diff --git a/configs/default.yaml b/configs/default.yaml
index 55ccad5..e7dc8d6 100644
--- a/configs/default.yaml
+++ b/configs/default.yaml
@@ -16,6 +16,8 @@ remove_unused_columns: false
 # w/o this the trainer stores logits of all sample in memory...
 batch_eval_metrics: true
 
+per_device_train_batch_size: 128
+per_device_eval_batch_size: 128
 optim: schedule_free_adamw
 learning_rate: 0.0001
 lr_scheduler_type: "constant_with_warmup"
diff --git a/hyperlora/model_loading.py b/hyperlora/model_loading.py
index 145c72c..d03bef6 100644
--- a/hyperlora/model_loading.py
+++ b/hyperlora/model_loading.py
@@ -33,6 +33,7 @@ def get_model_and_tokenizer(
     )
     tokenizer = get_tokenizer(model_name_or_path, tokenizer_kwargs, peft_config, train)
     model.config.pad_token_id = tokenizer.pad_token_id
+    model.generation_config.pad_token_id = tokenizer.pad_token_id
     return model, tokenizer