Enable TruncationStrategy override for pipelines (#9432)

* Enable TruncationStrategy override for pipelines

* Update isort.

* Fixing test

* Fixing text_generation pipeline.

* Using same DummyTok as other PR  for easier merge later.

* Some more import guards.

* Remove bogus file.

* Do not pass `generate_kwargs` to `_parse_and_tokenize`.
@patrickvonplaten

* Removed DummyTok.

* Doc quality.

tests/test_pipelines_summarization.py

@@ -14,15 +14,72 @@
 
 import unittest
 
-from transformers import pipeline
+from transformers import AutoTokenizer, is_torch_available, pipeline
 from transformers.testing_utils import require_torch, slow, torch_device
+from transformers.tokenization_utils import TruncationStrategy
 
 from .test_pipelines_common import MonoInputPipelineCommonMixin
 
 
+if is_torch_available():
+    import torch
+
+    from transformers.models.bart import BartConfig, BartForConditionalGeneration
+
 DEFAULT_DEVICE_NUM = -1 if torch_device == "cpu" else 0
 
 
+class SimpleSummarizationPipelineTests(unittest.TestCase):
+    @require_torch
+    def test_input_too_long(self):
+        torch.manual_seed(0)
+        config = BartConfig(
+            vocab_size=257,
+            d_model=32,
+            encoder_layers=1,
+            decoder_layers=1,
+            encoder_ffn_dim=32,
+            decoder_ffn_dim=32,
+            # So any text > 4 should raise an exception
+            max_position_embeddings=4,
+            encoder_attention_heads=1,
+            decoder_attention_heads=1,
+            max_length=4,
+            min_length=1,
+        )
+        model = BartForConditionalGeneration(config)
+        # Bias output towards L
+        V, C = model.lm_head.weight.shape
+
+        bias = torch.zeros(V, requires_grad=True)
+        bias[76] = 10
+
+        model.lm_head.bias = torch.nn.Parameter(bias)
+
+        # # Generated with:
+        # import tempfile
+        # from tokenizers import Tokenizer, models
+        # from transformers import PreTrainedTokenizerFast
+        # model_max_length = 4
+        # vocab = [(chr(i), i) for i in range(256)]
+        # tokenizer = Tokenizer(models.Unigram(vocab))
+        # with tempfile.NamedTemporaryFile() as f:
+        #     tokenizer.save(f.name)
+        #     real_tokenizer = PreTrainedTokenizerFast(tokenizer_file=f.name, model_max_length=model_max_length)
+        # real_tokenizer._tokenizer.save("tokenizer.json")
+        # # + add missing config.json with albert as model_type
+        tokenizer = AutoTokenizer.from_pretrained("Narsil/small_summarization_test")
+        nlp = pipeline(task="summarization", model=model, tokenizer=tokenizer)
+
+        with self.assertLogs("transformers", level="WARNING"):
+            with self.assertRaises(IndexError):
+                _ = nlp("This is a test")
+
+        output = nlp("This is a test", truncation=TruncationStrategy.ONLY_FIRST)
+        # 2 is default BOS from Bart.
+        self.assertEqual(output, [{"summary_text": "\x02 L L L"}])
+
+
 class SummarizationPipelineTests(MonoInputPipelineCommonMixin, unittest.TestCase):
     pipeline_task = "summarization"
     pipeline_running_kwargs = {"num_beams": 2, "min_length": 2, "max_length": 5}