[BugFix] Corner case that could cause out-of-sync with external launcher mode and dp >1

vllm/v1/worker/gpu_model_runner.py

@@ -2532,6 +2532,18 @@ def execute_model(
                         return make_empty_encoder_model_runner_output(scheduler_output)
 
                 if not num_scheduled_tokens:
+                    if (
+                        self.parallel_config.distributed_executor_backend
+                        == "external_launcher"
+                        and self.parallel_config.data_parallel_size > 1
+                    ):
+                        # this is a corner case when both external launcher
+                        # and DP are enabled, num_scheduled_tokens could be
+                        # 0, and has_unfinished_requests in the outer loop
+                        # returns True. before returning early here we call
+                        # dummy run to ensure coordinate_batch_across_dp
+                        # is called into to avoid out of sync issues.
+                        self._dummy_run(1)
                     if not has_kv_transfer_group():
                         # Return empty ModelRunnerOutput if no work to do.
                         return EMPTY_MODEL_RUNNER_OUTPUT