Fix TF Causal LM models' returned logits (#15256)

* Fix TF Causal LM models' returned logits

* Fix expected shape in the tests

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

src/transformers/models/bert/modeling_tf_bert.py

@@ -1542,9 +1542,9 @@ class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss):
 
         if inputs["labels"] is not None:
             # shift labels to the left and cut last logit token
-            logits = logits[:, :-1]
+            shifted_logits = logits[:, :-1]
             labels = inputs["labels"][:, 1:]
-            loss = self.hf_compute_loss(labels=labels, logits=logits)
+            loss = self.hf_compute_loss(labels=labels, logits=shifted_logits)
 
         if not inputs["return_dict"]:
             output = (logits,) + outputs[2:]

src/transformers/models/ctrl/modeling_tf_ctrl.py

@@ -735,9 +735,9 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
         loss = None
         if inputs["labels"] is not None:
             # shift labels to the left and cut last logit token
-            logits = logits[:, :-1]
+            shifted_logits = logits[:, :-1]
             labels = inputs["labels"][:, 1:]
-            loss = self.hf_compute_loss(labels, logits)
+            loss = self.hf_compute_loss(labels, shifted_logits)
 
         if not inputs["return_dict"]:
             output = (logits,) + transformer_outputs[1:]

src/transformers/models/gpt2/modeling_tf_gpt2.py

@@ -949,9 +949,9 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel, TFCausalLanguageModelingLoss):
         loss = None
         if inputs["labels"] is not None:
             # shift labels to the left and cut last logit token
-            logits = logits[:, :-1]
+            shifted_logits = logits[:, :-1]
             labels = inputs["labels"][:, 1:]
-            loss = self.hf_compute_loss(labels, logits)
+            loss = self.hf_compute_loss(labels, shifted_logits)
 
         if not inputs["return_dict"]:
             output = (logits,) + transformer_outputs[1:]

src/transformers/models/openai/modeling_tf_openai.py

@@ -656,9 +656,9 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel, TFCausalLanguageModelin
         loss = None
         if inputs["labels"] is not None:
             # shift labels to the left and cut last logit token
-            logits = logits[:, :-1]
+            shifted_logits = logits[:, :-1]
             labels = inputs["labels"][:, 1:]
-            loss = self.hf_compute_loss(labels, logits)
+            loss = self.hf_compute_loss(labels, shifted_logits)
 
         if not inputs["return_dict"]:
             output = (logits,) + transformer_outputs[1:]

src/transformers/models/rembert/modeling_tf_rembert.py

@@ -1275,9 +1275,9 @@ class TFRemBertForCausalLM(TFRemBertPreTrainedModel, TFCausalLanguageModelingLos
 
         if inputs["labels"] is not None:
             # shift labels to the left and cut last logit token
-            logits = logits[:, :-1]
+            shifted_logits = logits[:, :-1]
             labels = inputs["labels"][:, 1:]
-            loss = self.hf_compute_loss(labels=labels, logits=logits)
+            loss = self.hf_compute_loss(labels=labels, logits=shifted_logits)
 
         if not inputs["return_dict"]:
             output = (logits,) + outputs[2:]

src/transformers/models/roberta/modeling_tf_roberta.py

@@ -1310,9 +1310,9 @@ class TFRobertaForCausalLM(TFRobertaPreTrainedModel, TFCausalLanguageModelingLos
 
         if inputs["labels"] is not None:
             # shift labels to the left and cut last logit token
-            logits = logits[:, :-1]
+            shifted_logits = logits[:, :-1]
             labels = inputs["labels"][:, 1:]
-            loss = self.hf_compute_loss(labels=labels, logits=logits)
+            loss = self.hf_compute_loss(labels=labels, logits=shifted_logits)
 
         if not inputs["return_dict"]:
             output = (logits,) + outputs[2:]

src/transformers/models/roformer/modeling_tf_roformer.py

@@ -1035,9 +1035,9 @@ class TFRoFormerForCausalLM(TFRoFormerPreTrainedModel, TFCausalLanguageModelingL
 
         if inputs["labels"] is not None:
             # shift labels to the left and cut last logit token
-            logits = logits[:, :-1]
+            shifted_logits = logits[:, :-1]
             labels = inputs["labels"][:, 1:]
-            loss = self.hf_compute_loss(labels=labels, logits=logits)
+            loss = self.hf_compute_loss(labels=labels, logits=shifted_logits)
 
         if not inputs["return_dict"]:
             output = (logits,) + outputs[2:]

templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_tf_{{cookiecutter.lowercase_modelname}}.py

@@ -1262,9 +1262,9 @@ class TF{{cookiecutter.camelcase_modelname}}ForCausalLM(TF{{cookiecutter.camelca
 
         if inputs["labels"] is not None:
             # shift labels to the left and cut last logit token
-            logits = logits[:, :-1]
+            shifted_logits = logits[:, :-1]
             labels = inputs["labels"][:, 1:]
-            loss = self.hf_compute_loss(labels=labels, logits=logits)
+            loss = self.hf_compute_loss(labels=labels, logits=shifted_logits)
 
         if not inputs["return_dict"]:
             output = (logits,) + outputs[2:]

tests/test_modeling_tf_encoder_decoder.py

@@ -240,7 +240,7 @@ class TFEncoderDecoderMixin:
         assert "loss" in outputs_encoder_decoder
 
         batch_size, seq_len = decoder_input_ids.shape
-        expected_shape = (batch_size, seq_len - 1, decoder_config.vocab_size)
+        expected_shape = (batch_size, seq_len, decoder_config.vocab_size)
         self.assertEqual(outputs_encoder_decoder["logits"].shape, expected_shape)
         self.assertEqual(
             outputs_encoder_decoder["encoder_last_hidden_state"].shape, (input_ids.shape + (config.hidden_size,))

tests/test_modeling_tf_vision_encoder_decoder.py

@@ -231,7 +231,7 @@ class TFVisionEncoderDecoderMixin:
         self.assertIn("loss", outputs_encoder_decoder)
 
         batch_size, seq_len = decoder_input_ids.shape
-        expected_shape = (batch_size, seq_len - 1, decoder_config.vocab_size)
+        expected_shape = (batch_size, seq_len, decoder_config.vocab_size)
         self.assertEqual(outputs_encoder_decoder["logits"].shape, expected_shape)
         self.assertEqual(outputs_encoder_decoder["encoder_last_hidden_state"].shape[0], pixel_values.shape[0])
         self.assertEqual(outputs_encoder_decoder["encoder_last_hidden_state"].shape[-1], config.hidden_size)