diff --git a/examples/extras/galore/adamw.sh b/examples/extras/galore/adamw.sh
index cad03879..1fd2aaf0 100644
--- a/examples/extras/galore/adamw.sh
+++ b/examples/extras/galore/adamw.sh
@@ -15,7 +15,7 @@ CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
     --preprocessing_num_workers 16 \
     --per_device_train_batch_size 1 \
     --per_device_eval_batch_size 1 \
-    --gradient_accumulation_steps 8 \
+    --gradient_accumulation_steps 2 \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
     --warmup_steps 20 \
diff --git a/examples/extras/galore/adamw_8bit_bf16.sh b/examples/extras/galore/adamw_8bit_bf16.sh
index 9599bf00..01f4e8de 100644
--- a/examples/extras/galore/adamw_8bit_bf16.sh
+++ b/examples/extras/galore/adamw_8bit_bf16.sh
@@ -16,7 +16,7 @@ CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
     --preprocessing_num_workers 16 \
     --per_device_train_batch_size 1 \
     --per_device_eval_batch_size 1 \
-    --gradient_accumulation_steps 8 \
+    --gradient_accumulation_steps 2 \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
     --warmup_steps 20 \
diff --git a/examples/extras/galore/galore_adamw.sh b/examples/extras/galore/galore_adamw.sh
index 28ce72bb..83be6a51 100644
--- a/examples/extras/galore/galore_adamw.sh
+++ b/examples/extras/galore/galore_adamw.sh
@@ -18,7 +18,7 @@ CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
     --preprocessing_num_workers 16 \
     --per_device_train_batch_size 1 \
     --per_device_eval_batch_size 1 \
-    --gradient_accumulation_steps 8 \
+    --gradient_accumulation_steps 2 \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
     --warmup_steps 20 \
diff --git a/examples/extras/galore/galore_adamw_8bit_bf16.sh b/examples/extras/galore/galore_adamw_8bit_bf16.sh
index 0578856c..ddddcb33 100644
--- a/examples/extras/galore/galore_adamw_8bit_bf16.sh
+++ b/examples/extras/galore/galore_adamw_8bit_bf16.sh
@@ -10,7 +10,7 @@ CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
     --finetuning_type full \
     --use_galore \
     --galore_target mlp,self_attn \
-    --galore_rank 32 \
+    --galore_rank 16 \
     --optim adamw_8bit \
     --output_dir ../../../saves/LLaMA2-7B/galore/sft \
     --overwrite_cache \
@@ -19,7 +19,7 @@ CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
     --preprocessing_num_workers 16 \
     --per_device_train_batch_size 1 \
     --per_device_eval_batch_size 1 \
-    --gradient_accumulation_steps 8 \
+    --gradient_accumulation_steps 2 \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
     --warmup_steps 20 \