vllm.v1.engine.core

class DPEngineCoreActor(DPEngineCoreProc):
    """
    Ray actor for running EngineCore in a data parallel context
    """

    def __init__(
        self,
        vllm_config: VllmConfig,
        local_client: bool,
        addresses: EngineZmqAddresses,
        executor_class: type[Executor],
        log_stats: bool,
        dp_rank: int = 0,
        local_dp_rank: int = 0,
    ):
        self.addresses = addresses
        vllm_config.parallel_config.data_parallel_rank = dp_rank
        vllm_config.parallel_config.data_parallel_rank_local = \
            local_dp_rank

        # Set CUDA_VISIBLE_DEVICES as early as possible in actor life cycle
        # NOTE: in MP we set CUDA_VISIBLE_DEVICES at process creation time,
        # and this cannot be done in the same way for Ray because:
        # 1) Ray manages life cycle of all ray workers (including
        # DPEngineCoreActor)
        # 2) Ray sets CUDA_VISIBLE_DEVICES based on num_gpus configuration
        # To bypass 2, we need to also set
        # RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES, but vLLM workers created
        # thereafter would have CUDA_VISIBLE_DEVICES set, which is sticky:
        # https://github.com/ray-project/ray/blob/e752fc319ddedd9779a0989b6d3613909bad75c9/python/ray/_private/worker.py#L456 # noqa: E501
        # This is problematic because when the vLLM worker (a Ray actor)
        # executes a task, it indexes into the sticky CUDA_VISIBLE_DEVICES
        # rather than directly using the GPU ID, potentially resulting in
        # index out of bounds error. See:
        # https://github.com/ray-project/ray/pull/40461/files#diff-31e8159767361e4bc259b6d9883d9c0d5e5db780fcea4a52ead4ee3ee4a59a78R1860 # noqa: E501
        # and get_accelerator_ids_for_accelerator_resource() in worker.py
        # of ray.
        self._set_cuda_visible_devices(vllm_config, local_dp_rank)

        super().__init__(vllm_config, local_client, "", executor_class,
                         log_stats)

    def _set_cuda_visible_devices(self, vllm_config: VllmConfig,
                                  local_dp_rank: int):
        from vllm.platforms import current_platform
        device_control_env_var = current_platform.device_control_env_var
        world_size = vllm_config.parallel_config.world_size
        # Set CUDA_VISIBLE_DEVICES or equivalent.
        try:
            os.environ[device_control_env_var] = ",".join(
                str(current_platform.device_id_to_physical_device_id(i))
                for i in range(local_dp_rank *
                               world_size, (local_dp_rank + 1) * world_size))
        except IndexError as e:
            raise Exception(
                f"Error setting {device_control_env_var}: "
                f"local range: [{local_dp_rank * world_size}, "
                f"{(local_dp_rank + 1) * world_size}) "
                f"base value: \"{os.getenv(device_control_env_var)}\"") from e

    @contextmanager
    def _perform_handshakes(self, handshake_address: str, identity: bytes,
                            local_client: bool, vllm_config: VllmConfig,
                            client_handshake_address: Optional[str]):
        """
        For Ray, we don't need to actually perform handshake.
        All addresses information is known before the actor creation.
        Therefore, we simply yield these addresses.
        """
        yield self.addresses

    def wait_for_init(self):
        """
        Wait until the engine core is initialized.

        This is just an empty method. When ray.get() on this method
        (or any other method of the actor) returns, it is guaranteed
        that actor creation (i.e., __init__) is complete.
        """
        pass

    def run(self):
        """
        Run the engine core busy loop.
        """
        try:
            self.run_busy_loop()
        except SystemExit:
            logger.debug("EngineCore exiting.")
            raise
        except Exception:
            logger.exception("EngineCore encountered a fatal error.")
            raise
        finally:
            self.shutdown()

class DPEngineCoreProc(EngineCoreProc):
    """ZMQ-wrapper for running EngineCore in background process
    in a data parallel context."""

    def __init__(
        self,
        vllm_config: VllmConfig,
        local_client: bool,
        handshake_address: str,
        executor_class: type[Executor],
        log_stats: bool,
        client_handshake_address: Optional[str] = None,
    ):
        # Counts forward-passes of the model so that we can synchronize
        # finished with DP peers every N steps.
        self.step_counter = 0
        self.current_wave = 0
        self.last_counts = (0, 0)

        # Initialize the engine.
        dp_rank = vllm_config.parallel_config.data_parallel_rank
        super().__init__(vllm_config, local_client, handshake_address,
                         executor_class, log_stats, client_handshake_address,
                         dp_rank)

    def _init_data_parallel(self, vllm_config: VllmConfig):

        # Configure GPUs and stateless process group for data parallel.
        dp_rank = vllm_config.parallel_config.data_parallel_rank
        dp_size = vllm_config.parallel_config.data_parallel_size
        local_dp_rank = vllm_config.parallel_config.data_parallel_rank_local

        assert dp_size > 1
        assert 0 <= local_dp_rank <= dp_rank < dp_size

        if vllm_config.kv_transfer_config is not None:
            # modify the engine_id and append the local_dp_rank to it to ensure
            # that the kv_transfer_config is unique for each DP rank.
            vllm_config.kv_transfer_config.engine_id = (
                f"{vllm_config.kv_transfer_config.engine_id}_dp{local_dp_rank}"
            )
            logger.debug("Setting kv_transfer_config.engine_id to %s",
                         vllm_config.kv_transfer_config.engine_id)

        self.dp_rank = dp_rank
        self.dp_group = vllm_config.parallel_config.stateless_init_dp_group()

    def shutdown(self):
        super().shutdown()
        if dp_group := getattr(self, "dp_group", None):
            stateless_destroy_torch_distributed_process_group(dp_group)

    def add_request(self, request: Request, request_wave: int = 0):
        if self.has_coordinator and request_wave != self.current_wave:
            if request_wave > self.current_wave:
                self.current_wave = request_wave
            elif not self.engines_running:
                # Request received for an already-completed wave, notify
                # front-end that we need to start the next one.
                self.output_queue.put_nowait(
                    (-1, EngineCoreOutputs(start_wave=self.current_wave)))

        super().add_request(request, request_wave)

    def _handle_client_request(self, request_type: EngineCoreRequestType,
                               request: Any) -> None:
        if request_type == EngineCoreRequestType.START_DP_WAVE:
            new_wave, exclude_eng_index = request
            if exclude_eng_index != self.engine_index and (
                    new_wave >= self.current_wave):
                self.current_wave = new_wave
                if not self.engines_running:
                    logger.debug("EngineCore starting idle loop for wave %d.",
                                 new_wave)
                    self.engines_running = True
        else:
            super()._handle_client_request(request_type, request)

    def _maybe_publish_request_counts(self):
        if not self.publish_dp_lb_stats:
            return

        # Publish our request counts (if they've changed).
        counts = self.scheduler.get_request_counts()
        if counts != self.last_counts:
            self.last_counts = counts
            stats = SchedulerStats(*counts,
                                   step_counter=self.step_counter,
                                   current_wave=self.current_wave)
            self.output_queue.put_nowait(
                (-1, EngineCoreOutputs(scheduler_stats=stats)))

    def run_busy_loop(self):
        """Core busy loop of the EngineCore for data parallel case."""

        # Loop until process is sent a SIGINT or SIGTERM
        while True:
            # 1) Poll the input queue until there is work to do.
            self._process_input_queue()

            # 2) Step the engine core.
            executed = self._process_engine_step()
            self._maybe_publish_request_counts()

            local_unfinished_reqs = self.scheduler.has_unfinished_requests()
            if not executed:
                if not local_unfinished_reqs and not self.engines_running:
                    # All engines are idle.
                    continue

                # We are in a running state and so must execute a dummy pass
                # if the model didn't execute any ready requests.
                self.execute_dummy_batch()

            # 3) All-reduce operation to determine global unfinished reqs.
            self.engines_running = self._has_global_unfinished_reqs(
                local_unfinished_reqs)

            if not self.engines_running:
                if self.dp_rank == 0 or not self.has_coordinator:
                    # Notify client that we are pausing the loop.
                    logger.debug("Wave %d finished, pausing engine loop.",
                                 self.current_wave)
                    # In the coordinator case, dp rank 0 sends updates to the
                    # coordinator. Otherwise (offline spmd case), each rank
                    # sends the update to its colocated front-end process.
                    client_index = -1 if self.has_coordinator else 0
                    self.output_queue.put_nowait(
                        (client_index,
                         EngineCoreOutputs(wave_complete=self.current_wave)))
                # Increment wave count and reset step counter.
                self.current_wave += 1
                self.step_counter = 0

    def _has_global_unfinished_reqs(self, local_unfinished: bool) -> bool:

        # Optimization - only perform finish-sync all-reduce every 32 steps.
        self.step_counter += 1
        if self.step_counter % 32 != 0:
            return True

        return ParallelConfig.has_unfinished_dp(self.dp_group,
                                                local_unfinished)

    def reinitialize_distributed(
            self, reconfig_request: ReconfigureDistributedRequest) -> None:
        stateless_destroy_torch_distributed_process_group(self.dp_group)
        self.shutdown()

        parallel_config = self.vllm_config.parallel_config
        old_dp_size = parallel_config.data_parallel_size
        parallel_config.data_parallel_size = \
            reconfig_request.new_data_parallel_size
        if reconfig_request.new_data_parallel_rank != -1:
            parallel_config.data_parallel_rank = \
                reconfig_request.new_data_parallel_rank
        # local rank specifies device visibility, it should not be changed
        assert reconfig_request.new_data_parallel_rank_local == \
            ReconfigureRankType.KEEP_CURRENT_RANK
        parallel_config.data_parallel_master_ip = \
            reconfig_request.new_data_parallel_master_ip
        parallel_config.data_parallel_master_port = \
            reconfig_request.new_data_parallel_master_port
        if reconfig_request.new_data_parallel_rank != -2:
            self.dp_rank = parallel_config.data_parallel_rank
            self.dp_group = parallel_config.stateless_init_dp_group()
        reconfig_request.new_data_parallel_master_port = \
            parallel_config.data_parallel_master_port

        self.model_executor.reinitialize_distributed(reconfig_request)
        if reconfig_request.new_data_parallel_size > old_dp_size:
            assert self.available_gpu_memory_for_kv_cache > 0
            # pass available_gpu_memory_for_kv_cache from existing
            # engine-cores to new engine-cores so they can directly
            # use it in _initialize_kv_caches() rather than profiling.
            ParallelConfig.sync_kv_cache_memory_size(
                self.dp_group, self.available_gpu_memory_for_kv_cache)
            # NOTE(yongji): newly joined workers require dummy_run even
            # CUDA graph is not used
            self.model_executor.collective_rpc("compile_or_warm_up_model")
        if reconfig_request.new_data_parallel_rank == \
        ReconfigureRankType.SHUTDOWN_CURRENT_RANK:
            self.shutdown()
            logger.info("DPEngineCoreProc %s shutdown", self.dp_rank)
        else:
            logger.info("Distributed environment reinitialized for DP rank %s",
                        self.dp_rank)

class EngineCore:
    """Inner loop of vLLM's Engine."""

    def __init__(self,
                 vllm_config: VllmConfig,
                 executor_class: type[Executor],
                 log_stats: bool,
                 executor_fail_callback: Optional[Callable] = None):

        # plugins need to be loaded at the engine/scheduler level too
        from vllm.plugins import load_general_plugins
        load_general_plugins()

        self.vllm_config = vllm_config
        logger.info("Initializing a V1 LLM engine (v%s) with config: %s",
                    VLLM_VERSION, vllm_config)

        self.log_stats = log_stats

        # Setup Model.
        self.model_executor = executor_class(vllm_config)
        if executor_fail_callback is not None:
            self.model_executor.register_failure_callback(
                executor_fail_callback)

        self.available_gpu_memory_for_kv_cache = -1

        # Setup KV Caches and update CacheConfig after profiling.
        num_gpu_blocks, num_cpu_blocks, kv_cache_config = \
            self._initialize_kv_caches(vllm_config)

        vllm_config.cache_config.num_gpu_blocks = num_gpu_blocks
        vllm_config.cache_config.num_cpu_blocks = num_cpu_blocks
        self.collective_rpc("initialize_cache",
                            args=(num_gpu_blocks, num_cpu_blocks))

        self.structured_output_manager = StructuredOutputManager(vllm_config)

        # Setup scheduler.
        if isinstance(vllm_config.scheduler_config.scheduler_cls, str):
            Scheduler = resolve_obj_by_qualname(
                vllm_config.scheduler_config.scheduler_cls)
        else:
            Scheduler = vllm_config.scheduler_config.scheduler_cls

        # This warning can be removed once the V1 Scheduler interface is
        # finalized and we can maintain support for scheduler classes that
        # implement it
        if Scheduler is not V1Scheduler:
            logger.warning(
                "Using configured V1 scheduler class %s. "
                "This scheduler interface is not public and "
                "compatibility may not be maintained.",
                vllm_config.scheduler_config.scheduler_cls)

        if len(kv_cache_config.kv_cache_groups) == 0:
            # Encoder models without KV cache don't support
            # chunked prefill. But do SSM models?
            logger.info("Disabling chunked prefill for model without KVCache")
            vllm_config.scheduler_config.chunked_prefill_enabled = False

        self.scheduler: SchedulerInterface = Scheduler(
            vllm_config=vllm_config,
            kv_cache_config=kv_cache_config,
            structured_output_manager=self.structured_output_manager,
            include_finished_set=vllm_config.parallel_config.data_parallel_size
            > 1,
            log_stats=self.log_stats,
        )
        self.use_spec_decode = vllm_config.speculative_config is not None

        self.mm_input_cache_server = MultiModalInputCacheServer(
            vllm_config.model_config, MULTIMODAL_REGISTRY)

        # Setup batch queue for pipeline parallelism.
        # Batch queue for scheduled batches. This enables us to asynchronously
        # schedule and execute batches, and is required by pipeline parallelism
        # to eliminate pipeline bubbles.
        self.batch_queue_size = self.model_executor.max_concurrent_batches
        self.batch_queue: Optional[queue.Queue[tuple[Future[ModelRunnerOutput],
                                                     SchedulerOutput]]] = None
        if self.batch_queue_size > 1:
            logger.info("Batch queue is enabled with size %d",
                        self.batch_queue_size)
            self.batch_queue = queue.Queue(self.batch_queue_size)

        self.request_block_hasher: Optional[Callable[[Request],
                                                     list[BlockHash]]] = None
        if (self.vllm_config.cache_config.enable_prefix_caching
                or self.scheduler.get_kv_connector() is not None):

            block_size = vllm_config.cache_config.block_size
            caching_hash_fn = get_hash_fn_by_name(
                vllm_config.cache_config.prefix_caching_hash_algo)
            init_none_hash(caching_hash_fn)

            self.request_block_hasher = get_request_block_hasher(
                block_size, caching_hash_fn)

    def _initialize_kv_caches(
            self, vllm_config: VllmConfig) -> tuple[int, int, KVCacheConfig]:
        start = time.time()

        # Get all kv cache needed by the model
        kv_cache_specs = self.model_executor.get_kv_cache_specs()

        has_kv_cache = any(kv_cache_spec for kv_cache_spec in kv_cache_specs)
        if has_kv_cache:
            if os.environ.get("VLLM_ELASTIC_EP_SCALE_UP_LAUNCH") == "1":
                dp_group = getattr(self, "dp_group", None)
                assert dp_group is not None
                self.available_gpu_memory_for_kv_cache = \
                    ParallelConfig.sync_kv_cache_memory_size(dp_group, -1)
                available_gpu_memory = [
                    self.available_gpu_memory_for_kv_cache
                ] * len(kv_cache_specs)
            else:
                # Profiles the peak memory usage of the model to determine how
                # much memory can be allocated for kv cache.
                available_gpu_memory = (
                    self.model_executor.determine_available_memory())
                self.available_gpu_memory_for_kv_cache = \
                    available_gpu_memory[0]
        else:
            # Attention free models don't need memory for kv cache
            available_gpu_memory = [0] * len(kv_cache_specs)

        assert len(kv_cache_specs) == len(available_gpu_memory)
        # Get the kv cache tensor size
        kv_cache_configs = [
            get_kv_cache_config(vllm_config, kv_cache_spec_one_worker,
                                available_gpu_memory_one_worker)
            for kv_cache_spec_one_worker, available_gpu_memory_one_worker in
            zip(kv_cache_specs, available_gpu_memory)
        ]

        # Since we use a shared centralized controller, we need the
        # `kv_cache_config` to be consistent across all workers to make sure
        # all the memory operators can be applied to all workers.
        unify_kv_cache_configs(kv_cache_configs)

        # All workers have the same kv_cache_config except layer names, so use
        # an arbitrary one to initialize the scheduler.
        assert all([
            cfg.num_blocks == kv_cache_configs[0].num_blocks
            for cfg in kv_cache_configs
        ])
        num_gpu_blocks = kv_cache_configs[0].num_blocks
        num_cpu_blocks = 0
        scheduler_kv_cache_config = kv_cache_configs[0]

        # Initialize kv cache and warmup the execution
        self.model_executor.initialize_from_config(kv_cache_configs)

        elapsed = time.time() - start
        logger.info(("init engine (profile, create kv cache, "
                     "warmup model) took %.2f seconds"), elapsed)
        return num_gpu_blocks, num_cpu_blocks, scheduler_kv_cache_config

    def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        return self.model_executor.supported_tasks

    def add_request(self, request: Request, request_wave: int = 0):
        """Add request to the scheduler.

        `request_wave`: indicate which wave of requests this is expected to
        belong to in DP case
        """
        # Validate the request_id type.
        if not isinstance(request.request_id, str):
            raise TypeError(
                f"request_id must be a string, got {type(request.request_id)}")

        if pooling_params := request.pooling_params:
            supported_pooling_tasks = [
                task for task in self.get_supported_tasks()
                if task in POOLING_TASKS
            ]

            if pooling_params.task not in supported_pooling_tasks:
                raise ValueError(f"Unsupported task: {pooling_params.task!r} "
                                 f"Supported tasks: {supported_pooling_tasks}")

        if request.kv_transfer_params is not None and (
                not self.scheduler.get_kv_connector()):
            logger.warning("Got kv_transfer_params, but no KVConnector found. "
                           "Disabling KVTransfer for this request.")

        self.scheduler.add_request(request)

    def abort_requests(self, request_ids: list[str]):
        """Abort requests from the scheduler."""

        # TODO: The scheduler doesn't really need to know the
        # specific finish reason, TBD whether we propagate that
        # (i.e. client-aborted vs stop criteria met).
        self.scheduler.finish_requests(request_ids,
                                       RequestStatus.FINISHED_ABORTED)

    def execute_model_with_error_logging(
        self,
        model_fn: Callable[[SchedulerOutput], ModelRunnerOutput],
        scheduler_output: SchedulerOutput,
    ) -> ModelRunnerOutput:
        """Execute the model and log detailed info on failure."""
        try:
            return model_fn(scheduler_output)
        except Exception as err:
            # We do not want to catch BaseException here since we're only
            # interested in dumping info when the exception is due to an
            # error from execute_model itself.

            # NOTE: This method is exception-free
            dump_engine_exception(self.vllm_config, scheduler_output,
                                  self.scheduler.make_stats())
            raise err

    def step(self) -> tuple[dict[int, EngineCoreOutputs], bool]:
        """Schedule, execute, and make output.

        Returns tuple of outputs and a flag indicating whether the model
        was executed.
        """

        # Check for any requests remaining in the scheduler - unfinished,
        # or finished and not yet removed from the batch.
        if not self.scheduler.has_requests():
            return {}, False
        scheduler_output = self.scheduler.schedule()
        model_output = self.execute_model_with_error_logging(
            self.model_executor.execute_model,  # type: ignore
            scheduler_output)
        engine_core_outputs = self.scheduler.update_from_output(
            scheduler_output, model_output)  # type: ignore

        return (engine_core_outputs,
                scheduler_output.total_num_scheduled_tokens > 0)

    def post_step(self, model_executed: bool) -> None:
        if self.use_spec_decode and model_executed:
            # Take the draft token ids.
            draft_token_ids = self.model_executor.take_draft_token_ids()
            if draft_token_ids is not None:
                self.scheduler.update_draft_token_ids(draft_token_ids)

    def step_with_batch_queue(
            self) -> tuple[Optional[dict[int, EngineCoreOutputs]], bool]:
        """Schedule and execute batches with the batch queue.
        Note that if nothing to output in this step, None is returned.

        The execution flow is as follows:
        1. Try to schedule a new batch if the batch queue is not full.
        If a new batch is scheduled, directly return an empty engine core
        output. In other words, fulfilling the batch queue has a higher priority
        than getting model outputs.
        2. If there is no new scheduled batch, meaning that the batch queue
        is full or no other requests can be scheduled, we block until the first
        batch in the job queue is finished.
        3. Update the scheduler from the output.
        """
        assert self.batch_queue is not None

        engine_core_outputs = None
        scheduler_output = None
        # Try to schedule a new batch if the batch queue is not full, but
        # the scheduler may return an empty batch if all requests are scheduled.
        # Note that this is not blocking.
        if not self.batch_queue.full():
            scheduler_output = self.scheduler.schedule()
            if scheduler_output.total_num_scheduled_tokens > 0:
                future = self.model_executor.execute_model(scheduler_output)
                self.batch_queue.put_nowait(
                    (future, scheduler_output))  # type: ignore

        scheduled_batch = (scheduler_output is not None
                           and scheduler_output.total_num_scheduled_tokens > 0)

        # If no more requests can be scheduled and the job queue is not empty,
        # block until the first batch in the job queue is finished.
        # TODO(comaniac): Ideally we should peek the first batch in the
        # job queue to check if it's finished before scheduling a new batch,
        # but peeking the first element in a queue is not thread-safe,
        # so we need more work.
        if not scheduled_batch and not self.batch_queue.empty():
            future, scheduler_output = self.batch_queue.get_nowait()

            # Blocking until the first result is available.
            model_output = self.execute_model_with_error_logging(
                lambda _: future.result(), scheduler_output)

            self.batch_queue.task_done()
            engine_core_outputs = (self.scheduler.update_from_output(
                scheduler_output, model_output))

        return engine_core_outputs, scheduled_batch

    def shutdown(self):
        self.structured_output_manager.clear_backend()
        if self.model_executor:
            self.model_executor.shutdown()
        if self.scheduler:
            self.scheduler.shutdown()

    def profile(self, is_start: bool = True):
        self.model_executor.profile(is_start)

    def reset_mm_cache(self):
        # NOTE: Since this is mainly for debugging, we don't attempt to
        # re-sync the internal caches (P0 processor, P0 mirror, P1 mirror)
        if self.scheduler.has_unfinished_requests():
            logger.warning("Resetting the multi-modal cache when requests are "
                           "in progress may lead to desynced internal caches.")

        self.mm_input_cache_server.reset()

    def reset_prefix_cache(self):
        self.scheduler.reset_prefix_cache()

    def sleep(self, level: int = 1):
        self.model_executor.sleep(level)

    def wake_up(self, tags: Optional[list[str]] = None):
        self.model_executor.wake_up(tags)

    def is_sleeping(self) -> bool:
        return self.model_executor.is_sleeping

    def execute_dummy_batch(self):
        self.model_executor.collective_rpc("execute_dummy_batch")

    def add_lora(self, lora_request: LoRARequest) -> bool:
        return self.model_executor.add_lora(lora_request)

    def remove_lora(self, lora_id: int) -> bool:
        return self.model_executor.remove_lora(lora_id)

    def list_loras(self) -> set[int]:
        return self.model_executor.list_loras()

    def pin_lora(self, lora_id: int) -> bool:
        return self.model_executor.pin_lora(lora_id)

    def save_sharded_state(
        self,
        path: str,
        pattern: Optional[str] = None,
        max_size: Optional[int] = None,
    ) -> None:
        self.model_executor.save_sharded_state(path=path,
                                               pattern=pattern,
                                               max_size=max_size)

    def collective_rpc(self,
                       method: Union[str, Callable[..., _R]],
                       timeout: Optional[float] = None,
                       args: tuple = (),
                       kwargs: Optional[dict[str, Any]] = None) -> list[_R]:
        return self.model_executor.collective_rpc(method, timeout, args,
                                                  kwargs)

    def save_tensorized_model(
        self,
        tensorizer_config,
    ) -> None:
        self.model_executor.save_tensorized_model(
            tensorizer_config=tensorizer_config, )

    def preprocess_add_request(
            self, request: EngineCoreRequest) -> tuple[Request, int]:
        """Preprocess the request.

        This function could be directly used in input processing thread to allow
        request initialization running in parallel with Model forward
        """
        if request.mm_hashes is not None:
            assert request.mm_kwargs is not None

            # Note on thread safety: no race condition.
            # `mm_input_cache_server` is reset at the end of LLMEngine init,
            # and will only accessed in the input processing thread afterwards.
            request.mm_kwargs = self.mm_input_cache_server.get_and_update(
                request.mm_kwargs, request.mm_hashes)

        req = Request.from_engine_core_request(request,
                                               self.request_block_hasher)
        if req.use_structured_output:
            # Note on thread safety: no race condition.
            # `grammar_init` is only invoked in input processing thread. For
            # `structured_output_manager`, each request is independent and
            # grammar compilation is async. Scheduler always checks grammar
            # compilation status before scheduling request.
            self.structured_output_manager.grammar_init(req)
        return req, request.current_wave