[RL] Release weight-transfer communicators without stopping inference (#61060)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Signed-off-by: fh3 <a@b>
This commit is contained in:
aoshen02
2026-10-11 21:15:14 +08:00
committed by GitHub
parent c85616efc6
commit bd5bbce6e4
22 changed files with 181 additions and 14 deletions
+9 -1
View File
@@ -60,7 +60,7 @@ class HTTPVLLMWeightSyncClient:
Mirrors `vllm/entrypoints/rl/online/api_router.py`:
`/init_weight_transfer_engine`, `/start_weight_update`, `/update_weights`,
`/finish_weight_update`.
`/finish_weight_update`, `/shutdown_weight_transfer_engine`.
"""
def __init__(self, base_url: str, timeout: float = 300) -> None:
@@ -78,6 +78,9 @@ class HTTPVLLMWeightSyncClient:
def init_weight_transfer_engine(self, init_info: dict[str, Any]) -> None:
self._post("init_weight_transfer_engine", {"init_info": init_info})
def shutdown_weight_transfer_engine(self) -> None:
self._post("shutdown_weight_transfer_engine")
def start_weight_update(self) -> None:
self._post("start_weight_update")
@@ -109,6 +112,11 @@ class RayVLLMWeightSyncClient:
request = WeightTransferInitRequest(init_info=init_info)
ray.get([h.init_weight_transfer_engine.remote(request) for h in self.handles])
def shutdown_weight_transfer_engine(self) -> None:
import ray
ray.get([h.shutdown_weight_transfer_engine.remote() for h in self.handles])
def start_weight_update(self) -> None:
import ray