Merge branch 'transformers-4.47' of https://github.com/huggingface/optimum-intel into transformers-4.47

Author: IlyasMoutawwakil

optimum/exporters/openvino/convert.py

@@ -365,7 +365,7 @@ def export_pytorch(
     import torch
     from torch.utils._pytree import tree_map
 
-    from optimum.exporters.onnx.convert import check_dummy_inputs_are_allowed
+    from optimum.exporters.utils import check_dummy_inputs_are_allowed
 
     logger.info(f"Using framework PyTorch: {torch.__version__}")
     output = Path(output)

optimum/exporters/openvino/model_configs.py

@@ -1783,6 +1783,23 @@ def generate(self, input_name: str, framework: str = "pt", int_dtype: str = "int
         return super().generate(input_name, framework, int_dtype, float_dtype)
 
 
+class DummyUnetVisionInputGenerator(DummyVisionInputGenerator):
+    def generate(self, input_name: str, framework: str = "pt", int_dtype: str = "int64", float_dtype: str = "fp32"):
+        if input_name not in ["sample", "latent_sample"]:
+            return super().generate(input_name, framework, int_dtype, float_dtype)
+        # add height and width discount for enable any resolution generation
+        return self.random_float_tensor(
+            shape=[self.batch_size, self.num_channels, self.height - 1, self.width - 1],
+            framework=framework,
+            dtype=float_dtype,
+        )
+
+
+@register_in_tasks_manager("unet", *["semantic-segmentation"], library_name="diffusers")
+class UnetOpenVINOConfig(UNetOnnxConfig):
+    DUMMY_INPUT_GENERATOR_CLASSES = (DummyUnetVisionInputGenerator,) + UNetOnnxConfig.DUMMY_INPUT_GENERATOR_CLASSES[1:]
+
+
 @register_in_tasks_manager("sd3-transformer", *["semantic-segmentation"], library_name="diffusers")
 class SD3TransformerOpenVINOConfig(UNetOnnxConfig):
     DUMMY_INPUT_GENERATOR_CLASSES = (

optimum/intel/openvino/quantization.py

@@ -523,7 +523,7 @@ def _quantize_torchmodel(
 
         quantization_config = ov_config.quantization_config
         if isinstance(quantization_config, OVWeightQuantizationConfig):
-            from optimum.exporters.onnx.convert import check_dummy_inputs_are_allowed
+            from optimum.exporters.utils import check_dummy_inputs_are_allowed
 
             if stateful:
                 # patch model before weight compression

tests/openvino/test_diffusion.py

@@ -144,6 +144,17 @@ def test_compare_to_diffusers_pipeline(self, model_arch: str):
 
             np.testing.assert_allclose(ov_output, diffusers_output, atol=6e-3, rtol=1e-2)
 
+        # test on inputs nondivisible on 64
+        height, width, batch_size = 96, 96, 1
+
+        for output_type in ["latent", "np", "pt"]:
+            inputs["output_type"] = output_type
+
+            ov_output = ov_pipeline(**inputs, generator=get_generator("pt", SEED)).images
+            diffusers_output = diffusers_pipeline(**inputs, generator=get_generator("pt", SEED)).images
+
+            np.testing.assert_allclose(ov_output, diffusers_output, atol=6e-3, rtol=1e-2)
+
     @parameterized.expand(CALLBACK_SUPPORT_ARCHITECTURES)
     @require_diffusers
     def test_callback(self, model_arch: str):
@@ -541,6 +552,20 @@ def test_compare_to_diffusers_pipeline(self, model_arch: str):
 
             np.testing.assert_allclose(ov_output, diffusers_output, atol=6e-3, rtol=1e-2)
 
+        # test generation when input resolution nondevisible on 64
+        height, width, batch_size = 96, 96, 1
+
+        inputs = self.generate_inputs(height=height, width=width, batch_size=batch_size, model_type=model_arch)
+
+        for output_type in ["latent", "np", "pt"]:
+            print(output_type)
+            inputs["output_type"] = output_type
+
+            ov_output = ov_pipeline(**inputs, generator=get_generator("pt", SEED)).images
+            diffusers_output = diffusers_pipeline(**inputs, generator=get_generator("pt", SEED)).images
+
+            np.testing.assert_allclose(ov_output, diffusers_output, atol=6e-3, rtol=1e-2)
+
     @parameterized.expand(SUPPORTED_ARCHITECTURES)
     @require_diffusers
     def test_image_reproducibility(self, model_arch: str):
@@ -777,6 +802,18 @@ def test_compare_to_diffusers_pipeline(self, model_arch: str):
 
             np.testing.assert_allclose(ov_output, diffusers_output, atol=6e-3, rtol=1e-2)
 
+        # test generation when input resolution nondevisible on 64
+        height, width, batch_size = 96, 96, 1
+        inputs = self.generate_inputs(height=height, width=width, batch_size=batch_size)
+
+        for output_type in ["latent", "np", "pt"]:
+            inputs["output_type"] = output_type
+
+            ov_output = ov_pipeline(**inputs, generator=get_generator("pt", SEED)).images
+            diffusers_output = diffusers_pipeline(**inputs, generator=get_generator("pt", SEED)).images
+
+            np.testing.assert_allclose(ov_output, diffusers_output, atol=6e-3, rtol=1e-2)
+
     @parameterized.expand(SUPPORTED_ARCHITECTURES)
     @require_diffusers
     def test_image_reproducibility(self, model_arch: str):