fix: the position of torch.cuda.empty_cache()

Author: rank-Yu

verl/utils/fsdp_utils.py

@@ -178,4 +178,5 @@ def summon_lora_params(fsdp_module) -> dict[str, torch.Tensor]:
         else param.detach().cpu()
         for name, param in lora_params.items()
     }
+    torch.cuda.empty_cache()
     return lora_params

verl/workers/sharding_manager/fsdp_vllm.py

@@ -112,7 +112,7 @@ def _collect_lora_params(self):
                 if isinstance(param, DTensor):
                     param = param.to(cuda_device).full_tensor()
                 lora_params[name] = param.detach().cpu()
-        torch.cuda.empty_cache()
+            torch.cuda.empty_cache()
         return lora_params
 
     def _sync_weight_to_vllm(self):