Enabling `imageGPT` auto feature extractor. (#16871)

* Enablign `imageGPT` auto feature extractor.
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Small updates.

* Update after rebase to use `input_ids` instead of `pixel_values`.
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

src/transformers/models/auto/feature_extraction_auto.py

@@ -50,6 +50,7 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
         ("flava", "FlavaFeatureExtractor"),
         ("glpn", "GLPNFeatureExtractor"),
         ("hubert", "Wav2Vec2FeatureExtractor"),
+        ("imagegpt", "ImageGPTFeatureExtractor"),
         ("layoutlmv2", "LayoutLMv2FeatureExtractor"),
         ("layoutlmv3", "LayoutLMv3FeatureExtractor"),
         ("maskformer", "MaskFormerFeatureExtractor"),

src/transformers/pipelines/base.py

@@ -75,14 +75,19 @@ def _pad(items, key, padding_value, padding_side):
         # Others include `attention_mask` etc...
         shape = items[0][key].shape
         dim = len(shape)
-        if dim == 4:
+        if key == "pixel_values":
             # This is probable image so padding shouldn't be necessary
             # B, C, H, W
             return torch.cat([item[key] for item in items], dim=0)
         max_length = max(item[key].shape[1] for item in items)
+        min_length = min(item[key].shape[1] for item in items)
         dtype = items[0][key].dtype
 
         if dim == 2:
+            if max_length == min_length:
+                # Bypass for `ImageGPT` which doesn't provide a padding value, yet
+                # we can consistently pad since the size should be matching
+                return torch.cat([item[key] for item in items], dim=0)
             tensor = torch.zeros((batch_size, max_length), dtype=dtype) + padding_value
         elif dim == 3:
             tensor = torch.zeros((batch_size, max_length, shape[-1]), dtype=dtype) + padding_value
@@ -146,7 +151,11 @@ def pad_collate_fn(tokenizer, feature_extractor):
         padded = {}
         for key in keys:
             if key in {"input_ids"}:
-                _padding_value = t_padding_value
+                # ImageGPT uses a feature extractor
+                if feature_extractor is not None:
+                    _padding_value = f_padding_value
+                else:
+                    _padding_value = t_padding_value
             elif key in {"input_values", "pixel_values", "input_features"}:
                 _padding_value = f_padding_value
             elif key in {"p_mask", "special_tokens_mask"}:

tests/models/imagegpt/test_modeling_imagegpt.py

@@ -171,6 +171,12 @@ class ImageGPTModelTester:
             reorder_and_upcast_attn=reorder_and_upcast_attn,
         )
 
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 513
+        config.max_position_embeddings = 1024
+        return config
+
     def prepare_config_and_inputs_for_decoder(self):
         (
             config,