diff --git a/comfy/model_management.py b/comfy/model_management.py
index 813b927be..98ffcc1f1 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -1227,6 +1227,9 @@ def cast_to_gathered(tensors, r, non_blocking=False, stream=None):
                 continue
             if comfy.memory_management.read_tensor_file_slice_into(tensor, dest_view):
                 continue
+            storage = tensor.untyped_storage()
+            if hasattr(storage, "_comfy_tensor_mmap_touched"):
+                storage._comfy_tensor_mmap_touched = True
             dest_view.copy_(tensor, non_blocking=non_blocking)
 
 
diff --git a/comfy/utils.py b/comfy/utils.py
index c30366f8e..795350830 100644
--- a/comfy/utils.py
+++ b/comfy/utils.py
@@ -104,9 +104,11 @@ def load_safetensors(ckpt):
                 #We are working with read-only RAM by design
                 warnings.filterwarnings("ignore", message="The given buffer is not writable")
                 tensor = torch.frombuffer(mv[start:end], dtype=_TYPES[info["dtype"]]).view(info["shape"])
-                setattr(tensor.untyped_storage(),
+                storage = tensor.untyped_storage()
+                setattr(storage,
                         "_comfy_tensor_file_slice",
                         comfy.memory_management.TensorFileSlice(f, threading.get_ident(), data_base_offset + start, end - start))
+                setattr(storage, "_comfy_tensor_mmap_touched", False)
                 sd[name] = tensor
 
     return sd, header.get("__metadata__", {}),