diff --git a/tests/distributed/test_pipeline_parallel.py b/tests/distributed/test_pipeline_parallel.py
index d666b8a1d44bd..5ff39ddfbf996 100644
--- a/tests/distributed/test_pipeline_parallel.py
+++ b/tests/distributed/test_pipeline_parallel.py
@@ -1,3 +1,10 @@
+"""
+WARNING: This test runs in both single-node (4 GPUs) and multi-node
+ (2 node with 2 GPUs each) modes. If the test only uses 2 GPUs, it is
+ important to set the distributed backend to "mp" to avoid Ray scheduling
+ all workers in a node other than the head node, which can cause the test
+ to fail.
+"""
 import os
 
 import pytest
@@ -78,7 +85,7 @@ def test_pp_cudagraph(PP_SIZE, MODEL_NAME, ATTN_BACKEND):
         "--pipeline-parallel-size",
         str(PP_SIZE),
         "--distributed-executor-backend",
-        "ray",
+        "mp",
     ]
     os.environ["VLLM_ATTENTION_BACKEND"] = ATTN_BACKEND