[Bugfix][KV Transfer][NIXL] Notify P node on pre-admission rejection to free stranded KV blocks (#41269)
This commit is contained in:
@@ -108,6 +108,20 @@ class EngineClient(ABC):
|
||||
"""
|
||||
...
|
||||
|
||||
@abstractmethod
|
||||
async def notify_kv_transfer_request_rejected(
|
||||
self,
|
||||
request_id: str,
|
||||
kv_transfer_params: dict[str, Any],
|
||||
*,
|
||||
data_parallel_rank: int | None = None,
|
||||
) -> None:
|
||||
"""Notify the engine that a KV-transfer request was rejected before
|
||||
engine admission, so connector-side cleanup can run (e.g. free
|
||||
prefill blocks pinned on the P node).
|
||||
"""
|
||||
...
|
||||
|
||||
@abstractmethod
|
||||
async def is_tracing_enabled(self) -> bool: ...
|
||||
|
||||
|
||||
Reference in New Issue
Block a user