fix: Use nargs for custom_bash_cmds

scripts/performance/argument_parser.py

@@ -422,9 +422,10 @@ def parse_cli_args():
     slurm_args.add_argument(
         "-cb",
         "--custom_bash_cmds",
-        type=list_of_strings,
-        help="Comma separated string of bash commands",
-        default=[],
+        nargs="*",
+        action="append",
+        help="List of bash commands to execute before the main command",
+        default=None,
     )
     slurm_args.add_argument(
         "--gres",

scripts/performance/setup_experiment.py

@@ -206,7 +206,7 @@ def main(
     custom_mounts: List[str],
     custom_env_vars: Dict[str, str],
     custom_srun_args: List[str],
-    custom_bash_cmds: List[str],
+    custom_bash_cmds: List[List[str]],
     nccl_ub: bool,
     pretrained_checkpoint: Optional[str],
     num_gpus: int,

scripts/performance/utils/executors.py

@@ -64,7 +64,7 @@ def slurm_executor(
     nemo_home: str = DEFAULT_NEMO_HOME,
     wandb_key: str = None,
     network: str = None,
-    custom_bash_cmds: List[str] = None,
+    custom_bash_cmds: List[List[str]] = None,
     additional_slurm_params: Dict[str, Any] = None,
     gres: Optional[str] = None,
 ) -> run.SlurmExecutor:
@@ -79,7 +79,7 @@ def slurm_executor(
                 #SBATCH --nodelist=node001,node002
                 #SBATCH --constraint=gpu
     """
-    custom_bash_cmds = [] if custom_bash_cmds is None else custom_bash_cmds
+    custom_bash_cmds = [] if custom_bash_cmds is None else [" ".join(cmd) for cmd in custom_bash_cmds]
     mounts = []
     # Explicitly request GPU resources to ensure proper allocation
     # Without --gres=gpu:N, some clusters only allocate 1 GPU regardless of ntasks_per_node