Add onnx export cuda support (huggingface#17183)

Co-authored-by: Lysandre Debut <[email protected]>

Co-authored-by: lewtun <[email protected]>

Author: 2 people

src/transformers/models/big_bird/modeling_big_bird.py

@@ -3099,7 +3099,7 @@ def forward(
             # setting lengths logits to `-inf`
             logits_mask = self.prepare_question_mask(question_lengths, seqlen)
             if token_type_ids is None:
-                token_type_ids = torch.ones(logits_mask.size(), dtype=int) - logits_mask
+                token_type_ids = torch.ones(logits_mask.size(), dtype=int, device=logits_mask.device) - logits_mask
             logits_mask = logits_mask
             logits_mask[:, 0] = False
             logits_mask.unsqueeze_(2)

src/transformers/onnx/convert.py

@@ -86,6 +86,7 @@ def export_pytorch(
     opset: int,
     output: Path,
     tokenizer: "PreTrainedTokenizer" = None,
+    device: str = "cpu",
 ) -> Tuple[List[str], List[str]]:
     """
     Export a PyTorch model to an ONNX Intermediate Representation (IR)
@@ -101,6 +102,8 @@ def export_pytorch(
             The version of the ONNX operator set to use.
         output (`Path`):
             Directory to store the exported ONNX model.
+        device (`str`, *optional*, defaults to `cpu`):
+            The device on which the ONNX model will be exported. Either `cpu` or `cuda`.
 
     Returns:
         `Tuple[List[str], List[str]]`: A tuple with an ordered list of the model's inputs, and the named inputs from
@@ -137,6 +140,10 @@ def export_pytorch(
             # Ensure inputs match
             # TODO: Check when exporting QA we provide "is_pair=True"
             model_inputs = config.generate_dummy_inputs(preprocessor, framework=TensorType.PYTORCH)
+            device = torch.device(device)
+            if device.type == "cuda" and torch.cuda.is_available():
+                model.to(device)
+                model_inputs = dict((k, v.to(device)) for k, v in model_inputs.items())
             inputs_match, matched_inputs = ensure_model_and_config_inputs_match(model, model_inputs.keys())
             onnx_outputs = list(config.outputs.keys())
 
@@ -268,6 +275,7 @@ def export(
     opset: int,
     output: Path,
     tokenizer: "PreTrainedTokenizer" = None,
+    device: str = "cpu",
 ) -> Tuple[List[str], List[str]]:
     """
     Export a Pytorch or TensorFlow model to an ONNX Intermediate Representation (IR)
@@ -283,6 +291,9 @@ def export(
             The version of the ONNX operator set to use.
         output (`Path`):
             Directory to store the exported ONNX model.
+        device (`str`, *optional*, defaults to `cpu`):
+            The device on which the ONNX model will be exported. Either `cpu` or `cuda`. Only PyTorch is supported for
+            export on CUDA devices.
 
     Returns:
         `Tuple[List[str], List[str]]`: A tuple with an ordered list of the model's inputs, and the named inputs from
@@ -294,6 +305,9 @@ def export(
             "Please install torch or tensorflow first."
         )
 
+    if is_tf_available() and isinstance(model, TFPreTrainedModel) and device == "cuda":
+        raise RuntimeError("`tf2onnx` does not support export on CUDA device.")
+
     if isinstance(preprocessor, PreTrainedTokenizerBase) and tokenizer is not None:
         raise ValueError("You cannot provide both a tokenizer and a preprocessor to export the model.")
     if tokenizer is not None:
@@ -318,7 +332,7 @@ def export(
             )
 
     if is_torch_available() and issubclass(type(model), PreTrainedModel):
-        return export_pytorch(preprocessor, model, config, opset, output, tokenizer=tokenizer)
+        return export_pytorch(preprocessor, model, config, opset, output, tokenizer=tokenizer, device=device)
     elif is_tf_available() and issubclass(type(model), TFPreTrainedModel):
         return export_tensorflow(preprocessor, model, config, opset, output, tokenizer=tokenizer)
 
@@ -359,6 +373,8 @@ def validate_model_outputs(
     session = InferenceSession(onnx_model.as_posix(), options, providers=["CPUExecutionProvider"])
 
     # Compute outputs from the reference model
+    if is_torch_available() and issubclass(type(reference_model), PreTrainedModel):
+        reference_model.to("cpu")
     ref_outputs = reference_model(**reference_model_inputs)
     ref_outputs_dict = {}
 

tests/onnx/test_onnx_v2.py

@@ -242,7 +242,7 @@ class OnnxExportTestCaseV2(TestCase):
     Integration tests ensuring supported models are correctly exported
     """
 
-    def _onnx_export(self, test_name, name, model_name, feature, onnx_config_class_constructor):
+    def _onnx_export(self, test_name, name, model_name, feature, onnx_config_class_constructor, device="cpu"):
         from transformers.onnx import export
 
         model_class = FeaturesManager.get_model_class_for_feature(feature)
@@ -273,7 +273,7 @@ def _onnx_export(self, test_name, name, model_name, feature, onnx_config_class_c
         with NamedTemporaryFile("w") as output:
             try:
                 onnx_inputs, onnx_outputs = export(
-                    preprocessor, model, onnx_config, onnx_config.default_onnx_opset, Path(output.name)
+                    preprocessor, model, onnx_config, onnx_config.default_onnx_opset, Path(output.name), device=device
                 )
                 validate_model_outputs(
                     onnx_config,
@@ -294,6 +294,14 @@ def _onnx_export(self, test_name, name, model_name, feature, onnx_config_class_c
     def test_pytorch_export(self, test_name, name, model_name, feature, onnx_config_class_constructor):
         self._onnx_export(test_name, name, model_name, feature, onnx_config_class_constructor)
 
+    @parameterized.expand(_get_models_to_test(PYTORCH_EXPORT_MODELS))
+    @slow
+    @require_torch
+    @require_vision
+    @require_rjieba
+    def test_pytorch_export_on_cuda(self, test_name, name, model_name, feature, onnx_config_class_constructor):
+        self._onnx_export(test_name, name, model_name, feature, onnx_config_class_constructor, device="cuda")
+
     @parameterized.expand(_get_models_to_test(PYTORCH_EXPORT_WITH_PAST_MODELS))
     @slow
     @require_torch