scheduler.py

import enum
import time
from typing import Dict, List, Optional, Tuple

from vllm.config import CacheConfig, SchedulerConfig
from vllm.core.block_manager import BlockSpaceManager
from vllm.core.policy import PolicyFactory
from vllm.logger import init_logger
from vllm.sequence import (Sequence, SequenceData, SequenceGroup,
                           SequenceGroupMetadata, SequenceOutputs,
                           SequenceStatus)

logger = init_logger(__name__)

_LOGGING_INTERVAL_SEC = 5


class PreemptionMode(enum.Enum):
    """Preemption modes.

    1. Swapping: Swap out the blocks of the preempted sequences to CPU memory
    and swap them back in when the sequences are resumed.
    2. Recomputation: Discard the blocks of the preempted sequences and
    recompute them when the sequences are resumed, treating the sequences as
    new prompts.
    """
    SWAP = enum.auto()
    RECOMPUTE = enum.auto()


class SchedulerOutputs:

    def __init__(
        self,
        blocks_to_swap_in: Dict[int, int],
        blocks_to_swap_out: Dict[int, int],
        blocks_to_copy: Dict[int, List[int]],
    ) -> None:
        self.blocks_to_swap_in = blocks_to_swap_in
        self.blocks_to_swap_out = blocks_to_swap_out
        self.blocks_to_copy = blocks_to_copy
        # Swap in and swap out should never happen at the same time.
        assert not (blocks_to_swap_in and blocks_to_swap_out)

    def is_empty(self) -> bool:
        return (not self.blocks_to_swap_in
                and not self.blocks_to_swap_out
                and not self.blocks_to_copy)


class Scheduler:

    def __init__(
        self,
        scheduler_config: SchedulerConfig,
        cache_config: CacheConfig,
        log_stats: bool,
    ) -> None:
        self.scheduler_config = scheduler_config
        self.cache_config = cache_config
        self.log_stats = log_stats

        # Instantiate the scheduling policy.
        self.policy = PolicyFactory.get_policy(policy_name='fcfs')
        # Create the block space manager.
        self.block_manager = BlockSpaceManager(
            block_size=self.cache_config.block_size,
            num_gpu_blocks=self.cache_config.num_gpu_blocks,
            num_cpu_blocks=self.cache_config.num_cpu_blocks,
        )

        # Sequence groups in the WAITING state.
        self.waiting: List[SequenceGroup] = []
        # Sequence groups in the RUNNING state.
        self.running: List[SequenceGroup] = []
        # Sequence groups in the SWAPPED state.
        self.swapped: List[SequenceGroup] = []

        self.last_logging_time: float = 0.0
        # List[timestamp, num_tokens]
        self.num_input_tokens: List[Tuple[float, int]] = []

    def add_seq_group(self, seq_group: SequenceGroup) -> None:
        # Add sequence groups to the waiting queue.
        self.waiting.append(seq_group)

    def abort_seq_group(self, request_id: str) -> None:
        for state_queue in [self.waiting, self.running, self.swapped]:
            for seq_group in state_queue:
                if seq_group.request_id == request_id:
                    # Remove the sequence group from the state queue.
                    state_queue.remove(seq_group)
                    for seq in seq_group.seqs:
                        if seq.is_finished():
                            continue
                        self.free_seq(seq, SequenceStatus.FINISHED_ABORTED)
                    return

    def has_unfinished_seqs(self) -> bool:
        return self.waiting or self.running or self.swapped

    def get_num_unfinished_seq_groups(self) -> int:
        return len(self.waiting) + len(self.running) + len(self.swapped)

    def _schedule(self) -> Tuple[SchedulerOutputs, List[str]]:
        # Blocks that need to be swaped or copied before model execution.
        blocks_to_swap_in: Dict[int, int] = {}
        blocks_to_swap_out: Dict[int, int] = {}
        blocks_to_copy: Dict[int, List[int]] = {}

        # Fix the current time.
        now = time.time()

        # NOTE(woosuk): We prioritize the sequence groups in the RUNNING state
        # in order to minimize the preemption overheads.
        # Preemption happens only when there is no available slot to keep all
        # the sequence groups in the RUNNING state.
        # In this case, the policy is responsible for deciding which sequence
        # groups to preempt.
        self.running = self.policy.sort_by_priority(now, self.running)

        # Reserve new token slots for the running sequence groups.
        running: List[SequenceGroup] = []
        preempted: List[SequenceGroup] = []
        while self.running:
            seq_group = self.running.pop(0)
            while not self.block_manager.can_append_slot(seq_group):
                if self.running:
                    # Preempt the lowest-priority sequence groups.
                    victim_seq_group = self.running.pop(-1)
                    self._preempt(victim_seq_group, blocks_to_swap_out)
                    preempted.append(victim_seq_group)
                else:
                    # No other sequence groups can be preempted.
                    # Preempt the current sequence group.
                    self._preempt(seq_group, blocks_to_swap_out)
                    preempted.append(seq_group)
                    break
            else:
                # Append new slots to the sequence group.
                self._append_slot(seq_group, blocks_to_copy)
                running.append(seq_group)
        self.running = running

        # Swap in the sequence groups in the SWAPPED state if possible.
        self.swapped = self.policy.sort_by_priority(now, self.swapped)
        while self.swapped and not blocks_to_swap_out:
            seq_group = self.swapped[0]
            # If the sequence group has been preempted in this step, stop.
            if seq_group in preempted:
                break
            # If the sequence group cannot be swapped in, stop.
            if not self.block_manager.can_swap_in(seq_group):
                break

            # The total number of sequences in the RUNNING state should not
            # exceed the maximum number of sequences.
            num_new_seqs = seq_group.num_seqs(status=SequenceStatus.SWAPPED)
            num_curr_seqs = sum(
                seq_group.num_seqs(status=SequenceStatus.RUNNING)
                for seq_group in self.running)
            if num_curr_seqs + num_new_seqs > self.scheduler_config.max_num_seqs:
                break

            seq_group = self.swapped.pop(0)
            self._swap_in(seq_group, blocks_to_swap_in)
            self._append_slot(seq_group, blocks_to_copy)
            self.running.append(seq_group)

        num_batched_tokens = sum(
            seq_group.num_seqs(status=SequenceStatus.RUNNING)
            for seq_group in self.running
        )

        # Join waiting sequences if possible.
        prompt_group_ids: List[str] = []
        # NOTE(woosuk): The sequence groups in the SWAPPED state are strictly
        # prioritized over the sequence groups in the WAITING state.
        # This is because we want to bound the amount of CPU memory taken by
        # the swapped sequence groups.
        if not self.swapped:
            # Optimization: We do not sort the waiting queue since the preempted
            # sequence groups are added to the front and the new sequence groups
            # are added to the back.
            while self.waiting:
                seq_group = self.waiting[0]
                # If the sequence group has been preempted in this step, stop.
                if seq_group in preempted:
                    break
                # If the sequence group cannot be allocated, stop.
                if not self.block_manager.can_allocate(seq_group):
                    break

                # If the number of batched tokens exceeds the limit, stop.
                num_prompt_tokens = seq_group.get_seqs()[0].get_len()
                if (num_batched_tokens + num_prompt_tokens
                    > self.scheduler_config.max_num_batched_tokens):
                    break

                # The total number of sequences in the RUNNING state should not
                # exceed the maximum number of sequences.
                num_new_seqs = seq_group.num_seqs(status=SequenceStatus.WAITING)
                num_curr_seqs = sum(
                    seq_group.num_seqs(status=SequenceStatus.RUNNING)
                    for seq_group in self.running)
                if num_curr_seqs + num_new_seqs > self.scheduler_config.max_num_seqs:
                    break

                seq_group = self.waiting.pop(0)
                self._allocate(seq_group)
                self.running.append(seq_group)
                num_batched_tokens += num_prompt_tokens
                prompt_group_ids.append(seq_group.request_id)

        scheduler_outputs = SchedulerOutputs(
            blocks_to_swap_in=blocks_to_swap_in,
            blocks_to_swap_out=blocks_to_swap_out,
            blocks_to_copy=blocks_to_copy,
        )
        if not self.log_stats:
            return scheduler_outputs, prompt_group_ids

        # TODO(woosuk): Move the below code to the engine.
        now = time.time()
        if num_batched_tokens > 0:
            self.num_input_tokens.append((now, num_batched_tokens))
        elapsed_time = now - self.last_logging_time
        if elapsed_time > _LOGGING_INTERVAL_SEC:
            self.last_logging_time = now
            self.num_input_tokens = [
                (t, n) for t, n in self.num_input_tokens
                if now - t < _LOGGING_INTERVAL_SEC
            ]
            if len(self.num_input_tokens) > 1:
                total_num_tokens = sum(n for _, n in self.num_input_tokens[:-1])
                window = now - self.num_input_tokens[0][0]
                avg_throughput = total_num_tokens / window
            else:
                avg_throughput = 0.0

            total_num_gpu_blocks = self.cache_config.num_gpu_blocks
            num_free_gpu_blocks = self.block_manager.get_num_free_gpu_blocks()
            num_used_gpu_blocks = total_num_gpu_blocks - num_free_gpu_blocks
            gpu_cache_usage = num_used_gpu_blocks / total_num_gpu_blocks

            total_num_cpu_blocks = self.cache_config.num_cpu_blocks
            if total_num_cpu_blocks > 0:
                num_free_cpu_blocks = self.block_manager.get_num_free_cpu_blocks()
                num_used_cpu_blocks = total_num_cpu_blocks - num_free_cpu_blocks
                cpu_cache_usage = num_used_cpu_blocks / total_num_cpu_blocks
            else:
                cpu_cache_usage = 0.0

            logger.info(
                f"Throughput: {avg_throughput:.1f} tokens/s, "
                f"Running: {len(self.running)} reqs, "
                f"Swapped: {len(self.swapped)} reqs, "
                f"Pending: {len(self.waiting)} reqs, "
                f"GPU KV cache usage: {gpu_cache_usage * 100:.1f}%, "
                f"CPU KV cache usage: {cpu_cache_usage * 100:.1f}%")
        return scheduler_outputs, prompt_group_ids

    def schedule(self) -> Tuple[List[SequenceGroupMetadata], SchedulerOutputs]:
        # Schedule sequence groups.
        # This function call changes the internal states of the scheduler
        # such as self.running, self.swapped, and self.waiting.
        scheduler_outputs, prompt_group_ids = self._schedule()

        # Create input data structures.
        seq_group_metadata_list: List[SequenceGroupMetadata] = []
        for seq_group in self.running:
            is_prompt = seq_group.request_id in prompt_group_ids

            seq_data: Dict[int, List[SequenceData]] = {}
            block_tables: Dict[int, List[int]] = {}
            for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
                seq_id = seq.seq_id
                seq_data[seq_id] = seq.data
                block_tables[seq_id] = self.block_manager.get_block_table(seq)

            seq_group_metadata = SequenceGroupMetadata(
                request_id=seq_group.request_id,
                is_prompt=is_prompt,
                seq_data=seq_data,
                sampling_params=seq_group.sampling_params,
                block_tables=block_tables,
            )
            seq_group_metadata_list.append(seq_group_metadata)
        return seq_group_metadata_list, scheduler_outputs

    def update(
        self,
        seq_outputs: Dict[int, SequenceOutputs],
    ) -> List[SequenceGroup]:
        # Update the running sequences and free blocks.
        for seq_group in self.running:
            # Process beam search results before processing the new tokens.
            for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
                output = seq_outputs[seq.seq_id]
                if seq.seq_id != output.parent_seq_id:
                    # The sequence is a fork of the parent sequence (beam search).
                    # Free the current sequence.
                    self.block_manager.free(seq)
                    # Fork the parent sequence.
                    parent_seq = seq_group.find(output.parent_seq_id)
                    parent_seq.fork(seq)
                    self.block_manager.fork(parent_seq, seq)

            # Process the new tokens.
            for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
                # Append a new token to the sequence.
                output = seq_outputs[seq.seq_id]
                seq.append_token_id(output.output_token, output.logprobs)
        # Return a shallow copy of the running queue to prevent the queue
        # from being modified by the caller.
        return self.running.copy()

    def free_seq(self, seq: Sequence, finish_status: SequenceStatus) -> None:
        seq.status = finish_status
        self.block_manager.free(seq)

    def free_finished_seq_groups(self) -> None:
        self.running = [
            seq_group for seq_group in self.running
            if not seq_group.is_finished()
        ]

    def _allocate(self, seq_group: SequenceGroup) -> None:
        self.block_manager.allocate(seq_group)
        for seq in seq_group.get_seqs():
            seq.status = SequenceStatus.RUNNING

    def _append_slot(
        self,
        seq_group: SequenceGroup,
        blocks_to_copy: Dict[int, List[int]],
    ) -> None:
        for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
            ret = self.block_manager.append_slot(seq)
            if ret is not None:
                src_block, dst_block = ret
                if src_block in blocks_to_copy:
                    blocks_to_copy[src_block].append(dst_block)
                else:
                    blocks_to_copy[src_block] = [dst_block]

    def _preempt(
        self,
        seq_group: SequenceGroup,
        blocks_to_swap_out: Dict[int, int],
        preemption_mode: Optional[PreemptionMode] = None,
    ) -> None:
        # If preemption mode is not specified, we determine the mode as follows:
        # We use recomputation by default since it incurs lower overhead than
        # swapping. However, when the sequence group has multiple sequences
        # (e.g., beam search), recomputation is not supported. In such a case,
        # we use swapping instead.
        # FIXME(woosuk): This makes our scheduling policy a bit bizarre.
        # As swapped sequences are prioritized over waiting sequences,
        # sequence groups with multiple sequences are implicitly prioritized
        # over sequence groups with a single sequence.
        # TODO(woosuk): Support recomputation for sequence groups with multiple
        # sequences. This may require a more sophisticated CUDA kernel.
        if preemption_mode is None:
            seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
            if len(seqs) == 1:
                preemption_mode = PreemptionMode.RECOMPUTE
            else:
                preemption_mode = PreemptionMode.SWAP
        if preemption_mode == PreemptionMode.RECOMPUTE:
            self._preempt_by_recompute(seq_group)
        elif preemption_mode == PreemptionMode.SWAP:
            self._preempt_by_swap(seq_group, blocks_to_swap_out)
        else:
            assert False, 'Invalid preemption mode.'

    def _preempt_by_recompute(
        self,
        seq_group: SequenceGroup,
    ) -> None:
        seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
        assert len(seqs) == 1
        for seq in seqs:
            seq.status = SequenceStatus.WAITING
            self.block_manager.free(seq)
        # NOTE: For FCFS, we insert the preempted sequence group to the front
        # of the waiting queue.
        self.waiting.insert(0, seq_group)

    def _preempt_by_swap(
        self,
        seq_group: SequenceGroup,
        blocks_to_swap_out: Dict[int, int],
    ) -> None:
        seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
        for seq in seqs:
            seq.status = SequenceStatus.SWAPPED
        self._swap_out(seq_group, blocks_to_swap_out)
        self.swapped.append(seq_group)

    def _swap_in(
        self,
        seq_group: SequenceGroup,
        blocks_to_swap_in: Dict[int, int],
    ) -> None:
        mapping = self.block_manager.swap_in(seq_group)
        blocks_to_swap_in.update(mapping)
        for seq in seq_group.get_seqs(status=SequenceStatus.SWAPPED):
            seq.status = SequenceStatus.RUNNING

    def _swap_out(
        self,
        seq_group: SequenceGroup,
        blocks_to_swap_out: Dict[int, int],
    ) -> None:
        if not self.block_manager.can_swap_out(seq_group):
            # FIXME(woosuk): Abort the sequence group instead of aborting the
            # entire engine.
            raise RuntimeError(
                "Aborted due to the lack of CPU swap space. Please increase "
                "the swap space to avoid this error.")
        mapping = self.block_manager.swap_out(seq_group)
        blocks_to_swap_out.update(mapping)
        for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
            seq.status = SequenceStatus.SWAPPED