Loose the criteria tolerance appropriately for Intel XPU devices (huggingface#12460)

* Loose the criteria tolerance appropriately for Intel XPU devices

Signed-off-by: Liu, Kaixuan <[email protected]>

* change back the atol value

Signed-off-by: Liu, Kaixuan <[email protected]>

* use expectations

Signed-off-by: Liu, Kaixuan <[email protected]>

* Update tests/pipelines/kandinsky2_2/test_kandinsky_controlnet.py

---------

Signed-off-by: Liu, Kaixuan <[email protected]>
Co-authored-by: Ilyas Moutawwakil <[email protected]>

Author: kaixuanliu

tests/pipelines/kandinsky2_2/test_kandinsky_controlnet.py

@@ -29,6 +29,7 @@
 )
 
 from ...testing_utils import (
+    Expectations,
     backend_empty_cache,
     enable_full_determinism,
     floats_tensor,
@@ -290,4 +291,11 @@ def test_kandinsky_controlnet(self):
 
         assert image.shape == (512, 512, 3)
         max_diff = numpy_cosine_similarity_distance(expected_image.flatten(), image.flatten())
-        assert max_diff < 2e-4
+        expected_max_diffs = Expectations(
+            {
+                ("xpu", 3): 2e-3,
+                ("cuda", 7): 2e-4,
+            }
+        )
+        expected_max_diff = expected_max_diffs.get_expectation()
+        assert max_diff < expected_max_diff