Support transformers 4.43 (#1971)

* fix bt bark test

* setup

* patch clip models for sd

* infer ort model dtype property from inputs dtypes

* patch all clip variants

* device setter

* bigger model for now

* fix device attribution

* onnx opset for owlvit and owlv2

* model dtype

* revert

* use model part dtype instead

* no need for dtype with diffusion pipelines

* revert

* fix clip text model with projection not outputting hidden states

* whisper generation

* fix whisper, support cache_position, and using transformers whisper generation loop

* style

* create cache position for merged decoder and fix test for non whisper speech to text

* typo

* conditioned cache position argument

* update whisper min transformers version

* compare whisper ort generation with transformers

* fix generation length for speech to text model type

* cache position in whisper only with dynamic axis decoder_sequence_length

* use minimal prepare_inputs_for_generation in ORTModelForSpeechSeq2Seq

* remove version restrictions on whisper

* comment

* fix

* simpler

---------

Co-authored-by: Ella Charlaix <80481427+echarlaix@users.noreply.github.com>

optimum/exporters/onnx/config.py

@@ -289,12 +289,10 @@ def inputs(self) -> Dict[str, Dict[int, str]]:
         if self._behavior is not ConfigBehavior.ENCODER:
             if self.use_past_in_inputs:
                 common_inputs["decoder_input_ids"] = {0: "batch_size"}
+                self.add_past_key_values(common_inputs, direction="inputs")
             else:
                 common_inputs["decoder_input_ids"] = {0: "batch_size", 1: "decoder_sequence_length"}
 
-            if self.use_past_in_inputs:
-                self.add_past_key_values(common_inputs, direction="inputs")
-
         if self._behavior is ConfigBehavior.DECODER:
             common_inputs["encoder_outputs"] = {0: "batch_size", 1: "encoder_sequence_length"}
 

optimum/exporters/onnx/model_configs.py

@@ -53,6 +53,7 @@
     NormalizedTextConfig,
     NormalizedTextConfigWithGQA,
     NormalizedVisionConfig,
+    check_if_transformers_greater,
     is_diffusers_available,
     logging,
 )
@@ -71,6 +72,7 @@
 )
 from .constants import ONNX_DECODER_MERGED_NAME, ONNX_DECODER_NAME, ONNX_DECODER_WITH_PAST_NAME
 from .model_patcher import (
+    CLIPModelPatcher,
     FalconModelPatcher,
     MistralModelPatcher,
     MusicgenModelPatcher,
@@ -913,10 +915,16 @@ def outputs(self) -> Dict[str, Dict[int, str]]:
 
         return common_outputs
 
+    def patch_model_for_export(
+        self,
+        model: Union["PreTrainedModel", "TFPreTrainedModel", "ModelMixin"],
+        model_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> "ModelPatcher":
+        return CLIPModelPatcher(self, model, model_kwargs=model_kwargs)
+
 
 class CLIPOnnxConfig(TextAndVisionOnnxConfig):
     NORMALIZED_CONFIG_CLASS = CLIPNormalizedConfig
-    DEFAULT_ONNX_OPSET = 14
 
     @property
     def inputs(self) -> Dict[str, Dict[int, str]]:
@@ -935,6 +943,13 @@ def outputs(self) -> Dict[str, Dict[int, str]]:
             "image_embeds": {0: "image_batch_size"},
         }
 
+    def patch_model_for_export(
+        self,
+        model: Union["PreTrainedModel", "TFPreTrainedModel", "ModelMixin"],
+        model_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> "ModelPatcher":
+        return CLIPModelPatcher(self, model, model_kwargs=model_kwargs)
+
 
 class SentenceTransformersCLIPOnnxConfig(CLIPOnnxConfig):
     @property
@@ -980,6 +995,13 @@ def outputs(self) -> Dict[str, Dict[int, str]]:
 
         return common_outputs
 
+    def patch_model_for_export(
+        self,
+        model: Union["PreTrainedModel", "TFPreTrainedModel", "ModelMixin"],
+        model_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> "ModelPatcher":
+        return CLIPModelPatcher(self, model, model_kwargs=model_kwargs)
+
 
 class CLIPTextOnnxConfig(CLIPTextWithProjectionOnnxConfig):
     @property
@@ -997,12 +1019,20 @@ def outputs(self) -> Dict[str, Dict[int, str]]:
     def generate_dummy_inputs(self, framework: str = "pt", **kwargs):
         dummy_inputs = super().generate_dummy_inputs(framework=framework, **kwargs)
 
+        # TODO: fix should be by casting inputs during inference and not export
         if framework == "pt":
             import torch
 
             dummy_inputs["input_ids"] = dummy_inputs["input_ids"].to(dtype=torch.int32)
         return dummy_inputs
 
+    def patch_model_for_export(
+        self,
+        model: Union["PreTrainedModel", "TFPreTrainedModel", "ModelMixin"],
+        model_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> "ModelPatcher":
+        return CLIPModelPatcher(self, model, model_kwargs=model_kwargs)
+
 
 class UNetOnnxConfig(VisionOnnxConfig):
     ATOL_FOR_VALIDATION = 1e-3
@@ -1135,6 +1165,9 @@ class OwlViTOnnxConfig(CLIPOnnxConfig):
     ATOL_FOR_VALIDATION = 1e-4
     MIN_TORCH_VERSION = version.parse("2.1")
 
+    # needs einsum operator support, available since opset 12
+    DEFAULT_ONNX_OPSET = 12
+
     def __init__(
         self,
         config: "PretrainedConfig",
@@ -1438,7 +1471,12 @@ def inputs(self) -> Dict[str, Dict[int, str]]:
             if self._behavior is not ConfigBehavior.DECODER:
                 common_inputs["input_features"] = {0: "batch_size"}  # Remove unnecessary dynamic axis.
 
-            if self._behavior is ConfigBehavior.DECODER and self.use_past_in_inputs is False:
+            if self._behavior is not ConfigBehavior.ENCODER and self.use_past_in_inputs:
+                if check_if_transformers_greater("4.43.0"):
+                    # since https://github.com/huggingface/transformers/pull/31166
+                    common_inputs["cache_position"] = {0: "decoder_sequence_length"}
+
+            if self._behavior is ConfigBehavior.DECODER and not self.use_past_in_inputs:
                 common_inputs["encoder_outputs"][1] = f"{common_inputs['encoder_outputs'][1]} / 2"
         return common_inputs
 

optimum/exporters/onnx/model_patcher.py

@@ -1138,3 +1138,20 @@ def __init__(
                 self._update_causal_mask_original = self._model.model._update_causal_mask
             else:
                 self._update_causal_mask_original = self._model._update_causal_mask
+
+
+class CLIPModelPatcher(ModelPatcher):
+    def __enter__(self):
+        super().__enter__()
+
+        if _transformers_version >= version.parse("4.43"):
+            from transformers.models.clip.modeling_clip import CLIPAttention, CLIPSdpaAttention
+
+            self.original_sdpa_forward, CLIPSdpaAttention.forward = CLIPSdpaAttention.forward, CLIPAttention.forward
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        super().__exit__(exc_type, exc_value, traceback)
+        if _transformers_version >= version.parse("4.43"):
+            from transformers.models.clip.modeling_clip import CLIPSdpaAttention
+
+            CLIPSdpaAttention.forward = self.original_sdpa_forward

optimum/exporters/utils.py

@@ -96,7 +96,7 @@ def _get_submodels_for_export_diffusion(
         pipeline, (StableDiffusionPipeline, StableDiffusionImg2ImgPipeline, StableDiffusionInpaintPipeline)
     )
     is_stable_diffusion_xl = isinstance(
-        pipeline, (StableDiffusionXLImg2ImgPipeline, StableDiffusionXLInpaintPipeline, StableDiffusionXLPipeline)
+        pipeline, (StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline, StableDiffusionXLInpaintPipeline)
     )
     is_latent_consistency_model = isinstance(
         pipeline, (LatentConsistencyModelPipeline, LatentConsistencyModelImg2ImgPipeline)
@@ -117,10 +117,11 @@ def _get_submodels_for_export_diffusion(
     models_for_export = {}
 
     # Text encoder
-    if pipeline.text_encoder is not None:
+    text_encoder = getattr(pipeline, "text_encoder", None)
+    if text_encoder is not None:
         if is_stable_diffusion_xl:
-            pipeline.text_encoder.config.output_hidden_states = True
-        models_for_export["text_encoder"] = pipeline.text_encoder
+            text_encoder.config.output_hidden_states = True
+        models_for_export["text_encoder"] = text_encoder
 
     # U-NET
     # ONNX export of torch.nn.functional.scaled_dot_product_attention not supported for < v2.1.0
@@ -151,6 +152,7 @@ def _get_submodels_for_export_diffusion(
     text_encoder_2 = getattr(pipeline, "text_encoder_2", None)
     if text_encoder_2 is not None:
         text_encoder_2.config.output_hidden_states = True
+        text_encoder_2.text_model.config.output_hidden_states = True
         models_for_export["text_encoder_2"] = text_encoder_2
 
     return models_for_export

optimum/onnxruntime/base.py

@@ -24,6 +24,7 @@
 
 from ..utils import NormalizedConfigManager
 from ..utils.logging import warn_once
+from .io_binding import TypeHelper
 from .modeling_ort import ORTModel
 from .utils import get_ordered_input_names, logging
 
@@ -62,6 +63,20 @@ def __init__(
     def device(self):
         return self.parent_model.device
 
+    @property
+    def dtype(self):
+        for dtype in self.input_dtypes.values():
+            torch_dtype = TypeHelper.ort_type_to_torch_type(dtype)
+            if torch_dtype.is_floating_point:
+                return torch_dtype
+
+        for dtype in self.output_dtypes.values():
+            torch_dtype = TypeHelper.ort_type_to_torch_type(dtype)
+            if torch_dtype.is_floating_point:
+                return torch_dtype
+
+        return None
+
     @abstractmethod
     def forward(self, *args, **kwargs):
         pass
@@ -220,6 +235,7 @@ def forward(
         encoder_attention_mask: Optional[torch.LongTensor] = None,
         past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
         labels: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.Tensor] = None,
         use_cache_branch: None = None,
     ) -> Seq2SeqLMOutput:
         # Adding use_cache_branch in the signature here is just a hack for IO Binding
@@ -236,8 +252,8 @@ def forward(
         # no-ops if merged decoder is not used
         use_merged_no_cache = past_key_values is None and self.parent_model.use_merged
         use_merged_cache = past_key_values is not None and self.parent_model.use_merged
-        use_cache_branch_tensor, past_key_values = self.prepare_inputs_for_merged(
-            input_ids, past_key_values, use_torch=use_torch
+        use_cache_branch_tensor, past_key_values, cache_position = self.prepare_inputs_for_merged(
+            input_ids, past_key_values, cache_position, use_torch=use_torch
         )
 
         if self.parent_model.use_io_binding:
@@ -274,6 +290,9 @@ def forward(
             if use_cache_branch_tensor is not None:
                 model_inputs.append(use_cache_branch_tensor)
 
+            if "cache_position" in self.input_names:
+                model_inputs.append(cache_position)
+
             io_binding, output_shapes, output_buffers = self.parent_model._prepare_io_binding(
                 self.session,
                 *model_inputs,
@@ -346,6 +365,7 @@ def forward(
                 "decoder_attention_mask": decoder_attention_mask,
                 "encoder_attention_mask": encoder_attention_mask,
                 "use_cache_branch": use_cache_branch_tensor,
+                "cache_position": cache_position,
                 "labels": labels,
             }
             if past_key_values is not None:
@@ -405,20 +425,20 @@ def forward(
 
     def prepare_inputs_for_merged(
         self,
-        input_ids: Union[None, torch.LongTensor, np.ndarray],
-        past_key_values: Union[None, Tuple[torch.FloatTensor], Tuple[np.ndarray]],
+        input_ids: Optional[Union[torch.LongTensor, np.ndarray]],
+        past_key_values: Optional[Tuple[Union[torch.FloatTensor, np.ndarray]]],
+        cache_position: Optional[Union[torch.Tensor, np.ndarray]],
         use_torch: bool,
     ):
+        constructor = torch if use_torch is True else np
+
         if self.parent_model.use_merged:
-            constructor = torch if use_torch is True else np
             # Uses without/with branch of a merged decoder depending on whether real past key values are passed
-            use_cache_branch = constructor.full((1,), past_key_values is not None)
+            use_cache_branch_tensor = constructor.full((1,), past_key_values is not None)
+            if use_torch and use_cache_branch_tensor is not None:
+                use_cache_branch_tensor = use_cache_branch_tensor.to(self.device)
         else:
-            # Uses separate decoders
-            use_cache_branch = None
-
-        if use_torch and use_cache_branch is not None:
-            use_cache_branch = use_cache_branch.to(self.device)
+            use_cache_branch_tensor = None
 
         # Generate dummy past for the first forward if uses a merged decoder
         if self.parent_model.use_merged and past_key_values is None:
@@ -434,7 +454,13 @@ def prepare_inputs_for_merged(
 
             past_key_values = tuple(key_or_value for _ in range(len(self.key_value_input_names)))
 
-        return use_cache_branch, past_key_values
+        # Generate dummy position cache for the first forward if uses a merged decoder
+        if self.parent_model.use_merged and cache_position is None:
+            cache_position = constructor.zeros((1,), dtype=constructor.int64)
+            if use_torch is True:
+                cache_position = cache_position.to(self.device)
+
+        return use_cache_branch_tensor, past_key_values, cache_position
 
 
 class ORTDecoder(ORTDecoderForSeq2Seq):

optimum/onnxruntime/modeling_diffusion.py

@@ -452,10 +452,14 @@ def to(self, device: Union[torch.device, str, int]):
         Returns:
             `ORTModel`: the model placed on the requested device.
         """
+
         device, provider_options = parse_device(device)
         provider = get_provider_for_device(device)
         validate_provider_availability(provider)  # raise error if the provider is not available
-        self.device = device
+
+        if device.type == "cuda" and self.providers[0] == "TensorrtExecutionProvider":
+            return self
+
         self.vae_decoder.session.set_providers([provider], provider_options=[provider_options])
         self.text_encoder.session.set_providers([provider], provider_options=[provider_options])
         self.unet.session.set_providers([provider], provider_options=[provider_options])
@@ -464,6 +468,8 @@ def to(self, device: Union[torch.device, str, int]):
             self.vae_encoder.session.set_providers([provider], provider_options=[provider_options])
 
         self.providers = self.vae_decoder.session.get_providers()
+        self._device = device
+
         return self
 
     @classmethod

optimum/onnxruntime/modeling_ort.py

@@ -276,7 +276,24 @@ def __init__(
 
         self._ordered_input_names = get_ordered_input_names(self.input_names.keys(), func=self.forward)
 
-    # TODO: why do we make device a property since we are only access the value, and do not do any check when setting the value?
+    @property
+    def dtype(self) -> torch.dtype:
+        """
+        `torch.dtype`: The dtype of the model.
+        """
+
+        for dtype in self.input_dtypes.values():
+            torch_dtype = TypeHelper.ort_type_to_torch_type(dtype)
+            if torch_dtype.is_floating_point:
+                return torch_dtype
+
+        for dtype in self.output_dtypes.values():
+            torch_dtype = TypeHelper.ort_type_to_torch_type(dtype)
+            if torch_dtype.is_floating_point:
+                return torch_dtype
+
+        return None
+
     @property
     def device(self) -> torch.device:
         """
@@ -286,8 +303,8 @@ def device(self) -> torch.device:
         return self._device
 
     @device.setter
-    def device(self, value: torch.device):
-        self._device = value
+    def device(self, **kwargs):
+        raise AttributeError("The device attribute is read-only, please use the `to` method to change the device.")
 
     @property
     def use_io_binding(self):
@@ -309,13 +326,13 @@ def to(self, device: Union[torch.device, str, int]):
         Returns:
             `ORTModel`: the model placed on the requested device.
         """
+
         device, provider_options = parse_device(device)
 
         if device.type == "cuda" and self.providers[0] == "TensorrtExecutionProvider":
             return self
 
-        self.device = device
-        provider = get_provider_for_device(self.device)
+        provider = get_provider_for_device(device)
         validate_provider_availability(provider)  # raise error if the provider is not available
 
         # IOBinding is only supported for CPU and CUDA Execution Providers.
@@ -331,6 +348,7 @@ def to(self, device: Union[torch.device, str, int]):
 
         self.model.set_providers([provider], provider_options=[provider_options])
         self.providers = self.model.get_providers()
+        self._device = device
 
         return self