[GPT2] Fix gradient checkpointing bug

src/transformers/models/decision_transformer/modeling_decision_transformer.py

@@ -607,6 +607,13 @@ def forward(
 
         output_shape = input_shape + (hidden_states.size(-1),)
 
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
         presents = () if use_cache else None
         all_self_attentions = () if output_attentions else None
         all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
@@ -627,11 +634,6 @@ def forward(
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             if self.gradient_checkpointing and self.training:
-                if use_cache:
-                    logger.warning(
-                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                    )
-                    use_cache = False
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):

src/transformers/models/gpt2/modeling_gpt2.py

@@ -851,6 +851,13 @@ def forward(
 
         output_shape = input_shape + (hidden_states.size(-1),)
 
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
         presents = () if use_cache else None
         all_self_attentions = () if output_attentions else None
         all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
@@ -871,11 +878,6 @@ def forward(
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             if self.gradient_checkpointing and self.training:
-                if use_cache:
-                    logger.warning(
-                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                    )
-                    use_cache = False
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):

src/transformers/models/prophetnet/modeling_prophetnet.py

@@ -1569,6 +1569,14 @@ def forward(
         all_main_stream_attns = () if output_attentions else None
         all_ngram_stream_attns = () if output_attentions else None
         all_cross_attns = () if output_attentions and self.config.add_cross_attention else None
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
         present_key_values = () if use_cache else None
 
         # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
@@ -1588,11 +1596,6 @@ def forward(
             past_key_value = past_key_values[idx] if past_key_values is not None else None
 
             if self.gradient_checkpointing and self.training:
-                if use_cache:
-                    logger.warning(
-                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                    )
-                    use_cache = False
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):

src/transformers/models/xlm_prophetnet/modeling_xlm_prophetnet.py

@@ -1592,6 +1592,14 @@ def forward(
         all_main_stream_attns = () if output_attentions else None
         all_ngram_stream_attns = () if output_attentions else None
         all_cross_attns = () if output_attentions and self.config.add_cross_attention else None
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
         present_key_values = () if use_cache else None
 
         # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
@@ -1611,11 +1619,6 @@ def forward(
             past_key_value = past_key_values[idx] if past_key_values is not None else None
 
             if self.gradient_checkpointing and self.training:
-                if use_cache:
-                    logger.warning(
-                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                    )
-                    use_cache = False
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):