diff --git a/torch/autograd/_functions/tensor.py b/torch/autograd/_functions/tensor.py
index 457a11bb41b..8e334d5dc40 100644
--- a/torch/autograd/_functions/tensor.py
+++ b/torch/autograd/_functions/tensor.py
@@ -1,4 +1,5 @@
 from functools import reduce
+import warnings
 import torch
 import torch._utils
 from ..function import Function
@@ -8,6 +9,8 @@ class Type(Function):
 
     @staticmethod
     def forward(ctx, i, dest_type):
+        warnings.warn("torch.autograd._functions.Type is deprecated as of PyTorch 2.1, please use "
+                      "torch.tensor.to(dtype=dtype) instead.")
         ctx.input_type = type(i)
         ctx.input_device = -1 if not i.is_cuda else i.get_device()
         return i.type(dest_type)
diff --git a/torch/autograd/graph.py b/torch/autograd/graph.py
index dc657e18827..182a5e8f56f 100644
--- a/torch/autograd/graph.py
+++ b/torch/autograd/graph.py
@@ -258,16 +258,17 @@ class save_on_cpu(saved_tensors_hooks):
         >>> # all intermediary tensors are released (deleted) after the call to backward
 
     """
-    def __init__(self, pin_memory=False):
+    def __init__(self, pin_memory=False, device_type="cuda"):
+        device_module = getattr(torch, device_type, torch.cuda)
+
         def pack_to_cpu(tensor):
             if not pin_memory:
                 return (tensor.device, tensor.cpu())
-
             packed = torch.empty(
                 tensor.size(),
                 dtype=tensor.dtype,
                 layout=tensor.layout,
-                pin_memory=(torch.cuda.is_available() and not tensor.is_sparse))
+                pin_memory=(device_module.is_available() and not tensor.is_sparse))
             packed.copy_(tensor)
             return (tensor.device, packed)