Skip to content

vllm_omni.diffusion.request

DUMMY_DIFFUSION_REQUEST_ID module-attribute 

DUMMY_DIFFUSION_REQUEST_ID = 'dummy_req_id'

OmniDiffusionRequest dataclass

Input payload for a single diffusion request.

This dataclass contains the prompt and sampling parameters for the diffusion pipeline execution. It also contains a request_id for other components to trace this request and its outputs. The runner wraps one or more requests into a DiffusionRequestBatch before pipeline execution.

batch_compatibility_key class-attribute instance-attribute 

batch_compatibility_key: tuple[Any, ...] | None = None

diffusion_kv_requests class-attribute instance-attribute 

diffusion_kv_requests: (
    tuple[DiffusionKVRequest, ...] | None
) = None

kv_computed_tokens class-attribute instance-attribute 

kv_computed_tokens: tuple[int, ...] = ()

kv_recv_ms class-attribute instance-attribute 

kv_recv_ms: float = 0.0

kv_sender_info class-attribute instance-attribute 

kv_sender_info: dict[str, Any] | None = None

kv_transfer_params class-attribute instance-attribute 

kv_transfer_params: dict[str, Any] | None = None

payload_sender_info class-attribute instance-attribute 

payload_sender_info: dict[str, Any] | None = None

prepared_layout class-attribute instance-attribute 

prepared_layout: Any | None = None

prompt instance-attribute 

prompt: OmniPromptType

request_id instance-attribute 

request_id: str

sampling_params instance-attribute 

sampling_params: OmniDiffusionSamplingParams

scheduler_queue_wait_ms class-attribute instance-attribute 

scheduler_queue_wait_ms: float | None = None

use_step_execution class-attribute instance-attribute 

use_step_execution: bool = True

is_dummy_run 

is_dummy_run() -> bool

is_dummy_run_request_id classmethod 

is_dummy_run_request_id(request_id: str | None) -> bool

resolve_video_num_frames 

resolve_video_num_frames(
    num_frames: int | None,
    *,
    default_num_frames: int,
    is_dummy_run: bool,
) -> int

Resolve the shared image-model frame sentinel for a video pipeline.

OmniDiffusionSamplingParams defaults num_frames to one for image models, so an omitted video API field reaches model code as 1. Video pipelines with a different default must materialize it at their boundary. Startup profiling intentionally requests one frame, however, and must stay lightweight.