Untie the lifetime of returns and the lifetime of module in pybindings

Summary: In cpp these are tied to not force a clone on everyone due to performance overhead and there are ways around it with customizing your own output buffer. In python which is just for testing really its fine to always clone and avoid unintuitive behavior

Reviewed By: dulinriley

Differential Revision: D47933661

fbshipit-source-id: 84178bfd5e9e055078b2d7cfd2c6739df2b3e1ea

Author: JacobSzwejbka

extension/pybindings/module.cpp

@@ -270,12 +270,13 @@ py::object pyFromEValue(const EValue& v, KeepAlive& keep_alive) {
     return py::cast(v.toBool());
   } else if (Tag::Tensor == v.tag) {
 #ifdef USE_ATEN_LIB
-    return py::cast(v.toTensor());
+    return py::cast(v.toTensor().clone());
 #else
-    return py::cast(
-        torch::util::atTensorFromETensor(
-            v.toTensor().unsafeGetTensorImpl(), keep_alive.sizes),
-        py::return_value_policy::reference);
+    // Clone so the outputs in python do not share a lifetime with the module
+    // object
+    return py::cast(torch::util::atTensorFromETensor(
+                        v.toTensor().unsafeGetTensorImpl(), keep_alive.sizes)
+                        .clone());
 #endif
   }
   ET_ASSERT_UNREACHABLE();

extension/pybindings/test/test.py

@@ -157,3 +157,14 @@ def test_multiple_entry(self):
 
         executorch_output2 = executorch_module.run_method("forward2", inputs)[0]
         self.assertTrue(torch.allclose(executorch_output2, torch.ones(2, 2) * 3))
+
+    def test_output_lifespan(self):
+        def lower_function_call():
+            program, inputs = create_program(ModuleMulti())
+            executorch_module = _load_for_executorch_from_buffer(program.buffer)
+
+            return executorch_module.forward(inputs)
+            # executorch_module is destructed here and all of its memory is freed
+
+        outputs = lower_function_call()
+        self.assertTrue(torch.allclose(outputs[0], torch.ones(2, 2) * 2))