1')

Bases: object

Top-level V-JEPA 2 config (shared across the HF and upstream ports).

Field names match HF VJEPA2Config so a plain JSON round-trip works.

Parameters:

• patch_size (int)

• crop_size (int)

• frames_per_clip (int)

• tubelet_size (int)

• hidden_size (int)

• in_chans (int)

• num_attention_heads (int)

• num_hidden_layers (int)

• drop_path_rate (float)

• mlp_ratio (float)

• layer_norm_eps (float)

• qkv_bias (bool)

• attention_probs_dropout_prob (float)

• hidden_act (str)

• pred_hidden_size (int)

• pred_num_attention_heads (int)

• pred_num_hidden_layers (int)

• pred_num_mask_tokens (int)

• pred_zero_init_mask_tokens (bool)

• pred_mlp_ratio (float)

• predictor_kind (str)

• ac_predictor (VJepa2ACPredictorConfig | None)

• max_rollout_horizon (int)

• rollout_num_output_frames (int)

• rollout_frames_per_second (int)

• rollout_anticipation_seconds (float)

• mpc_cost_fn (str)

ac_predictor: VJepa2ACPredictorConfig | None = None

attention_probs_dropout_prob: float = 0.0

crop_size: int = 256

drop_path_rate: float = 0.0

frames_per_clip: int = 64

classmethod from_hf_config(config_dict)

Parameters:

config_dict (dict[str, Any])

Return type:

VJepa2Config

property grid_depth: int

property grid_size: int

hidden_act: str = 'gelu'

hidden_size: int = 1024

in_chans: int = 3

layer_norm_eps: float = 1e-06

max_rollout_horizon: int = 16

mlp_ratio: float = 4.0

mpc_cost_fn: str = 'l1'

num_attention_heads: int = 16

num_hidden_layers: int = 24

property num_patches: int

patch_size: int = 16

pred_hidden_size: int = 384

pred_mlp_ratio: float = 4.0

pred_num_attention_heads: int = 12

pred_num_hidden_layers: int = 12

pred_num_mask_tokens: int = 10

pred_zero_init_mask_tokens: bool = True

predictor_kind: str = 'masked'

qkv_bias: bool = True

rollout_anticipation_seconds: float = 1.0

rollout_frames_per_second: int = 4

rollout_num_output_frames: int = 2

tubelet_size: int = 2