Fix compatibility with 1.12 (huggingface#17925)

* Fix compatibility with 1.12

* Remove pin from examples requirements

* Update torch scatter version

* Fix compatibility with 1.12

* Remove pin from examples requirements

* Update torch scatter version

* fix torch.onnx.symbolic_opset12 import

* Reject bad version

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

.circleci/config.yml

@@ -83,7 +83,7 @@ jobs:
             - run: git lfs install
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,tf-cpu,torch,testing,sentencepiece,torch-speech,vision]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.12.0+cpu.html
             - run: pip install tensorflow_probability
             - run: pip install https://github.com/kpu/kenlm/archive/master.zip
             - run: pip install git+https://github.com/huggingface/accelerate
@@ -124,7 +124,7 @@ jobs:
             - run: git lfs install
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,tf-cpu,torch,testing,sentencepiece,torch-speech,vision]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.12.0+cpu.html
             - run: pip install tensorflow_probability
             - run: pip install https://github.com/kpu/kenlm/archive/master.zip
             - run: pip install git+https://github.com/huggingface/accelerate
@@ -159,7 +159,7 @@ jobs:
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,flax,torch,testing,sentencepiece,torch-speech,vision]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.12.0+cpu.html
             - run: pip install https://github.com/kpu/kenlm/archive/master.zip
             - run: pip install git+https://github.com/huggingface/accelerate
             - save_cache:
@@ -198,7 +198,7 @@ jobs:
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,flax,torch,testing,sentencepiece,torch-speech,vision]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.12.0+cpu.html
             - run: pip install https://github.com/kpu/kenlm/archive/master.zip
             - run: pip install git+https://github.com/huggingface/accelerate
             - save_cache:
@@ -231,7 +231,7 @@ jobs:
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng time
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,torch,testing,sentencepiece,torch-speech,vision,timm]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.12.0+cpu.html
             - run: pip install https://github.com/kpu/kenlm/archive/master.zip
             - run: pip install git+https://github.com/huggingface/accelerate
             - save_cache:
@@ -269,7 +269,7 @@ jobs:
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,torch,testing,sentencepiece,torch-speech,vision,timm]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.12.0+cpu.html
             - run: pip install https://github.com/kpu/kenlm/archive/master.zip
             - run: pip install git+https://github.com/huggingface/accelerate
             - save_cache:
@@ -439,7 +439,7 @@ jobs:
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,torch,testing,sentencepiece,torch-speech,vision,timm]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.12.0+cpu.html
             - run: pip install https://github.com/kpu/kenlm/archive/master.zip
             - save_cache:
                   key: v0.5-torch-{{ checksum "setup.py" }}
@@ -477,7 +477,7 @@ jobs:
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,torch,testing,sentencepiece,torch-speech,vision,timm]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.12.0+cpu.html
             - run: pip install https://github.com/kpu/kenlm/archive/master.zip
             - save_cache:
                   key: v0.5-torch-{{ checksum "setup.py" }}

examples/pytorch/_tests_requirements.txt

@@ -22,5 +22,4 @@ protobuf
 torchvision
 jiwer
 librosa
-torch < 1.12
 evaluate >= 0.2.0

setup.py

@@ -162,7 +162,7 @@
     "timeout-decorator",
     "timm",
     "tokenizers>=0.11.1,!=0.11.3,<0.13",
-    "torch>=1.0,<1.12",
+    "torch>=1.0,!=0.12.0",
     "torchaudio",
     "pyctcdecode>=0.3.0",
     "tqdm>=4.27",

src/transformers/dependency_versions_table.py

@@ -68,7 +68,7 @@
     "timeout-decorator": "timeout-decorator",
     "timm": "timm",
     "tokenizers": "tokenizers>=0.11.1,!=0.11.3,<0.13",
-    "torch": "torch>=1.0,<1.12",
+    "torch": "torch>=1.0,!=0.12.0",
     "torchaudio": "torchaudio",
     "pyctcdecode": "pyctcdecode>=0.3.0",
     "tqdm": "tqdm>=4.27",

src/transformers/models/deberta/modeling_deberta.py

@@ -187,6 +187,8 @@ def backward(ctx, grad_output):
 
     @staticmethod
     def symbolic(g: torch._C.Graph, input: torch._C.Value, local_ctx: Union[float, DropoutContext]) -> torch._C.Value:
+        from torch.onnx import symbolic_opset12
+
         dropout_p = local_ctx
         if isinstance(local_ctx, DropoutContext):
             dropout_p = local_ctx.dropout
@@ -198,7 +200,7 @@ def symbolic(g: torch._C.Graph, input: torch._C.Value, local_ctx: Union[float, D
         # Once https://github.com/pytorch/pytorch/issues/78391 is fixed, do something like:
         # if opset_version < 12:
         #   return torch.onnx.symbolic_opset9.dropout(g, input, dropout_p, train)
-        return torch.onnx.symbolic_opset12.dropout(g, input, dropout_p, train)
+        return symbolic_opset12.dropout(g, input, dropout_p, train)
 
 
 class StableDropout(nn.Module):

src/transformers/models/deberta_v2/modeling_deberta_v2.py

@@ -193,6 +193,8 @@ def backward(ctx, grad_output):
 
     @staticmethod
     def symbolic(g: torch._C.Graph, input: torch._C.Value, local_ctx: Union[float, DropoutContext]) -> torch._C.Value:
+        from torch.onnx import symbolic_opset12
+
         dropout_p = local_ctx
         if isinstance(local_ctx, DropoutContext):
             dropout_p = local_ctx.dropout
@@ -204,7 +206,7 @@ def symbolic(g: torch._C.Graph, input: torch._C.Value, local_ctx: Union[float, D
         # Once https://github.com/pytorch/pytorch/issues/78391 is fixed, do something like:
         # if opset_version < 12:
         #   return torch.onnx.symbolic_opset9.dropout(g, input, dropout_p, train)
-        return torch.onnx.symbolic_opset12.dropout(g, input, dropout_p, train)
+        return symbolic_opset12.dropout(g, input, dropout_p, train)
 
 
 # Copied from transformers.models.deberta.modeling_deberta.StableDropout

src/transformers/models/sew_d/modeling_sew_d.py

@@ -597,6 +597,8 @@ def backward(ctx, grad_output):
 
     @staticmethod
     def symbolic(g: torch._C.Graph, input: torch._C.Value, local_ctx: Union[float, DropoutContext]) -> torch._C.Value:
+        from torch.onnx import symbolic_opset12
+
         dropout_p = local_ctx
         if isinstance(local_ctx, DropoutContext):
             dropout_p = local_ctx.dropout
@@ -608,7 +610,7 @@ def symbolic(g: torch._C.Graph, input: torch._C.Value, local_ctx: Union[float, D
         # Once https://github.com/pytorch/pytorch/issues/78391 is fixed, do something like:
         # if opset_version < 12:
         #   return torch.onnx.symbolic_opset9.dropout(g, input, dropout_p, train)
-        return torch.onnx.symbolic_opset12.dropout(g, input, dropout_p, train)
+        return symbolic_opset12.dropout(g, input, dropout_p, train)
 
 
 # Copied from transformers.models.deberta.modeling_deberta.StableDropout