diff --git a/modules_forge/initialization.py b/modules_forge/initialization.py
index 02914592..7d475a26 100644
--- a/modules_forge/initialization.py
+++ b/modules_forge/initialization.py
@@ -7,9 +7,19 @@ def initialize_forge():
                                          "Use this when you ara on MAC or have more than 20GB VRAM like RTX4096.")
 
     args_parser.args = args_parser.parser.parse_known_args()[0]
-    args_parser.args.always_offload_from_vram = not args_parser.args.disable_offload_from_vram
 
     import ldm_patched.modules.model_management as model_management
+
+    if args_parser.args.disable_offload_from_vram:
+        print('User disabled VRAM offload.')
+        model_management.ALWAYS_VRAM_OFFLOAD = False
+    elif model_management.total_vram > 20 * 1024:
+        print('Automatically disable VRAM offload since user have more than 20GB VRAM.')
+        model_management.ALWAYS_VRAM_OFFLOAD = False
+    else:
+        print('Always offload models from VRAM.')
+        model_management.ALWAYS_VRAM_OFFLOAD = True
+
     import torch
 
     device = model_management.get_torch_device()