vllm_omni.diffusion.offloader.component_utils ¶
Shared component-plan helpers for diffusion offload backends.
clear_encoder_layerwise_state ¶
Clear encoder layerwise state after its backend hooks are removed.
encoder_component_type ¶
encoder_component_type(
name: str, plan: OffloadPlan | None
) -> str | None
Map a discovered encoder path to its public offload component.
get_encoder_block_groups ¶
get_encoder_block_groups(
module: Module,
name: str,
plan: OffloadPlan | None,
*,
strict: bool = False,
) -> list[ModuleList]
Resolve the streamable block lists declared for one encoder.
iter_streamable_dits ¶
iter_streamable_dits(
resolved: ResolvedOffloadPlan, device: device
) -> Iterator[tuple[ResolvedComponent, BlockStack]]
Yield selected DiTs that resolved to a streamable block stack.
move_non_block_state_to_device ¶
move_non_block_state_to_device(
module: Module,
block_groups: Sequence[Sequence[Module]],
device: device,
) -> None
Keep component state outside streamed block lists resident on device.
prepare_component ¶
prepare_component(
module: Module,
name: str,
*,
device: device,
stage_on_demand: bool,
blockwise: bool,
staged_components: list[Module],
) -> None
Stage a selected component or keep its non-streamed form resident.
prepare_pipeline_components ¶
prepare_pipeline_components(
resolved: ResolvedOffloadPlan,
*,
device: device,
staged_components: list[Module],
enable_encoder_blocks: Callable[
[ResolvedComponent], bool
],
) -> None
Apply the shared encoder/VAE/resident placement policy.