[KV Connector][Offloading] Keep per-layer KV registration when canonical_layout is requested (#51688)

Signed-off-by: Itay Etelis <itay.etelis@ibm.com>
Co-authored-by: Itay Etelis <itay.etelis@ibm.com>
This commit is contained in:
Itay Etelis
2026-08-11 06:03:59 +03:00
committed by GitHub
co-authored by Itay Etelis
parent 419b51b385
commit 0f2ea973be
2 changed files with 31 additions and 1 deletions
@@ -49,7 +49,8 @@ from vllm.v1.request import Request
class OffloadingConnector(KVConnectorBase_V1, SupportsHMA):
@property
def prefer_cross_layer_blocks(self) -> bool:
return True
# Cross-layer slabs have no per-layer refs to certify canonical mappings
return not self._canonical_layout
@property
def requires_kv_delivery(self) -> bool:
@@ -66,6 +67,7 @@ class OffloadingConnector(KVConnectorBase_V1, SupportsHMA):
super().__init__(vllm_config, role, kv_cache_config)
offloading_config = build_offloading_config(vllm_config, kv_cache_config)
self._canonical_layout = offloading_config.canonical_layout
spec = OffloadingSpecFactory.create_spec(offloading_config)
self.connector_scheduler: OffloadingConnectorScheduler | None = None