Merge pull request #3 from ModelTC/liangkh/partial-dev-memory

bugfix: fix empty cache

Author: yqyao

verl/workers/fsdp_workers.py

@@ -411,6 +411,8 @@ def init_model(self):
                 lr_scheduler=self.actor_lr_scheduler,
                 processing_class=self.processor if self.processor is not None else self.tokenizer)
 
+        torch.cuda.empty_cache()
+
     @register(dispatch_mode=Dispatch.DP_COMPUTE_PROTO)
     def update_actor(self, data: DataProto):
         # Support all hardwares