From 8415bf2cdbdd00c1786abb766df568d5ae19e2c0 Mon Sep 17 00:00:00 2001 From: Martin Hickey Date: Mon, 11 May 2026 13:05:41 +0100 Subject: [PATCH] [kv_offload] Set offloading connector to prefer HND layout (#41928) Signed-off-by: Martin Hickey --- .../kv_transfer/kv_connector/v1/offloading_connector.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py b/vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py index 4ba59e2b1e..daa8042c8b 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py @@ -163,6 +163,10 @@ class OffloadingConnector(KVConnectorBase_V1, SupportsHMA): assert self.connector_scheduler is not None return self.connector_scheduler.take_events() + @classmethod + def get_required_kvcache_layout(cls, vllm_config: VllmConfig) -> str | None: + return "HND" + def get_kv_connector_stats(self) -> KVConnectorStats | None: if self.connector_worker is None: return None # We only emit stats from the worker-side