[RL] Return a failed weight update when the cache flush is refused (#43758)

This commit is contained in:
Liangsheng Yin
2026-10-11 21:57:58 -07:00
committed by GitHub
parent a297cd3da1
commit 2240049194
@@ -117,12 +117,20 @@ class SchedulerWeightUpdaterManager:
time.perf_counter() - t0, source
)
def flush_cache_after_weight_update(self, recv_req) -> None:
if recv_req.flush_cache:
flush_cache_success = self.flush_cache(
empty_cache=recv_req.torch_empty_cache
def flush_cache_after_weight_update(
self, recv_req, success: bool, message: str
) -> Tuple[bool, str]:
"""Flush the cache the old weights produced. A flush refused because
requests are still running or queued fails the update call."""
if recv_req.flush_cache and not self.flush_cache(
empty_cache=recv_req.torch_empty_cache
):
return False, (
"Weights were updated, but the cache was not flushed because "
"requests are still running or queued. Retract or finish them, "
"then call /flush_cache."
)
assert flush_cache_success, "Cache flush failed after updating weights"
return success, message
def record_weight_version_after_update(self, weight_version: Optional[str]) -> None:
self.scheduler.record_weight_version_change(new_version=weight_version)
@@ -141,12 +149,14 @@ class SchedulerWeightUpdaterManager:
if not success:
break
target_updated |= role == "target"
# the served weights changed even if a draft runner failed afterwards
if target_updated:
self.flush_cache_after_weight_update(recv_req)
if success:
self.record_weight_version_after_update(recv_req.weight_version)
else:
# the served weights changed even if a draft runner failed afterwards
if target_updated:
success, message = self.flush_cache_after_weight_update(
recv_req, success, message
)
if not success:
logger.error(message)
return UpdateWeightFromDiskReqOutput(
success=success, message=message, num_paused_requests=0
@@ -211,8 +221,10 @@ class SchedulerWeightUpdaterManager:
self._session = msgspec.structs.replace(
self._session, loaded_weights=True
)
self.flush_cache_after_weight_update(recv_req)
self.record_weight_version_after_update(recv_req.weight_version)
success, message = self.flush_cache_after_weight_update(
recv_req, success, message
)
return UpdateWeightsFromDistributedReqOutput(
success=success, message=message
)
@@ -241,9 +253,11 @@ class SchedulerWeightUpdaterManager:
self._session = msgspec.structs.replace(
self._session, loaded_weights=True
)
self.flush_cache_after_weight_update(recv_req)
self.record_weight_version_after_update(recv_req.weight_version)
else:
success, message = self.flush_cache_after_weight_update(
recv_req, success, message
)
if not success:
logger.error(message)
torch.distributed.barrier(group=self.tp_cpu_group)
return UpdateWeightsFromTensorReqOutput(success=success, message=message)
@@ -260,12 +274,14 @@ class SchedulerWeightUpdaterManager:
if not success:
break
target_updated |= role == "target"
# the served weights changed even if a draft runner failed afterwards
if target_updated:
self.flush_cache_after_weight_update(recv_req)
if success:
self.record_weight_version_after_update(recv_req.weight_version)
else:
# the served weights changed even if a draft runner failed afterwards
if target_updated:
success, message = self.flush_cache_after_weight_update(
recv_req, success, message
)
if not success:
logger.error(message)
torch.distributed.barrier(group=self.tp_cpu_group)
return UpdateWeightsFromIPCReqOutput(success=success, message=message)