enable QA bf16 pipeline (huggingface#34483)

* enable QA bf16 pipeline

* add tests

Author: jiqing-feng

src/transformers/pipelines/question_answering.py

@@ -540,8 +540,14 @@ def postprocess(
         min_null_score = 1000000  # large and positive
         answers = []
         for output in model_outputs:
-            start_ = output["start"]
-            end_ = output["end"]
+            if self.framework == "pt" and output["start"].dtype == torch.bfloat16:
+                start_ = output["start"].to(torch.float32)
+            else:
+                start_ = output["start"]
+            if self.framework == "pt" and output["start"].dtype == torch.bfloat16:
+                end_ = output["end"].to(torch.float32)
+            else:
+                end_ = output["end"]
             example = output["example"]
             p_mask = output["p_mask"]
             attention_mask = (

tests/pipelines/test_pipelines_question_answering.py

@@ -27,13 +27,18 @@
 from transformers.testing_utils import (
     compare_pipeline_output_to_hub_spec,
     is_pipeline_test,
+    is_torch_available,
     nested_simplify,
     require_tf,
     require_torch,
     require_torch_or_tf,
     slow,
 )
 
+
+if is_torch_available():
+    import torch
+
 from .test_pipelines_common import ANY
 
 
@@ -165,6 +170,34 @@ def test_small_model_pt(self):
 
         self.assertEqual(nested_simplify(outputs), {"score": 0.01, "start": 0, "end": 11, "answer": "HuggingFace"})
 
+    @require_torch
+    def test_small_model_pt_fp16(self):
+        question_answerer = pipeline(
+            "question-answering",
+            model="sshleifer/tiny-distilbert-base-cased-distilled-squad",
+            torch_dtype=torch.float16,
+        )
+
+        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_bf16(self):
+        question_answerer = pipeline(
+            "question-answering",
+            model="sshleifer/tiny-distilbert-base-cased-distilled-squad",
+            torch_dtype=torch.bfloat16,
+        )
+
+        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_iterator(self):
         # https://github.com/huggingface/transformers/issues/18510