diff --git a/torchx/schedulers/slurm_scheduler.py b/torchx/schedulers/slurm_scheduler.py
index f74dd2448..fde0fbf96 100644
--- a/torchx/schedulers/slurm_scheduler.py
+++ b/torchx/schedulers/slurm_scheduler.py
@@ -210,6 +210,7 @@ def from_role(
                     sbatch_opts.setdefault("gpus-per-node", str(resource.gpu))
                 else:
                     sbatch_opts.setdefault("gpus-per-task", str(resource.gpu))
+                    sbatch_opts.setdefault("ntasks", "1")
 
         srun_opts = {
             "output": f"slurm-{macros.app_id}-{name}.out",
diff --git a/torchx/schedulers/test/slurm_scheduler_test.py b/torchx/schedulers/test/slurm_scheduler_test.py
index 480f02bc8..ef7f3383e 100644
--- a/torchx/schedulers/test/slurm_scheduler_test.py
+++ b/torchx/schedulers/test/slurm_scheduler_test.py
@@ -128,6 +128,7 @@ def test_replica_request(self, mock_version: MagicMock) -> None:
                 "--cpus-per-task=2",
                 "--mem=10",
                 "--gpus-per-task=3",
+                "--ntasks=1",
             ],
         )
         self.assertEqual(
@@ -163,6 +164,7 @@ def test_replica_request_nomem(self, mock_version: MagicMock) -> None:
                 "--ntasks-per-node=1",
                 "--cpus-per-task=2",
                 "--gpus-per-task=3",
+                "--ntasks=1",
             ],
         )