From 9d6252e63bac241e5c6191eb47adb51b84a5d782 Mon Sep 17 00:00:00 2001 From: Volpeon Date: Tue, 21 Feb 2023 11:50:11 +0100 Subject: Don't rely on Accelerate for gradient accumulation --- train_lora.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'train_lora.py') diff --git a/train_lora.py b/train_lora.py index a06591d..d7c2de0 100644 --- a/train_lora.py +++ b/train_lora.py @@ -399,7 +399,6 @@ def main(): accelerator = Accelerator( log_with=LoggerType.TENSORBOARD, logging_dir=f"{output_dir}", - gradient_accumulation_steps=args.gradient_accumulation_steps, mixed_precision=args.mixed_precision ) @@ -561,6 +560,7 @@ def main(): optimizer=optimizer, lr_scheduler=lr_scheduler, num_train_epochs=args.num_train_epochs, + gradient_accumulation_steps=args.gradient_accumulation_steps, sample_frequency=args.sample_frequency, # -- tokenizer=tokenizer, -- cgit v1.2.3-54-g00ecf