make narrow_tensor_symint DDE-free (pytorch#166379)

 pytorch#158081

Pull Request resolved: pytorch#166379
Approved by: https://github.com/Lucaskabela
ghstack dependencies: pytorch#166361

Author: laithsakka

aten/src/ATen/native/TensorShape.cpp

@@ -1784,8 +1784,8 @@ Tensor narrow_tensor_symint(
       start.dim() == 0 &&
           isIntegralType(start.scalar_type(), /*includeBool=*/false),
       "start must be an 0-dim integral Tensor.");
-  int64_t st = start.item<int64_t>();
-  return at::narrow_symint(self, dim, c10::SymInt(st), std::move(length));
+  c10::SymInt st = start.item().toSymInt();
+  return at::narrow_symint(self, dim, std::move(st), std::move(length));
 }
 
 std::

test/functorch/test_aotdispatch.py

@@ -8126,7 +8126,7 @@ def fn(x):
     xfail("corrcoef"),
     xfail("quantile"),
     xfail("nanquantile"),
-    xfail("narrow"),
+    skip("narrow"),
     xfail("istft"),
     xfail("linalg.eig"),
     skip("as_strided_scatter"),

test/test_dynamic_shapes.py

@@ -4452,6 +4452,19 @@ def test_narrow_unbacked_start_cpp_wrapper(self):
         """Test narrow with unbacked start with cpp_wrapper"""
         self.test_narrow_unbacked_start()
 
+    @torch._dynamo.config.patch(capture_scalar_outputs=True)
+    def test_narrow_with_tensor_start(self):
+        @torch.compile(backend="inductor", fullgraph=True)
+        def f(x, start, end):
+            return torch.narrow(x, 0, start, end)
+
+        x = torch.tensor(
+            [False], device="cuda:0" if torch.cuda.is_available() else "cpu"
+        )
+        start = torch.tensor(0)
+        res = f(x, start, 0)
+        self.assertEqual(res.shape, torch.Size([0]))
+
 
 instantiate_parametrized_tests(TestUnbacked)
 

test/test_proxy_tensor.py

@@ -1987,7 +1987,6 @@ def f(t):
 }
 
 only_fake_tensor_failures = {
-    xfail('narrow'),
     xfail('tensor_split'),
 }