fix gradient issue for qlora on seq2seq. (#897)

Signed-off-by: Ye, Xinyu <xinyu.ye@intel.com>
intel · Dec 9, 2023 · ff04655 · ff04655
1 parent 60942ea
commit ff04655
Showing 1 changed file with 2 additions and 0 deletions.
diff --git a/intel_extension_for_transformers/llm/finetuning/finetuning.py b/intel_extension_for_transformers/llm/finetuning/finetuning.py
@@ -821,6 +821,8 @@ def preprocess_logits_for_metrics(logits, labels):
                 model = prepare_model_for_kbit_training(
                     model, use_gradient_checkpointing=training_args.gradient_checkpointing
                 )
+            if training_args.gradient_checkpointing:
+                model.gradient_checkpointing_enable()
 
             if not finetune_args.full_finetune:
                 # PEFT settings