diff --git a/verl/utils/debug/performance.py b/verl/utils/debug/performance.py
index 615475a66a5..869fdb498e7 100644
--- a/verl/utils/debug/performance.py
+++ b/verl/utils/debug/performance.py
@@ -19,8 +19,11 @@
 
 def log_gpu_memory_usage(head: str, logger: logging.Logger = None, level=logging.DEBUG, rank: int = 0):
     if (not dist.is_initialized()) or (rank is None) or (dist.get_rank() == rank):
-        memory_allocated = torch.cuda.memory_allocated() / 1024**3
-        memory_reserved = torch.cuda.memory_reserved() / 1024**3
+        # take care of vllm sleep mem, see https://github.com/volcengine/verl/pull/1101
+        free_mem, total_mem = torch.cuda.mem_get_info()
+        vllm_sleep_mem = torch.cuda.memory_reserved() - (total_mem - free_mem)
+        memory_allocated = (torch.cuda.memory_allocated() - vllm_sleep_mem) / 1024**3
+        memory_reserved = (torch.cuda.memory_reserved() - vllm_sleep_mem) / 1024**3
 
         message = f'{head}, memory allocated (GB): {memory_allocated}, memory reserved (GB): {memory_reserved}'