increaaase bs=128 and change dtype to fp32

run.sh

@@ -9,8 +9,8 @@ python3 run_mlm_flax.py \
     --max_seq_length="128" \
     --weight_decay="0.01" \
     --preprocessing_num_workers="64" \
-    --per_device_train_batch_size="32" \
-    --per_device_eval_batch_size="32" \
+    --per_device_train_batch_size="128" \
+    --per_device_eval_batch_size="128" \
     --learning_rate="3e-5" \
     --warmup_steps="5000" \
     --overwrite_output_dir \
@@ -21,5 +21,4 @@ python3 run_mlm_flax.py \
     --logging_steps="500" \
     --save_steps="5000" \
     --eval_steps="20000" \
-    --dtype="bfloat16" \
     --push_to_hub