Fx with meta (#16836)

* Add meta proxy

* Uses meta data to trace data dependent control-flow

* Remove commented class

* Handles torch creating functions

* Added type annotation to fix tracing

* Tracing works for everything but T5 and GPT-J

* Almost all previously supported models pass

* All architectures can be traced except T5

* Intermediate commit to have a trace of the comparison operators for HFProxy

* Everything works, except loss computation

* Everything works

* Removed unused import

* Overriden methods do not use underlying ops (linear and torch.matmul), and model attributes are copied to the traced version

* Fix torch_matmul_override

* Change attributes reference to deepcopy

* Remove breakpoint and add torch_index_override

* Small fix

* Fix typo

* Replace asserts by explicit exceptions

src/transformers/models/electra/modeling_electra.py

@@ -850,7 +850,7 @@ class ElectraModel(ElectraPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, BaseModelOutputWithCrossAttentions]:
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithCrossAttentions]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
@@ -985,7 +985,7 @@ class ElectraForSequenceClassification(ElectraPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, SequenceClassifierOutput]:
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
             Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
@@ -1075,7 +1075,7 @@ class ElectraForPreTraining(ElectraPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, ElectraForPreTrainingOutput]:
+    ) -> Union[Tuple[torch.Tensor], ElectraForPreTrainingOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for computing the ELECTRA loss. Input should be a sequence of tokens (see `input_ids` docstring)
@@ -1197,7 +1197,7 @@ class ElectraForMaskedLM(ElectraPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, MaskedLMOutput]:
+    ) -> Union[Tuple[torch.Tensor], MaskedLMOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
@@ -1283,7 +1283,7 @@ class ElectraForTokenClassification(ElectraPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, TokenClassifierOutput]:
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
@@ -1368,7 +1368,7 @@ class ElectraForQuestionAnswering(ElectraPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+    ) -> Union[Tuple[torch.Tensor], QuestionAnsweringModelOutput]:
         r"""
         start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
             Labels for position (index) of the start of the labelled span for computing the token classification loss.
@@ -1469,7 +1469,7 @@ class ElectraForMultipleChoice(ElectraPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, MultipleChoiceModelOutput]:
+    ) -> Union[Tuple[torch.Tensor], MultipleChoiceModelOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
             Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
@@ -1564,7 +1564,7 @@ class ElectraForCausalLM(ElectraPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
+    ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
         r"""
         encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
             Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if

src/transformers/models/gpt_neo/modeling_gpt_neo.py

@@ -508,7 +508,7 @@ class GPTNeoModel(GPTNeoPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
@@ -727,7 +727,7 @@ class GPTNeoForCausalLM(GPTNeoPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
+    ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
@@ -842,7 +842,7 @@ class GPTNeoForSequenceClassification(GPTNeoPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutputWithPast]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
             Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,

src/transformers/models/roberta/modeling_roberta.py

@@ -919,7 +919,7 @@ class RobertaForCausalLM(RobertaPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
+    ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
         r"""
         encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
             Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
@@ -1080,7 +1080,7 @@ class RobertaForMaskedLM(RobertaPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, MaskedLMOutput]:
+    ) -> Union[Tuple[torch.Tensor], MaskedLMOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
@@ -1193,7 +1193,7 @@ class RobertaForSequenceClassification(RobertaPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, SequenceClassifierOutput]:
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
             Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
@@ -1290,7 +1290,7 @@ class RobertaForMultipleChoice(RobertaPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, MultipleChoiceModelOutput]:
+    ) -> Union[Tuple[torch.Tensor], MultipleChoiceModelOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
             Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
@@ -1390,7 +1390,7 @@ class RobertaForTokenClassification(RobertaPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, TokenClassifierOutput]:
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
@@ -1496,7 +1496,7 @@ class RobertaForQuestionAnswering(RobertaPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+    ) -> Union[Tuple[torch.Tensor], QuestionAnsweringModelOutput]:
         r"""
         start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
             Labels for position (index) of the start of the labelled span for computing the token classification loss.