vllm_omni.diffusion.models.internvla_a1.config ¶

action_expert_variant: str = 'qwen3_28l'

attn_implementation: str = 'eager'

chunk_size: int = 50

compile_mode: str = 'max-autotune'

compile_model: bool = False

device: str = 'cuda'

dtype: str = 'bfloat16'

empty_cameras: int = 0

enable_regional_compile: bool = False

enable_suffix_static_context_optimization: bool = False

freeze_vision_encoder: bool = False

gradient_checkpointing: bool = False

hidden_size: int = 128

image_history: int = 2

image_resolution: tuple[int, int] = (224, 224)

input_features: dict[str, Any] = field(default_factory=dict)

intermediate_size: int = 256

lambda_gen: float = 0.01

max_action_dim: int = 32

max_period: float = 4.0

max_state_dim: int = 32

min_period: float = 0.004

n_action_steps: int = 50

num_attention_heads: int = 4

num_cameras: int = 3

num_hidden_layers: int = 2

num_inference_steps: int = 10

output_features: dict[str, Any] = field(
    default_factory=dict
)

pixel_feature_dim: int = 48

qwen3_vl_variant: str = 'qwen3_vl_28l'

regional_compile_dynamic: bool = True

scale_factor: int = 8

time_sampling_beta_alpha: float = 1.5

time_sampling_beta_beta: float = 1.0

time_sampling_offset: float = 0.001

time_sampling_scale: float = 0.999

tokenizer_max_length: int = 48

train_expert_only: bool = False

train_vlm_only: bool = False

type: str = 'internvla_a1'

vocab_size: int = 256

action_mode: str = 'delta'

processor_model_name: str = DEFAULT_QWEN3_VL_MODEL