Skip to content

vllm_omni.diffusion.worker.request_batch

Request-level batch abstraction for diffusion runner.

DiffusionRequestBatch dataclass

Request-level batch wrapping original diffusion requests.

Each :class:~vllm_omni.diffusion.request.OmniDiffusionRequest represents one logical diffusion request with one prompt. The scheduler and runner use this wrapper to present a compatible request batch to pipeline forward() methods without reintroducing list-shaped request payloads.

This is distinct from InputBatch (aliased as StepInputBatch), which manages step/tensor-level data for stepwise execution.

Parameters:

Name Type Description Default
requests list[OmniDiffusionRequest]

Independent diffusion requests scheduled together for request-mode execution.

 required

Attributes:

Name Type Description
requests list[OmniDiffusionRequest]

Original request objects in scheduler order.
num_reqs int

Number of requests in the batch.
request_ids list[str]

Request IDs in the same order as requests.
prompts list[OmniPromptType]

Prompt list assembled from each request's single prompt.
sampling_params_list list[OmniDiffusionSamplingParams]

Per-request sampling parameters in scheduler order. Request-batch pipelines read request-local values here.
sampling_params OmniDiffusionSamplingParams

Sampling parameters for single-request legacy paths.
request_id str

First request ID, kept as a compatibility convenience for code paths that handle a single-request batch.
kv_sender_info dict | None

KV-transfer metadata from the first request.

kv_sender_info property 

kv_sender_info: dict | None

num_reqs property 

num_reqs: int

prompts property 

prompts: list[OmniPromptType]

request_id property 

request_id: str

request_ids property 

request_ids: list[str]

requests instance-attribute 

requests: list[OmniDiffusionRequest]

sampling_params property 

sampling_params: OmniDiffusionSamplingParams

sampling_params_list property 

sampling_params_list: list[OmniDiffusionSamplingParams]

collate_prompt_field_map staticmethod 

collate_prompt_field_map(
    prompts: list[OmniPromptType],
    field_defaults: Mapping[str, Tensor | None],
    field_aliases: Mapping[str, Sequence[str]]
    | None = None,
) -> dict[str, Tensor | None]

collate_prompt_fields staticmethod 

collate_prompt_fields(
    prompts: list[OmniPromptType],
    name: str,
    default_tensor: Tensor | None,
) -> Tensor | None

collate_prompt_tensors staticmethod 

collate_prompt_tensors(
    tensors: list[Tensor | None],
    name: str,
    default_tensor: Tensor | None,
) -> Tensor | None

collate_request_generators 

collate_request_generators(
    num_outputs_per_prompt: int,
    default_generator: Generator | list[Generator] | None,
) -> Generator | list[Generator] | None

collate_request_tensors 

collate_request_tensors(
    attr: str, default_tensor: Tensor | None
) -> Tensor | None

collate_sampling_param_generators staticmethod 

collate_sampling_param_generators(
    sampling_params_list: list[Any],
    num_outputs_per_prompt: int,
    default_generator: Generator | list[Generator] | None,
) -> Generator | list[Generator] | None

collate_tensors staticmethod 

collate_tensors(
    tensors: list[Tensor | None],
    name: str,
    default_tensor: Tensor | None,
) -> Tensor | None

get 

get(request_id: str) -> OmniDiffusionRequest | None

get_prompt_field staticmethod 

get_prompt_field(prompt: OmniPromptType, name: str) -> Any

get_prompt_field_with_aliases staticmethod 

get_prompt_field_with_aliases(
    prompt: OmniPromptType, names: Sequence[str]
) -> Any

is_dummy_run 

is_dummy_run() -> bool

split_diffusion_output_by_request 

split_diffusion_output_by_request(
    result: DiffusionOutput,
    req: DiffusionRequestBatch,
    *,
    num_outputs_per_prompt: int,
) -> list[DiffusionOutput]

Split a batched DiffusionOutput into one output per request.