Generate - update cookie cutters to not initialize cache with trainin…

…g and gradient checkpointing (huggingface#21759)
justin-barton · Feb 24, 2023 · 440f397 · 440f397
1 parent 087436c
commit 440f397
Showing 1 changed file with 13 additions and 15 deletions.
diff --git a/...tter-template-{{cookiecutter.modelname}}/modeling_{{cookiecutter.lowercase_modelname}}.py b/...tter-template-{{cookiecutter.modelname}}/modeling_{{cookiecutter.lowercase_modelname}}.py
@@ -184,7 +184,7 @@ def forward(
                 token_type_ids = buffered_token_type_ids_expanded
             else:
                 token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
-                
+
         if inputs_embeds is None:
             inputs_embeds = self.word_embeddings(input_ids)
         token_type_embeddings = self.token_type_embeddings(token_type_ids)
@@ -525,11 +525,17 @@ def forward(
         output_hidden_states=False,
         return_dict=True,
     ):
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+            )
+            use_cache = False
+
         all_hidden_states = () if output_hidden_states else None
         all_self_attentions = () if output_attentions else None
         all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
-
         next_decoder_cache = () if use_cache else None
+
         for i, layer_module in enumerate(self.layer):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
@@ -538,13 +544,6 @@ def forward(
             past_key_value = past_key_values[i] 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):
                         return module(*inputs, past_key_value, output_attentions)
@@ -2024,7 +2023,7 @@ def _init_weights(self, module):
             module.weight.data.normal_(mean=0.0, std=std)
             if module.padding_idx is not None:
                 module.weight.data[module.padding_idx].zero_()
-    
+
     def _set_gradient_checkpointing(self, module, value=False):
         if isinstance(module, ({{cookiecutter.camelcase_modelname}}Decoder, {{cookiecutter.camelcase_modelname}}Encoder)):
             module.gradient_checkpointing = value
@@ -2525,6 +2524,10 @@ def forward(
         hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
 
         # decoder layers
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning("`use_cache = True` is incompatible with gradient checkpointing`. Setting `use_cache = False`...")
+            use_cache = False
+
         all_hidden_states = () if output_hidden_states else None
         all_self_attns = () if output_attentions else None
         all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
@@ -2547,11 +2550,6 @@ 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):
                         # None for past_key_value