Copy evaluation engine GPU counts under the controller lock

Materialize M48 G30 from the frozen revision-003 patch and source map 86efe0f5. Preserve the reviewed patch boundary and keep unit-test migrations in the final GU operation.
This commit is contained in:
Tom
2026-10-01 15:43:36 +08:00
parent acb267b22f
commit ffd9bb40de
2 changed files with 5 additions and 4 deletions
+4 -3
View File
@@ -87,11 +87,12 @@ class InferenceControllerEvalFleet:
self.args = args
self._srv = srv
@property
def info(self) -> EvalFleetInfo:
async def info(self) -> EvalFleetInfo:
async with self._srv.context_lock:
engine_gpu_counts = list(self._srv.engine_gpu_counts)
return EvalFleetInfo(
router=HostAndPort(host=self._srv.router_ip, port=self._srv.router_port),
engine_gpu_counts=self._srv.engine_gpu_counts,
engine_gpu_counts=engine_gpu_counts,
)
async def pin(self, checkpoint_dir: str, weight_version: str) -> EvalFleetPin:
+1 -1
View File
@@ -304,7 +304,7 @@ class InferenceController:
@lock_exempt
async def get_eval_fleet_info(self) -> EvalFleetInfo | None:
return self._eval_fleet.info if self._eval_fleet is not None else None
return await self._eval_fleet.info() if self._eval_fleet is not None else None
@lock_exempt
async def pin_eval_fleet(self, checkpoint_dir: str, weight_version: str) -> EvalFleetPin: