Skip to content

vllm_omni.diffusion.models.gr00t ¶

Modules:

Name	Description
`configs`
`dataio`
`modeling`
`pipeline_gr00t`
`policy`

Gr00tN1d7Pipeline ¶

Bases: Module

GR00T N1.7 policy pipeline backed by vLLM-Omni's local GR00T port.

vLLM-Omni owns the serving integration: OpenPI observations arrive through sampling_params.extra_args["robot_obs"], this pipeline runs GR00T policy inference, and actions are returned through DiffusionOutput.output["actions"].

device `instance-attribute` ¶

device = 'cuda' if torch.cuda.is_available() else 'cpu'

embodiment_tag `instance-attribute` ¶

embodiment_tag = str(
    model_config.get("embodiment_tag")
    or "OXE_DROID_RELATIVE_EEF_RELATIVE_JOINT"
)

model_path `instance-attribute` ¶

model_path = od_config.model

policy `instance-attribute` ¶

policy = Gr00tPolicy(
    model_path=self.model_path,
    embodiment_tag=self.embodiment_tag,
    device=self.device,
    strict=self.strict,
)

strict `instance-attribute` ¶

strict = bool(model_config.get('strict', True))

weights_sources `property` ¶

weights_sources: tuple[Any, ...]

forward ¶

forward(
    req: OmniDiffusionRequest, **kwargs
) -> DiffusionOutput

load_weights ¶

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]

reset ¶

reset() -> dict[str, Any]