[WIP ] Proposal to address precision issues in CI (#4775)

* proposal for flaky tests

* clean up

Author: DN6

src/diffusers/utils/testing_utils.py

@@ -19,6 +19,7 @@
 import PIL.Image
 import PIL.ImageOps
 import requests
+from numpy.linalg import norm
 from packaging import version
 
 from .import_utils import (
@@ -72,6 +73,13 @@ def torch_all_close(a, b, *args, **kwargs):
     return True
 
 
+def numpy_cosine_similarity_distance(a, b):
+    similarity = np.dot(a, b) / (norm(a) * norm(b))
+    distance = 1.0 - similarity.mean()
+
+    return distance
+
+
 def print_tensor_test(tensor, filename="test_corrections.txt", expected_tensor_name="expected_slice"):
     test_name = os.environ.get("PYTEST_CURRENT_TEST")
     if not torch.is_tensor(tensor):

tests/pipelines/test_pipelines_common.py

@@ -22,7 +22,12 @@
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import logging
 from diffusers.utils.import_utils import is_accelerate_available, is_accelerate_version, is_xformers_available
-from diffusers.utils.testing_utils import CaptureLogger, require_torch, torch_device
+from diffusers.utils.testing_utils import (
+    CaptureLogger,
+    numpy_cosine_similarity_distance,
+    require_torch,
+    torch_device,
+)
 
 from ..others.test_utils import TOKEN, USER, is_staging_test
 
@@ -543,7 +548,7 @@ def test_float16_inference(self, expected_max_diff=1e-2):
         output = pipe(**self.get_dummy_inputs(torch_device))[0]
         output_fp16 = pipe_fp16(**self.get_dummy_inputs(torch_device))[0]
 
-        max_diff = np.abs(to_np(output) - to_np(output_fp16)).max()
+        max_diff = numpy_cosine_similarity_distance(to_np(output).flatten(), to_np(output_fp16).flatten())
         self.assertLess(max_diff, expected_max_diff, "The outputs of the fp16 and fp32 pipelines are too different.")
 
     @unittest.skipIf(torch_device != "cuda", reason="float16 requires CUDA")