Use f32 inference for some OpenVINO stable diffusion/training tests (h…

…uggingface#518)

* Use f32 precision for compare-to-diffusers tests

* Add f32 ov_config to training/stable diffusion tests

tests/openvino/test_stable_diffusion.py

@@ -55,6 +55,9 @@
 from optimum.utils.import_utils import _diffusers_version
 
 
+F32_CONFIG = {"INFERENCE_PRECISION_HINT": "f32"}
+
+
 def _generate_inputs(batch_size=1):
     inputs = {
         "prompt": ["sailing ship in storm by Leonardo da Vinci"] * batch_size,
@@ -170,7 +173,7 @@ class OVStableDiffusionImg2ImgPipelineTest(OVStableDiffusionPipelineBaseTest):
     @parameterized.expand(SUPPORTED_ARCHITECTURES)
     def test_compare_diffusers_pipeline(self, model_arch: str):
         model_id = MODEL_NAMES[model_arch]
-        pipeline = self.MODEL_CLASS.from_pretrained(model_id, export=True)
+        pipeline = self.MODEL_CLASS.from_pretrained(model_id, export=True, ov_config=F32_CONFIG)
         height, width, batch_size = 128, 128, 1
         inputs = self.generate_inputs(height=height, width=width, batch_size=batch_size)
         inputs["prompt"] = "A painting of a squirrel eating a burger"
@@ -208,7 +211,7 @@ class OVStableDiffusionPipelineTest(unittest.TestCase):
     @parameterized.expand(SUPPORTED_ARCHITECTURES)
     def test_compare_to_diffusers(self, model_arch: str):
         model_id = MODEL_NAMES[model_arch]
-        ov_pipeline = self.MODEL_CLASS.from_pretrained(model_id, export=True)
+        ov_pipeline = self.MODEL_CLASS.from_pretrained(model_id, export=True, ov_config=F32_CONFIG)
         self.assertIsInstance(ov_pipeline.text_encoder, OVModelTextEncoder)
         self.assertIsInstance(ov_pipeline.vae_encoder, OVModelVaeEncoder)
         self.assertIsInstance(ov_pipeline.vae_decoder, OVModelVaeDecoder)
@@ -301,7 +304,7 @@ class OVStableDiffusionInpaintPipelineTest(OVStableDiffusionPipelineBaseTest):
     @parameterized.expand(SUPPORTED_ARCHITECTURES)
     def test_compare_diffusers_pipeline(self, model_arch: str):
         model_id = MODEL_NAMES[model_arch]
-        pipeline = self.MODEL_CLASS.from_pretrained(model_id, export=True)
+        pipeline = self.MODEL_CLASS.from_pretrained(model_id, export=True, ov_config=F32_CONFIG)
         batch_size, num_images, height, width = 1, 1, 64, 64
         latents = pipeline.prepare_latents(
             batch_size * num_images,
@@ -361,7 +364,7 @@ class OVtableDiffusionXLPipelineTest(unittest.TestCase):
 
     @parameterized.expand(SUPPORTED_ARCHITECTURES)
     def test_compare_to_diffusers(self, model_arch: str):
-        ov_pipeline = self.MODEL_CLASS.from_pretrained(MODEL_NAMES[model_arch], export=True)
+        ov_pipeline = self.MODEL_CLASS.from_pretrained(MODEL_NAMES[model_arch], export=True, ov_config=F32_CONFIG)
         self.assertIsInstance(ov_pipeline.text_encoder, OVModelTextEncoder)
         self.assertIsInstance(ov_pipeline.text_encoder_2, OVModelTextEncoder)
         self.assertIsInstance(ov_pipeline.vae_encoder, OVModelVaeEncoder)
@@ -447,11 +450,11 @@ class OVStableDiffusionXLImg2ImgPipelineTest(unittest.TestCase):
 
     def test_inference(self):
         model_id = "hf-internal-testing/tiny-stable-diffusion-xl-pipe"
-        pipeline = self.MODEL_CLASS.from_pretrained(model_id)
+        pipeline = self.MODEL_CLASS.from_pretrained(model_id, ov_config=F32_CONFIG)
 
         with tempfile.TemporaryDirectory() as tmp_dir:
             pipeline.save_pretrained(tmp_dir)
-            pipeline = self.MODEL_CLASS.from_pretrained(tmp_dir)
+            pipeline = self.MODEL_CLASS.from_pretrained(tmp_dir, ov_config=F32_CONFIG)
 
         batch_size, height, width = 1, 128, 128
         inputs = self.generate_inputs(height=height, width=width, batch_size=batch_size)
@@ -488,7 +491,7 @@ class OVLatentConsistencyModelPipelineTest(unittest.TestCase):
     @parameterized.expand(SUPPORTED_ARCHITECTURES)
     @unittest.skipIf(parse(_diffusers_version) <= Version("0.21.4"), "not supported with this diffusers version")
     def test_compare_to_diffusers(self, model_arch: str):
-        ov_pipeline = self.MODEL_CLASS.from_pretrained(MODEL_NAMES[model_arch], export=True)
+        ov_pipeline = self.MODEL_CLASS.from_pretrained(MODEL_NAMES[model_arch], export=True, ov_config=F32_CONFIG)
         self.assertIsInstance(ov_pipeline.text_encoder, OVModelTextEncoder)
         self.assertIsInstance(ov_pipeline.vae_encoder, OVModelVaeEncoder)
         self.assertIsInstance(ov_pipeline.vae_decoder, OVModelVaeDecoder)

tests/openvino/test_training.py

@@ -56,6 +56,9 @@
 from optimum.intel.openvino.utils import OV_XML_FILE_NAME
 
 
+F32_CONFIG = {"INFERENCE_PRECISION_HINT": "f32"}
+
+
 def initialize_movement_sparsifier_parameters_by_sparsity(
     movement_controller: MovementSparsityController,
     sparsity: float = 0.95,
@@ -250,7 +253,7 @@ def get_ov_config(self, nncf_compression_config: Union[List[Dict], Dict, None])
 
     def get_ov_model(self, model_id=None) -> OVModel:
         model_id = model_id or self.output_dir
-        return self.ovmodel_cls.from_pretrained(model_id)
+        return self.ovmodel_cls.from_pretrained(model_id, ov_config=F32_CONFIG)
 
     def get_nncf_config_with_overflow_fix_override(
         self, nncf_compression_config: Union[List[Dict], Dict, None], value: str = "enable"
@@ -628,7 +631,7 @@ def get_ov_model(self, model_id=None) -> OVModel:
         # image models, e.g. swin, may require a determined image size
         model_id = model_id or self.output_dir
         size = (self.feature_extractor.size["height"], self.feature_extractor.size["width"])
-        ovmodel = self.ovmodel_cls.from_pretrained(model_id, compile=False)
+        ovmodel = self.ovmodel_cls.from_pretrained(model_id, compile=False, ov_config=F32_CONFIG)
         # dynamic batch size for tiny-swin does not work in OpenVINO 2023.0
         batch_size = 1 if self.is_swin else -1
         ovmodel.reshape(batch_size, 3, *size)