diff --git a/examples/offline_inference/torchrun_example.py b/examples/offline_inference/torchrun_example.py
index 35df6011550f..7a57f29a07fa 100644
--- a/examples/offline_inference/torchrun_example.py
+++ b/examples/offline_inference/torchrun_example.py
@@ -23,10 +23,14 @@
 
 # Use `distributed_executor_backend="external_launcher"` so that
 # this llm engine/instance only creates one worker.
+# it is important to set an explicit seed to make sure that
+# all ranks have the same random seed, so that sampling can be
+# deterministic across ranks.
 llm = LLM(
     model="facebook/opt-125m",
     tensor_parallel_size=2,
     distributed_executor_backend="external_launcher",
+    seed=0,
 )
 
 outputs = llm.generate(prompts, sampling_params)
diff --git a/vllm/config.py b/vllm/config.py
index 1255d716a2e4..69cde4e362c8 100644
--- a/vllm/config.py
+++ b/vllm/config.py
@@ -761,6 +761,12 @@ def verify_with_parallel_config(
         self,
         parallel_config: "ParallelConfig",
     ) -> None:
+
+        if parallel_config.distributed_executor_backend == "external_launcher":
+            assert self.seed is not None, (
+                "Seed must be set when using external launcher backend to "
+                "make sure sampling results are the same across workers.")
+
         total_num_attention_heads = getattr(self.hf_text_config,
                                             "num_attention_heads", 0)
         tensor_parallel_size = parallel_config.tensor_parallel_size