[bugfix] add seed in torchrun_example.py (vllm-project#15980)

Signed-off-by: youkaichao <[email protected]>
Signed-off-by: Louis Ulmer <[email protected]>

Author: youkaichao

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)

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