[Diffusion] Implement paged AR↔DiT KV connector ("v1", reusing vllm's native mooncake) (#6310)

Signed-off-by: asukaqaq-s <1311722138@qq.com>
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Co-authored-by: asukaqaq-s <1311722138@qq.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
authored and GitHub committed 2026-09-06 05:41:54 +08:00
1 parent ea3cf99894
commit f26f9c3a7f
18 files changed
+517 -13

No files matched your search

+5
View File
@@ -12,6 +12,8 @@ from vllm_omni.diffusion.diffusion_kv.metadata import DiffusionKVMetadata
from vllm_omni.diffusion.request import OmniDiffusionRequest
if TYPE_CHECKING:
from vllm.distributed.kv_transfer.kv_connector.v1.base import KVConnectorMetadata
from vllm_omni.diffusion.diffusion_kv.request import DiffusionKVRequest
@@ -251,6 +253,9 @@ class DiffusionSchedulerOutput:
num_waiting_reqs: int
# next request to background-prefetch KV
kv_prefetch_job: KVPrefetchJob | None = None
# Opaque metadata emitted by a future Scheduler-role connector. PR0 keeps
# the input port but does not build or consume it.
kv_connector_metadata: KVConnectorMetadata | None = None
@cached_property
def scheduled_request_ids(self) -> list[str]: