Byebye torch 1.10 (#28207)

* fix

* fix

---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

src/transformers/training_args.py

@@ -1439,15 +1439,6 @@ class TrainingArguments:
                         raise ValueError(
                             "Your setup doesn't support bf16/gpu. You need torch>=1.10, using Ampere GPU with cuda>=11.0"
                         )
-                    elif is_torch_npu_available():
-                        # npu
-                        from .pytorch_utils import is_torch_greater_or_equal_than_1_11
-
-                        if not is_torch_greater_or_equal_than_1_11:
-                            raise ValueError(
-                                "Your setup doesn't support bf16/npu. You need torch>=1.11, using Ascend NPU with "
-                                "`torch_npu` installed"
-                            )
                     elif not is_torch_xpu_available():
                         # xpu
                         from .pytorch_utils import is_torch_greater_or_equal_than_1_12

src/transformers/utils/import_utils.py

@@ -64,6 +64,7 @@ USE_JAX = os.environ.get("USE_FLAX", "AUTO").upper()
 
 FORCE_TF_AVAILABLE = os.environ.get("FORCE_TF_AVAILABLE", "AUTO").upper()
 
+# `transformers` requires `torch>=1.11` but this variable is exposed publicly, and we can't simply remove it.
 # This is the version of torch required to run torch.fx features and torch.onnx with dictionary inputs.
 TORCH_FX_REQUIRED_VERSION = version.parse("1.10")
 

tests/models/longt5/test_modeling_longt5.py

@@ -40,7 +40,6 @@ if is_torch_available():
         LongT5Model,
     )
     from transformers.models.longt5.modeling_longt5 import LONGT5_PRETRAINED_MODEL_ARCHIVE_LIST
-    from transformers.pytorch_utils import is_torch_less_than_1_11
 
 
 class LongT5ModelTester:
@@ -595,10 +594,6 @@ class LongT5ModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMix
             model = LongT5Model.from_pretrained(model_name)
             self.assertIsNotNone(model)
 
-    @unittest.skipIf(
-        not is_torch_available() or is_torch_less_than_1_11,
-        "Test failed with torch < 1.11 with an exception in a C++ file.",
-    )
     @slow
     def test_export_to_onnx(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()

tests/models/pix2struct/test_image_processing_pix2struct.py

@@ -28,10 +28,6 @@ from ...test_image_processing_common import ImageProcessingTestMixin, prepare_im
 if is_torch_available():
     import torch
 
-    from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_11
-else:
-    is_torch_greater_or_equal_than_1_11 = False
-
 if is_vision_available():
     from PIL import Image
 
@@ -85,10 +81,6 @@ class Pix2StructImageProcessingTester(unittest.TestCase):
         )
 
 
-@unittest.skipIf(
-    not is_torch_greater_or_equal_than_1_11,
-    reason="`Pix2StructImageProcessor` requires `torch>=1.11.0`.",
-)
 @require_torch
 @require_vision
 class Pix2StructImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
@@ -290,10 +282,6 @@ class Pix2StructImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase)
             )
 
 
-@unittest.skipIf(
-    not is_torch_greater_or_equal_than_1_11,
-    reason="`Pix2StructImageProcessor` requires `torch>=1.11.0`.",
-)
 @require_torch
 @require_vision
 class Pix2StructImageProcessingTestFourChannels(ImageProcessingTestMixin, unittest.TestCase):

tests/models/pix2struct/test_modeling_pix2struct.py

@@ -49,9 +49,6 @@ if is_torch_available():
         Pix2StructVisionModel,
     )
     from transformers.models.pix2struct.modeling_pix2struct import PIX2STRUCT_PRETRAINED_MODEL_ARCHIVE_LIST
-    from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_11
-else:
-    is_torch_greater_or_equal_than_1_11 = False
 
 
 if is_vision_available():
@@ -746,10 +743,6 @@ def prepare_img():
     return im
 
 
-@unittest.skipIf(
-    not is_torch_greater_or_equal_than_1_11,
-    reason="`Pix2StructImageProcessor` requires `torch>=1.11.0`.",
-)
 @require_vision
 @require_torch
 @slow

tests/models/pix2struct/test_processor_pix2struct.py

@@ -19,14 +19,9 @@ import numpy as np
 import pytest
 
 from transformers.testing_utils import require_torch, require_vision
-from transformers.utils import is_torch_available, is_vision_available
+from transformers.utils import is_vision_available
 
 
-if is_torch_available():
-    from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_11
-else:
-    is_torch_greater_or_equal_than_1_11 = False
-
 if is_vision_available():
     from PIL import Image
 
@@ -39,10 +34,6 @@ if is_vision_available():
     )
 
 
-@unittest.skipIf(
-    not is_torch_greater_or_equal_than_1_11,
-    reason="`Pix2StructImageProcessor` requires `torch>=1.11.0`.",
-)
 @require_vision
 @require_torch
 class Pix2StructProcessorTest(unittest.TestCase):

tests/models/pop2piano/test_modeling_pop2piano.py

@@ -45,10 +45,6 @@ if is_torch_available():
 
     from transformers import Pop2PianoForConditionalGeneration
     from transformers.models.pop2piano.modeling_pop2piano import POP2PIANO_PRETRAINED_MODEL_ARCHIVE_LIST
-    from transformers.pytorch_utils import is_torch_1_8_0
-
-else:
-    is_torch_1_8_0 = False
 
 
 @require_torch
@@ -616,10 +612,6 @@ class Pop2PianoModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTester
             self.assertIsNotNone(model)
 
     @require_onnx
-    @unittest.skipIf(
-        is_torch_1_8_0,
-        reason="Test has a segmentation fault on torch 1.8.0",
-    )
     def test_export_to_onnx(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         model = Pop2PianoForConditionalGeneration(config_and_inputs[0]).to(torch_device)

tests/models/unispeech_sat/test_modeling_unispeech_sat.py

@@ -906,7 +906,6 @@ class UniSpeechSatModelIntegrationTest(unittest.TestCase):
         )
         self.assertEqual(labels[0, :, 0].sum(), 270)
         self.assertEqual(labels[0, :, 1].sum(), 647)
-        # TODO: update the tolerance after the CI moves to torch 1.10
         self.assertTrue(torch.allclose(outputs.logits[:, :4], expected_logits, atol=1e-2))
 
     def test_inference_speaker_verification(self):
@@ -931,5 +930,4 @@ class UniSpeechSatModelIntegrationTest(unittest.TestCase):
         # id10002 vs id10004
         self.assertAlmostEqual(cosine_sim(embeddings[2], embeddings[3]).item(), 0.5616, 3)
 
-        # TODO: update the tolerance after the CI moves to torch 1.10
         self.assertAlmostEqual(outputs.loss.item(), 18.5925, 2)

tests/models/wav2vec2/test_modeling_wav2vec2.py

@@ -1928,7 +1928,6 @@ class Wav2Vec2ModelIntegrationTest(unittest.TestCase):
         )
         self.assertEqual(labels[0, :, 0].sum(), 555)
         self.assertEqual(labels[0, :, 1].sum(), 299)
-        # TODO: update the tolerance after the CI moves to torch 1.10
         self.assertTrue(torch.allclose(outputs.logits[:, :4], expected_logits, atol=1e-2))
 
     def test_inference_speaker_verification(self):
@@ -1953,7 +1952,6 @@ class Wav2Vec2ModelIntegrationTest(unittest.TestCase):
         # id10002 vs id10004
         self.assertAlmostEqual(cosine_sim(embeddings[2], embeddings[3]).numpy(), 0.7594, 3)
 
-        # TODO: update the tolerance after the CI moves to torch 1.10
         self.assertAlmostEqual(outputs.loss.item(), 17.7963, 2)
 
     @require_torchaudio

tests/models/wavlm/test_modeling_wavlm.py

@@ -515,7 +515,6 @@ class WavLMModelIntegrationTest(unittest.TestCase):
         EXPECTED_HIDDEN_STATES_SLICE = torch.tensor(
             [[[0.0577, 0.1161], [0.0579, 0.1165]], [[0.0199, 0.1237], [0.0059, 0.0605]]]
         )
-        # TODO: update the tolerance after the CI moves to torch 1.10
         self.assertTrue(torch.allclose(hidden_states_slice, EXPECTED_HIDDEN_STATES_SLICE, atol=5e-2))
 
     def test_inference_large(self):
@@ -567,7 +566,6 @@ class WavLMModelIntegrationTest(unittest.TestCase):
         )
         self.assertEqual(labels[0, :, 0].sum(), 258)
         self.assertEqual(labels[0, :, 1].sum(), 647)
-        # TODO: update the tolerance after the CI moves to torch 1.10
         self.assertTrue(torch.allclose(outputs.logits[:, :4], expected_logits, atol=1e-2))
 
     def test_inference_speaker_verification(self):
@@ -592,5 +590,4 @@ class WavLMModelIntegrationTest(unittest.TestCase):
         # id10002 vs id10004
         self.assertAlmostEqual(cosine_sim(embeddings[2], embeddings[3]).item(), 0.4780, 3)
 
-        # TODO: update the tolerance after the CI moves to torch 1.10
         self.assertAlmostEqual(outputs.loss.item(), 18.4154, 2)

tests/pipelines/test_pipelines_image_to_text.py

@@ -20,7 +20,6 @@ from transformers import MODEL_FOR_VISION_2_SEQ_MAPPING, TF_MODEL_FOR_VISION_2_S
 from transformers.pipelines import pipeline
 from transformers.testing_utils import (
     is_pipeline_test,
-    is_torch_available,
     require_tf,
     require_torch,
     require_vision,
@@ -30,12 +29,6 @@ from transformers.testing_utils import (
 from .test_pipelines_common import ANY
 
 
-if is_torch_available():
-    from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_11
-else:
-    is_torch_greater_or_equal_than_1_11 = False
-
-
 if is_vision_available():
     from PIL import Image
 else:
@@ -217,9 +210,6 @@ class ImageToTextPipelineTests(unittest.TestCase):
         with self.assertRaises(ValueError):
             outputs = pipe([image, image], prompt=[prompt, prompt])
 
-    @unittest.skipIf(
-        not is_torch_greater_or_equal_than_1_11, reason="`Pix2StructImageProcessor` requires `torch>=1.11.0`."
-    )
     @slow
     @require_torch
     def test_conditional_generation_pt_pix2struct(self):