xm.save() should not set sync_xla_data=True when sync'ing.

Setting sync_xla_data=True performs tensor graph sync as if
it's a mark step, which triggers buffer aliasing to be performed.
However, it's not safe to do so unless all live tensors are
being sync'd.

Also fix torch_xla.utils.serialization.save() which has the same
issue.

This fixes #8422

test/test_input_output_aliases.py

@@ -1,3 +1,4 @@
+import os
 import sys
 
 import torch
@@ -162,6 +163,28 @@ def test_separate_graphs(self):
 
     self.assertEqual(t1.item(), 3)
 
+  def test_xm_save_no_aliasing(self):
+    """
+    Test that xm.save() does not perform aliasing.
+    """
+    xla_device = xm.xla_device()
+    t0 = torch.tensor([1], device=xla_device)
+    t1 = torch.tensor([2], device=xla_device)
+    xm.mark_step()
+
+    t2 = t0 + t1
+    t1.add_(1)
+
+    # Save the new value of t1 should not result in the old value
+    # being donated...
+    xm.save(t1, os.devnull)
+
+    # otherwise this mark_step could crash, or compute the wrong value
+    # for t2.
+    xm.mark_step()
+
+    self.assertEqual(t2.item(), 3)
+
 
 if __name__ == '__main__':
   test = unittest.main()

torch_xla/core/xla_model.py

@@ -1293,7 +1293,7 @@ def _maybe_convert_to_cpu(data: Any, convert: bool = True) -> ToXlaTensorArena:
 
   def convert_fn(tensors):
     torch_xla._XLAC._xla_sync_multi(
-        tensors, devices=[], wait=True, sync_xla_data=True)
+        tensors, devices=[], wait=True, sync_xla_data=False)
     if not convert:
       return tensors
     return torch_xla._XLAC._xla_get_cpu_tensors(tensors)

torch_xla/utils/serialization.py

@@ -25,7 +25,7 @@ def _rewrite_data(path, data, save_tensors):
 
   def convert_fn(tensors):
     torch_xla._XLAC._xla_sync_multi(
-        tensors, devices=[], wait=True, sync_xla_data=True)
+        tensors, devices=[], wait=True, sync_xla_data=False)
     rewritten_tensors = []
     for i, t in enumerate(tensors):
       if save_tensors: