fix(config): passing gradient_checkpoint_kwargs

README.md

@@ -843,7 +843,7 @@ group_by_length: false
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing
 # gradient_checkpointing_kwargs:
-#   use_reentrant: false
+#   use_reentrant: true
 
 # Stop training after this many evaluation losses have increased in a row
 # https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback

src/axolotl/core/trainer_builder.py

@@ -837,10 +837,6 @@ def build(self, total_num_steps):
                 training_arguments_kwargs[
                     "gradient_checkpointing_kwargs"
                 ] = self.cfg.gradient_checkpointing_kwargs
-            else:
-                training_arguments_kwargs["gradient_checkpointing_kwargs"] = {
-                    "use_reentrant": False
-                }
         if self.cfg.fsdp:
             training_arguments_kwargs["fsdp"] = self.cfg.fsdp
             if self.cfg.fsdp_config:

src/axolotl/utils/models.py

@@ -888,7 +888,9 @@ def load_and_quantize_parallel(name_param, model, **kwargs):
 
     if cfg.adapter in ["lora", "qlora"]:
         if cfg.gradient_checkpointing:
-            model.gradient_checkpointing_enable()
+            model.gradient_checkpointing_enable(
+                gradient_checkpointing_kwargs=cfg.gradient_checkpointing_kwargs
+            )
         if (
             cfg.load_in_8bit or cfg.load_in_4bit
         ) and not skip_prepare_model_for_kbit_training: