Skip to content

vllm_omni.diffusion.diffusion_kv.manager ¶

DiffusionKVAdmissionError ¶

Bases: ValueError

A request-specific KV admission failure safe to return to the caller.

DiffusionKVCacheManager ¶

Atomic public-request facade over native vLLM KVCacheManager.

enable_prefix_caching `instance-attribute` ¶

enable_prefix_caching = enable_prefix_caching

hash_block_size `instance-attribute` ¶

hash_block_size = hash_block_size

max_model_len `instance-attribute` ¶

max_model_len = max_model_len

native_manager `instance-attribute` ¶

native_manager = KVCacheManager(
    kv_cache_config=kv_cache_config,
    max_model_len=max_model_len,
    scheduler_block_size=scheduler_block_size,
    hash_block_size=hash_block_size,
    max_in_flight_tokens=max_in_flight_tokens,
    enable_caching=enable_prefix_caching,
)

close ¶

close() -> None

free_request ¶

free_request(public_request_id: str) -> None

get_metadata ¶

get_metadata(public_request_id: str) -> DiffusionKVMetadata

has_request ¶

has_request(public_request_id: str) -> bool

publish_request ¶

publish_request(public_request_id: str) -> None

Publish successfully materialized, block-aligned stable prefixes.

reserve_request ¶

reserve_request(
    public_request_id: str,
    kv_requests: Sequence[DiffusionKVRequest],
) -> DiffusionKVMetadata | None

Reserve all CFG sequences or roll back the complete public request.