Fix the issue that run out of requests make no parallelism in PP mode in ray backend.

Signed-off-by: Weiliang Liu <weiliangl@nvidia.com>

Author: weireweire

vllm/v1/core/sched/scheduler.py

@@ -397,7 +397,12 @@ def schedule(self) -> SchedulerOutput:
             while self.waiting and token_budget > 0:
                 if len(self.running) == self.max_num_running_reqs:
                     break
-
+                if len(scheduled_resumed_reqs) + len(scheduled_new_reqs) >= max(
+                    1,
+                    self.max_num_running_reqs
+                    // self.parallel_config.pipeline_parallel_size,
+                ):
+                    break
                 request = self.waiting.peek_request()
 
                 # KVTransfer: skip request if still waiting for remote kvs.