Add sleep function for disagg gen-only benchmarking (NVIDIA#5398)

Signed-off-by: Xianjie <5410381+qiaoxj07@users.noreply.github.com>

Author: qiaoxj07

tensorrt_llm/_torch/pyexecutor/py_executor.py

@@ -236,6 +236,8 @@ def __init__(self,
         self.ctx_in_transmission_requests = []
         self.previous_batch: Optional[BatchState] = None
         self.num_scheduled_requests: int = 0
+        self.benchmark_req_queues_size = int(
+            os.environ.get("TLLM_BENCHMARK_REQ_QUEUES_SIZE", 0))
 
         # list of requests in each PP micro batch
         self.num_micro_batches = self.dist.pp_size
@@ -996,6 +998,13 @@ def _prepare_draft_requests(self):
 
     def _executor_loop_overlap(self):
         torch.cuda.set_device(self.device_id)
+        if self.dist.rank == 0 and not self.is_warmup and self.benchmark_req_queues_size > 0 and self.kv_cache_transceiver:
+            while self.request_queue.qsize() < self.benchmark_req_queues_size:
+                logger.info(
+                    f"sleep 5 seconds, num_request_queue: {self.request_queue.qsize()}"
+                )
+                time.sleep(5)
+
         with self._profiler() as profile_step:
             iter_start_time = time.time()
             iter_stats = None