diff --git a/src/transformers/models/mt5/modeling_mt5.py b/src/transformers/models/mt5/modeling_mt5.py
index 951a68cb7678..e2235fc69055 100644
--- a/src/transformers/models/mt5/modeling_mt5.py
+++ b/src/transformers/models/mt5/modeling_mt5.py
@@ -978,6 +978,13 @@ def forward(
         else:
             encoder_extended_attention_mask = None
 
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
         # Prepare head mask if needed
         head_mask = self.get_head_mask(head_mask, self.config.num_layers)
         cross_attn_head_mask = self.get_head_mask(cross_attn_head_mask, self.config.num_layers)
@@ -1015,11 +1022,6 @@ def forward(
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             if self.gradient_checkpointing and self.training:
-                if use_cache:
-                    logger.warning_once(
-                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                    )
-                    use_cache = False
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):
diff --git a/src/transformers/models/t5/modeling_t5.py b/src/transformers/models/t5/modeling_t5.py
index 976899863119..19cb83dac352 100644
--- a/src/transformers/models/t5/modeling_t5.py
+++ b/src/transformers/models/t5/modeling_t5.py
@@ -1007,6 +1007,13 @@ def forward(
         else:
             encoder_extended_attention_mask = None
 
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
         # Prepare head mask if needed
         head_mask = self.get_head_mask(head_mask, self.config.num_layers)
         cross_attn_head_mask = self.get_head_mask(cross_attn_head_mask, self.config.num_layers)
@@ -1044,11 +1051,6 @@ def forward(
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             if self.gradient_checkpointing and self.training:
-                if use_cache:
-                    logger.warning_once(
-                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                    )
-                    use_cache = False
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):