Make interpolate_bilinear deterministic using decomposition (#101115)

Signed-off-by: Edward Z. Yang <ezyang@meta.com>

Pull Request resolved: https://github.com/pytorch/pytorch/pull/101115
Approved by: https://github.com/ngimel

test/test_torch.py

@@ -1439,6 +1439,24 @@ else:
             'upsample_bilinear2d_backward_out_cuda',
             torch.device(device).type == 'cuda')
 
+    @skipIfTorchInductor("aot-autograd issue")
+    def test_deterministic_interpolate_bilinear(self, device):
+        input = torch.randn(1, 2, 4, 4, device=device, requires_grad=True)
+        grad = None
+        with DeterministicGuard(True):
+            for _ in range(5):
+                res = torch.nn.functional.interpolate(
+                    input,
+                    size=12,
+                    mode='bilinear',
+                    align_corners=False)
+                res.backward(torch.ones_like(res))
+                if grad is None:
+                    grad = input.grad
+                else:
+                    self.assertEqual(grad, input.grad, atol=0, rtol=0)
+                input.grad = None
+
     @skipIfMps
     @skipIfTorchInductor("aot-autograd issue")
     def test_nondeterministic_alert_interpolate_bicubic(self, device):