根据不同场景，更新默认调度和分离调度的选择

vllm/model_executor/layers/rotary_embedding.py

@@ -937,8 +937,9 @@ class DeepseekScalingRotaryEmbedding(RotaryEmbedding):
         if self.cos_sin_cache.device != positions.device:
             self.cos_sin_cache: torch.Tensor = self.cos_sin_cache.to(
                 positions.device)
-        cos_sin = self.cos_sin_cache[torch.add(positions, offsets)
-                                     if offsets is not None else positions]
+        if not envs.VLLM_USE_LIGHTOP:
+            cos_sin = self.cos_sin_cache[torch.add(positions, offsets)
+                                        if offsets is not None else positions]
         if query.device.type == 'cuda' and not self.is_neox_style \
             and not self.reference:
             assert len(query.shape) == 3

vllm/v1/core/sched/scheduler.py

@@ -1028,10 +1028,15 @@ class Scheduler(SchedulerInterface):
         return scheduler_output
     
     def schedule(self) -> SchedulerOutput:
-        if envs.VLLM_USE_PD_SPLIT or (self.full_cuda_graph and self.use_mla and self.num_spec_tokens > 0) :
+        if envs.VLLM_USE_PD_SPLIT: 
             return self.schedule_split_pd()
         else:
-            return self.schedule_default()
+            if self.connector is not None:
+                return self.schedule_default()
+            if self.full_cuda_graph and self.use_mla and self.num_spec_tokens > 0 :
+                return self.schedule_split_pd()
+            else:
+                return self.schedule_default()
 
     def _update_after_schedule(
         self,