[`ViTHybrid`] Fix `accelerate` slow tests (#20679)

* fix failing `accelerate` tests

* make fixup

* smaller values

* even lower

src/transformers/models/vit_hybrid/configuration_vit_hybrid.py

@@ -71,6 +71,8 @@ class ViTHybridConfig(PretrainedConfig):
             Whether to add a bias to the queries, keys and values.
         backbone_config (`Union[Dict[str, Any], PretrainedConfig]`, *optional*, defaults to `None`):
             The configuration of the backbone in a dictionary or the config object of the backbone.
+        backbone_featmap_shape (`List[int]`, *optional*, defaults to `[1, 1024, 24, 24]`):
+            Used only for the `hybrid` embedding type. The shape of the feature maps of the backbone.
 
     Example:
 
@@ -103,6 +105,7 @@ class ViTHybridConfig(PretrainedConfig):
         image_size=224,
         patch_size=1,
         num_channels=3,
+        backbone_featmap_shape=[1, 1024, 24, 24],
         qkv_bias=True,
         **kwargs
     ):
@@ -128,6 +131,7 @@ class ViTHybridConfig(PretrainedConfig):
                 backbone_config_class = BitConfig
             backbone_config = backbone_config_class(**backbone_config)
 
+        self.backbone_featmap_shape = backbone_featmap_shape
         self.backbone_config = backbone_config
         self.hidden_size = hidden_size
         self.num_hidden_layers = num_hidden_layers

src/transformers/models/vit_hybrid/modeling_vit_hybrid.py

@@ -166,11 +166,10 @@ class ViTHybridPatchEmbeddings(nn.Module):
         feature_dim = self.backbone.channels[-1]
 
         if feature_size is None:
-            dummy_image = torch.zeros(1, num_channels, image_size[0], image_size[1])
-            with torch.no_grad():
-                feature_map = self.backbone(dummy_image).feature_maps[-1]
-            feature_size = feature_map.shape[-2:]
-            feature_dim = feature_map.shape[1]
+            feature_map = config.backbone_featmap_shape
+
+            feature_size = feature_map[-2:]
+            feature_dim = feature_map[1]
         else:
             feature_size = (
                 feature_size if isinstance(feature_size, collections.abc.Iterable) else (feature_size, feature_size)

tests/models/vit_hybrid/test_modeling_vit_hybrid.py

@@ -19,7 +19,7 @@ import inspect
 import unittest
 
 from transformers import ViTHybridConfig
-from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+from transformers.testing_utils import require_accelerate, require_torch, require_vision, slow, torch_device
 from transformers.utils import cached_property, is_torch_available, is_vision_available
 
 from ...test_configuration_common import ConfigTester
@@ -57,6 +57,7 @@ class ViTHybridModelTester:
         attention_probs_dropout_prob=0.1,
         type_sequence_label_size=10,
         initializer_range=0.02,
+        backbone_featmap_shape=[1, 16, 4, 4],
         scope=None,
     ):
         self.parent = parent
@@ -76,6 +77,7 @@ class ViTHybridModelTester:
         self.type_sequence_label_size = type_sequence_label_size
         self.initializer_range = initializer_range
         self.scope = scope
+        self.backbone_featmap_shape = backbone_featmap_shape
 
         # in ViT hybrid, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
         # the number of patches is based on the feature map of the backbone, which by default uses an output stride
@@ -95,6 +97,16 @@ class ViTHybridModelTester:
         return config, pixel_values, labels
 
     def get_config(self):
+        backbone_config = {
+            "global_padding": "same",
+            "layer_type": "bottleneck",
+            "depths": [3, 4, 9],
+            "out_features": ["stage1", "stage2", "stage3"],
+            "embedding_dynamic_padding": True,
+            "hidden_sizes": [4, 8, 16, 32],
+            "num_groups": 2,
+        }
+
         return ViTHybridConfig(
             image_size=self.image_size,
             patch_size=self.patch_size,
@@ -108,6 +120,8 @@ class ViTHybridModelTester:
             attention_probs_dropout_prob=self.attention_probs_dropout_prob,
             is_decoder=False,
             initializer_range=self.initializer_range,
+            backbone_featmap_shape=self.backbone_featmap_shape,
+            backbone_config=backbone_config,
         )
 
     def create_and_check_model(self, config, pixel_values, labels):
@@ -229,3 +243,19 @@ class ViTModelIntegrationTest(unittest.TestCase):
         expected_slice = torch.tensor([-1.9090, -0.4993, -0.2389]).to(torch_device)
 
         self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
+
+    @slow
+    @require_accelerate
+    def test_accelerate_inference(self):
+        feature_extractor = ViTHybridImageProcessor.from_pretrained("google/vit-hybrid-base-bit-384")
+        model = ViTHybridForImageClassification.from_pretrained("google/vit-hybrid-base-bit-384", device_map="auto")
+
+        image = prepare_img()
+
+        inputs = feature_extractor(images=image, return_tensors="pt")
+        outputs = model(**inputs)
+        logits = outputs.logits
+        # model predicts one of the 1000 ImageNet classes
+        predicted_class_idx = logits.argmax(-1).item()
+
+        self.assertTrue(model.config.id2label[predicted_class_idx], "tabby, tabby cat")