[V1][Scheduler] Avoid calling `_try_schedule_encoder_inputs` for every request (#15778)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

vllm/v1/core/sched/scheduler.py

@@ -171,19 +171,23 @@ class Scheduler(SchedulerInterface):
             assert num_new_tokens > 0
 
             # Schedule encoder inputs.
-            encoder_inputs_to_schedule, num_new_tokens, new_encoder_budget = (
-                self._try_schedule_encoder_inputs(request,
-                                                  request.num_computed_tokens,
-                                                  num_new_tokens,
-                                                  encoder_budget))
+            if request.has_encoder_inputs:
+                (encoder_inputs_to_schedule, num_new_tokens,
+                 new_encoder_budget) = self._try_schedule_encoder_inputs(
+                     request, request.num_computed_tokens, num_new_tokens,
+                     encoder_budget)
                 if num_new_tokens == 0:
                     # The request cannot be scheduled because the encoder budget
                     # or the encoder cache is exhausted.
-                # NOTE(woosuk): Here, by doing `continue` instead of `break`,
-                # we do not strictly follow the FCFS scheduling policy and
-                # allow the lower-priority requests to be scheduled.
+                    # NOTE(woosuk): By using `continue` instead of `break` here,
+                    # we intentionally relax the strict FCFS scheduling policy
+                    # to allow lower-priority requests to be scheduled when a
+                    # higher-priority request is blocked by encoder constraints.
                     req_index += 1
                     continue
+            else:
+                encoder_inputs_to_schedule = None
+                new_encoder_budget = encoder_budget
 
             while True:
                 new_blocks = self.kv_cache_manager.allocate_slots(
@@ -318,6 +322,7 @@ class Scheduler(SchedulerInterface):
                 assert num_new_tokens > 0
 
                 # Schedule encoder inputs.
+                if request.has_encoder_inputs:
                     (encoder_inputs_to_schedule, num_new_tokens,
                      new_encoder_budget) = self._try_schedule_encoder_inputs(
                          request, num_computed_tokens, num_new_tokens,
@@ -325,6 +330,9 @@ class Scheduler(SchedulerInterface):
                     if num_new_tokens == 0:
                         # The request cannot be scheduled.
                         break
+                else:
+                    encoder_inputs_to_schedule = None
+                    new_encoder_budget = encoder_budget
 
                 new_blocks = self.kv_cache_manager.allocate_slots(
                     request, num_new_tokens, computed_blocks)
@@ -506,9 +514,6 @@ class Scheduler(SchedulerInterface):
         limitations, the method adjusts `num_new_tokens` to schedule only the
         decoder tokens up to just before the unschedulable encoder input.
         """
-        if not request.has_encoder_inputs():
-            return [], num_new_tokens, encoder_budget
-
         encoder_inputs_to_schedule: list[int] = []
         mm_positions = request.mm_positions
         assert mm_positions is not None

vllm/v1/request.py

@@ -59,6 +59,8 @@ class Request:
         self.mm_positions = multi_modal_placeholders or []
         self.mm_inputs = multi_modal_inputs or []
         self.mm_hashes: list[str] = multi_modal_hashes or []
+        self.num_encoder_inputs = len(self.mm_inputs)
+        self.has_encoder_inputs = self.num_encoder_inputs > 0
 
         # Sanity check
         assert len(self.mm_inputs) == len(self.mm_positions)
@@ -117,13 +119,6 @@ class Request:
     def get_finished_reason(self) -> Union[FinishReason, None]:
         return RequestStatus.get_finished_reason(self.status)
 
-    def has_encoder_inputs(self) -> bool:
-        return len(self.mm_inputs) > 0
-
-    @property
-    def num_encoder_inputs(self) -> int:
-        return len(self.mm_positions)
-
     def get_num_encoder_tokens(self, input_id: int) -> int:
         assert input_id < len(self.mm_positions)
         num_tokens = self.mm_positions[input_id]["length"]