resolve shen cmt on "Rollback ddp_tutorial fix"

Summary:

#1618 was merged unintentionally before all the cmts were resolved.
Also did a minior fix to skip demo parallel when ngpus is < 4

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

[ghstack-poisoned]

intermediate_source/ddp_tutorial.rst

@@ -285,8 +285,10 @@ either the application or the model ``forward()`` method.
 
     if __name__ == "__main__":
         n_gpus = torch.cuda.device_count()
-        assert n_gpus >= 8, f"Requires at least 2 GPUs to run, but got {n_gpus}"
-        world_size = n_gpus
-        run_demo(demo_basic, world_size)
-        run_demo(demo_checkpoint, world_size)
-        run_demo(demo_model_parallel, world_size)
+        assert n_gpus >= 2, f"Requires at least 2 GPUs to run, but got {n_gpus}"
+        run_demo(demo_basic, n_gpus)
+        run_demo(demo_checkpoint, n_gpus)
+        if n_gpus < 4:
+            print("Skipped demo_model_parallel since it requires >= 4 GPUs.")
+        else:
+            run_demo(demo_model_parallel, world_size)