From ff04655ee5838d63da43776e7f0590fb2b19be1d Mon Sep 17 00:00:00 2001 From: XinyuYe-Intel Date: Sat, 9 Dec 2023 10:33:27 +0800 Subject: [PATCH] fix gradient issue for qlora on seq2seq. (#897) Signed-off-by: Ye, Xinyu --- intel_extension_for_transformers/llm/finetuning/finetuning.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/intel_extension_for_transformers/llm/finetuning/finetuning.py b/intel_extension_for_transformers/llm/finetuning/finetuning.py index b08d3b71842..151aca21df7 100644 --- 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