Add ONNX support for MarianMT models (#14586)

* First commit to add MarianMT to ONNX

* Now MarianModel.forward() automatically generates decoder_input_ids, like BartModel.forward()

* Adjusted MarianOnnxConfig.inputs and outputs to work with seq2seq-lm feature

* Style fix

* Added support for other features for already supported models

* Partial support for causal and seq2seq models

* Partial support for causal and seq2seq models

* Add default task for MarianMT ONNX

* Remove automatic creation of decoder_input_ids

* Extend inputs and outputs for MarianMT ONNX config

* Add MarianMT to ONNX unit tests

* Refactor

* OnnxSeq2SeqConfigWithPast to support seq2seq models

* Parameterized the onnx tests

* Restored run_mlm.py

* Restored run_mlm.py

* [WIP] BART update

* BART and MBART

* Add past_key_values and fix dummy decoder inputs

Using a sequence length of 1 in generate_dummy_outputs() produces large discrepancies, presumably due to some hidden optimisations.

* Refactor MarianOnnxConfig to remove custom past_key_values logic

* Fix quality

* Revert "Revert "Added support for other features for already supported models (#14358)" (#14679)"

This reverts commit 0f4e39c.

* is_torch_available test to avoid failing imports

* sorting parameterize parameters to solve ERROR gw0 gw1

* tests fix

* tests fix

* GPT2 with past fix

* Fixed stateful class attribute change that was breaking things when converting multiple models sequentially

* Removed onnx file

* Refactor Marian export to account for base changes

* Fix copies

* Implemented suggestions

* Extend support for causal LM

* Revert "Revert "Added support for other features for already supported models (#14358)" (#14679)"

This reverts commit 0f4e39c.

* is_torch_available test to avoid failing imports

* sorting parameterize parameters to solve ERROR gw0 gw1

* tests fix

* tests fix

* GPT2 with past fix

* Fixed stateful class attribute change that was breaking things when converting multiple models sequentially

* Removed onnx file

* Implemented suggestions

* Fixed __init__ to resolve conflict with master

* Revert "Revert "Added support for other features for already supported models (#14358)" (#14679)"

This reverts commit 0f4e39c.

* is_torch_available test to avoid failing imports

* sorting parameterize parameters to solve ERROR gw0 gw1

* tests fix

* tests fix

* GPT2 with past fix

* Fixed stateful class attribute change that was breaking things when converting multiple models sequentially

* Removed onnx file

* Implemented suggestions

* Fixed __init__ to resolve conflict with master

* Remove commented import

* Remove ONNX model

* Remove redundant class method

* Tidy up imports

* Fix quality

* Refactor dummy input function

* Add copied from statements to Marian config functions

* Remove false copied from comments

* Fix copy from comment

Co-authored-by: Massimiliano Bruni <massimiliano.bruni@hcl.com>
Co-authored-by: Michael Benayoun <mickbenayoun@gmail.com>

docs/source/serialization.mdx

@@ -42,6 +42,7 @@ Ready-made configurations include the following models:
 - GPT Neo
 - LayoutLM
 - Longformer
+- Marian
 - mBART
 - OpenAI GPT-2
 - RoBERTa

src/transformers/models/marian/__init__.py

@@ -28,7 +28,7 @@
 
 
 _import_structure = {
-    "configuration_marian": ["MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP", "MarianConfig"],
+    "configuration_marian": ["MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP", "MarianConfig", "MarianOnnxConfig"],
 }
 
 if is_sentencepiece_available():
@@ -49,7 +49,7 @@
 if is_flax_available():
     _import_structure["modeling_flax_marian"] = ["FlaxMarianModel", "FlaxMarianMTModel", "FlaxMarianPreTrainedModel"]
 if TYPE_CHECKING:
-    from .configuration_marian import MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP, MarianConfig
+    from .configuration_marian import MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP, MarianConfig, MarianOnnxConfig
 
     if is_sentencepiece_available():
         from .tokenization_marian import MarianTokenizer

src/transformers/models/marian/configuration_marian.py

@@ -13,8 +13,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """ Marian model configuration """
+from collections import OrderedDict
+from typing import Any, Mapping, Optional
 
+from ... import PreTrainedTokenizer
 from ...configuration_utils import PretrainedConfig
+from ...file_utils import TensorType, is_torch_available
+from ...onnx import OnnxConfig, OnnxConfigWithPast, OnnxSeq2SeqConfigWithPast
+from ...onnx.utils import compute_effective_axis_dimension
 from ...utils import logging
 
 
@@ -160,3 +166,226 @@ def __init__(
             forced_eos_token_id=forced_eos_token_id,
             **kwargs,
         )
+
+
+class MarianOnnxConfig(OnnxSeq2SeqConfigWithPast):
+    @property
+    # Copied from transformers.models.bart.configuration_bart.BartOnnxConfig.inputs
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        if self.task in ["default", "seq2seq-lm"]:
+            common_inputs = OrderedDict(
+                [
+                    ("input_ids", {0: "batch", 1: "encoder_sequence"}),
+                    ("attention_mask", {0: "batch", 1: "encoder_sequence"}),
+                ]
+            )
+
+            if self.use_past:
+                common_inputs["decoder_input_ids"] = {0: "batch"}
+                common_inputs["decoder_attention_mask"] = {0: "batch", 1: "past_decoder_sequence + sequence"}
+            else:
+                common_inputs["decoder_input_ids"] = {0: "batch", 1: "decoder_sequence"}
+                common_inputs["decoder_attention_mask"] = {0: "batch", 1: "decoder_sequence"}
+
+            if self.use_past:
+                self.fill_with_past_key_values_(common_inputs, direction="inputs")
+        elif self.task == "causal-lm":
+            # TODO: figure this case out.
+            common_inputs = OrderedDict(
+                [
+                    ("input_ids", {0: "batch", 1: "encoder_sequence"}),
+                    ("attention_mask", {0: "batch", 1: "encoder_sequence"}),
+                ]
+            )
+            if self.use_past:
+                num_encoder_layers, _ = self.num_layers
+                for i in range(num_encoder_layers):
+                    common_inputs[f"past_key_values.{i}.key"] = {0: "batch", 2: "past_sequence + sequence"}
+                    common_inputs[f"past_key_values.{i}.value"] = {0: "batch", 2: "past_sequence + sequence"}
+        else:
+            common_inputs = OrderedDict(
+                [
+                    ("input_ids", {0: "batch", 1: "encoder_sequence"}),
+                    ("attention_mask", {0: "batch", 1: "encoder_sequence"}),
+                    ("decoder_input_ids", {0: "batch", 1: "decoder_sequence"}),
+                    ("decoder_attention_mask", {0: "batch", 1: "decoder_sequence"}),
+                ]
+            )
+
+        return common_inputs
+
+    @property
+    # Copied from transformers.models.bart.configuration_bart.BartOnnxConfig.outputs
+    def outputs(self) -> Mapping[str, Mapping[int, str]]:
+        if self.task in ["default", "seq2seq-lm"]:
+            common_outputs = super().outputs
+        else:
+            common_outputs = super(OnnxConfigWithPast, self).outputs
+            if self.use_past:
+                num_encoder_layers, _ = self.num_layers
+                for i in range(num_encoder_layers):
+                    common_outputs[f"present.{i}.key"] = {0: "batch", 2: "past_sequence + sequence"}
+                    common_outputs[f"present.{i}.value"] = {0: "batch", 2: "past_sequence + sequence"}
+        return common_outputs
+
+    def _generate_dummy_inputs_for_default_and_seq2seq_lm(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        encoder_inputs = self._generate_dummy_inputs_for_encoder_and_decoder(
+            tokenizer, batch_size, seq_length, is_pair, framework
+        )
+
+        # Generate decoder inputs
+        decoder_seq_length = seq_length if not self.use_past else 1
+        decoder_inputs = self._generate_dummy_inputs_for_encoder_and_decoder(
+            tokenizer, batch_size, decoder_seq_length, is_pair, framework
+        )
+        decoder_inputs = {f"decoder_{name}": tensor for name, tensor in decoder_inputs.items()}
+        common_inputs = dict(**encoder_inputs, **decoder_inputs)
+
+        if self.use_past:
+            if not is_torch_available():
+                raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.")
+            else:
+                import torch
+            batch, encoder_seq_length = common_inputs["input_ids"].shape
+            decoder_seq_length = common_inputs["decoder_input_ids"].shape[1]
+            num_encoder_attention_heads, num_decoder_attention_heads = self.num_attention_heads
+            encoder_shape = (
+                batch,
+                num_encoder_attention_heads,
+                encoder_seq_length,
+                self._config.hidden_size // num_encoder_attention_heads,
+            )
+            decoder_past_length = decoder_seq_length + 3
+            decoder_shape = (
+                batch,
+                num_decoder_attention_heads,
+                decoder_past_length,
+                self._config.hidden_size // num_decoder_attention_heads,
+            )
+
+            common_inputs["decoder_attention_mask"] = torch.cat(
+                [common_inputs["decoder_attention_mask"], torch.ones(batch, decoder_past_length)], dim=1
+            )
+
+            common_inputs["past_key_values"] = []
+            # If the number of encoder and decoder layers are present in the model configuration, both are considered
+            num_encoder_layers, num_decoder_layers = self.num_layers
+            min_num_layers = min(num_encoder_layers, num_decoder_layers)
+            max_num_layers = max(num_encoder_layers, num_decoder_layers) - min_num_layers
+            remaining_side_name = "encoder" if num_encoder_layers > num_decoder_layers else "decoder"
+
+            for _ in range(min_num_layers):
+                common_inputs["past_key_values"].append(
+                    (
+                        torch.zeros(decoder_shape),
+                        torch.zeros(decoder_shape),
+                        torch.zeros(encoder_shape),
+                        torch.zeros(encoder_shape),
+                    )
+                )
+            # TODO: test this.
+            shape = encoder_shape if remaining_side_name == "encoder" else decoder_shape
+            for _ in range(min_num_layers, max_num_layers):
+                common_inputs["past_key_values"].append((torch.zeros(shape), torch.zeros(shape)))
+        return common_inputs
+
+    def _generate_dummy_inputs_for_causal_lm(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        common_inputs = self._generate_dummy_inputs_for_encoder_and_decoder(
+            tokenizer, batch_size, seq_length, is_pair, framework
+        )
+
+        if self.use_past:
+            if not is_torch_available():
+                raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.")
+            else:
+                import torch
+            batch, seqlen = common_inputs["input_ids"].shape
+            # Not using the same length for past_key_values
+            past_key_values_length = seqlen + 2
+            num_encoder_layers, _ = self.num_layers
+            num_encoder_attention_heads, _ = self.num_attention_heads
+            past_shape = (
+                batch,
+                num_encoder_attention_heads,
+                past_key_values_length,
+                self._config.hidden_size // num_encoder_attention_heads,
+            )
+
+            common_inputs["attention_mask"] = torch.cat(
+                [common_inputs["attention_mask"], torch.ones(batch, past_key_values_length)], dim=1
+            )
+            common_inputs["past_key_values"] = [
+                (torch.zeros(past_shape), torch.zeros(past_shape)) for _ in range(num_encoder_layers)
+            ]
+        return common_inputs
+
+    # Copied from BartOnnxConfig._generate_dummy_inputs_for_sequence_classification_and_question_answering
+    # We renamed this function because Marian models do not have a sequence classification or question answering head
+    def _generate_dummy_inputs_for_encoder_and_decoder(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        # Copied from OnnxConfig.generate_dummy_inputs
+        # Did not use super(OnnxConfigWithPast, self).generate_dummy_inputs for code clarity.
+        # If dynamic axis (-1) we forward with a fixed dimension of 2 samples to avoid optimizations made by ONNX
+        batch_size = compute_effective_axis_dimension(
+            batch_size, fixed_dimension=OnnxConfig.DEFAULT_FIXED_BATCH, num_token_to_add=0
+        )
+
+        # If dynamic axis (-1) we forward with a fixed dimension of 8 tokens to avoid optimizations made by ONNX
+        token_to_add = tokenizer.num_special_tokens_to_add(is_pair)
+        seq_length = compute_effective_axis_dimension(
+            seq_length, fixed_dimension=OnnxConfig.DEFAULT_FIXED_SEQUENCE, num_token_to_add=token_to_add
+        )
+
+        # Generate dummy inputs according to compute batch and sequence
+        dummy_input = [" ".join([tokenizer.unk_token]) * seq_length] * batch_size
+        common_inputs = dict(tokenizer(dummy_input, return_tensors=framework))
+        return common_inputs
+
+    def generate_dummy_inputs(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        if self.task in ["default", "seq2seq-lm"]:
+            common_inputs = self._generate_dummy_inputs_for_default_and_seq2seq_lm(
+                tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework
+            )
+
+        else:
+            common_inputs = self._generate_dummy_inputs_for_causal_lm(
+                tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework
+            )
+
+        return common_inputs
+
+    # Copied from transformers.models.bart.configuration_bart.BartOnnxConfig._flatten_past_key_values_
+    def _flatten_past_key_values_(self, flattened_output, name, idx, t):
+        if self.task in ["default", "seq2seq-lm"]:
+            flattened_output = super()._flatten_past_key_values_(flattened_output, name, idx, t)
+        else:
+            flattened_output = super(OnnxSeq2SeqConfigWithPast, self)._flatten_past_key_values_(
+                flattened_output, name, idx, t
+            )

src/transformers/models/marian/tokenization_marian.py

@@ -310,7 +310,7 @@ def __setstate__(self, d: Dict) -> None:
         self.current_spm = self.spm_source
         self._setup_normalizer()
 
-    def num_special_tokens_to_add(self, **unused):
+    def num_special_tokens_to_add(self, *args, **kwargs):
         """Just EOS"""
         return 1
 

src/transformers/onnx/features.py

@@ -11,6 +11,7 @@
 from ..models.gpt_neo import GPTNeoOnnxConfig
 from ..models.layoutlm import LayoutLMOnnxConfig
 from ..models.longformer import LongformerOnnxConfig
+from ..models.marian import MarianOnnxConfig
 from ..models.mbart import MBartOnnxConfig
 from ..models.roberta import RobertaOnnxConfig
 from ..models.t5 import T5OnnxConfig
@@ -152,6 +153,15 @@ class FeaturesManager:
             "question-answering",
             onnx_config_cls=LongformerOnnxConfig,
         ),
+        "marian": supported_features_mapping(
+            "default",
+            "default-with-past",
+            "seq2seq-lm",
+            "seq2seq-lm-with-past",
+            "causal-lm",
+            "causal-lm-with-past",
+            onnx_config_cls=MarianOnnxConfig,
+        ),
         "roberta": supported_features_mapping(
             "default",
             "masked-lm",

tests/test_onnx_v2.py

@@ -188,6 +188,7 @@ def test_values_override(self):
     ("bart", "facebook/bart-base"),
     ("mbart", "sshleifer/tiny-mbart"),
     ("t5", "t5-small"),
+    ("marian", "Helsinki-NLP/opus-mt-en-de"),
 }