[ModelRunner V2][BugFix] Free all model refs on shutdown (#47483)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Nick Hill committed
Jul 3, 2026 at 03:32 UTC
276b837dc4d6a15ec7a82099dccd4c997eec916b
1 file changed
+4
vllm/v1/worker/gpu/model_runner.py
+4
@@ -1555,6 +1555,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
1555
if hasattr(self, "kv_cache_config"):
1556
del self.kv_cache_config
1557
free_before_shutdown(self.vllm_config)
1558
+ if hasattr(self, "model_state"):
1559
+ del self.model_state
1560
+ if getattr(self, "speculator", None) is not None:
1561
+ self.speculator = None
1562
if hasattr(self, "model"):
1563
del self.model
1564