enable low-precision pipeline (#31625)

* enable low-precision pipeline

* fix parameter for ASR

* reformat

* fix asr bug

* fix bug for zero-shot

* add dtype check

* rm useless comments

* add np.float16 check

* Update src/transformers/pipelines/image_classification.py

Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>

* Update src/transformers/pipelines/token_classification.py

Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>

* fix comments

* fix asr check

* make fixup

* No more need for is_torch_available()

---------

Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>
Co-authored-by: Matt <rocketknight1@gmail.com>

src/transformers/pipelines/automatic_speech_recognition.py

@@ -565,7 +565,10 @@ class AutomaticSpeechRecognitionPipeline(ChunkPipeline):
         key = "logits" if self.type == "ctc_with_lm" else "tokens"
         stride = None
         for outputs in model_outputs:
-            items = outputs[key].numpy()
+            if self.framework == "pt" and outputs[key].dtype in (torch.bfloat16, torch.float16):
+                items = outputs[key].to(torch.float32).numpy()
+            else:
+                items = outputs[key].numpy()
             stride = outputs.get("stride", None)
             if stride is not None and self.type in {"ctc", "ctc_with_lm"}:
                 total_n, left, right = stride

src/transformers/pipelines/token_classification.py

@@ -19,6 +19,8 @@ if is_tf_available():
 
     from ..models.auto.modeling_tf_auto import TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES
 if is_torch_available():
+    import torch
+
     from ..models.auto.modeling_auto import MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES
 
 
@@ -299,7 +301,11 @@ class TokenClassificationPipeline(ChunkPipeline):
             ignore_labels = ["O"]
         all_entities = []
         for model_outputs in all_outputs:
-            logits = model_outputs["logits"][0].numpy()
+            if self.framework == "pt" and model_outputs["logits"][0].dtype in (torch.bfloat16, torch.float16):
+                logits = model_outputs["logits"][0].to(torch.float32).numpy()
+            else:
+                logits = model_outputs["logits"][0].numpy()
+
             sentence = all_outputs[0]["sentence"]
             input_ids = model_outputs["input_ids"][0]
             offset_mapping = (

src/transformers/testing_utils.py

@@ -2143,7 +2143,7 @@ def nested_simplify(obj, decimals=3):
         return nested_simplify(obj.numpy().tolist())
     elif isinstance(obj, float):
         return round(obj, decimals)
-    elif isinstance(obj, (np.int32, np.float32)):
+    elif isinstance(obj, (np.int32, np.float32, np.float16)):
         return nested_simplify(obj.item(), decimals)
     else:
         raise Exception(f"Not supported: {type(obj)}")

tests/pipelines/test_pipelines_automatic_speech_recognition.py

@@ -167,6 +167,48 @@ class AutomaticSpeechRecognitionPipelineTests(unittest.TestCase):
         ):
             _ = speech_recognizer(waveform, return_timestamps="char")
 
+    @require_torch
+    def test_small_model_pt_fp16(self):
+        speech_recognizer = pipeline(
+            task="automatic-speech-recognition",
+            model="facebook/s2t-small-mustc-en-fr-st",
+            tokenizer="facebook/s2t-small-mustc-en-fr-st",
+            framework="pt",
+            torch_dtype=torch.float16,
+        )
+        waveform = np.tile(np.arange(1000, dtype=np.float32), 34)
+        output = speech_recognizer(waveform)
+        self.assertEqual(output, {"text": "(Applaudissements)"})
+        output = speech_recognizer(waveform, chunk_length_s=10)
+        self.assertEqual(output, {"text": "(Applaudissements)"})
+
+        # Non CTC models cannot use return_timestamps
+        with self.assertRaisesRegex(
+            ValueError, "^We cannot return_timestamps yet on non-CTC models apart from Whisper!$"
+        ):
+            _ = speech_recognizer(waveform, return_timestamps="char")
+
+    @require_torch
+    def test_small_model_pt_bf16(self):
+        speech_recognizer = pipeline(
+            task="automatic-speech-recognition",
+            model="facebook/s2t-small-mustc-en-fr-st",
+            tokenizer="facebook/s2t-small-mustc-en-fr-st",
+            framework="pt",
+            torch_dtype=torch.bfloat16,
+        )
+        waveform = np.tile(np.arange(1000, dtype=np.float32), 34)
+        output = speech_recognizer(waveform)
+        self.assertEqual(output, {"text": "(Applaudissements)"})
+        output = speech_recognizer(waveform, chunk_length_s=10)
+        self.assertEqual(output, {"text": "(Applaudissements)"})
+
+        # Non CTC models cannot use return_timestamps
+        with self.assertRaisesRegex(
+            ValueError, "^We cannot return_timestamps yet on non-CTC models apart from Whisper!$"
+        ):
+            _ = speech_recognizer(waveform, return_timestamps="char")
+
     @slow
     @require_torch_accelerator
     def test_whisper_fp16(self):

tests/pipelines/test_pipelines_token_classification.py

@@ -27,6 +27,7 @@ from transformers import (
 from transformers.pipelines import AggregationStrategy, TokenClassificationArgumentHandler
 from transformers.testing_utils import (
     is_pipeline_test,
+    is_torch_available,
     nested_simplify,
     require_tf,
     require_torch,
@@ -38,6 +39,10 @@ from transformers.testing_utils import (
 from .test_pipelines_common import ANY
 
 
+if is_torch_available():
+    import torch
+
+
 VALID_INPUTS = ["A simple string", ["list of strings", "A simple string that is quite a bit longer"]]
 
 # These 2 model types require different inputs than those of the usual text models.
@@ -841,6 +846,36 @@ class TokenClassificationPipelineTests(unittest.TestCase):
             ],
         )
 
+    @require_torch
+    def test_small_model_pt_fp16(self):
+        model_name = "hf-internal-testing/tiny-bert-for-token-classification"
+        token_classifier = pipeline(
+            task="token-classification", model=model_name, framework="pt", torch_dtype=torch.float16
+        )
+        outputs = token_classifier("This is a test !")
+        self.assertEqual(
+            nested_simplify(outputs),
+            [
+                {"entity": "I-MISC", "score": 0.115, "index": 1, "word": "this", "start": 0, "end": 4},
+                {"entity": "I-MISC", "score": 0.115, "index": 2, "word": "is", "start": 5, "end": 7},
+            ],
+        )
+
+    @require_torch
+    def test_small_model_pt_bf16(self):
+        model_name = "hf-internal-testing/tiny-bert-for-token-classification"
+        token_classifier = pipeline(
+            task="token-classification", model=model_name, framework="pt", torch_dtype=torch.bfloat16
+        )
+        outputs = token_classifier("This is a test !")
+        self.assertEqual(
+            nested_simplify(outputs),
+            [
+                {"entity": "I-MISC", "score": 0.115, "index": 1, "word": "this", "start": 0, "end": 4},
+                {"entity": "I-MISC", "score": 0.115, "index": 2, "word": "is", "start": 5, "end": 7},
+            ],
+        )
+
     @require_torch
     def test_pt_ignore_subwords_slow_tokenizer_raises(self):
         model_name = "sshleifer/tiny-dbmdz-bert-large-cased-finetuned-conll03-english"

tests/pipelines/test_pipelines_zero_shot.py

@@ -21,11 +21,22 @@ from transformers import (
     ZeroShotClassificationPipeline,
     pipeline,
 )
-from transformers.testing_utils import is_pipeline_test, nested_simplify, require_tf, require_torch, slow
+from transformers.testing_utils import (
+    is_pipeline_test,
+    is_torch_available,
+    nested_simplify,
+    require_tf,
+    require_torch,
+    slow,
+)
 
 from .test_pipelines_common import ANY
 
 
+if is_torch_available():
+    import torch
+
+
 # These 2 model types require different inputs than those of the usual text models.
 _TO_SKIP = {"LayoutLMv2Config", "LayoutLMv3Config"}
 
@@ -176,6 +187,48 @@ class ZeroShotClassificationPipelineTests(unittest.TestCase):
             },
         )
 
+    @require_torch
+    def test_small_model_pt_fp16(self):
+        zero_shot_classifier = pipeline(
+            "zero-shot-classification",
+            model="sshleifer/tiny-distilbert-base-cased-distilled-squad",
+            framework="pt",
+            torch_dtype=torch.float16,
+        )
+        outputs = zero_shot_classifier(
+            "Who are you voting for in 2020?", candidate_labels=["politics", "public health", "science"]
+        )
+
+        self.assertEqual(
+            nested_simplify(outputs),
+            {
+                "sequence": "Who are you voting for in 2020?",
+                "labels": ["science", "public health", "politics"],
+                "scores": [0.333, 0.333, 0.333],
+            },
+        )
+
+    @require_torch
+    def test_small_model_pt_bf16(self):
+        zero_shot_classifier = pipeline(
+            "zero-shot-classification",
+            model="sshleifer/tiny-distilbert-base-cased-distilled-squad",
+            framework="pt",
+            torch_dtype=torch.bfloat16,
+        )
+        outputs = zero_shot_classifier(
+            "Who are you voting for in 2020?", candidate_labels=["politics", "public health", "science"]
+        )
+
+        self.assertEqual(
+            nested_simplify(outputs),
+            {
+                "sequence": "Who are you voting for in 2020?",
+                "labels": ["science", "public health", "politics"],
+                "scores": [0.333, 0.333, 0.333],
+            },
+        )
+
     @require_tf
     def test_small_model_tf(self):
         zero_shot_classifier = pipeline(