vllm.v1.engine.core

HANDSHAKE_TIMEOUT_MINS module-attribute

HANDSHAKE_TIMEOUT_MINS = 5

POLLING_TIMEOUT_S module-attribute

POLLING_TIMEOUT_S = 2.5

_R module-attribute

_R = TypeVar('_R')

logger module-attribute

logger = init_logger(__name__)

DPEngineCoreProc

Bases: EngineCoreProc

ZMQ-wrapper for running EngineCore in background process in a data parallel context.

Source code in vllm/v1/engine/core.py

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

    def __init__(
        self,
        vllm_config: VllmConfig,
        on_head_node: bool,
        input_address: str,
        executor_class: type[Executor],
        log_stats: bool,
    ):
        # Add process-specific prefix to stdout and stderr before
        # we initialize the engine.
        from multiprocessing import current_process
        process_name = current_process().name
        pid = os.getpid()
        _add_prefix(sys.stdout, process_name, pid)
        _add_prefix(sys.stderr, process_name, pid)

        # Counts forward-passes of the model so that we can synchronize
        # finished with DP peers every N steps.
        self.counter = 0

        # Initialize the engine.
        dp_rank = vllm_config.parallel_config.data_parallel_rank
        super().__init__(vllm_config, on_head_node, input_address,
                         executor_class, log_stats, 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

        from vllm.platforms import current_platform
        device_control_env_var = current_platform.device_control_env_var
        world_size = vllm_config.parallel_config.world_size
        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))

        self.dp_rank = dp_rank
        self.dp_group = vllm_config.parallel_config.stateless_init_dp_group()
        self.current_wave = 0

    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: EngineCoreRequest):
        if request.current_wave != self.current_wave:
            if request.current_wave > self.current_wave:
                self.current_wave = request.current_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(
                    EngineCoreOutputs(start_wave=self.current_wave))

        super().add_request(request)

    def _handle_client_request(self, request_type: EngineCoreRequestType,
                               request: Any) -> None:
        if request_type == EngineCoreRequestType.START_DP_WAVE:
            new_wave: int = request
            if 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 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()

            local_unfinished_reqs = self.scheduler.has_unfinished_requests()

            if local_unfinished_reqs:
                # 2) Step the engine core.
                self._process_engine_step()

                # Check if we have now finished all requests.
                local_unfinished_reqs = (
                    self.scheduler.has_unfinished_requests())
            else:
                if self.scheduler.has_finished_requests():
                    # There are no unfinished requests, but there are some
                    # finished requests remaining to be removed from the
                    # batch state. This engine step won't perform a forward
                    # pass but will flush the finished requests to ensure
                    # up-to-date state is returned in the engine outputs.
                    self._process_engine_step()

                if not self.engines_running:
                    # All engines are idle.
                    continue

                # There must be unfinished requests in DP peers, run a
                # dummy forward pass.
                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:
                    # Notify client that we are pausing the loop.
                    logger.debug("Wave %d finished, pausing engine loop.",
                                 self.current_wave)
                    self.output_queue.put_nowait(
                        EngineCoreOutputs(wave_complete=self.current_wave))
                self.current_wave += 1

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

        # Optimization - only perform finish-sync all-reduce every 24 steps.
        self.counter += 1
        if self.counter != 24:
            return True
        self.counter = 0

        return ParallelConfig.has_unfinished_dp(self.dp_group,
                                                local_unfinished)

counter instance-attribute

counter = 0

__init__

__init__(
    vllm_config: VllmConfig,
    on_head_node: bool,
    input_address: str,
    executor_class: type[Executor],
    log_stats: bool,
)

Source code in vllm/v1/engine/core.py

def __init__(
    self,
    vllm_config: VllmConfig,
    on_head_node: bool,
    input_address: str,
    executor_class: type[Executor],
    log_stats: bool,
):
    # Add process-specific prefix to stdout and stderr before
    # we initialize the engine.
    from multiprocessing import current_process
    process_name = current_process().name
    pid = os.getpid()
    _add_prefix(sys.stdout, process_name, pid)
    _add_prefix(sys.stderr, process_name, pid)

    # Counts forward-passes of the model so that we can synchronize
    # finished with DP peers every N steps.
    self.counter = 0

    # Initialize the engine.
    dp_rank = vllm_config.parallel_config.data_parallel_rank
    super().__init__(vllm_config, on_head_node, input_address,
                     executor_class, log_stats, dp_rank)

_handle_client_request

_handle_client_request(
    request_type: EngineCoreRequestType, request: Any
) -> None

Source code in vllm/v1/engine/core.py

def _handle_client_request(self, request_type: EngineCoreRequestType,
                           request: Any) -> None:
    if request_type == EngineCoreRequestType.START_DP_WAVE:
        new_wave: int = request
        if 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)

_has_global_unfinished_reqs

_has_global_unfinished_reqs(local_unfinished: bool) -> bool

Source code in vllm/v1/engine/core.py

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

    # Optimization - only perform finish-sync all-reduce every 24 steps.
    self.counter += 1
    if self.counter != 24:
        return True
    self.counter = 0

    return ParallelConfig.has_unfinished_dp(self.dp_group,
                                            local_unfinished)

_init_data_parallel

_init_data_parallel(vllm_config: VllmConfig)

Source code in vllm/v1/engine/core.py

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

    from vllm.platforms import current_platform
    device_control_env_var = current_platform.device_control_env_var
    world_size = vllm_config.parallel_config.world_size
    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))

    self.dp_rank = dp_rank
    self.dp_group = vllm_config.parallel_config.stateless_init_dp_group()
    self.current_wave = 0

add_request

add_request(request: EngineCoreRequest)

Source code in vllm/v1/engine/core.py

def add_request(self, request: EngineCoreRequest):
    if request.current_wave != self.current_wave:
        if request.current_wave > self.current_wave:
            self.current_wave = request.current_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(
                EngineCoreOutputs(start_wave=self.current_wave))

    super().add_request(request)

run_busy_loop

run_busy_loop()

Core busy loop of the EngineCore for data parallel case.

Source code in vllm/v1/engine/core.py

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()

        local_unfinished_reqs = self.scheduler.has_unfinished_requests()

        if local_unfinished_reqs:
            # 2) Step the engine core.
            self._process_engine_step()

            # Check if we have now finished all requests.
            local_unfinished_reqs = (
                self.scheduler.has_unfinished_requests())
        else:
            if self.scheduler.has_finished_requests():
                # There are no unfinished requests, but there are some
                # finished requests remaining to be removed from the
                # batch state. This engine step won't perform a forward
                # pass but will flush the finished requests to ensure
                # up-to-date state is returned in the engine outputs.
                self._process_engine_step()

            if not self.engines_running:
                # All engines are idle.
                continue

            # There must be unfinished requests in DP peers, run a
            # dummy forward pass.
            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:
                # Notify client that we are pausing the loop.
                logger.debug("Wave %d finished, pausing engine loop.",
                             self.current_wave)
                self.output_queue.put_nowait(
                    EngineCoreOutputs(wave_complete=self.current_wave))
            self.current_wave += 1

shutdown

shutdown()

Source code in vllm/v1/engine/core.py

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

EngineCore

Inner loop of vLLM's Engine.

Source code in vllm/v1/engine/core.py

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):
        assert vllm_config.model_config.runner_type != "pooling"

        # 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)

        # 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.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)

        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,
        )

        # Setup MM Input Mapper.
        self.mm_input_cache_server = MirroredProcessingCache(
            vllm_config.model_config)

        # 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.vllm_config = vllm_config

    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()

        # 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()

        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 add_request(self, request: EngineCoreRequest):
        """Add request to the scheduler."""

        if request.mm_hashes is not None:
            # Here, if hash exists for a multimodal input, then it will be
            # fetched from the cache, else it will be added to the cache.
            # Note that the cache here is mirrored with the client cache, so
            # anything that has a hash must have a HIT cache entry here
            # as well.
            assert request.mm_inputs is not None
            request.mm_inputs = self.mm_input_cache_server.get_and_update_p1(
                request.mm_inputs, request.mm_hashes)

        req = Request.from_engine_core_request(request)
        if req.use_structured_output:
            # Start grammar compilation asynchronously
            self.structured_output_manager.grammar_init(req)

        if req.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(req)

    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(self, scheduler_output: SchedulerOutput):
        try:
            return self.model_executor.execute_model(scheduler_output)
        except BaseException as err:
            # NOTE: This method is exception-free
            dump_engine_exception(self.vllm_config, scheduler_output,
                                  self.scheduler.make_stats())
            # Re-raise exception
            raise err

    def step(self) -> EngineCoreOutputs:
        """Schedule, execute, and make output."""

        # 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 EngineCoreOutputs(
                outputs=[],
                scheduler_stats=self.scheduler.make_stats(),
            )
        scheduler_output = self.scheduler.schedule()
        model_output = self.execute_model(scheduler_output)
        engine_core_outputs = self.scheduler.update_from_output(
            scheduler_output, model_output)  # type: ignore

        return engine_core_outputs

    def step_with_batch_queue(self) -> Optional[EngineCoreOutputs]:
        """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 = future.result()
            self.batch_queue.task_done()
            engine_core_outputs = self.scheduler.update_from_output(
                scheduler_output, model_output)

        return engine_core_outputs

    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, )

batch_queue instance-attribute

batch_queue: Optional[
    Queue[tuple[Future[ModelRunnerOutput], SchedulerOutput]]
] = None

batch_queue_size instance-attribute

batch_queue_size = max_concurrent_batches

log_stats instance-attribute

log_stats = log_stats

mm_input_cache_server instance-attribute

mm_input_cache_server = MirroredProcessingCache(
    model_config
)

model_executor instance-attribute

model_executor = executor_class(vllm_config)

scheduler instance-attribute

scheduler: SchedulerInterface = Scheduler(
    vllm_config=vllm_config,
    kv_cache_config=kv_cache_config,
    structured_output_manager=structured_output_manager,
    include_finished_set=data_parallel_size > 1,
    log_stats=log_stats,
)

structured_output_manager instance-attribute

structured_output_manager = StructuredOutputManager(
    vllm_config
)

vllm_config instance-attribute

vllm_config = vllm_config

__init__

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

Source code in vllm/v1/engine/core.py

def __init__(self,
             vllm_config: VllmConfig,
             executor_class: type[Executor],
             log_stats: bool,
             executor_fail_callback: Optional[Callable] = None):
    assert vllm_config.model_config.runner_type != "pooling"

    # 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)

    # 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.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)

    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,
    )

    # Setup MM Input Mapper.
    self.mm_input_cache_server = MirroredProcessingCache(
        vllm_config.model_config)

    # 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.vllm_config = vllm_config

_initialize_kv_caches

_initialize_kv_caches(
    vllm_config: VllmConfig,
) -> tuple[int, int, KVCacheConfig]

Source code in vllm/v1/engine/core.py

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()

    # 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()

    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

abort_requests

abort_requests(request_ids: list[str])

Abort requests from the scheduler.

Source code in vllm/v1/engine/core.py

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)

add_lora

add_lora(lora_request: LoRARequest) -> bool

Source code in vllm/v1/engine/core.py

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

add_request

add_request(request: EngineCoreRequest)

Add request to the scheduler.

Source code in vllm/v1/engine/core.py

def add_request(self, request: EngineCoreRequest):
    """Add request to the scheduler."""

    if request.mm_hashes is not None:
        # Here, if hash exists for a multimodal input, then it will be
        # fetched from the cache, else it will be added to the cache.
        # Note that the cache here is mirrored with the client cache, so
        # anything that has a hash must have a HIT cache entry here
        # as well.
        assert request.mm_inputs is not None
        request.mm_inputs = self.mm_input_cache_server.get_and_update_p1(
            request.mm_inputs, request.mm_hashes)

    req = Request.from_engine_core_request(request)
    if req.use_structured_output:
        # Start grammar compilation asynchronously
        self.structured_output_manager.grammar_init(req)

    if req.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(req)

collective_rpc

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

Source code in vllm/v1/engine/core.py

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)

execute_dummy_batch

execute_dummy_batch()

Source code in vllm/v1/engine/core.py

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

execute_model

execute_model(scheduler_output: SchedulerOutput)

Source code in vllm/v1/engine/core.py

def execute_model(self, scheduler_output: SchedulerOutput):
    try:
        return self.model_executor.execute_model(scheduler_output)
    except BaseException as err:
        # NOTE: This method is exception-free
        dump_engine_exception(self.vllm_config, scheduler_output,
                              self.scheduler.make_stats())
        # Re-raise exception
        raise err

is_sleeping

is_sleeping() -> bool

Source code in vllm/v1/engine/core.py

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

list_loras

list_loras() -> set[int]

Source code in vllm/v1/engine/core.py

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

pin_lora

pin_lora(lora_id: int) -> bool

Source code in vllm/v1/engine/core.py

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

profile

profile(is_start: bool = True)

Source code in vllm/v1/engine/core.py

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

remove_lora

remove_lora(lora_id: int) -> bool

Source code in vllm/v1/engine/core.py

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

reset_mm_cache

reset_mm_cache()

Source code in vllm/v1/engine/core.py

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()

reset_prefix_cache

reset_prefix_cache()

Source code in vllm/v1/engine/core.py

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

save_sharded_state

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

Source code in vllm/v1/engine/core.py

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)

save_tensorized_model

save_tensorized_model(tensorizer_config) -> None

Source code in vllm/v1/engine/core.py

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

shutdown

shutdown()

Source code in vllm/v1/engine/core.py

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

sleep

sleep(level: int = 1)

Source code in vllm/v1/engine/core.py

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

step

step() -> EngineCoreOutputs

Schedule, execute, and make output.

Source code in vllm/v1/engine/core.py

def step(self) -> EngineCoreOutputs:
    """Schedule, execute, and make output."""

    # 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 EngineCoreOutputs(
            outputs=[],
            scheduler_stats=self.scheduler.make_stats(),
        )
    scheduler_output = self.scheduler.schedule()
    model_output = self.execute_model(scheduler_output)
    engine_core_outputs = self.scheduler.update_from_output(
        scheduler_output, model_output)  # type: ignore

    return engine_core_outputs

step_with_batch_queue

step_with_batch_queue() -> Optional[EngineCoreOutputs]

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.

Source code in vllm/v1/engine/core.py

def step_with_batch_queue(self) -> Optional[EngineCoreOutputs]:
    """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 = future.result()
        self.batch_queue.task_done()
        engine_core_outputs = self.scheduler.update_from_output(
            scheduler_output, model_output)

    return engine_core_outputs

wake_up

wake_up(tags: Optional[list[str]] = None)

Source code in vllm/v1/engine/core.py

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

EngineCoreProc

Bases: EngineCore

ZMQ-wrapper for running EngineCore in background process.

Source code in vllm/v1/engine/core.py

class EngineCoreProc(EngineCore):
    """ZMQ-wrapper for running EngineCore in background process."""

    ENGINE_CORE_DEAD = b'ENGINE_CORE_DEAD'

    def __init__(
        self,
        vllm_config: VllmConfig,
        on_head_node: bool,
        input_address: str,
        executor_class: type[Executor],
        log_stats: bool,
        engine_index: int = 0,
    ):
        input_queue = queue.Queue[tuple[EngineCoreRequestType, Any]]()

        executor_fail_callback = lambda: input_queue.put_nowait(
            (EngineCoreRequestType.EXECUTOR_FAILED, b''))

        # Create input socket.
        input_ctx = zmq.Context()
        identity = engine_index.to_bytes(length=2, byteorder="little")
        input_socket = make_zmq_socket(input_ctx,
                                       input_address,
                                       zmq.DEALER,
                                       identity=identity,
                                       bind=False)
        try:
            # Register engine with front-end.
            output_address = self.startup_handshake(
                input_socket, on_head_node, vllm_config.parallel_config)

            # Update config which may have changed from the handshake.
            vllm_config.__post_init__()

            # Set up data parallel environment.
            self._init_data_parallel(vllm_config)

            # Initialize engine core and model.
            super().__init__(vllm_config, executor_class, log_stats,
                             executor_fail_callback)

            self.step_fn = (self.step if self.batch_queue is None else
                            self.step_with_batch_queue)
            self.engines_running = False

            # Send ready message.
            num_gpu_blocks = vllm_config.cache_config.num_gpu_blocks
            input_socket.send(
                msgspec.msgpack.encode({
                    "status": "READY",
                    "local": on_head_node,
                    "num_gpu_blocks": num_gpu_blocks,
                }))

            # Background Threads and Queues for IO. These enable us to
            # overlap ZMQ socket IO with GPU since they release the GIL,
            # and to overlap some serialization/deserialization with the
            # model forward pass.
            # Threads handle Socket <-> Queues and core_busy_loop uses Queue.
            self.input_queue = input_queue
            self.output_queue = queue.Queue[Union[EngineCoreOutputs, bytes]]()
            threading.Thread(target=self.process_input_socket,
                             args=(input_socket, ),
                             daemon=True).start()
            input_socket = None
            self.output_thread = threading.Thread(
                target=self.process_output_socket,
                args=(output_address, engine_index),
                daemon=True)
            self.output_thread.start()
        finally:
            if input_socket is not None:
                input_socket.close(linger=0)

    @staticmethod
    def startup_handshake(input_socket: zmq.Socket, on_head_node: bool,
                          parallel_config: ParallelConfig) -> str:

        # Send registration message.
        input_socket.send(
            msgspec.msgpack.encode({
                "status": "HELLO",
                "local": on_head_node,
            }))

        # Receive initialization message.
        logger.info("Waiting for init message from front-end.")
        if not input_socket.poll(timeout=HANDSHAKE_TIMEOUT_MINS * 60 * 1000):
            raise RuntimeError("Did not receive response from front-end "
                               f"process within {HANDSHAKE_TIMEOUT_MINS} "
                               f"minutes")
        init_bytes = input_socket.recv()
        init_message = msgspec.msgpack.decode(init_bytes)
        logger.debug("Received init message: %s", init_message)

        output_socket_address = init_message["output_socket_address"]
        #TBD(nick) maybe replace IP with configured head node address

        received_parallel_config = init_message["parallel_config"]
        for key, value in received_parallel_config.items():
            setattr(parallel_config, key, value)

        return output_socket_address

    @staticmethod
    def run_engine_core(*args,
                        dp_rank: int = 0,
                        local_dp_rank: int = 0,
                        **kwargs):
        """Launch EngineCore busy loop in background process."""

        # Signal handler used for graceful termination.
        # SystemExit exception is only raised once to allow this and worker
        # processes to terminate without error
        shutdown_requested = False

        # Ensure we can serialize transformer config after spawning
        maybe_register_config_serialize_by_value()

        def signal_handler(signum, frame):
            nonlocal shutdown_requested
            if not shutdown_requested:
                shutdown_requested = True
                raise SystemExit()

        # Either SIGTERM or SIGINT will terminate the engine_core
        signal.signal(signal.SIGTERM, signal_handler)
        signal.signal(signal.SIGINT, signal_handler)

        engine_core: Optional[EngineCoreProc] = None
        try:
            parallel_config: ParallelConfig = kwargs[
                "vllm_config"].parallel_config
            if parallel_config.data_parallel_size > 1 or dp_rank > 0:
                # Set data parallel rank for this engine process.
                parallel_config.data_parallel_rank = dp_rank
                parallel_config.data_parallel_rank_local = local_dp_rank
                engine_core = DPEngineCoreProc(*args, **kwargs)
            else:
                engine_core = EngineCoreProc(*args, **kwargs)

            engine_core.run_busy_loop()

        except SystemExit:
            logger.debug("EngineCore exiting.")
            raise
        except Exception as e:
            if engine_core is None:
                logger.exception("EngineCore failed to start.")
            else:
                logger.exception("EngineCore encountered a fatal error.")
                engine_core._send_engine_dead()
            raise e
        finally:
            if engine_core is not None:
                engine_core.shutdown()

    def _init_data_parallel(self, vllm_config: VllmConfig):
        pass

    def run_busy_loop(self):
        """Core busy loop of the EngineCore."""

        # 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 and return the outputs.
            self._process_engine_step()

    def _process_input_queue(self):
        """Exits when an engine step needs to be performed."""

        waited = False
        while not self.engines_running and not (self.scheduler.has_requests()):
            if logger.isEnabledFor(DEBUG) and self.input_queue.empty():
                logger.debug("EngineCore waiting for work.")
                waited = True
            req = self.input_queue.get()
            self._handle_client_request(*req)

        if waited:
            logger.debug("EngineCore loop active.")

        # Handle any more client requests.
        while not self.input_queue.empty():
            req = self.input_queue.get_nowait()
            self._handle_client_request(*req)

    def _process_engine_step(self):
        """Called only when there are unfinished local requests."""

        # Step the engine core.
        outputs = self.step_fn()
        # Put EngineCoreOutputs into the output queue.
        if outputs is not None:
            self.output_queue.put_nowait(outputs)

    def _handle_client_request(self, request_type: EngineCoreRequestType,
                               request: Any) -> None:
        """Dispatch request from client."""

        if request_type == EngineCoreRequestType.ADD:
            self.add_request(request)
        elif request_type == EngineCoreRequestType.ABORT:
            self.abort_requests(request)
        elif request_type == EngineCoreRequestType.UTILITY:
            call_id, method_name, args = request
            output = UtilityOutput(call_id)
            try:
                method = getattr(self, method_name)
                output.result = method(
                    *self._convert_msgspec_args(method, args))
            except BaseException as e:
                logger.exception("Invocation of %s method failed", method_name)
                output.failure_message = (f"Call to {method_name} method"
                                          f" failed: {str(e)}")
            self.output_queue.put_nowait(
                EngineCoreOutputs(utility_output=output))
        elif request_type == EngineCoreRequestType.EXECUTOR_FAILED:
            raise RuntimeError("Executor failed.")
        else:
            logger.error("Unrecognized input request type encountered: %s",
                         request_type)

    @staticmethod
    def _convert_msgspec_args(method, args):
        """If a provided arg type doesn't match corresponding target method
         arg type, try converting to msgspec object."""
        if not args:
            return args
        arg_types = signature(method).parameters.values()
        assert len(args) <= len(arg_types)
        return tuple(
            msgspec.convert(v, type=p.annotation) if isclass(p.annotation)
            and issubclass(p.annotation, msgspec.Struct)
            and not isinstance(v, p.annotation) else v
            for v, p in zip(args, arg_types))

    def _send_engine_dead(self):
        """Send EngineDead status to the EngineCoreClient."""

        # Put ENGINE_CORE_DEAD in the queue.
        self.output_queue.put_nowait(EngineCoreProc.ENGINE_CORE_DEAD)

        # Wait until msg sent by the daemon before shutdown.
        self.output_thread.join(timeout=5.0)
        if self.output_thread.is_alive():
            logger.fatal("vLLM shutdown signal from EngineCore failed "
                         "to send. Please report this issue.")

    def process_input_socket(self, input_socket: zmq.Socket):
        """Input socket IO thread."""

        # Msgpack serialization decoding.
        add_request_decoder = MsgpackDecoder(EngineCoreRequest)
        generic_decoder = MsgpackDecoder()

        while True:
            # (RequestType, RequestData)
            type_frame, *data_frames = input_socket.recv_multipart(copy=False)
            request_type = EngineCoreRequestType(bytes(type_frame.buffer))

            # Deserialize the request data.
            decoder = add_request_decoder if (
                request_type == EngineCoreRequestType.ADD) else generic_decoder
            request = decoder.decode(data_frames)

            # Push to input queue for core busy loop.
            self.input_queue.put_nowait((request_type, request))

    def process_output_socket(self, output_path: str, engine_index: int):
        """Output socket IO thread."""

        # Msgpack serialization encoding.
        encoder = MsgpackEncoder()
        # Send buffers to reuse.
        reuse_buffers: list[bytearray] = []
        # Keep references to outputs and buffers until zmq is finished
        # with them (outputs may contain tensors/np arrays whose
        # backing buffers were extracted for zero-copy send).
        pending = deque[tuple[zmq.MessageTracker, Any, bytearray]]()

        # We must set linger to ensure the ENGINE_CORE_DEAD
        # message is sent prior to closing the socket.
        with zmq_socket_ctx(output_path, zmq.constants.PUSH,
                            linger=4000) as socket:
            while True:
                outputs = self.output_queue.get()
                if outputs == EngineCoreProc.ENGINE_CORE_DEAD:
                    socket.send(outputs, copy=False)
                    break
                assert not isinstance(outputs, bytes)
                outputs.engine_index = engine_index

                # Reclaim buffers that zmq is finished with.
                while pending and pending[-1][0].done:
                    reuse_buffers.append(pending.pop()[2])

                buffer = reuse_buffers.pop() if reuse_buffers else bytearray()
                buffers = encoder.encode_into(outputs, buffer)
                tracker = socket.send_multipart(buffers,
                                                copy=False,
                                                track=True)
                if not tracker.done:
                    ref = outputs if len(buffers) > 1 else None
                    pending.appendleft((tracker, ref, buffer))
                elif len(reuse_buffers) < 2:
                    # Keep at most 2 buffers to reuse.
                    reuse_buffers.append(buffer)

ENGINE_CORE_DEAD class-attribute instance-attribute

ENGINE_CORE_DEAD = b'ENGINE_CORE_DEAD'

engines_running instance-attribute

engines_running = False

input_queue instance-attribute

input_queue = input_queue

output_queue instance-attribute

output_queue = Queue[Union[EngineCoreOutputs, bytes]]()

output_thread instance-attribute

output_thread = Thread(
    target=process_output_socket,
    args=(output_address, engine_index),
    daemon=True,
)

step_fn instance-attribute

step_fn = (
    step if batch_queue is None else step_with_batch_queue
)

__init__

__init__(
    vllm_config: VllmConfig,
    on_head_node: bool,
    input_address: str,
    executor_class: type[Executor],
    log_stats: bool,
    engine_index: int = 0,
)

Source code in vllm/v1/engine/core.py

def __init__(
    self,
    vllm_config: VllmConfig,
    on_head_node: bool,
    input_address: str,
    executor_class: type[Executor],
    log_stats: bool,
    engine_index: int = 0,
):
    input_queue = queue.Queue[tuple[EngineCoreRequestType, Any]]()

    executor_fail_callback = lambda: input_queue.put_nowait(
        (EngineCoreRequestType.EXECUTOR_FAILED, b''))

    # Create input socket.
    input_ctx = zmq.Context()
    identity = engine_index.to_bytes(length=2, byteorder="little")
    input_socket = make_zmq_socket(input_ctx,
                                   input_address,
                                   zmq.DEALER,
                                   identity=identity,
                                   bind=False)
    try:
        # Register engine with front-end.
        output_address = self.startup_handshake(
            input_socket, on_head_node, vllm_config.parallel_config)

        # Update config which may have changed from the handshake.
        vllm_config.__post_init__()

        # Set up data parallel environment.
        self._init_data_parallel(vllm_config)

        # Initialize engine core and model.
        super().__init__(vllm_config, executor_class, log_stats,
                         executor_fail_callback)

        self.step_fn = (self.step if self.batch_queue is None else
                        self.step_with_batch_queue)
        self.engines_running = False

        # Send ready message.
        num_gpu_blocks = vllm_config.cache_config.num_gpu_blocks
        input_socket.send(
            msgspec.msgpack.encode({
                "status": "READY",
                "local": on_head_node,
                "num_gpu_blocks": num_gpu_blocks,
            }))

        # Background Threads and Queues for IO. These enable us to
        # overlap ZMQ socket IO with GPU since they release the GIL,
        # and to overlap some serialization/deserialization with the
        # model forward pass.
        # Threads handle Socket <-> Queues and core_busy_loop uses Queue.
        self.input_queue = input_queue
        self.output_queue = queue.Queue[Union[EngineCoreOutputs, bytes]]()
        threading.Thread(target=self.process_input_socket,
                         args=(input_socket, ),
                         daemon=True).start()
        input_socket = None
        self.output_thread = threading.Thread(
            target=self.process_output_socket,
            args=(output_address, engine_index),
            daemon=True)
        self.output_thread.start()
    finally:
        if input_socket is not None:
            input_socket.close(linger=0)

_convert_msgspec_args staticmethod

_convert_msgspec_args(method, args)

If a provided arg type doesn't match corresponding target method arg type, try converting to msgspec object.

Source code in vllm/v1/engine/core.py

@staticmethod
def _convert_msgspec_args(method, args):
    """If a provided arg type doesn't match corresponding target method
     arg type, try converting to msgspec object."""
    if not args:
        return args
    arg_types = signature(method).parameters.values()
    assert len(args) <= len(arg_types)
    return tuple(
        msgspec.convert(v, type=p.annotation) if isclass(p.annotation)
        and issubclass(p.annotation, msgspec.Struct)
        and not isinstance(v, p.annotation) else v
        for v, p in zip(args, arg_types))

_handle_client_request

_handle_client_request(
    request_type: EngineCoreRequestType, request: Any
) -> None

Dispatch request from client.

Source code in vllm/v1/engine/core.py

def _handle_client_request(self, request_type: EngineCoreRequestType,
                           request: Any) -> None:
    """Dispatch request from client."""

    if request_type == EngineCoreRequestType.ADD:
        self.add_request(request)
    elif request_type == EngineCoreRequestType.ABORT:
        self.abort_requests(request)
    elif request_type == EngineCoreRequestType.UTILITY:
        call_id, method_name, args = request
        output = UtilityOutput(call_id)
        try:
            method = getattr(self, method_name)
            output.result = method(
                *self._convert_msgspec_args(method, args))
        except BaseException as e:
            logger.exception("Invocation of %s method failed", method_name)
            output.failure_message = (f"Call to {method_name} method"
                                      f" failed: {str(e)}")
        self.output_queue.put_nowait(
            EngineCoreOutputs(utility_output=output))
    elif request_type == EngineCoreRequestType.EXECUTOR_FAILED:
        raise RuntimeError("Executor failed.")
    else:
        logger.error("Unrecognized input request type encountered: %s",
                     request_type)

_init_data_parallel

_init_data_parallel(vllm_config: VllmConfig)

Source code in vllm/v1/engine/core.py

def _init_data_parallel(self, vllm_config: VllmConfig):
    pass

_process_engine_step

_process_engine_step()

Called only when there are unfinished local requests.

Source code in vllm/v1/engine/core.py

def _process_engine_step(self):
    """Called only when there are unfinished local requests."""

    # Step the engine core.
    outputs = self.step_fn()
    # Put EngineCoreOutputs into the output queue.
    if outputs is not None:
        self.output_queue.put_nowait(outputs)

_process_input_queue

_process_input_queue()

Exits when an engine step needs to be performed.

Source code in vllm/v1/engine/core.py

def _process_input_queue(self):
    """Exits when an engine step needs to be performed."""

    waited = False
    while not self.engines_running and not (self.scheduler.has_requests()):
        if logger.isEnabledFor(DEBUG) and self.input_queue.empty():
            logger.debug("EngineCore waiting for work.")
            waited = True
        req = self.input_queue.get()
        self._handle_client_request(*req)

    if waited:
        logger.debug("EngineCore loop active.")

    # Handle any more client requests.
    while not self.input_queue.empty():
        req = self.input_queue.get_nowait()
        self._handle_client_request(*req)

_send_engine_dead

_send_engine_dead()

Send EngineDead status to the EngineCoreClient.

Source code in vllm/v1/engine/core.py

def _send_engine_dead(self):
    """Send EngineDead status to the EngineCoreClient."""

    # Put ENGINE_CORE_DEAD in the queue.
    self.output_queue.put_nowait(EngineCoreProc.ENGINE_CORE_DEAD)

    # Wait until msg sent by the daemon before shutdown.
    self.output_thread.join(timeout=5.0)
    if self.output_thread.is_alive():
        logger.fatal("vLLM shutdown signal from EngineCore failed "
                     "to send. Please report this issue.")

process_input_socket

process_input_socket(input_socket: Socket)

Input socket IO thread.

Source code in vllm/v1/engine/core.py

def process_input_socket(self, input_socket: zmq.Socket):
    """Input socket IO thread."""

    # Msgpack serialization decoding.
    add_request_decoder = MsgpackDecoder(EngineCoreRequest)
    generic_decoder = MsgpackDecoder()

    while True:
        # (RequestType, RequestData)
        type_frame, *data_frames = input_socket.recv_multipart(copy=False)
        request_type = EngineCoreRequestType(bytes(type_frame.buffer))

        # Deserialize the request data.
        decoder = add_request_decoder if (
            request_type == EngineCoreRequestType.ADD) else generic_decoder
        request = decoder.decode(data_frames)

        # Push to input queue for core busy loop.
        self.input_queue.put_nowait((request_type, request))

process_output_socket

process_output_socket(output_path: str, engine_index: int)

Output socket IO thread.

Source code in vllm/v1/engine/core.py

def process_output_socket(self, output_path: str, engine_index: int):
    """Output socket IO thread."""

    # Msgpack serialization encoding.
    encoder = MsgpackEncoder()
    # Send buffers to reuse.
    reuse_buffers: list[bytearray] = []
    # Keep references to outputs and buffers until zmq is finished
    # with them (outputs may contain tensors/np arrays whose
    # backing buffers were extracted for zero-copy send).
    pending = deque[tuple[zmq.MessageTracker, Any, bytearray]]()

    # We must set linger to ensure the ENGINE_CORE_DEAD
    # message is sent prior to closing the socket.
    with zmq_socket_ctx(output_path, zmq.constants.PUSH,
                        linger=4000) as socket:
        while True:
            outputs = self.output_queue.get()
            if outputs == EngineCoreProc.ENGINE_CORE_DEAD:
                socket.send(outputs, copy=False)
                break
            assert not isinstance(outputs, bytes)
            outputs.engine_index = engine_index

            # Reclaim buffers that zmq is finished with.
            while pending and pending[-1][0].done:
                reuse_buffers.append(pending.pop()[2])

            buffer = reuse_buffers.pop() if reuse_buffers else bytearray()
            buffers = encoder.encode_into(outputs, buffer)
            tracker = socket.send_multipart(buffers,
                                            copy=False,
                                            track=True)
            if not tracker.done:
                ref = outputs if len(buffers) > 1 else None
                pending.appendleft((tracker, ref, buffer))
            elif len(reuse_buffers) < 2:
                # Keep at most 2 buffers to reuse.
                reuse_buffers.append(buffer)

run_busy_loop

run_busy_loop()

Core busy loop of the EngineCore.

Source code in vllm/v1/engine/core.py

def run_busy_loop(self):
    """Core busy loop of the EngineCore."""

    # 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 and return the outputs.
        self._process_engine_step()

run_engine_core staticmethod

run_engine_core(
    *args,
    dp_rank: int = 0,
    local_dp_rank: int = 0,
    **kwargs,
)

Launch EngineCore busy loop in background process.

Source code in vllm/v1/engine/core.py

@staticmethod
def run_engine_core(*args,
                    dp_rank: int = 0,
                    local_dp_rank: int = 0,
                    **kwargs):
    """Launch EngineCore busy loop in background process."""

    # Signal handler used for graceful termination.
    # SystemExit exception is only raised once to allow this and worker
    # processes to terminate without error
    shutdown_requested = False

    # Ensure we can serialize transformer config after spawning
    maybe_register_config_serialize_by_value()

    def signal_handler(signum, frame):
        nonlocal shutdown_requested
        if not shutdown_requested:
            shutdown_requested = True
            raise SystemExit()

    # Either SIGTERM or SIGINT will terminate the engine_core
    signal.signal(signal.SIGTERM, signal_handler)
    signal.signal(signal.SIGINT, signal_handler)

    engine_core: Optional[EngineCoreProc] = None
    try:
        parallel_config: ParallelConfig = kwargs[
            "vllm_config"].parallel_config
        if parallel_config.data_parallel_size > 1 or dp_rank > 0:
            # Set data parallel rank for this engine process.
            parallel_config.data_parallel_rank = dp_rank
            parallel_config.data_parallel_rank_local = local_dp_rank
            engine_core = DPEngineCoreProc(*args, **kwargs)
        else:
            engine_core = EngineCoreProc(*args, **kwargs)

        engine_core.run_busy_loop()

    except SystemExit:
        logger.debug("EngineCore exiting.")
        raise
    except Exception as e:
        if engine_core is None:
            logger.exception("EngineCore failed to start.")
        else:
            logger.exception("EngineCore encountered a fatal error.")
            engine_core._send_engine_dead()
        raise e
    finally:
        if engine_core is not None:
            engine_core.shutdown()

startup_handshake staticmethod

startup_handshake(
    input_socket: Socket,
    on_head_node: bool,
    parallel_config: ParallelConfig,
) -> str

Source code in vllm/v1/engine/core.py

@staticmethod
def startup_handshake(input_socket: zmq.Socket, on_head_node: bool,
                      parallel_config: ParallelConfig) -> str:

    # Send registration message.
    input_socket.send(
        msgspec.msgpack.encode({
            "status": "HELLO",
            "local": on_head_node,
        }))

    # Receive initialization message.
    logger.info("Waiting for init message from front-end.")
    if not input_socket.poll(timeout=HANDSHAKE_TIMEOUT_MINS * 60 * 1000):
        raise RuntimeError("Did not receive response from front-end "
                           f"process within {HANDSHAKE_TIMEOUT_MINS} "
                           f"minutes")
    init_bytes = input_socket.recv()
    init_message = msgspec.msgpack.decode(init_bytes)
    logger.debug("Received init message: %s", init_message)

    output_socket_address = init_message["output_socket_address"]
    #TBD(nick) maybe replace IP with configured head node address

    received_parallel_config = init_message["parallel_config"]
    for key, value in received_parallel_config.items():
        setattr(parallel_config, key, value)

    return output_socket_address