llm_engine.py

import copy
from collections import defaultdict
import os
import time
import pickle
import importlib
from typing import (TYPE_CHECKING, Any, Dict, Iterable, List, Optional, Tuple,
                    Union)

import vllm
from vllm.lora.request import LoRARequest
from vllm.config import (CacheConfig, DeviceConfig, ModelConfig,
                         ParallelConfig, SchedulerConfig, LoRAConfig)
from vllm.core.scheduler import Scheduler, SchedulerOutputs
from vllm.engine.arg_utils import EngineArgs
from vllm.engine.metrics import StatLogger, Stats
from vllm.engine.ray_utils import RayWorkerVllm, initialize_cluster, ray
from vllm.logger import init_logger
from vllm.outputs import RequestOutput
from vllm.sampling_params import SamplingParams
from vllm.sequence import (Logprob, SamplerOutput, Sequence, SequenceGroup,
                           SequenceGroupOutput, SequenceOutput, SequenceStatus)
from vllm.transformers_utils.tokenizer import (detokenize_incrementally,
                                               TokenizerGroup)
from vllm.utils import (Counter, set_cuda_visible_devices, get_ip,
                        get_open_port, get_distributed_init_method)

if ray:
    from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy

if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

logger = init_logger(__name__)
_LOCAL_LOGGING_INTERVAL_SEC = 5

# A map between the device type (in device config) to its worker module.
DEVICE_TO_WORKER_MODULE_MAP = {
    "cuda": "vllm.worker.worker",
    "neuron": "vllm.worker.neuron_worker",
}

# If the env var is set, it uses the Ray's compiled DAG API
# which optimizes the control plane overhead.
# Run VLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.
USE_RAY_COMPILED_DAG = bool(os.getenv("VLLM_USE_RAY_COMPILED_DAG", 0))


class LLMEngine:
    """An LLM engine that receives requests and generates texts.

    This is the main class for the vLLM engine. It receives requests
    from clients and generates texts from the LLM. It includes a tokenizer, a
    language model (possibly distributed across multiple GPUs), and GPU memory
    space allocated for intermediate states (aka KV cache). This class utilizes
    iteration-level scheduling and efficient memory management to maximize the
    serving throughput.

    The `LLM` class wraps this class for offline batched inference and the
    `AsyncLLMEngine` class wraps this class for online serving.

    NOTE: The config arguments are derived from the `EngineArgs` class. For the
    comprehensive list of arguments, see `EngineArgs`.

    Args:
        model_config: The configuration related to the LLM model.
        cache_config: The configuration related to the KV cache memory
            management.
        parallel_config: The configuration related to distributed execution.
        scheduler_config: The configuration related to the request scheduler.
        device_config: The configuration related to the device.
        placement_group: Ray placement group for distributed execution.
            Required for distributed execution.
        log_stats: Whether to log statistics.
    """

    def __init__(
        self,
        model_config: ModelConfig,
        cache_config: CacheConfig,
        parallel_config: ParallelConfig,
        scheduler_config: SchedulerConfig,
        device_config: DeviceConfig,
        lora_config: Optional[LoRAConfig],
        placement_group: Optional["PlacementGroup"],
        log_stats: bool,
    ) -> None:
        logger.info(
            f"Initializing an LLM engine (v{vllm.__version__}) with config: "
            f"model={model_config.model!r}, "
            f"tokenizer={model_config.tokenizer!r}, "
            f"tokenizer_mode={model_config.tokenizer_mode}, "
            f"revision={model_config.revision}, "
            f"tokenizer_revision={model_config.tokenizer_revision}, "
            f"trust_remote_code={model_config.trust_remote_code}, "
            f"dtype={model_config.dtype}, "
            f"max_seq_len={model_config.max_model_len}, "
            f"download_dir={model_config.download_dir!r}, "
            f"load_format={model_config.load_format}, "
            f"tensor_parallel_size={parallel_config.tensor_parallel_size}, "
            f"disable_custom_all_reduce={parallel_config.disable_custom_all_reduce}, "
            f"quantization={model_config.quantization}, "
            f"enforce_eager={model_config.enforce_eager}, "
            f"kv_cache_dtype={cache_config.cache_dtype}, "
            f"device_config={device_config.device}, "
            f"seed={model_config.seed})")
        # TODO(woosuk): Print more configs in debug mode.

        self.model_config = model_config
        self.cache_config = cache_config
        self.lora_config = lora_config
        self.parallel_config = parallel_config
        self.scheduler_config = scheduler_config
        self.device_config = device_config
        self.log_stats = log_stats
        self._verify_args()

        self._init_tokenizer()
        self.seq_counter = Counter()

        # Create the parallel GPU workers.
        if self.parallel_config.worker_use_ray:
            # Disable Ray usage stats collection.
            ray_usage = os.environ.get("RAY_USAGE_STATS_ENABLED", "0")
            if ray_usage != "1":
                os.environ["RAY_USAGE_STATS_ENABLED"] = "0"
            # Pass additional arguments to initialize the worker
            additional_ray_args = {}
            if self.parallel_config.ray_workers_use_nsight:
                logger.info("Configuring Ray workers to use nsight.")
                additional_ray_args = {
                    "runtime_env": {
                        "nsight": {
                            "t": "cuda,cudnn,cublas",
                            "o": "'worker_process_%p'",
                            "cuda-graph-trace": "node",
                        }
                    }
                }
            self._init_workers_ray(placement_group, **additional_ray_args)
        else:
            self._init_workers()

        # Profile the memory usage and initialize the cache.
        self._init_cache()

        # Create the scheduler.
        self.scheduler = Scheduler(scheduler_config, cache_config, lora_config)

        # Metric Logging.
        if self.log_stats:
            self.stat_logger = StatLogger(
                local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
                labels=dict(model_name=model_config.model))
            self.stat_logger.info("cache_config", self.cache_config)

        self.forward_dag = None
        if USE_RAY_COMPILED_DAG:
            self.forward_dag = self._compiled_ray_dag()

    def get_tokenizer_for_seq(self, sequence: Sequence):
        return self.tokenizer.get_lora_tokenizer(sequence.lora_request)

    def _dispatch_worker(self):
        worker_module = DEVICE_TO_WORKER_MODULE_MAP[
            self.device_config.device_type]
        imported_worker = importlib.import_module(worker_module)
        Worker = imported_worker.Worker
        return Worker

    def _init_workers(self):
        # Lazy import the Worker to avoid importing torch.cuda/xformers
        # before CUDA_VISIBLE_DEVICES is set in the Worker
        Worker = self._dispatch_worker()

        assert self.parallel_config.world_size == 1, (
            "Ray is required if parallel_config.world_size > 1.")

        self.workers: List[Worker] = []
        distributed_init_method = get_distributed_init_method(
            get_ip(), get_open_port())
        self.driver_worker = Worker(
            self.model_config,
            self.parallel_config,
            self.scheduler_config,
            self.device_config,
            local_rank=0,
            rank=0,
            distributed_init_method=distributed_init_method,
            lora_config=self.lora_config,
            kv_cache_dtype=self.cache_config.cache_dtype,
            is_driver_worker=True,
        )
        self._run_workers("init_model")
        self._run_workers("load_model")

    def _init_tokenizer(self, **tokenizer_init_kwargs):
        init_kwargs = dict(
            enable_lora=bool(self.lora_config),
            max_num_seqs=self.scheduler_config.max_num_seqs,
            max_input_length=None,
            tokenizer_mode=self.model_config.tokenizer_mode,
            trust_remote_code=self.model_config.trust_remote_code,
            revision=self.model_config.tokenizer_revision)
        init_kwargs.update(tokenizer_init_kwargs)
        self.tokenizer: TokenizerGroup = TokenizerGroup(
            self.model_config.tokenizer, **init_kwargs)

    def _init_workers_ray(self, placement_group: "PlacementGroup",
                          **ray_remote_kwargs):
        if self.parallel_config.tensor_parallel_size == 1:
            num_gpus = self.cache_config.gpu_memory_utilization
        else:
            num_gpus = 1

        self.driver_dummy_worker: RayWorkerVllm = None
        self.workers: List[RayWorkerVllm] = []

        driver_ip = get_ip()
        for bundle_id, bundle in enumerate(placement_group.bundle_specs):
            if not bundle.get("GPU", 0):
                continue
            scheduling_strategy = PlacementGroupSchedulingStrategy(
                placement_group=placement_group,
                placement_group_capture_child_tasks=True,
                placement_group_bundle_index=bundle_id,
            )
            worker = ray.remote(
                num_cpus=0,
                num_gpus=num_gpus,
                scheduling_strategy=scheduling_strategy,
                **ray_remote_kwargs,
            )(RayWorkerVllm).remote(self.model_config.trust_remote_code)

            worker_ip = ray.get(worker.get_node_ip.remote())
            if worker_ip == driver_ip and self.driver_dummy_worker is None:
                # If the worker is on the same node as the driver, we use it
                # as the resource holder for the driver process.
                self.driver_dummy_worker = worker
            else:
                self.workers.append(worker)

        if self.driver_dummy_worker is None:
            raise ValueError(
                "Ray does not allocate any GPUs on the driver node. Consider "
                "adjusting the Ray placement group or running the driver on a "
                "GPU node.")

        driver_node_id, driver_gpu_ids = ray.get(
            self.driver_dummy_worker.get_node_and_gpu_ids.remote())
        worker_node_and_gpu_ids = ray.get(
            [worker.get_node_and_gpu_ids.remote() for worker in self.workers])

        node_workers = defaultdict(list)
        node_gpus = defaultdict(list)

        node_workers[driver_node_id].append(0)
        node_gpus[driver_node_id].extend(driver_gpu_ids)
        for i, (node_id, gpu_ids) in enumerate(worker_node_and_gpu_ids,
                                               start=1):
            node_workers[node_id].append(i)
            node_gpus[node_id].extend(gpu_ids)
        for node_id, gpu_ids in node_gpus.items():
            node_gpus[node_id] = sorted(gpu_ids)

        # Set CUDA_VISIBLE_DEVICES for the driver.
        set_cuda_visible_devices(node_gpus[driver_node_id])
        for worker, (node_id, _) in zip(self.workers, worker_node_and_gpu_ids):
            worker.set_cuda_visible_devices.remote(node_gpus[node_id])

        distributed_init_method = get_distributed_init_method(
            driver_ip, get_open_port())

        # Lazy import the Worker to avoid importing torch.cuda/xformers
        # before CUDA_VISIBLE_DEVICES is set in the Worker
        Worker = self._dispatch_worker()

        # Initialize torch distributed process group for the workers.
        model_config = copy.deepcopy(self.model_config)
        parallel_config = copy.deepcopy(self.parallel_config)
        scheduler_config = copy.deepcopy(self.scheduler_config)
        device_config = copy.deepcopy(self.device_config)

        for rank, (worker, (node_id,
                            _)) in enumerate(zip(self.workers,
                                                 worker_node_and_gpu_ids),
                                             start=1):
            local_rank = node_workers[node_id].index(rank)
            worker.init_worker.remote(
                lambda rank=rank, local_rank=local_rank: Worker(
                    model_config,
                    parallel_config,
                    scheduler_config,
                    device_config,
                    local_rank,
                    rank,
                    distributed_init_method,
                    lora_config=self.lora_config,
                    kv_cache_dtype=self.cache_config.cache_dtype,
                ))

        driver_rank = 0
        driver_local_rank = node_workers[driver_node_id].index(driver_rank)
        self.driver_worker = Worker(
            model_config,
            parallel_config,
            scheduler_config,
            device_config,
            driver_local_rank,
            driver_rank,
            distributed_init_method,
            lora_config=self.lora_config,
            kv_cache_dtype=self.cache_config.cache_dtype,
            is_driver_worker=True,
        )

        # don't use cupy for eager mode
        self._run_workers("init_model",
                          cupy_port=get_open_port()
                          if not model_config.enforce_eager else None)
        self._run_workers(
            "load_model",
            max_concurrent_workers=self.parallel_config.
            max_parallel_loading_workers,
        )

    def _verify_args(self) -> None:
        self.model_config.verify_with_parallel_config(self.parallel_config)
        self.cache_config.verify_with_parallel_config(self.parallel_config)
        if self.lora_config:
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)

    def _init_cache(self) -> None:
        """Profiles the memory usage and initializes the KV cache.

        The engine will first conduct a profiling of the existing memory usage.
        Then, it calculate the maximum possible number of GPU and CPU blocks
        that can be allocated with the remaining free memory.
        More details can be found in the
        :meth:`~vllm.worker.worker.Worker.profile_num_available_blocks` method
        from class :class:`~vllm.worker.Worker`.

        Afterwards, as there may be multiple workers,
        we take the minimum number of blocks across all workers
        to ensure this can be applied to all of them.

        Finally, the engine will initialize the KV cache
        with the calculated number of blocks.

        .. tip::
            You may limit the usage of GPU memory
            by adjusting the `gpu_memory_utilization` parameters.
        """
        # Get the maximum number of blocks that can be allocated on GPU and CPU.
        num_blocks = self._run_workers(
            "profile_num_available_blocks",
            block_size=self.cache_config.block_size,
            gpu_memory_utilization=self.cache_config.gpu_memory_utilization,
            cpu_swap_space=self.cache_config.swap_space_bytes,
            cache_dtype=self.cache_config.cache_dtype,
        )

        # Since we use a shared centralized controller, we take the minimum
        # number of blocks across all workers to make sure all the memory
        # operators can be applied to all workers.
        num_gpu_blocks = min(b[0] for b in num_blocks)
        num_cpu_blocks = min(b[1] for b in num_blocks)
        # FIXME(woosuk): Change to debug log.
        logger.info(f"# GPU blocks: {num_gpu_blocks}, "
                    f"# CPU blocks: {num_cpu_blocks}")

        if num_gpu_blocks <= 0:
            raise ValueError("No available memory for the cache blocks. "
                             "Try increasing `gpu_memory_utilization` when "
                             "initializing the engine.")
        max_seq_len = self.cache_config.block_size * num_gpu_blocks
        if self.model_config.max_model_len > max_seq_len:
            raise ValueError(
                f"The model's max seq len ({self.model_config.max_model_len}) "
                "is larger than the maximum number of tokens that can be "
                f"stored in KV cache ({max_seq_len}). Try increasing "
                "`gpu_memory_utilization` or decreasing `max_model_len` when "
                "initializing the engine.")

        self.cache_config.num_gpu_blocks = num_gpu_blocks
        self.cache_config.num_cpu_blocks = num_cpu_blocks

        # Initialize the cache.
        self._run_workers("init_cache_engine", cache_config=self.cache_config)
        # Warm up the model. This includes capturing the model into CUDA graph
        # if enforce_eager is False.
        self._run_workers("warm_up_model")

    @classmethod
    def from_engine_args(cls, engine_args: EngineArgs) -> "LLMEngine":
        """Creates an LLM engine from the engine arguments."""
        # Create the engine configs.
        engine_configs = engine_args.create_engine_configs()
        parallel_config = engine_configs[2]
        # Initialize the cluster.
        placement_group = initialize_cluster(parallel_config)
        # Create the LLM engine.
        engine = cls(*engine_configs,
                     placement_group,
                     log_stats=not engine_args.disable_log_stats)
        return engine

    def encode_request(
        self,
        request_id: str,  # pylint: disable=unused-argument
        prompt: Optional[str],
        prompt_token_ids: Optional[List[int]] = None,
        lora_request: Optional[LoRARequest] = None,
    ):
        if prompt_token_ids is None:
            assert prompt is not None
            prompt_token_ids = self.tokenizer.encode(request_id=request_id,
                                                     prompt=prompt,
                                                     lora_request=lora_request)
        return prompt_token_ids

    def add_request(
        self,
        request_id: str,
        prompt: Optional[str],
        sampling_params: SamplingParams,
        prompt_token_ids: Optional[List[int]] = None,
        arrival_time: Optional[float] = None,
        lora_request: Optional[LoRARequest] = None,
    ) -> None:
        """Add a request to the engine's request pool.

        The request is added to the request pool and will be processed by the
        scheduler as `engine.step()` is called. The exact scheduling policy is
        determined by the scheduler.

        Args:
            request_id: The unique ID of the request.
            prompt: The prompt string. Can be None if prompt_token_ids is
                provided.
            sampling_params: The sampling parameters for text generation.
            prompt_token_ids: The token IDs of the prompt. If None, we
                use the tokenizer to convert the prompts to token IDs.
            arrival_time: The arrival time of the request. If None, we use
                the current monotonic time.

        Details:
            - Set arrival_time to the current time if it is None.
            - Set prompt_token_ids to the encoded prompt if it is None.
            - Create `best_of` number of :class:`~vllm.Sequence` objects.
            - Create a :class:`~vllm.SequenceGroup` object
              from the list of :class:`~vllm.Sequence`.
            - Add the :class:`~vllm.SequenceGroup` object to the scheduler.

        Example:
            >>> # initialize engine
            >>> engine = LLMEngine.from_engine_args(engine_args)
            >>> # set request arguments
            >>> example_prompt = "Who is the president of the United States?"
            >>> sampling_params = SamplingParams(temperature=0.0)
            >>> request_id = 0
            >>>
            >>> # add the request to the engine
            >>> engine.add_request(
            >>>    str(request_id),
            >>>    example_prompt,
            >>>    SamplingParams(temperature=0.0))
            >>> # continue the request processing
            >>> ...
        """
        if lora_request is not None and not self.lora_config:
            raise ValueError(f"Got lora_request {lora_request} but LoRA is "
                             "not enabled!")
        max_logprobs = self.get_model_config().max_logprobs
        if (sampling_params.logprobs
                and sampling_params.logprobs > max_logprobs) or (
                    sampling_params.prompt_logprobs
                    and sampling_params.prompt_logprobs > max_logprobs):
            raise ValueError(f"Cannot request more than "
                             f"{max_logprobs} logprobs.")
        if arrival_time is None:
            arrival_time = time.monotonic()
        prompt_token_ids = self.encode_request(
            request_id=request_id,
            prompt=prompt,
            prompt_token_ids=prompt_token_ids,
            lora_request=lora_request)

        # Create the sequences.
        block_size = self.cache_config.block_size
        seq_id = next(self.seq_counter)
        eos_token_id = self.tokenizer.get_lora_tokenizer(
            lora_request).eos_token_id
        seq = Sequence(seq_id, prompt, prompt_token_ids, block_size,
                       eos_token_id, lora_request)

        # Defensive copy of SamplingParams, which are used by the sampler,
        # this doesn't deep-copy LogitsProcessor objects
        sampling_params = sampling_params.clone()

        # Create the sequence group.
        seq_group = SequenceGroup(request_id, [seq], sampling_params,
                                  arrival_time, lora_request)

        # Add the sequence group to the scheduler.
        self.scheduler.add_seq_group(seq_group)

    def abort_request(self, request_id: Union[str, Iterable[str]]) -> None:
        """Aborts a request(s) with the given ID.

        Args:
            request_id: The ID(s) of the request to abort.

        Details:
            - Refer to the
              :meth:`~vllm.core.scheduler.Scheduler.abort_seq_group`
              from class :class:`~vllm.core.scheduler.Scheduler`.

        Example:
            >>> # initialize engine and add a request with request_id
            >>> request_id = str(0)
            >>> # abort the request
            >>> engine.abort_request(request_id)
        """
        self.scheduler.abort_seq_group(request_id)

    def get_model_config(self) -> ModelConfig:
        """Gets the model configuration."""
        return self.model_config

    def get_num_unfinished_requests(self) -> int:
        """Gets the number of unfinished requests."""
        return self.scheduler.get_num_unfinished_seq_groups()

    def has_unfinished_requests(self) -> bool:
        """Returns True if there are unfinished requests."""
        return self.scheduler.has_unfinished_seqs()

    def _check_beam_search_early_stopping(
        self,
        early_stopping: Union[bool, str],
        sampling_params: SamplingParams,
        best_running_seq: Sequence,
        current_worst_seq: Sequence,
    ) -> bool:
        assert sampling_params.use_beam_search
        length_penalty = sampling_params.length_penalty
        if early_stopping is True:
            return True

        current_worst_score = current_worst_seq.get_beam_search_score(
            length_penalty=length_penalty,
            eos_token_id=current_worst_seq.eos_token_id)
        if early_stopping is False:
            highest_attainable_score = best_running_seq.get_beam_search_score(
                length_penalty=length_penalty,
                eos_token_id=best_running_seq.eos_token_id)
        else:
            assert early_stopping == "never"
            if length_penalty > 0.0:
                # If length_penalty > 0.0, beam search will prefer longer
                # sequences. The highest attainable score calculation is
                # based on the longest possible sequence length in this case.
                max_possible_length = max(
                    best_running_seq.get_prompt_len() +
                    sampling_params.max_tokens,
                    self.scheduler_config.max_model_len)
                highest_attainable_score = (
                    best_running_seq.get_beam_search_score(
                        length_penalty=length_penalty,
                        eos_token_id=best_running_seq.eos_token_id,
                        seq_len=max_possible_length))
            else:
                # Otherwise, beam search will prefer shorter sequences. The
                # highest attainable score calculation is based on the current
                # sequence length.
                highest_attainable_score = (
                    best_running_seq.get_beam_search_score(
                        length_penalty=length_penalty,
                        eos_token_id=best_running_seq.eos_token_id))
        return current_worst_score >= highest_attainable_score

    def _process_sequence_group_outputs(self, seq_group: SequenceGroup,
                                        outputs: SequenceGroupOutput) -> None:

        # Process prompt logprobs
        prompt_logprobs = outputs.prompt_logprobs
        if prompt_logprobs is not None:
            # We can pick any sequence for the prompt.
            seq = next(iter(seq_group.seqs_dict.values()))
            all_token_ids = seq.get_token_ids()
            for i, prompt_logprobs_for_token in enumerate(prompt_logprobs):
                self._decode_logprobs(seq, seq_group.sampling_params,
                                      prompt_logprobs_for_token,
                                      all_token_ids[:i])
            seq_group.prompt_logprobs = prompt_logprobs

        # Process samples
        samples = outputs.samples
        parent_seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
        existing_finished_seqs = seq_group.get_finished_seqs()
        parent_child_dict = {
            parent_seq.seq_id: []
            for parent_seq in parent_seqs
        }
        for sample in samples:
            parent_child_dict[sample.parent_seq_id].append(sample)
        # List of (child, parent)
        child_seqs: List[Tuple[Sequence, Sequence]] = []

        # Process the child samples for each parent sequence
        for parent in parent_seqs:
            child_samples: List[SequenceOutput] = parent_child_dict[
                parent.seq_id]
            if len(child_samples) == 0:
                # This parent sequence has no children samples. Remove
                # the parent sequence from the sequence group since it will
                # not be used in the future iterations.
                parent.status = SequenceStatus.FINISHED_ABORTED
                seq_group.remove(parent.seq_id)
                self.scheduler.free_seq(parent)
                continue
            # Fork the parent sequence if there are multiple child samples.
            for child_sample in child_samples[:-1]:
                new_child_seq_id = next(self.seq_counter)
                child = parent.fork(new_child_seq_id)
                child.append_token_id(child_sample.output_token,
                                      child_sample.logprobs)
                child_seqs.append((child, parent))
            # Continue the parent sequence for the last child sample.
            # We reuse the parent sequence here to reduce redundant memory
            # copies, especially when using non-beam search sampling methods.
            last_child_sample = child_samples[-1]
            parent.append_token_id(last_child_sample.output_token,
                                   last_child_sample.logprobs)
            child_seqs.append((parent, parent))

        for seq, _ in child_seqs:
            self._decode_sequence(seq, seq_group.sampling_params)
            self._check_stop(seq, seq_group.sampling_params)

        # Non-beam search case
        if not seq_group.sampling_params.use_beam_search:
            # For newly created child sequences, add them to the sequence group
            # and fork them in block manager if they are not finished.
            for seq, parent in child_seqs:
                if seq is not parent:
                    seq_group.add(seq)
                    if not seq.is_finished():
                        self.scheduler.fork_seq(parent, seq)

            # Free the finished and selected parent sequences' memory in block
            # manager. Keep them in the sequence group as candidate output.
            # NOTE: we need to fork the new sequences before freeing the
            # old sequences.
            for seq, parent in child_seqs:
                if seq is parent and seq.is_finished():
                    self.scheduler.free_seq(seq)
            return

        # Beam search case
        # Select the child sequences to keep in the sequence group.
        selected_child_seqs = []
        unselected_child_seqs = []
        beam_width = seq_group.sampling_params.best_of
        length_penalty = seq_group.sampling_params.length_penalty

        # Select the newly finished sequences with the highest scores
        # to replace existing finished sequences.
        # Tuple of (seq, parent, is_new)
        existing_finished_seqs = [(seq, None, False)
                                  for seq in existing_finished_seqs]
        new_finished_seqs = [(seq, parent, True) for seq, parent in child_seqs
                             if seq.is_finished()]
        all_finished_seqs = existing_finished_seqs + new_finished_seqs
        # Sort the finished sequences by their scores.
        all_finished_seqs.sort(key=lambda x: x[0].get_beam_search_score(
            length_penalty=length_penalty, eos_token_id=x[0].eos_token_id),
                               reverse=True)
        for seq, parent, is_new in all_finished_seqs[:beam_width]:
            if is_new:
                # A newly generated child sequence finishes and has a high
                # score, so we will add it into the sequence group.
                selected_child_seqs.append((seq, parent))
        for seq, parent, is_new in all_finished_seqs[beam_width:]:
            if is_new:
                # A newly generated child sequence finishes but has a low
                # score, so we will not add it into the sequence group.
                # Additionally, if this sequence is a continuation of a
                # parent sequence, we will need remove the parent sequence
                # from the sequence group.
                unselected_child_seqs.append((seq, parent))
            else:
                # An existing finished sequence has a low score, so we will
                # remove it from the sequence group.
                seq_group.remove(seq.seq_id)

        # select the top beam_width sequences from the running
        # sequences for the next iteration to continue the beam
        # search.
        running_child_seqs = [(seq, parent) for seq, parent in child_seqs
                              if not seq.is_finished()]
        # Sort the running sequences by their scores.
        running_child_seqs.sort(key=lambda x: x[0].get_beam_search_score(
            length_penalty=length_penalty, eos_token_id=x[0].eos_token_id),
                                reverse=True)

        # Check if we can stop the beam search.
        if len(running_child_seqs) == 0:
            # No running sequences, stop the beam search.
            stop_beam_search = True
        elif len(all_finished_seqs) < beam_width:
            # Not enough finished sequences, continue the beam search.
            stop_beam_search = False
        else:
            # Check the early stopping criteria
            best_running_seq = running_child_seqs[0][0]
            current_worst_seq = all_finished_seqs[beam_width - 1][0]
            stop_beam_search = self._check_beam_search_early_stopping(
                seq_group.sampling_params.early_stopping,
                seq_group.sampling_params, best_running_seq, current_worst_seq)

        if stop_beam_search:
            # Stop the beam search and remove all the running sequences from
            # the sequence group.
            unselected_child_seqs.extend(running_child_seqs)
        else:
            # Continue the beam search and select the top beam_width sequences
            # to continue the beam search.
            selected_child_seqs.extend(running_child_seqs[:beam_width])
            # The remaining running sequences will not be used in the next
            # iteration. Again, if these sequences are continuations of
            # parent sequences, we will need to remove the parent sequences
            # from the sequence group.
            unselected_child_seqs.extend(running_child_seqs[beam_width:])

        # For newly created child sequences, add them to the sequence group
        # and fork them in block manager if they are not finished.
        for seq, parent in selected_child_seqs:
            if seq is not parent:
                seq_group.add(seq)
                if not seq.is_finished():
                    self.scheduler.fork_seq(parent, seq)

        # Free the finished and selected parent sequences' memory in block
        # manager. Keep them in the sequence group as candidate output.
        for seq, parent in selected_child_seqs:
            if seq is parent and seq.is_finished():
                self.scheduler.free_seq(seq)

        # Remove the unselected parent sequences from the sequence group and
        # free their memory in block manager.
        for seq, parent in unselected_child_seqs:
            if seq is parent:
                # Remove the parent sequence if it is not selected for next
                # iteration
                seq_group.remove(seq.seq_id)
                self.scheduler.free_seq(seq)

    def _process_model_outputs(
            self, output: SamplerOutput,
            scheduler_outputs: SchedulerOutputs) -> List[RequestOutput]:
        now = time.time()
        # Update the scheduled sequence groups with the model outputs.
        scheduled_seq_groups = scheduler_outputs.scheduled_seq_groups

        # If prefix caching is enabled, mark all blocks in the sequence groups
        # as completed so that future requests don't attempt to recompute them
        if self.cache_config.enable_prefix_caching:
            for seq_group in scheduled_seq_groups:
                self.scheduler.mark_blocks_as_computed(seq_group)

        for seq_group, outputs in zip(scheduled_seq_groups, output):
            self._process_sequence_group_outputs(seq_group, outputs)

        # Free the finished sequence groups.
        self.scheduler.free_finished_seq_groups()

        # Create the outputs.
        request_outputs: List[RequestOutput] = []
        for seq_group in scheduled_seq_groups:
            seq_group.maybe_set_first_token_time(now)
            request_output = RequestOutput.from_seq_group(seq_group)
            request_outputs.append(request_output)
        for seq_group in scheduler_outputs.ignored_seq_groups:
            request_output = RequestOutput.from_seq_group(seq_group)
            request_outputs.append(request_output)

        # Log stats.
        if self.log_stats:
            self.stat_logger.log(self._get_stats(scheduler_outputs))

        return request_outputs

    def step(self) -> List[RequestOutput]:
        """Performs one decoding iteration and returns newly generated results.

        .. figure:: https://i.imgur.com/sv2HssD.png
            :alt: Overview of the step function
            :align: center

            Overview of the step function.

        Details:
            - Step 1: Schedules the sequences to be executed in the next
              iteration and the token blocks to be swapped in/out/copy.

                - Depending on the scheduling policy,
                  sequences may be `preempted/reordered`.
                - A Sequence Group (SG) refer to a group of sequences
                  that are generated from the same prompt.

            - Step 2: Calls the workers to execute the model.
            - Step 3: Processes the model output. This mainly includes:

                - Decodes the relevant outputs.
                - Updates the scheduled sequence groups with model outputs
                  based on its `sampling parameters` (`use_beam_search` or not).
                - Frees the finished sequence groups.

            - Finally, it creates and returns the newly generated results.

        Example:
            >>> # Please see the example/ folder for more detailed examples.
            >>>
            >>> # initialize engine and request arguments
            >>> engine = LLMEngine.from_engine_args(engine_args)
            >>> example_inputs = [(0, "What is LLM?",
            >>>    SamplingParams(temperature=0.0))]
            >>>
            >>> # Start the engine with an event loop
            >>> while True:
            >>>     if example_inputs:
            >>>         req_id, prompt, sampling_params = example_inputs.pop(0)
            >>>         engine.add_request(str(req_id), prompt, sampling_params)
            >>>
            >>>     # continue the request processing
            >>>     request_outputs = engine.step()
            >>>     for request_output in request_outputs:
            >>>         if request_output.finished:
            >>>             # return or show the request output
            >>>
            >>>     if not (engine.has_unfinished_requests() or example_inputs):
            >>>         break
        """
        seq_group_metadata_list, scheduler_outputs = self.scheduler.schedule()

        if not scheduler_outputs.is_empty():
            # Execute the model.
            all_outputs = self._run_workers(
                "execute_model",
                driver_kwargs={
                    "seq_group_metadata_list": seq_group_metadata_list,
                    "blocks_to_swap_in": scheduler_outputs.blocks_to_swap_in,
                    "blocks_to_swap_out": scheduler_outputs.blocks_to_swap_out,
                    "blocks_to_copy": scheduler_outputs.blocks_to_copy,
                },
                use_ray_compiled_dag=USE_RAY_COMPILED_DAG)

            # Only the driver worker returns the sampling results.
            output = all_outputs[0]
        else:
            output = []

        return self._process_model_outputs(output, scheduler_outputs)

    def do_log_stats(self) -> None:
        """Forced log when no requests active."""
        if self.log_stats:
            self.stat_logger.log(self._get_stats(scheduler_outputs=None))

    def _get_stats(self,
                   scheduler_outputs: Optional[SchedulerOutputs]) -> Stats:
        """Get Stats to be Logged to Prometheus."""
        now = time.monotonic()

        # KV Cache Usage in %.
        num_total_gpu = self.cache_config.num_gpu_blocks
        num_free_gpu = self.scheduler.block_manager.get_num_free_gpu_blocks()
        gpu_cache_usage = 1.0 - (num_free_gpu / num_total_gpu)

        num_total_cpu = self.cache_config.num_cpu_blocks
        cpu_cache_usage = 0.
        if num_total_cpu > 0:
            num_free_cpu = self.scheduler.block_manager.get_num_free_cpu_blocks(
            )
            cpu_cache_usage = 1.0 - (num_free_cpu / num_total_cpu)

        # Scheduler State
        num_running = len(self.scheduler.running)
        num_swapped = len(self.scheduler.swapped)
        num_waiting = len(self.scheduler.waiting)

        # Iteration stats if we have scheduler output.
        num_prompt_tokens = 0
        num_generation_tokens = 0
        time_to_first_tokens = []
        time_per_output_tokens = []
        time_e2e_requests = []
        if scheduler_outputs is not None:
            prompt_run = scheduler_outputs.prompt_run

            # Number of Tokens.
            if prompt_run:
                num_prompt_tokens = sum(
                    len(seq_group.prompt_token_ids)
                    for seq_group in scheduler_outputs.scheduled_seq_groups)
                num_generation_tokens = sum(
                    seq_group.num_seqs()
                    for seq_group in scheduler_outputs.scheduled_seq_groups)
            else:
                num_generation_tokens = scheduler_outputs.num_batched_tokens

            # Latency Timings.
            time_last_iters = []
            for seq_group in scheduler_outputs.scheduled_seq_groups:
                # Time since last token. (n.b. updates seq_group.metrics.last_token_time)
                time_last_iters.append(seq_group.get_last_latency(now))
                # Time since arrival for all finished requests.
                if seq_group.is_finished():
                    time_e2e_requests.append(now -
                                             seq_group.metrics.arrival_time)

            time_to_first_tokens = time_last_iters if prompt_run else []
            time_per_output_tokens = [] if prompt_run else time_last_iters

        return Stats(
            now=now,
            num_running=num_running,
            num_swapped=num_swapped,
            num_waiting=num_waiting,
            gpu_cache_usage=gpu_cache_usage,
            cpu_cache_usage=cpu_cache_usage,
            num_prompt_tokens=num_prompt_tokens,
            num_generation_tokens=num_generation_tokens,
            time_to_first_tokens=time_to_first_tokens,
            time_per_output_tokens=time_per_output_tokens,
            time_e2e_requests=time_e2e_requests,
        )

    def _decode_logprobs(self, seq: Sequence, prms: SamplingParams,
                         logprobs: Dict[int, Logprob],
                         all_input_ids: List[int]) -> None:
        if not logprobs:
            return
        for token_id, sample_logprob in logprobs.items():
            if (sample_logprob.decoded_token is None and token_id != -1):
                all_input_ids_with_logprob = all_input_ids[:-1] + [token_id]
                _, new_text, prefix_offset, read_offset = detokenize_incrementally(
                    self.get_tokenizer_for_seq(seq),
                    all_input_ids=all_input_ids_with_logprob,
                    prev_tokens=seq.tokens,
                    prefix_offset=seq.prefix_offset,
                    read_offset=seq.read_offset,
                    skip_special_tokens=prms.skip_special_tokens,
                    spaces_between_special_tokens=prms.
                    spaces_between_special_tokens,
                )
                sample_logprob.decoded_token = new_text

    def _decode_sequence(self, seq: Sequence, prms: SamplingParams) -> None:
        """Decodes the new token for a sequence."""
        all_input_ids = seq.get_token_ids()
        self._decode_logprobs(seq, prms, seq.output_logprobs[-1],
                              all_input_ids)

        (new_tokens, new_output_text, prefix_offset,
         read_offset) = detokenize_incrementally(
             self.get_tokenizer_for_seq(seq),
             all_input_ids=all_input_ids,
             prev_tokens=seq.tokens,
             prefix_offset=seq.prefix_offset,
             read_offset=seq.read_offset,
             skip_special_tokens=prms.skip_special_tokens,
             spaces_between_special_tokens=prms.spaces_between_special_tokens,
         )
        if seq.tokens is None:
            seq.tokens = new_tokens
        else:
            seq.tokens.extend(new_tokens)
        seq.prefix_offset = prefix_offset
        seq.read_offset = read_offset
        seq.output_text += new_output_text

    def _check_stop(self, seq: Sequence,
                    sampling_params: SamplingParams) -> None:
        """Stop the finished sequences."""
        for stop_str in sampling_params.stop:
            if seq.output_text.endswith(stop_str):
                self._finalize_sequence(seq, sampling_params, stop_str)
                seq.status = SequenceStatus.FINISHED_STOPPED
                return
        if seq.get_last_token_id() in sampling_params.stop_token_ids:
            stop_str = self.get_tokenizer_for_seq(seq).convert_ids_to_tokens(
                seq.get_last_token_id())
            self._finalize_sequence(seq, sampling_params, stop_str)
            seq.status = SequenceStatus.FINISHED_STOPPED
            return

        # Check if the sequence has reached max_model_len.
        if seq.get_len() > self.scheduler_config.max_model_len:
            seq.status = SequenceStatus.FINISHED_LENGTH_CAPPED
            return

        # Check if the sequence has reached max_tokens.
        if seq.get_output_len() == sampling_params.max_tokens:
            seq.status = SequenceStatus.FINISHED_LENGTH_CAPPED
            return

        # Check if the sequence has generated the EOS token.
        if ((not sampling_params.ignore_eos)
                and seq.get_last_token_id() == seq.eos_token_id):
            seq.status = SequenceStatus.FINISHED_STOPPED
            return

    def _finalize_sequence(self, seq: Sequence,
                           sampling_params: SamplingParams,
                           stop_string: str) -> None:
        if sampling_params.include_stop_str_in_output:
            return

        if stop_string and seq.output_text.endswith(stop_string):
            # Truncate the output text so that the stop string is
            # not included in the output.
            seq.output_text = seq.output_text[:-len(stop_string)]

    def add_lora(self, lora_request: LoRARequest) -> bool:
        assert lora_request.lora_int_id > 0, "lora_id must be greater than 0."
        return self._run_workers(
            "add_lora",
            lora_request=lora_request,
        )

    def remove_lora(self, lora_id: int) -> bool:
        assert lora_id > 0, "lora_id must be greater than 0."
        return self._run_workers(
            "remove_lora",
            lora_id=lora_id,
        )

    def list_loras(self) -> List[int]:
        return self._run_workers("list_loras")

    def _run_workers(
        self,
        method: str,
        *args,
        driver_args: Optional[List[Any]] = None,
        driver_kwargs: Optional[Dict[str, Any]] = None,
        max_concurrent_workers: Optional[int] = None,
        use_ray_compiled_dag: bool = False,
        **kwargs,
    ) -> Any:
        """Runs the given method on all workers."""

        if max_concurrent_workers:
            raise NotImplementedError(
                "max_concurrent_workers is not supported yet.")

        if use_ray_compiled_dag:
            # Right now, compiled DAG can only accept a single
            # input. TODO(sang): Fix it.
            output_channels = self.forward_dag.execute(1)
        else:
            # Start the ray workers first.
            ray_worker_outputs = [
                worker.execute_method.remote(method, *args, **kwargs)
                for worker in self.workers
            ]

        if driver_args is None:
            driver_args = args
        if driver_kwargs is None:
            driver_kwargs = kwargs

        # Start the driver worker after all the ray workers.
        driver_worker_output = getattr(self.driver_worker,
                                       method)(*driver_args, **driver_kwargs)

        # Get the results of the ray workers.
        if self.workers:
            if use_ray_compiled_dag:
                try:
                    ray_worker_outputs = [
                        pickle.loads(chan.begin_read())
                        for chan in output_channels
                    ]
                finally:
                    # Has to call end_read in order to reuse the DAG.
                    for chan in output_channels:
                        chan.end_read()
            else:
                ray_worker_outputs = ray.get(ray_worker_outputs)

        return [driver_worker_output] + ray_worker_outputs

    def _compiled_ray_dag(self):
        import pkg_resources
        required_version = "2.9"
        current_version = pkg_resources.get_distribution("ray").version
        if current_version < required_version:
            raise ValueError(f"Ray version {required_version} or greater is "
                             f"required, but found {current_version}")

        from ray.dag import MultiOutputNode, InputNode
        assert self.parallel_config.worker_use_ray

        # Right now, compiled DAG requires at least 1 arg. We send
        # a dummy value for now. It will be fixed soon.
        with InputNode() as input_data:
            forward_dag = MultiOutputNode([
                worker.execute_model_compiled_dag_remote.bind(input_data)
                for worker in self.workers
            ])
        return forward_dag.experimental_compile()

    def check_health(self) -> None:
        """Raises an error if engine is unhealthy."""
        self._check_if_any_actor_is_dead()

    def _check_if_any_actor_is_dead(self):
        if not self.parallel_config.worker_use_ray:
            return

        if not self.workers:
            return

        dead_actors = []
        for actor in self.workers:
            actor_state = ray.state.actors(actor._ray_actor_id.hex())  # pylint: disable=protected-access
            if actor_state["State"] == "DEAD":
                dead_actors.append(actor)
        if dead_actors:
            raise RuntimeError("At least one Worker is dead. "
                               f"Dead Workers: {dead_actors}. ")