more cleanups

vllm/v1/core/scheduler.py

@@ -99,7 +99,6 @@ def schedule(self) -> "SchedulerOutput":
         req_to_new_block_ids: Dict[str, List[int]] = {}
         num_scheduled_tokens: Dict[str, int] = {}
         token_budget = self.max_num_scheduled_tokens
-
         # Encoder-related.
         scheduled_encoder_inputs: Dict[str, List[int]] = {}
         encoder_budget = self.max_num_encoder_input_tokens
@@ -344,7 +343,6 @@ def _try_schedule_encoder_inputs(
         assert mm_positions is not None
         assert len(mm_positions) > 0
         for i, pos_info in enumerate(mm_positions):
-
             start_pos = pos_info["offset"]
             num_encoder_tokens = pos_info["length"]
 
@@ -367,7 +365,6 @@ def _try_schedule_encoder_inputs(
                 # tokens just before the encoder input.
                 num_new_tokens = start_pos - num_computed_tokens
                 break
-
             if num_encoder_tokens > encoder_budget:
                 # The encoder budget is exhausted. We can only schedule the
                 # decoder tokens up until the encoder input.

vllm/v1/engine/core.py

@@ -83,11 +83,13 @@ def _initialize_kv_caches(self,
 
     def add_request(self, request: EngineCoreRequest):
         """Add request to the scheduler."""
-        req = Request.from_engine_core_request(request)
 
-        if req.mm_hashes is not None:
-            req.mm_inputs = self.mm_input_mapper_server.process_inputs(
-                req.mm_inputs, req.mm_hashes)
+        # Add doc
+        if request.mm_hashes is not None:
+            request.mm_inputs = self.mm_input_mapper_server.process_inputs(
+                request.mm_inputs, request.mm_hashes)
+
+        req = Request.from_engine_core_request(request)
 
         self.scheduler.add_request(req)
 

vllm/v1/request.py

@@ -20,7 +20,6 @@ def __init__(
         eos_token_id: Optional[int],
         arrival_time: float,
         lora_request: Optional[LoRARequest] = None,
-        mm_hashes: Optional[List[str]] = None,
     ) -> None:
         self.request_id = request_id
         self.inputs = SingletonInputsAdapter(inputs)
@@ -57,7 +56,6 @@ def __init__(
             self.mm_inputs = self.inputs.multi_modal_inputs
         else:
             self.mm_inputs: List[MultiModalKwargs] = []
-        self.mm_hashes = mm_hashes
 
     @classmethod
     def from_engine_core_request(cls, request: EngineCoreRequest) -> "Request":
@@ -75,7 +73,6 @@ def from_engine_core_request(cls, request: EngineCoreRequest) -> "Request":
             eos_token_id=request.eos_token_id,
             arrival_time=request.arrival_time,
             lora_request=request.lora_request,
-            mm_hashes=request.mm_hashes,
         )
 
     @property

vllm/v1/worker/gpu_model_runner.py

@@ -352,11 +352,9 @@ def _execute_encoder(self, scheduler_output: "SchedulerOutput"):
         req_input_ids: List[Tuple[int, int]] = []
         for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
             req_state = self.requests[req_id]
-
             for input_id in encoder_input_ids:
                 mm_inputs.append(req_state.mm_inputs[input_id])
                 req_input_ids.append((req_id, input_id))
-
         batched_mm_inputs = MultiModalKwargs.batch(mm_inputs)
         batched_mm_inputs = MultiModalKwargs.as_kwargs(batched_mm_inputs,
                                                        device=self.device)