Skip to content

vllm_omni.diffusion.diffusion_kv.metadata

Scheduler-to-Worker allocation payloads for diffusion KV.

Like native vLLM's SchedulerOutput DTOs, these classes carry trusted internal Scheduler results. Cache geometry and block-table validation stay with the native cache builders and Worker installation path.

DiffusionKVContextMetadata dataclass

Request-scoped Worker allocation for an independent K/V context.

context_id identifies this allocation within the enclosing :class:DiffusionKVMetadata.

block_ids instance-attribute 

block_ids: tuple[list[int], ...]

cache_role instance-attribute 

cache_role: str

context_id instance-attribute 

context_id: str

num_tokens instance-attribute 

num_tokens: int

DiffusionKVMetadata dataclass

Scheduler allocation sent with a newly scheduled public request.

Independent contexts live once at request scope so multiple sequences can reference the same allocation.

allocation_generation instance-attribute 

allocation_generation: int

contexts class-attribute instance-attribute 

contexts: tuple[DiffusionKVContextMetadata, ...] = ()

request_id instance-attribute 

request_id: str

sequences instance-attribute 

sequences: tuple[DiffusionKVSequenceMetadata, ...]

DiffusionKVSequenceMetadata dataclass

Worker allocation for one Scheduler-owned diffusion sequence.

context_ids reference request-scoped context allocations from the enclosing :class:DiffusionKVMetadata.

block_ids instance-attribute 

block_ids: tuple[list[int], ...]

cached_prefix_len class-attribute instance-attribute 

cached_prefix_len: int = 0

context_ids class-attribute instance-attribute 

context_ids: tuple[str, ...] = ()

num_computed_tokens class-attribute instance-attribute 

num_computed_tokens: int = 0

prefix_len instance-attribute 

prefix_len: int

seq_len instance-attribute 

seq_len: int

sequence_id instance-attribute 

sequence_id: int

target_len instance-attribute 

target_len: int