Refactor test

Signed-off-by: Shunkang <[email protected]>

Author: Shunkang

tensorrt_llm/_torch/pyexecutor/executor_request_queue.py

@@ -293,6 +293,8 @@ def _filter_out_unschedulable_requests(
                         unscheduled_requests.append(req_item)
                     else:
                         pending_requests.append(req_item)
+                else:
+                    unscheduled_requests.append(req_item)
             else:
                 unscheduled_requests.append(req_item)
 
@@ -312,7 +314,7 @@ def _schedule_attention_dp_requests(
         # Schedule the requests with attention dp rank and no relax
         for req_item in scheduled_requests:
             target_dp_rank = req_item.request.py_schedule_params.attention_dp_rank
-            assert self.all_ranks_num_active_requests[target_dp_rank] < self.max_num_active_requests, \
+            assert self.all_ranks_num_active_requests[target_dp_rank] <= self.max_num_active_requests, \
                 f"The number of active requests on rank {target_dp_rank} is {self.all_ranks_num_active_requests[target_dp_rank]}, " \
                 f"which is greater than the max_num_active_requests {self.max_num_active_requests}"
             self.all_ranks_num_active_requests[target_dp_rank] += 1