[MRV2] Release weight offloader on shutdown (#57834)
Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
Taneem Ibrahim committed
Sep 22, 2026 at 15:01 UTC
bc162b3f9280c2dc05dfd5aa282d29551f8f6d33
2 files changed
+6
-5
vllm/model_executor/offloader/base.py
+5
-5
@@ -129,14 +129,14 @@ def get_offloader() -> BaseOffloader:
129
return _instance
130
131
132
-def set_offloader(instance: BaseOffloader) -> None:
133
- """Set the global offloader instance."""
132
+def set_offloader(instance: BaseOffloader | None) -> None:
133
+ """Set or reset the global offloader instance."""
134
global _instance
135
- _instance = instance
136
- if isinstance(instance, NoopOffloader):
135
+ _instance = NoopOffloader() if instance is None else instance
136
+ if isinstance(_instance, NoopOffloader):
137
logger.debug_once("Offloader set to NoopOffloader (no offloading).")
138
else:
139
- logger.info_once("Offloader set to %s", type(instance).__name__)
139
+ logger.info_once("Offloader set to %s", type(_instance).__name__)
140
141
142
def create_offloader(offload_config: "OffloadConfig") -> BaseOffloader:
vllm/v1/worker/gpu/model_runner.py
+1
@@ -2248,6 +2248,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
2248
torch.accelerator.synchronize()
2249
if self.aux_output_connector is not None:
2250
self.aux_output_connector.close()
2251
+ set_offloader(None)
2252
self.cudagraph_manager = None
2253
self.fast_prefill = None
2254
self.pooling_runner = None