LogSumExp trick question_answering pipeline. (#17143)

* LogSumExp trick `question_answering` pipeline.

* Adding a failing test.

src/transformers/pipelines/question_answering.py

@@ -398,8 +398,11 @@ class QuestionAnsweringPipeline(ChunkPipeline):
             end_ = np.where(undesired_tokens_mask, -10000.0, end_)
 
             # Normalize logits and spans to retrieve the answer
-            start_ = np.exp(start_ - np.log(np.sum(np.exp(start_), axis=-1, keepdims=True)))
+            start_ = np.exp(start_ - start_.max(axis=-1, keepdims=True))
-            end_ = np.exp(end_ - np.log(np.sum(np.exp(end_), axis=-1, keepdims=True)))
+            start_ = start_ / start_.sum()
+
+            end_ = np.exp(end_ - end_.max(axis=-1, keepdims=True))
+            end_ = end_ / end_.sum()
 
             if handle_impossible_answer:
                 min_null_score = min(min_null_score, (start_[0, 0] * end_[0, 0]).item())

tests/pipelines/test_pipelines_question_answering.py

@@ -111,12 +111,47 @@ class QAPipelineTests(unittest.TestCase, metaclass=PipelineTestCaseMeta):
         question_answerer = pipeline(
             "question-answering", model="sshleifer/tiny-distilbert-base-cased-distilled-squad"
         )
+
         outputs = question_answerer(
             question="Where was HuggingFace founded ?", context="HuggingFace was founded in Paris."
         )
 
         self.assertEqual(nested_simplify(outputs), {"score": 0.01, "start": 0, "end": 11, "answer": "HuggingFace"})
 
+    @require_torch
+    def test_small_model_pt_softmax_trick(self):
+        question_answerer = pipeline(
+            "question-answering", model="sshleifer/tiny-distilbert-base-cased-distilled-squad"
+        )
+
+        real_postprocess = question_answerer.postprocess
+
+        # Tweak start and stop to make sure we encounter the softmax logits
+        # bug.
+        def ensure_large_logits_postprocess(
+            model_outputs,
+            top_k=1,
+            handle_impossible_answer=False,
+            max_answer_len=15,
+        ):
+            for output in model_outputs:
+                output["start"] = output["start"] * 1e6
+                output["end"] = output["end"] * 1e6
+            return real_postprocess(
+                model_outputs,
+                top_k=top_k,
+                handle_impossible_answer=handle_impossible_answer,
+                max_answer_len=max_answer_len,
+            )
+
+        question_answerer.postprocess = ensure_large_logits_postprocess
+
+        outputs = question_answerer(
+            question="Where was HuggingFace founded ?", context="HuggingFace was founded in Paris."
+        )
+
+        self.assertEqual(nested_simplify(outputs), {"score": 0.028, "start": 0, "end": 11, "answer": "HuggingFace"})
+
     @slow
     @require_torch
     def test_small_model_long_context_cls_slow(self):