Remove next sentence prediction from supported ONNX tasks (huggingface#17276)

Author: lewtun

src/transformers/onnx/features.py

@@ -45,7 +45,6 @@
         AutoModelForMaskedImageModeling,
         AutoModelForMaskedLM,
         AutoModelForMultipleChoice,
-        AutoModelForNextSentencePrediction,
         AutoModelForQuestionAnswering,
         AutoModelForSeq2SeqLM,
         AutoModelForSequenceClassification,
@@ -57,7 +56,6 @@
         TFAutoModelForCausalLM,
         TFAutoModelForMaskedLM,
         TFAutoModelForMultipleChoice,
-        TFAutoModelForNextSentencePrediction,
         TFAutoModelForQuestionAnswering,
         TFAutoModelForSeq2SeqLM,
         TFAutoModelForSequenceClassification,
@@ -112,7 +110,6 @@ class FeaturesManager:
             "question-answering": AutoModelForQuestionAnswering,
             "image-classification": AutoModelForImageClassification,
             "masked-im": AutoModelForMaskedImageModeling,
-            "next-sentence-prediction": AutoModelForNextSentencePrediction,
         }
     if is_tf_available():
         _TASKS_TO_TF_AUTOMODELS = {
@@ -124,7 +121,6 @@ class FeaturesManager:
             "token-classification": TFAutoModelForTokenClassification,
             "multiple-choice": TFAutoModelForMultipleChoice,
             "question-answering": TFAutoModelForQuestionAnswering,
-            "next-sentence-prediction": TFAutoModelForNextSentencePrediction,
         }
 
     # Set of model topologies we support associated to the features supported by each topology and the factory
@@ -159,7 +155,6 @@ class FeaturesManager:
             "multiple-choice",
             "token-classification",
             "question-answering",
-            "next-sentence-prediction",
             onnx_config_cls=BertOnnxConfig,
         ),
         "big-bird": supported_features_mapping(
@@ -326,7 +321,6 @@ class FeaturesManager:
         "mobilebert": supported_features_mapping(
             "default",
             "masked-lm",
-            "next-sentence-prediction",
             "sequence-classification",
             "multiple-choice",
             "token-classification",