[ML] Add support for MPNet PyTorch models

eland/ml/pytorch/transformers.py

@@ -48,6 +48,7 @@ SUPPORTED_TASK_TYPES = {
 SUPPORTED_TASK_TYPES_NAMES = ", ".join(sorted(SUPPORTED_TASK_TYPES))
 SUPPORTED_TOKENIZERS = (
     transformers.BertTokenizer,
+    transformers.MPNetTokenizer,
     transformers.DPRContextEncoderTokenizer,
     transformers.DPRQuestionEncoderTokenizer,
     transformers.DistilBertTokenizer,
@@ -96,7 +97,7 @@ class _DistilBertWrapper(nn.Module):  # type: ignore
         return self._model(input_ids=input_ids, attention_mask=attention_mask)
 
 
-class _SentenceTransformerWrapper(nn.Module):  # type: ignore
+class _SentenceTransformerWrapperModule(nn.Module):  # type: ignore
     """
     A wrapper around sentence-transformer models to provide pooling,
     normalization and other graph layers that are not defined in the base
@@ -108,6 +109,7 @@ class _SentenceTransformerWrapper(nn.Module):  # type: ignore
         self._hf_model = model
         self._st_model = SentenceTransformer(model.config.name_or_path)
         self._output_key = output_key
+        self.config = model.config
 
         self._remove_pooling_layer()
         self._replace_transformer_layer()
@@ -118,7 +120,10 @@ class _SentenceTransformerWrapper(nn.Module):  # type: ignore
     ) -> Optional[Any]:
         if model_id.startswith("sentence-transformers/"):
             model = AutoModel.from_pretrained(model_id, torchscript=True)
-            return _SentenceTransformerWrapper(model, output_key)
+            if isinstance(model.config, transformers.MPNetConfig):
+                return _MPNetSentenceTransformerWrapper(model, output_key)
+            else:
+                return _SentenceTransformerWrapper(model, output_key)
         else:
             return None
 
@@ -144,6 +149,11 @@ class _SentenceTransformerWrapper(nn.Module):  # type: ignore
 
         self._st_model._modules["0"].auto_model = self._hf_model
 
+
+class _SentenceTransformerWrapper(_SentenceTransformerWrapperModule):
+    def __init__(self, model: PreTrainedModel, output_key: str = DEFAULT_OUTPUT_KEY):
+        super().__init__(model=model, output_key=output_key)
+
     def forward(
         self,
         input_ids: Tensor,
@@ -167,6 +177,19 @@ class _SentenceTransformerWrapper(nn.Module):  # type: ignore
         return self._st_model(inputs)[self._output_key]
 
 
+class _MPNetSentenceTransformerWrapper(_SentenceTransformerWrapperModule):
+    def __init__(self, model: PreTrainedModel, output_key: str = DEFAULT_OUTPUT_KEY):
+        super().__init__(model=model, output_key=output_key)
+
+    def forward(self, input_ids: Tensor, attention_mask: Tensor) -> Tensor:
+        """Wrap the input and output to conform to the native process interface."""
+        inputs = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return self._st_model(inputs)[self._output_key]
+
+
 class _DPREncoderWrapper(nn.Module):  # type: ignore
     """
     AutoModel loading does not work for DPRContextEncoders, this only exists as
@@ -186,6 +209,7 @@ class _DPREncoderWrapper(nn.Module):  # type: ignore
     ):
         super().__init__()
         self._model = model
+        self.config = model.config
 
     @staticmethod
     def from_pretrained(model_id: str) -> Optional[Any]:
@@ -232,7 +256,7 @@ class _TraceableModel(ABC):
         tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
         model: Union[
             PreTrainedModel,
-            _SentenceTransformerWrapper,
+            _SentenceTransformerWrapperModule,
             _DPREncoderWrapper,
             _DistilBertWrapper,
         ],
@@ -250,13 +274,19 @@ class _TraceableModel(ABC):
         self._model.eval()
 
         inputs = self._prepare_inputs()
-        position_ids = torch.arange(inputs["input_ids"].size(1), dtype=torch.long)
 
         # Add params when not provided by the tokenizer (e.g. DistilBERT), to conform to BERT interface
         if "token_type_ids" not in inputs:
             inputs["token_type_ids"] = torch.zeros(
                 inputs["input_ids"].size(1), dtype=torch.long
             )
+        if isinstance(self._model.config, transformers.MPNetConfig):
+            return torch.jit.trace(
+                self._model,
+                (inputs["input_ids"], inputs["attention_mask"]),
+            )
+
+        position_ids = torch.arange(inputs["input_ids"].size(1), dtype=torch.long)
 
         return torch.jit.trace(
             self._model,
@@ -369,10 +399,15 @@ class TransformerModel:
         }
 
     def _create_config(self) -> Dict[str, Any]:
+        tokenizer_type = (
+            "mpnet"
+            if isinstance(self._tokenizer, transformers.MPNetTokenizer)
+            else "bert"
+        )
         inference_config: Dict[str, Dict[str, Any]] = {
             self._task_type: {
                 "tokenization": {
-                    "bert": {
+                    tokenizer_type: {
                         "do_lower_case": getattr(
                             self._tokenizer, "do_lower_case", False
                         ),
@@ -386,7 +421,7 @@ class TransformerModel:
                 self._model_id
             )
             if max_sequence_length:
-                inference_config[self._task_type]["tokenization"]["bert"][
+                inference_config[self._task_type]["tokenization"][tokenizer_type][
                     "max_sequence_length"
                 ] = max_sequence_length
 
@@ -427,7 +462,7 @@ class TransformerModel:
             return _TraceableTextClassificationModel(self._tokenizer, model)
 
         elif self._task_type == "text_embedding":
-            model = _SentenceTransformerWrapper.from_pretrained(self._model_id)
+            model = _SentenceTransformerWrapperModule.from_pretrained(self._model_id)
             if not model:
                 model = _DPREncoderWrapper.from_pretrained(self._model_id)
             if not model:

tests/ml/pytorch/test_pytorch_model_pytest.py

@@ -17,6 +17,7 @@
 import tempfile
 
 import pytest
+from elasticsearch import NotFoundError
 
 try:
     import sklearn  # noqa: F401
@@ -68,8 +69,11 @@ def setup_and_tear_down():
     yield
     for model_id, _, _, _ in TEXT_PREDICTION_MODELS:
         model = PyTorchModel(ES_TEST_CLIENT, model_id.replace("/", "__").lower()[:64])
-        model.stop()
-        model.delete()
+        try:
+            model.stop()
+            model.delete()
+        except NotFoundError:
+            pass
 
 
 def download_model_and_start_deployment(tmp_dir, quantize, model_id, task):
@@ -77,8 +81,11 @@ def download_model_and_start_deployment(tmp_dir, quantize, model_id, task):
     tm = TransformerModel(model_id, task, quantize)
     model_path, config_path, vocab_path = tm.save(tmp_dir)
     ptm = PyTorchModel(ES_TEST_CLIENT, tm.elasticsearch_model_id())
-    ptm.stop()
-    ptm.delete()
+    try:
+        ptm.stop()
+        ptm.delete()
+    except NotFoundError:
+        pass
     print(f"Importing model: {ptm.model_id}")
     ptm.import_model(model_path, config_path, vocab_path)
     ptm.start()