fix shift short attention

Former-commit-id: 9a49cce8e6f6b222f74a07bdab40efee6a77b0f1
2023-10-09 17:07:46 +08:00
parent 5c4248a29c
commit e387a50475
6 changed files with 46 additions and 52 deletions
--- a/src/llmtuner/tuner/sft/workflow.py
+++ b/src/llmtuner/tuner/sft/workflow.py
@@ -33,6 +33,7 @@ def run_sft(

    data_collator = DataCollatorForSeq2Seq(
        tokenizer=tokenizer,
+        pad_to_multiple_of=4, # for shift short attention
        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
    )