Fix torch version comparisons (helps with +cu*** or +cpu official builds)

examples/research_projects/wav2vec2/run_asr.py

@@ -30,7 +30,7 @@
 if is_apex_available():
     from apex import amp
 
-if version.parse(torch.__version__) >= version.parse("1.6"):
+if version.parse(version.parse(torch.__version__).base_version) >= version.parse("1.6"):
     _is_native_amp_available = True
     from torch.cuda.amp import autocast
 

examples/research_projects/wav2vec2/run_common_voice.py

@@ -33,7 +33,7 @@
     from apex import amp
 
 
-if version.parse(torch.__version__) >= version.parse("1.6"):
+if version.parse(version.parse(torch.__version__).base_version) >= version.parse("1.6"):
     _is_native_amp_available = True
     from torch.cuda.amp import autocast
 

examples/research_projects/wav2vec2/run_pretrain.py

@@ -26,7 +26,7 @@
 if is_apex_available():
     from apex import amp
 
-if version.parse(torch.__version__) >= version.parse("1.6"):
+if version.parse(version.parse(torch.__version__).base_version) >= version.parse("1.6"):
     _is_native_amp_available = True
     from torch.cuda.amp import autocast
 

src/transformers/activations.py

@@ -44,7 +44,7 @@ class GELUActivation(nn.Module):
 
     def __init__(self, use_gelu_python: bool = False):
         super().__init__()
-        if version.parse(torch.__version__) < version.parse("1.4") or use_gelu_python:
+        if version.parse(version.parse(torch.__version__).base_version) < version.parse("1.4") or use_gelu_python:
             self.act = self._gelu_python
         else:
             self.act = nn.functional.gelu
@@ -110,7 +110,7 @@ class SiLUActivation(nn.Module):
 
     def __init__(self):
         super().__init__()
-        if version.parse(torch.__version__) < version.parse("1.7"):
+        if version.parse(version.parse(torch.__version__).base_version) < version.parse("1.7"):
             self.act = self._silu_python
         else:
             self.act = nn.functional.silu
@@ -130,7 +130,7 @@ class MishActivation(nn.Module):
 
     def __init__(self):
         super().__init__()
-        if version.parse(torch.__version__) < version.parse("1.9"):
+        if version.parse(version.parse(torch.__version__).base_version) < version.parse("1.9"):
             self.act = self._mish_python
         else:
             self.act = nn.functional.mish

src/transformers/convert_graph_to_onnx.py

@@ -273,6 +273,8 @@ def convert_pytorch(nlp: Pipeline, opset: int, output: Path, use_external_format
     import torch
     from torch.onnx import export
 
+    from .pytorch_utils import is_torch_less_than_1_11
+
     print(f"Using framework PyTorch: {torch.__version__}")
 
     with torch.no_grad():
@@ -281,7 +283,7 @@ def convert_pytorch(nlp: Pipeline, opset: int, output: Path, use_external_format
 
         # PyTorch deprecated the `enable_onnx_checker` and `use_external_data_format` arguments in v1.11,
         # so we check the torch version for backwards compatibility
-        if parse(torch.__version__) <= parse("1.10.99"):
+        if is_torch_less_than_1_11:
             export(
                 nlp.model,
                 model_args,

src/transformers/models/albert/modeling_albert.py

@@ -20,7 +20,6 @@
 from typing import Dict, List, Optional, Tuple, Union
 
 import torch
-from packaging import version
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
@@ -35,7 +34,12 @@
     TokenClassifierOutput,
 )
 from ...modeling_utils import PreTrainedModel
-from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...pytorch_utils import (
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    is_torch_greater_than_1_6,
+    prune_linear_layer,
+)
 from ...utils import (
     ModelOutput,
     add_code_sample_docstrings,
@@ -212,7 +216,7 @@ def __init__(self, config: AlbertConfig):
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
+        if is_torch_greater_than_1_6:
             self.register_buffer(
                 "token_type_ids",
                 torch.zeros(self.position_ids.size(), dtype=torch.long),

src/transformers/models/bert/modeling_bert.py

@@ -24,7 +24,6 @@
 
 import torch
 import torch.utils.checkpoint
-from packaging import version
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
@@ -41,7 +40,12 @@
     TokenClassifierOutput,
 )
 from ...modeling_utils import PreTrainedModel
-from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...pytorch_utils import (
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    is_torch_greater_than_1_6,
+    prune_linear_layer,
+)
 from ...utils import (
     ModelOutput,
     add_code_sample_docstrings,
@@ -195,7 +199,7 @@ def __init__(self, config):
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
+        if is_torch_greater_than_1_6:
             self.register_buffer(
                 "token_type_ids",
                 torch.zeros(self.position_ids.size(), dtype=torch.long),

src/transformers/models/big_bird/modeling_big_bird.py

@@ -23,7 +23,6 @@
 import numpy as np
 import torch
 import torch.utils.checkpoint
-from packaging import version
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
@@ -38,7 +37,7 @@
     TokenClassifierOutput,
 )
 from ...modeling_utils import PreTrainedModel
-from ...pytorch_utils import apply_chunking_to_forward
+from ...pytorch_utils import apply_chunking_to_forward, is_torch_greater_than_1_6
 from ...utils import (
     ModelOutput,
     add_code_sample_docstrings,
@@ -260,7 +259,7 @@ def __init__(self, config):
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
+        if is_torch_greater_than_1_6:
             self.register_buffer(
                 "token_type_ids",
                 torch.zeros(self.position_ids.size(), dtype=torch.long),

src/transformers/models/convbert/modeling_convbert.py

@@ -22,7 +22,6 @@
 
 import torch
 import torch.utils.checkpoint
-from packaging import version
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
@@ -36,7 +35,12 @@
     TokenClassifierOutput,
 )
 from ...modeling_utils import PreTrainedModel, SequenceSummary
-from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...pytorch_utils import (
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    is_torch_greater_than_1_6,
+    prune_linear_layer,
+)
 from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
 from .configuration_convbert import ConvBertConfig
 
@@ -194,7 +198,7 @@ def __init__(self, config):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
+        if is_torch_greater_than_1_6:
             self.register_buffer(
                 "token_type_ids",
                 torch.zeros(self.position_ids.size(), dtype=torch.long),

src/transformers/models/data2vec/modeling_data2vec_text.py

@@ -19,7 +19,6 @@
 
 import torch
 import torch.utils.checkpoint
-from packaging import version
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
@@ -35,7 +34,12 @@
     TokenClassifierOutput,
 )
 from ...modeling_utils import PreTrainedModel
-from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...pytorch_utils import (
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    is_torch_greater_than_1_6,
+    prune_linear_layer,
+)
 from ...utils import (
     add_code_sample_docstrings,
     add_start_docstrings,
@@ -83,7 +87,7 @@ def __init__(self, config):
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
+        if is_torch_greater_than_1_6:
             self.register_buffer(
                 "token_type_ids",
                 torch.zeros(self.position_ids.size(), dtype=torch.long),

src/transformers/models/decision_transformer/modeling_decision_transformer.py

@@ -21,12 +21,16 @@
 
 import torch
 import torch.utils.checkpoint
-from packaging import version
 from torch import nn
 
 from ...activations import ACT2FN
 from ...modeling_utils import PreTrainedModel
-from ...pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
+from ...pytorch_utils import (
+    Conv1D,
+    find_pruneable_heads_and_indices,
+    is_torch_greater_or_equal_than_1_6,
+    prune_conv1d_layer,
+)
 from ...utils import (
     ModelOutput,
     add_start_docstrings,
@@ -36,7 +40,7 @@
 )
 
 
-if version.parse(torch.__version__) >= version.parse("1.6"):
+if is_torch_greater_or_equal_than_1_6:
     is_amp_available = True
     from torch.cuda.amp import autocast
 else:

src/transformers/models/distilbert/modeling_distilbert.py

@@ -23,7 +23,6 @@
 
 import numpy as np
 import torch
-from packaging import version
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
@@ -40,7 +39,12 @@
     TokenClassifierOutput,
 )
 from ...modeling_utils import PreTrainedModel
-from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...pytorch_utils import (
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    is_torch_greater_than_1_6,
+    prune_linear_layer,
+)
 from ...utils import (
     add_code_sample_docstrings,
     add_start_docstrings,
@@ -102,7 +106,7 @@ def __init__(self, config: PretrainedConfig):
 
         self.LayerNorm = nn.LayerNorm(config.dim, eps=1e-12)
         self.dropout = nn.Dropout(config.dropout)
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
+        if is_torch_greater_than_1_6:
             self.register_buffer(
                 "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
             )

src/transformers/models/electra/modeling_electra.py

@@ -21,7 +21,6 @@
 
 import torch
 import torch.utils.checkpoint
-from packaging import version
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
@@ -37,7 +36,12 @@
     TokenClassifierOutput,
 )
 from ...modeling_utils import PreTrainedModel, SequenceSummary
-from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...pytorch_utils import (
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    is_torch_greater_than_1_6,
+    prune_linear_layer,
+)
 from ...utils import (
     ModelOutput,
     add_code_sample_docstrings,
@@ -165,7 +169,7 @@ def __init__(self, config):
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
+        if is_torch_greater_than_1_6:
             self.register_buffer(
                 "token_type_ids",
                 torch.zeros(self.position_ids.size(), dtype=torch.long),

src/transformers/models/flaubert/modeling_flaubert.py

@@ -19,10 +19,10 @@
 from typing import Dict, Optional, Tuple, Union
 
 import torch
-from packaging import version
 from torch import nn
 
 from ...modeling_outputs import BaseModelOutput
+from ...pytorch_utils import is_torch_greater_than_1_6
 from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
 from ..xlm.modeling_xlm import (
     XLMForMultipleChoice,
@@ -139,7 +139,7 @@ def __init__(self, config):  # , dico, is_encoder, with_output):
         super().__init__(config)
         self.layerdrop = getattr(config, "layerdrop", 0.0)
         self.pre_norm = getattr(config, "pre_norm", False)
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
+        if is_torch_greater_than_1_6:
             self.register_buffer(
                 "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
             )

src/transformers/models/flava/modeling_flava.py

@@ -22,14 +22,14 @@
 
 import torch
 import torch.utils.checkpoint
-from packaging import version
 from torch import nn
 
 from transformers.utils.doc import add_code_sample_docstrings
 
 from ...activations import ACT2FN
 from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
 from ...modeling_utils import PreTrainedModel, find_pruneable_heads_and_indices, prune_linear_layer
+from ...pytorch_utils import is_torch_greater_than_1_6
 from ...utils import (
     ModelOutput,
     add_start_docstrings,
@@ -392,7 +392,7 @@ def __init__(self, config):
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
+        if is_torch_greater_than_1_6:
             self.register_buffer(
                 "token_type_ids",
                 torch.zeros(self.position_ids.size(), dtype=torch.long),

src/transformers/models/fnet/modeling_fnet.py

@@ -21,7 +21,6 @@
 
 import torch
 import torch.utils.checkpoint
-from packaging import version
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
@@ -44,7 +43,7 @@
     TokenClassifierOutput,
 )
 from ...modeling_utils import PreTrainedModel
-from ...pytorch_utils import apply_chunking_to_forward
+from ...pytorch_utils import apply_chunking_to_forward, is_torch_greater_than_1_6
 from ...utils import (
     add_code_sample_docstrings,
     add_start_docstrings,
@@ -118,7 +117,7 @@ def __init__(self, config):
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
 
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
+        if is_torch_greater_than_1_6:
             self.register_buffer(
                 "token_type_ids",
                 torch.zeros(self.position_ids.size(), dtype=torch.long),