Fix gradient checkpointing bug in gptneox (huggingface#21815)

* Fix gradient checkpointing bug in gptneox * Remove use_cache block
pioliverse · Feb 27, 2023 · 7006a85 · 7006a85
1 parent 38f6caa
commit 7006a85
Showing 1 changed file with 7 additions and 5 deletions.
diff --git a/src/transformers/models/gpt_neox/modeling_gpt_neox.py b/src/transformers/models/gpt_neox/modeling_gpt_neox.py
@@ -500,6 +500,13 @@ def forward(
 
         hidden_states = inputs_embeds
 
+        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_attentions = () if output_attentions else None
         all_hidden_states = () if output_hidden_states else None
@@ -508,11 +515,6 @@ 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):