diff --git a/README.md b/README.md
index 6ffbab6f7..de8ea060d 100644
--- a/README.md
+++ b/README.md
@@ -126,8 +126,13 @@ To use the Docker container as an interactive virtual environment, you can run a
       -v $HOME/algorithmic-efficiency:/algorithmic-efficiency \
       --gpus all \
       --ipc=host \
+<<<<<<< HEAD
+      <docker_image_name> \
+      -keep_container_alive true
+=======
       <docker_image_name> \
       --keep_container_alive true
+>>>>>>> ba5c6f6175a0ce12f23a7f035613d9d1edc0b74a
    ```
    Note: You may have to use double quotes around `algorithmic-efficiency` [path] in the mounting `-v` flag. If the above command fails try replacing the following line:
    ```bash
diff --git a/submission_runner.py b/submission_runner.py
index 656599a42..6d4cc98e2 100644
--- a/submission_runner.py
+++ b/submission_runner.py
@@ -149,6 +149,11 @@
     None,
     'Value of rng seed. If None, a random seed will'
     'be generated from hardware.')
+flags.DEFINE_boolean(
+    'set_pytorch_max_split_size', 
+    None,
+    'If true, set pytorch max_split_size_mb to 256'
+)
 FLAGS = flags.FLAGS
 USE_PYTORCH_DDP, RANK, DEVICE, N_GPUS = pytorch_setup()
 
@@ -601,6 +606,9 @@ def main(_):
   # Prevent OOM on librispeech conformer.
   if FLAGS.workload == 'librispeech_conformer':
     os.environ['XLA_PYTHON_CLIENT_MEM_FRACTION'] = '0.85'
+  
+  if FLAGS.set_pytorch_max_split_size is True:
+    os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:256'
 
   # Extend path according to framework.
   workload_metadata['workload_path'] = os.path.join(