Loosen output shape restrictions on GPT-style models (#25188)

* Loosen output shape restrictions on GPT-style models

* Use more self-explanatory variables

* Revert "Use more self-explanatory variables"

This reverts commit 5fd9ab3911.

src/transformers/models/decision_transformer/modeling_decision_transformer.py

@@ -603,7 +603,7 @@ class DecisionTransformerGPT2Model(DecisionTransformerGPT2PreTrainedModel):
 
         hidden_states = self.drop(hidden_states)
 
-        output_shape = input_shape + (hidden_states.size(-1),)
+        output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
 
         if self.gradient_checkpointing and self.training:
             if use_cache:

src/transformers/models/gpt2/modeling_gpt2.py

@@ -849,7 +849,7 @@ class GPT2Model(GPT2PreTrainedModel):
 
         hidden_states = self.drop(hidden_states)
 
-        output_shape = input_shape + (hidden_states.size(-1),)
+        output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
 
         if self.gradient_checkpointing and self.training:
             if use_cache:

src/transformers/models/gpt_neo/modeling_gpt_neo.py

@@ -588,7 +588,7 @@ class GPTNeoModel(GPTNeoPreTrainedModel):
 
         hidden_states = self.drop(hidden_states)
 
-        output_shape = input_shape + (hidden_states.size(-1),)
+        output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
 
         if self.gradient_checkpointing and self.training:
             if use_cache:

src/transformers/models/gptj/modeling_gptj.py

@@ -641,7 +641,7 @@ class GPTJModel(GPTJPreTrainedModel):
 
         hidden_states = self.drop(hidden_states)
 
-        output_shape = input_shape + (hidden_states.size(-1),)
+        output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
 
         if self.gradient_checkpointing and self.training:
             if use_cache: