--- a
+++ b/occupy.py
@@ -0,0 +1,17 @@
+import os; import torch
+
+def occumpy_mem(cuda_device):
+    def check_mem(cuda_device):
+        devices_info = os.popen('"/usr/bin/nvidia-smi" --query-gpu=memory.total,memory.used --format=csv,nounits,noheader').read().strip().split("\n")
+        total, used = devices_info[int(cuda_device)].split(',')
+        return total,used
+    total, used = check_mem(cuda_device)
+    total = int(total)
+    used = int(used)
+    max_mem = int(total * 0.85)
+    block_mem = max_mem - used
+    x = torch.FloatTensor(256,1024,block_mem).to(torch.device(f"cuda:{cuda_device}"))
+    del x
+
+occumpy_mem('0')
+occumpy_mem('1')
\ No newline at end of file