[Perf][Async Scheduling] Remove CPU->GPU sync in dummy_run (#27455)

Signed-off-by: Lehua Ding <lehuading@tencent.com>

vllm/v1/worker/gpu_model_runner.py

@@ -3492,7 +3492,10 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
             self.eplb_step(is_dummy=True, is_profile=is_profile)
 
         logit_indices = np.cumsum(num_scheduled_tokens) - 1
-        return hidden_states, hidden_states[logit_indices]
+        logit_indices_device = torch.from_numpy(logit_indices).to(
+            self.device, non_blocking=True
+        )
+        return hidden_states, hidden_states[logit_indices_device]
 
     @torch.inference_mode()
     def _dummy_sampler_run(