Running a pipeline of `float16`. (#17637)

When we're preparing the tensors for CPU for postprocessing, we need
to upgrade the `float16` to `float32` since CPUs don't have instructions
for `[b]float16`.

src/transformers/pipelines/base.py

@@ -869,6 +869,8 @@ class Pipeline(_ScikitCompat):
         elif isinstance(inputs, tuple):
             return tuple([self._ensure_tensor_on_device(item, device) for item in inputs])
         elif isinstance(inputs, torch.Tensor):
+            if device == torch.device("cpu") and inputs.dtype in {torch.float16, torch.bfloat16}:
+                inputs = inputs.float()
             return inputs.to(device)
         else:
             return inputs

tests/pipelines/test_pipelines_fill_mask.py

@@ -16,7 +16,14 @@ import unittest
 
 from transformers import MODEL_FOR_MASKED_LM_MAPPING, TF_MODEL_FOR_MASKED_LM_MAPPING, FillMaskPipeline, pipeline
 from transformers.pipelines import PipelineException
-from transformers.testing_utils import is_pipeline_test, nested_simplify, require_tf, require_torch, slow
+from transformers.testing_utils import (
+    is_pipeline_test,
+    nested_simplify,
+    require_tf,
+    require_torch,
+    require_torch_gpu,
+    slow,
+)
 
 from .test_pipelines_common import ANY, PipelineTestCaseMeta
 
@@ -130,6 +137,19 @@ class FillMaskPipelineTests(unittest.TestCase, metaclass=PipelineTestCaseMeta):
             ],
         )
 
+    @require_torch_gpu
+    def test_fp16_casting(self):
+        pipe = pipeline("fill-mask", model="hf-internal-testing/tiny-random-distilbert", device=0, framework="pt")
+
+        # convert model to fp16
+        pipe.model.half()
+
+        response = pipe("Paris is the [MASK] of France.")
+        # We actually don't care about the result, we just want to make sure
+        # it works, meaning the float16 tensor got casted back to float32
+        # for postprocessing.
+        self.assertIsInstance(response, list)
+
     @slow
     @require_torch
     def test_large_model_pt(self):