diff --git a/config/lora_llama2.yaml b/config/lora_llama2.yaml
index f4241f44e..cf6592153 100644
--- a/config/lora_llama2.yaml
+++ b/config/lora_llama2.yaml
@@ -9,7 +9,7 @@ trainer:
     project: "levanter-lora"
     tags: ["lora", "llama2"]
   num_train_steps: 5000  # tune to suit your needs
-  train_batch_size: 128
+  train_batch_size: 64
 
   # if using model parallelism, this is useful:
   tensor_parallel_axes: ["mlp", "heads"]