Iluvatar-mrv100 SDK 4.3.0

vllm/v1/core/sched/interface.py

@@ -0,0 +1,139 @@
+# SPDX-License-Identifier: Apache-2.0
+from abc import ABC, abstractmethod
+from collections.abc import Iterable
+from typing import TYPE_CHECKING, Optional, Union
+
+if TYPE_CHECKING:
+    from vllm.v1.core.sched.output import SchedulerOutput
+    from vllm.v1.engine import EngineCoreOutputs
+    from vllm.v1.metrics.stats import SchedulerStats
+    from vllm.v1.outputs import ModelRunnerOutput
+    from vllm.v1.request import Request, RequestStatus
+
+
+class SchedulerInterface(ABC):
+
+    @abstractmethod
+    def schedule(self) -> "SchedulerOutput":
+        """Schedule the requests to process in this scheduling step.
+
+        The scheduling decision is made at the iteration level. Each scheduling
+        step corresponds to a single forward pass of the model. Therefore, this
+        method is called repeatedly by a busy loop in the engine.
+
+        Essentially, the scheduler produces a dictionary of {req_id: num_tokens}
+        that specifies how many tokens to process for each request in this
+        scheduling step. For example, num_tokens can be as large as the number
+        of prompt tokens for new requests, or it can be 1 for the requests that
+        are auto-regressively generating new tokens one by one. Otherwise, it
+        can be somewhere in between in case of chunked prefills, prefix caching,
+        speculative decoding, etc.
+
+        Additionally, the scheduler also returns useful data about each request
+        or the batch as a whole. The model runner will use this information in
+        preparing inputs to the model.
+
+        Returns:
+            A SchedulerOutput object containing information about the scheduled
+            requests.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def update_from_output(
+        self,
+        scheduler_output: "SchedulerOutput",
+        model_runner_output: "ModelRunnerOutput",
+    ) -> "EngineCoreOutputs":
+        """Update the scheduler state based on the model runner output.
+
+        This method is called after the model runner has processed the scheduled
+        requests. The model runner output includes generated token ids, draft
+        token ids for next step, etc. The scheduler uses this information to
+        update its states, checks the finished requests, and returns the output
+        for each request.
+
+        Returns:
+            A EngineCoreOutputs object containing the outputs for each request.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def add_request(self, request: "Request") -> None:
+        """Add a new request to the scheduler's internal queue.
+        
+        Args:
+            request: The new request being added.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def finish_requests(
+        self,
+        request_ids: Union[str, Iterable[str]],
+        finished_status: "RequestStatus",
+    ) -> None:
+        """Finish the requests in the scheduler's internal queue. If the request
+        is not in the queue, this method will do nothing.
+
+        This method is called in two cases:
+        1. When the request is aborted by the client.
+        2. When the frontend process detects a stop string of the request after
+           de-tokenizing its generated tokens.
+           
+        Args:
+            request_ids: A single or a list of request IDs.
+            finished_status: The finished status of the given requests.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def get_num_unfinished_requests(self) -> int:
+        """Number of unfinished requests in the scheduler's internal queue."""
+        raise NotImplementedError
+
+    def has_unfinished_requests(self) -> bool:
+        """Returns True if there are unfinished requests in the scheduler's
+        internal queue."""
+        return self.get_num_unfinished_requests() > 0
+
+    @abstractmethod
+    def has_finished_requests(self) -> bool:
+        """Returns True if there are finished requests that need to be cleared.
+        NOTE: This is different from `not self.has_unfinished_requests()`.
+
+        The scheduler maintains an internal list of the requests finished in the
+        previous step. This list is returned from the next call to schedule(),
+        to be sent to the model runner in the next step to clear cached states
+        for these finished requests.
+
+        This method checks if this internal list of finished requests is
+        non-empty. This information is useful for DP attention.
+        """
+        raise NotImplementedError
+
+    def has_requests(self) -> bool:
+        """Returns True if there are unfinished requests, or finished requests
+        not yet returned in SchedulerOutputs."""
+        return self.has_unfinished_requests() or self.has_finished_requests()
+
+    @abstractmethod
+    def get_num_unscheduled_requests(self) -> int:
+        """Number of requests that are not being processed by the executor."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def reset_prefix_cache(self) -> bool:
+        """Reset the prefix cache for KV cache.
+
+        This is particularly required when the model weights are live-updated.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def make_stats(self) -> Optional["SchedulerStats"]:
+        """Make a SchedulerStats object for logging.
+
+        The SchedulerStats object is created for every scheduling step.
+        """
+        raise NotImplementedError

vllm/v1/core/sched/output.py

@@ -0,0 +1,123 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Optional
+
+if TYPE_CHECKING:
+    import numpy as np
+    import numpy.typing as npt
+
+    from vllm.lora.request import LoRARequest
+    from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
+    from vllm.sampling_params import SamplingParams
+    from vllm.v1.request import Request
+
+
+@dataclass
+class NewRequestData:
+
+    req_id: str
+    prompt_token_ids: list[int]
+    prompt: Optional[str]
+    mm_inputs: list[MultiModalKwargs]
+    mm_hashes: list[str]
+    mm_positions: list[PlaceholderRange]
+    sampling_params: SamplingParams
+    block_ids: list[int]
+    num_computed_tokens: int
+    lora_request: Optional[LoRARequest]
+
+    @classmethod
+    def from_request(
+        cls,
+        request: Request,
+        block_ids: list[int],
+    ) -> NewRequestData:
+        return cls(
+            req_id=request.request_id,
+            prompt_token_ids=request.prompt_token_ids,
+            prompt=request.prompt,
+            mm_inputs=request.mm_inputs,
+            mm_hashes=request.mm_hashes,
+            mm_positions=request.mm_positions,
+            sampling_params=request.sampling_params,
+            block_ids=block_ids,
+            num_computed_tokens=request.num_computed_tokens,
+            lora_request=request.lora_request,
+        )
+
+
+@dataclass
+class CachedRequestData:
+
+    req_id: str
+    # If resumed_from_preemption is False, new_block_ids will be appended to
+    # the request's block IDs. If True, new_block_ids will be used as the
+    # request's block IDs instead of appending to the existing block IDs.
+    resumed_from_preemption: bool
+    new_token_ids: list[int]
+    new_block_ids: list[int]
+    num_computed_tokens: int
+
+    @classmethod
+    def from_request(
+        cls,
+        request: Request,
+        resumed_from_preemption: bool,
+        new_token_ids: list[int],
+        new_block_ids: list[int],
+    ) -> CachedRequestData:
+        return cls(
+            req_id=request.request_id,
+            resumed_from_preemption=resumed_from_preemption,
+            new_token_ids=new_token_ids,
+            new_block_ids=new_block_ids,
+            num_computed_tokens=request.num_computed_tokens,
+        )
+
+
+@dataclass
+class SchedulerOutput:
+
+    # list of the requests that are scheduled for the first time.
+    # We cache the request's data in each worker process, so that we don't
+    # need to re-send it every scheduling step.
+    scheduled_new_reqs: list[NewRequestData]
+    # list of the requests that have been scheduled before.
+    # Since the request's data is already cached in the worker processes,
+    # we only send the diff to minimize the communication cost.
+    scheduled_cached_reqs: list[CachedRequestData]
+
+    # req_id -> num_scheduled_tokens
+    # Number of tokens scheduled for each request.
+    num_scheduled_tokens: dict[str, int]
+    # Total number of tokens scheduled for all requests.
+    # Equal to sum(num_scheduled_tokens.values())
+    total_num_scheduled_tokens: int
+    # req_id -> spec_token_ids
+    # If a request does not have any spec decode tokens, it will not be
+    # included in the dictionary.
+    scheduled_spec_decode_tokens: dict[str, list[int]]
+    # req_id -> encoder input indices that need processing.
+    # E.g., if a request has [0, 1], it could mean the vision encoder needs
+    # to process that the request's 0-th and 1-th images in the current step.
+    scheduled_encoder_inputs: dict[str, list[int]]
+    # Number of common prefix blocks for all requests.
+    # This can be used for cascade attention.
+    num_common_prefix_blocks: int
+
+    # Request IDs that are finished in between the previous and the current
+    # steps. This is used to notify the workers about the finished requests
+    # so that they can free the cached states for those requests.
+    finished_req_ids: set[str]
+    # list of (req_id, encoder_input_index) tuples.
+    # Used to free the encoder cache.
+    free_encoder_input_ids: list[tuple[str, int]]
+
+    # Dict of request ids to their index within the batch
+    # for filling the next token bitmask
+    structured_output_request_ids: dict[str, int]
+    # the bitmask for the whole batch
+    grammar_bitmask: Optional[npt.NDArray[np.int32]]

vllm/v1/core/sched/scheduler.py

@@ -0,0 +1,759 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import time
+from collections import deque
+from collections.abc import Iterable
+from typing import Optional, Union
+
+from vllm.config import CacheConfig, LoRAConfig, ModelConfig, SchedulerConfig
+from vllm.logger import init_logger
+from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
+from vllm.v1.core.encoder_cache_manager import (EncoderCacheManager,
+                                                compute_encoder_budget)
+from vllm.v1.core.kv_cache_manager import KVCacheManager
+from vllm.v1.core.sched.interface import SchedulerInterface
+from vllm.v1.core.sched.output import (CachedRequestData, NewRequestData,
+                                       SchedulerOutput)
+from vllm.v1.core.sched.utils import check_stop
+from vllm.v1.engine import (EngineCoreEventType, EngineCoreOutput,
+                            EngineCoreOutputs)
+from vllm.v1.kv_cache_interface import KVCacheConfig
+from vllm.v1.metrics.stats import SchedulerStats
+from vllm.v1.outputs import ModelRunnerOutput
+from vllm.v1.request import Request, RequestStatus
+from vllm.v1.spec_decode.metrics import SpecDecodingStats
+from vllm.v1.structured_output import StructuredOutputManager
+
+logger = init_logger(__name__)
+
+
+class Scheduler(SchedulerInterface):
+
+    def __init__(
+        self,
+        scheduler_config: SchedulerConfig,
+        model_config: ModelConfig,
+        cache_config: CacheConfig,
+        lora_config: Optional[LoRAConfig],
+        kv_cache_config: KVCacheConfig,
+        structured_output_manager: StructuredOutputManager,
+        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
+        include_finished_set: bool = False,
+        log_stats: bool = False,
+    ) -> None:
+        self.scheduler_config = scheduler_config
+        self.cache_config = cache_config
+        self.lora_config = lora_config
+        self.kv_cache_config = kv_cache_config
+        self.log_stats = log_stats
+        self.structured_output_manager = structured_output_manager
+
+        # include_finished_set controls whether a separate set of finished
+        # request ids should be included in the EngineCoreOutputs returned
+        # by update_from_outputs(). This is currently used in the multi-engine
+        # case to track request lifetimes efficiently.
+        self.include_finished_set = include_finished_set
+
+        # Scheduling constraints.
+        self.max_num_running_reqs = self.scheduler_config.max_num_seqs
+        self.max_num_scheduled_tokens = \
+            self.scheduler_config.max_num_batched_tokens
+        self.max_model_len = self.scheduler_config.max_model_len
+
+        # Create the KV cache manager.
+        self.kv_cache_manager = KVCacheManager(
+            kv_cache_config=kv_cache_config,
+            max_model_len=self.max_model_len,
+            enable_caching=cache_config.enable_prefix_caching,
+            caching_hash_algo=self.cache_config.prefix_caching_hash_algo,
+            log_stats=self.log_stats)
+        self.block_size = self.cache_config.block_size
+
+        # req_id -> Request
+        self.requests: dict[str, Request] = {}
+        # Priority queues for requests.
+        self.waiting: deque[Request] = deque()
+        self.running: list[Request] = []
+        # The requests that have been scheduled and are being executed
+        # by the executor.
+        self.scheduled_req_ids: set[str] = set()
+
+        # The request IDs that are finished in between the previous and the
+        # current steps. This is used to notify the workers about the finished
+        # requests so that they can free the cached states for those requests.
+        # This is flushed at the end of each scheduling step.
+        self.finished_req_ids: set[str] = set()
+
+        # OPTIMIZATION: Cache the CachedRequestData objects to avoid creating
+        # them at each scheduling step.
+        # Request id -> CachedRequestData
+        self._cached_reqs_data: dict[str, CachedRequestData] = {}
+
+        # Encoder-related.
+        # Calculate encoder cache size if applicable
+        # NOTE: For now we use the same budget for both compute and space.
+        # This can be changed when we make encoder cache for embedding caching
+        # across requests.
+        encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
+            model_config=model_config,
+            scheduler_config=scheduler_config,
+            mm_registry=mm_registry,
+        )
+
+        # NOTE(woosuk): Here, "encoder" includes the vision encoder (and
+        # projector if needed). Currently, we assume that the encoder also
+        # has the Transformer architecture (e.g., ViT).
+        self.max_num_encoder_input_tokens = encoder_compute_budget
+        # NOTE: For the models without encoder (e.g., text-only models),
+        # the encoder cache will not be initialized because cache size is 0
+        # for these models.
+        self.encoder_cache_manager = EncoderCacheManager(
+            cache_size=encoder_cache_size)
+
+    def schedule(self) -> SchedulerOutput:
+        # NOTE(woosuk) on the scheduling algorithm:
+        # There's no "decoding phase" nor "prefill phase" in the scheduler.
+        # Each request just has the num_computed_tokens and
+        # num_tokens_with_spec. num_tokens_with_spec =
+        # len(prompt_token_ids) + len(output_token_ids) + len(spec_token_ids).
+        # At each step, the scheduler tries to assign tokens to the requests
+        # so that each request's num_computed_tokens can catch up its
+        # num_tokens_with_spec. This is general enough to cover
+        # chunked prefills, prefix caching, speculative decoding,
+        # and the "jump decoding" optimization in the future.
+
+        scheduled_new_reqs: list[Request] = []
+        scheduled_resumed_reqs: list[Request] = []
+        scheduled_running_reqs: list[Request] = []
+        preempted_reqs: list[Request] = []
+
+        # NOTE: structured_output_request_ids maps
+        # a request's (request that uses structured output)
+        # request_id to the running request index.
+        # This will helps us determine to slice the grammar bitmask
+        # and only applies valid mask for requests that
+        # uses structured decoding.
+        structured_output_request_ids: dict[str, int] = {}
+
+        req_to_new_block_ids: dict[str, list[int]] = {}
+        num_scheduled_tokens: dict[str, int] = {}
+        token_budget = self.max_num_scheduled_tokens
+        # Encoder-related.
+        scheduled_encoder_inputs: dict[str, list[int]] = {}
+        encoder_budget = self.max_num_encoder_input_tokens
+        # Spec decode-related.
+        scheduled_spec_decode_tokens: dict[str, list[int]] = {}
+
+        # For logging.
+        scheduled_timestamp = time.monotonic()
+
+        # First, schedule the RUNNING requests.
+        req_index = 0
+        while req_index < len(self.running) and token_budget > 0:
+            request = self.running[req_index]
+            if request.request_id in self.scheduled_req_ids:
+                # This request has already been scheduled.
+                req_index += 1
+                continue
+
+            num_new_tokens = (request.num_tokens_with_spec -
+                              request.num_computed_tokens)
+            if (0 < self.scheduler_config.long_prefill_token_threshold <
+                    num_new_tokens):
+                num_new_tokens = (
+                    self.scheduler_config.long_prefill_token_threshold)
+            num_new_tokens = min(num_new_tokens, token_budget)
+            assert num_new_tokens > 0
+
+            # Schedule encoder inputs.
+            if request.has_encoder_inputs:
+                (encoder_inputs_to_schedule, num_new_tokens,
+                 new_encoder_budget) = self._try_schedule_encoder_inputs(
+                     request, request.num_computed_tokens, num_new_tokens,
+                     encoder_budget)
+                if num_new_tokens == 0:
+                    # The request cannot be scheduled because the encoder budget
+                    # or the encoder cache is exhausted.
+                    # NOTE(woosuk): By using `continue` instead of `break` here,
+                    # we intentionally relax the strict FCFS scheduling policy
+                    # to allow lower-priority requests to be scheduled when a
+                    # higher-priority request is blocked by encoder constraints.
+                    req_index += 1
+                    continue
+            else:
+                encoder_inputs_to_schedule = None
+                new_encoder_budget = encoder_budget
+
+            while True:
+                new_blocks = self.kv_cache_manager.allocate_slots(
+                    request, num_new_tokens)
+                if new_blocks is None:
+                    # The request cannot be scheduled.
+                    # Preempt the lowest-priority request.
+                    preempted_req = self.running.pop()
+                    self.kv_cache_manager.free(preempted_req)
+                    preempted_req.status = RequestStatus.PREEMPTED
+                    preempted_req.num_computed_tokens = 0
+                    if self.log_stats:
+                        preempted_req.record_event(
+                            EngineCoreEventType.PREEMPTED, scheduled_timestamp)
+
+                    self.waiting.appendleft(preempted_req)
+                    preempted_reqs.append(preempted_req)
+                    if preempted_req == request:
+                        # No more request to preempt.
+                        can_schedule = False
+                        break
+                else:
+                    # The request can be scheduled.
+                    can_schedule = True
+                    break
+            if not can_schedule:
+                break
+            assert new_blocks is not None
+
+            # Schedule the request.
+            scheduled_running_reqs.append(request)
+            self.scheduled_req_ids.add(request.request_id)
+            if request.use_structured_output:
+                # PERF: in case of chunked prefill,
+                # request might not include any new tokens.
+                # Therefore, we might introduce some additional
+                # cycle to fill in the bitmask, which could be a big no-op.
+                structured_output_request_ids[request.request_id] = req_index
+            req_to_new_block_ids[request.request_id] = [
+                b.block_id for b in new_blocks
+            ]
+            num_scheduled_tokens[request.request_id] = num_new_tokens
+            token_budget -= num_new_tokens
+            req_index += 1
+
+            # Speculative decode related.
+            if request.spec_token_ids:
+                num_scheduled_spec_tokens = (num_new_tokens +
+                                             request.num_computed_tokens -
+                                             request.num_tokens)
+                if num_scheduled_spec_tokens > 0:
+                    # Trim spec_token_ids list to num_scheduled_spec_tokens.
+                    del request.spec_token_ids[num_scheduled_spec_tokens:]
+                    scheduled_spec_decode_tokens[request.request_id] = (
+                        request.spec_token_ids)
+
+            # Encoder-related.
+            if encoder_inputs_to_schedule:
+                scheduled_encoder_inputs[request.request_id] = (
+                    encoder_inputs_to_schedule)
+                # Allocate the encoder cache.
+                for i in encoder_inputs_to_schedule:
+                    self.encoder_cache_manager.allocate(request, i)
+                encoder_budget = new_encoder_budget
+
+        # Record the LoRAs in scheduled_running_reqs
+        scheduled_loras: set[int] = set()
+        if self.lora_config:
+            scheduled_loras = set(
+                req.lora_request.lora_int_id for req in scheduled_running_reqs
+                if req.lora_request and req.lora_request.lora_int_id > 0)
+            assert len(scheduled_loras) <= self.lora_config.max_loras
+
+        # Use a temporary deque to collect requests that need to be skipped
+        # and put back at the head of the waiting queue later
+        skipped_waiting_requests: deque[Request] = deque()
+
+        # Next, schedule the WAITING requests.
+        if not preempted_reqs:
+            while self.waiting and token_budget > 0:
+                if len(self.running) == self.max_num_running_reqs:
+                    break
+
+                request = self.waiting[0]
+
+                # Skip request if the structured output request is still waiting
+                # for FSM compilation.
+                if request.status == RequestStatus.WAITING_FOR_FSM:
+                    structured_output_req = request.structured_output_request
+                    if structured_output_req and structured_output_req.grammar:
+                        request.status = RequestStatus.WAITING
+                    else:
+                        self.waiting.popleft()
+                        skipped_waiting_requests.appendleft(request)
+                        continue
+
+                # Check that adding the request still respects the max_loras
+                # constraint.
+                if self.lora_config and request.lora_request and (
+                        len(scheduled_loras) == self.lora_config.max_loras
+                        and request.lora_request.lora_int_id
+                        not in scheduled_loras):
+                    # Scheduling would exceed max_loras, skip.
+                    self.waiting.popleft()
+                    skipped_waiting_requests.appendleft(request)
+                    continue
+
+                # Get already-cached tokens.
+                computed_blocks, num_computed_tokens = \
+                    self.kv_cache_manager.get_computed_blocks(request)
+                # Number of tokens to be scheduled.
+                # We use `request.num_tokens` instead of
+                # `request.num_prompt_tokens` to consider the resumed requests,
+                # which have output tokens.
+                num_new_tokens = request.num_tokens - num_computed_tokens
+                if (0 < self.scheduler_config.long_prefill_token_threshold <
+                        num_new_tokens):
+                    num_new_tokens = (
+                        self.scheduler_config.long_prefill_token_threshold)
+                num_new_tokens = min(num_new_tokens, token_budget)
+                assert num_new_tokens > 0
+
+                # Schedule encoder inputs.
+                if request.has_encoder_inputs:
+                    (encoder_inputs_to_schedule, num_new_tokens,
+                     new_encoder_budget) = self._try_schedule_encoder_inputs(
+                         request, num_computed_tokens, num_new_tokens,
+                         encoder_budget)
+                    if num_new_tokens == 0:
+                        # The request cannot be scheduled.
+                        break
+                else:
+                    encoder_inputs_to_schedule = None
+                    new_encoder_budget = encoder_budget
+
+                new_blocks = self.kv_cache_manager.allocate_slots(
+                    request, num_new_tokens, computed_blocks)
+                if new_blocks is None:
+                    # The request cannot be scheduled.
+                    break
+
+                self.waiting.popleft()
+                if request.use_structured_output:
+                    structured_output_request_ids[
+                        request.request_id] = req_index
+                req_index += 1
+                self.running.append(request)
+                self.scheduled_req_ids.add(request.request_id)
+                if self.log_stats:
+                    request.record_event(EngineCoreEventType.SCHEDULED,
+                                         scheduled_timestamp)
+                if request.status == RequestStatus.WAITING:
+                    scheduled_new_reqs.append(request)
+                elif request.status == RequestStatus.PREEMPTED:
+                    scheduled_resumed_reqs.append(request)
+                else:
+                    raise RuntimeError(
+                        f"Invalid request status: {request.status}")
+
+                if self.lora_config and request.lora_request:
+                    scheduled_loras.add(request.lora_request.lora_int_id)
+                req_to_new_block_ids[request.request_id] = [
+                    b.block_id for b in computed_blocks + new_blocks
+                ]
+                num_scheduled_tokens[request.request_id] = num_new_tokens
+                token_budget -= num_new_tokens
+                request.status = RequestStatus.RUNNING
+                request.num_computed_tokens = num_computed_tokens
+
+                # Encoder-related.
+                if encoder_inputs_to_schedule:
+                    scheduled_encoder_inputs[request.request_id] = (
+                        encoder_inputs_to_schedule)
+                    # Allocate the encoder cache.
+                    for i in encoder_inputs_to_schedule:
+                        self.encoder_cache_manager.allocate(request, i)
+                    encoder_budget = new_encoder_budget
+
+        # Put back any skipped requests at the head of the waiting queue
+        if skipped_waiting_requests:
+            self.waiting.extendleft(skipped_waiting_requests)
+
+        # Check if the scheduling constraints are satisfied.
+        total_num_scheduled_tokens = sum(num_scheduled_tokens.values())
+        assert total_num_scheduled_tokens <= self.max_num_scheduled_tokens
+        assert token_budget >= 0
+        assert len(self.running) <= self.max_num_running_reqs
+        # Since some requests in the RUNNING queue may not be scheduled in
+        # this step, the total number of scheduled requests can be smaller than
+        # len(self.running).
+        assert (len(scheduled_new_reqs) + len(scheduled_resumed_reqs) +
+                len(scheduled_running_reqs) <= len(self.running))
+
+        # Get the longest common prefix among all requests in the running queue.
+        # This can be potentially used for cascade attention.
+        num_common_prefix_blocks = 0
+        if self.running:
+            any_request = self.running[0]
+            num_common_prefix_blocks = (
+                self.kv_cache_manager.get_num_common_prefix_blocks(
+                    any_request, len(self.running)))
+
+        grammar_bitmask = self.structured_output_manager.grammar_bitmask(
+            self.requests,
+            structured_output_request_ids,
+            len(self.running),
+        )
+        # Construct the scheduler output.
+        new_reqs_data = [
+            NewRequestData.from_request(req,
+                                        req_to_new_block_ids[req.request_id])
+            for req in scheduled_new_reqs
+        ]
+        resumed_reqs_data = [
+            self._make_cached_request_data(
+                req,
+                num_scheduled_tokens[req.request_id],
+                len(scheduled_spec_decode_tokens.get(req.request_id, ())),
+                req_to_new_block_ids[req.request_id],
+                resumed_from_preemption=True,
+            ) for req in scheduled_resumed_reqs
+        ]
+        running_reqs_data = [
+            self._make_cached_request_data(
+                req,
+                num_scheduled_tokens[req.request_id],
+                len(scheduled_spec_decode_tokens.get(req.request_id, ())),
+                req_to_new_block_ids[req.request_id],
+                resumed_from_preemption=False,
+            ) for req in scheduled_running_reqs
+        ]
+        scheduler_output = SchedulerOutput(
+            scheduled_new_reqs=new_reqs_data,
+            scheduled_cached_reqs=resumed_reqs_data + running_reqs_data,
+            num_scheduled_tokens=num_scheduled_tokens,
+            total_num_scheduled_tokens=total_num_scheduled_tokens,
+            scheduled_spec_decode_tokens=scheduled_spec_decode_tokens,
+            scheduled_encoder_inputs=scheduled_encoder_inputs,
+            num_common_prefix_blocks=num_common_prefix_blocks,
+            # finished_req_ids is an existing state in the scheduler,
+            # instead of being newly scheduled in this step.
+            # It contains the request IDs that are finished in between
+            # the previous and the current steps.
+            finished_req_ids=self.finished_req_ids,
+            free_encoder_input_ids=self.encoder_cache_manager.get_freed_ids(),
+            structured_output_request_ids=structured_output_request_ids,
+            grammar_bitmask=grammar_bitmask,
+        )
+
+        # Advance the number of computed tokens for the request AFTER
+        # the request is scheduled.
+        # 1. The scheduler_output of the current step has to include the
+        #    original number of scheduled tokens to determine input IDs.
+        # 2. Advance the number of computed tokens here allowing us to
+        #    schedule the prefill request again immediately in the next
+        #    scheduling step.
+        # 3. If some tokens (e.g. spec tokens) are rejected later, the number of
+        #    computed tokens will be adjusted in update_from_output.
+        for req_id, num_scheduled_token in num_scheduled_tokens.items():
+            self.requests[req_id].num_computed_tokens += num_scheduled_token
+
+        self.finished_req_ids = set()
+        return scheduler_output
+
+    def _make_cached_request_data(
+        self,
+        request: Request,
+        num_scheduled_tokens: int,
+        num_scheduled_spec_tokens: int,
+        new_block_ids: list[int],
+        resumed_from_preemption: bool,
+    ) -> CachedRequestData:
+        # OPTIMIZATION: Cache the CachedRequestData objects to avoid creating
+        # them at each scheduling step.
+        num_computed_tokens = request.num_computed_tokens
+        num_regular_tokens = num_scheduled_tokens - num_scheduled_spec_tokens
+        new_token_ids = request.all_token_ids[
+            num_computed_tokens:num_computed_tokens + num_regular_tokens]
+        req_data = self._cached_reqs_data.get(request.request_id)
+        if req_data is not None:
+            req_data.resumed_from_preemption = resumed_from_preemption
+            req_data.new_token_ids = new_token_ids
+            req_data.new_block_ids = new_block_ids
+            req_data.num_computed_tokens = num_computed_tokens
+        else:
+            req_data = CachedRequestData.from_request(request,
+                                                      resumed_from_preemption,
+                                                      new_token_ids,
+                                                      new_block_ids)
+            self._cached_reqs_data[request.request_id] = req_data
+        return req_data
+
+    def _try_schedule_encoder_inputs(
+        self,
+        request: Request,
+        num_computed_tokens: int,
+        num_new_tokens: int,
+        encoder_budget: int,
+    ) -> tuple[list[int], int, int]:
+        """
+        Determine which encoder inputs need to be scheduled in the current step,
+        and update `num_new_tokens` and encoder token budget accordingly.
+
+        An encoder input will be scheduled if:
+        - Its output tokens overlap with the range of tokens being computed
+        in this step, i.e.,
+        [num_computed_tokens, num_computed_tokens + num_new_tokens).
+        - It is not already computed and stored in the encoder cache.
+        - There is sufficient encoder token budget to process it.
+        - The encoder cache has space to store it.
+
+        If an encoder input cannot be scheduled due to cache or budget
+        limitations, the method adjusts `num_new_tokens` to schedule only the
+        decoder tokens up to just before the unschedulable encoder input.
+        """
+        encoder_inputs_to_schedule: list[int] = []
+        mm_positions = request.mm_positions
+        assert mm_positions is not None
+        assert len(mm_positions) > 0
+        for i, pos_info in enumerate(mm_positions):
+            start_pos = pos_info["offset"]
+            num_encoder_tokens = pos_info["length"]
+
+            # The encoder output is needed if the two ranges overlap:
+            # [num_computed_tokens, num_computed_tokens + num_new_tokens) and
+            # [start_pos, start_pos + num_encoder_tokens)
+            if start_pos >= num_computed_tokens + num_new_tokens:
+                # The encoder input is not needed in this step.
+                break
+            if start_pos + num_encoder_tokens <= num_computed_tokens:
+                # The encoder input is already computed and stored
+                # in the decoder's KV cache.
+                continue
+
+            if self.encoder_cache_manager.has_cache(request, i):
+                # The encoder input is already computed and cached.
+                continue
+            if (not self.encoder_cache_manager.can_allocate(request, i)
+                    or num_encoder_tokens > encoder_budget):
+                # The encoder cache is full or the encoder budget is exhausted.
+                # NOTE(woosuk): We assume that the encoder input tokens should
+                # be processed altogether, as the encoder usually uses
+                # bidirectional attention.
+                if num_computed_tokens < start_pos:
+                    # We only schedule the decoder tokens just before the
+                    # encoder input.
+                    num_new_tokens = start_pos - num_computed_tokens
+                else:
+                    # Because of prefix caching, num_computed_tokens is greater
+                    # than start_pos even though its encoder input is not
+                    # available. In this case, we can't schedule any token for
+                    # the request in this step.
+                    num_new_tokens = 0
+                break
+
+            encoder_budget -= num_encoder_tokens
+            encoder_inputs_to_schedule.append(i)
+        return encoder_inputs_to_schedule, num_new_tokens, encoder_budget
+
+    def update_from_output(
+        self,
+        scheduler_output: SchedulerOutput,
+        model_runner_output: ModelRunnerOutput,
+    ) -> EngineCoreOutputs:
+        sampled_token_ids = model_runner_output.sampled_token_ids
+        spec_token_ids = model_runner_output.spec_token_ids
+        logprobs = model_runner_output.logprobs
+        prompt_logprobs_dict = model_runner_output.prompt_logprobs_dict
+        num_scheduled_tokens = scheduler_output.num_scheduled_tokens
+
+        new_running: list[Request] = []
+        outputs: list[EngineCoreOutput] = []
+        spec_decoding_stats: Optional[SpecDecodingStats] = None
+
+        # NOTE(woosuk): As len(self.running) can be up to 1K or more, the below
+        # loop can be a performance bottleneck. We should do our best to avoid
+        # expensive operations inside the loop.
+        for request in self.running:
+            req_id = request.request_id
+            num_tokens_scheduled = num_scheduled_tokens.get(req_id, 0)
+            if num_tokens_scheduled == 0:
+                # The request was not scheduled in this step.
+                new_running.append(request)
+                continue
+
+            req_index = model_runner_output.req_id_to_index[req_id]
+            generated_token_ids = sampled_token_ids[req_index]
+
+            scheduled_spec_token_ids = (
+                scheduler_output.scheduled_spec_decode_tokens.get(req_id))
+            if scheduled_spec_token_ids:
+                # num_computed_tokens represents the number of tokens
+                # processed in the current step, considering scheduled
+                # tokens and rejections. If some tokens are rejected,
+                # num_computed_tokens is decreased by the number of rejected
+                # tokens, where is given by:
+                # len(scheduled_spec_token_ids) + 1 - len(generated_token_ids).
+                num_tokens_rejected = (len(scheduled_spec_token_ids) + 1 -
+                                       len(generated_token_ids))
+                request.num_computed_tokens -= num_tokens_rejected
+                spec_decoding_stats = self.make_spec_decoding_stats(
+                    spec_decoding_stats,
+                    num_draft_tokens=len(scheduled_spec_token_ids),
+                    num_accepted_tokens=len(generated_token_ids) - 1)
+
+            cached_encoder_input_ids = (
+                self.encoder_cache_manager.get_cached_input_ids(request))
+            # OPTIMIZATION: Avoid list(set) if the set is empty.
+            if cached_encoder_input_ids:
+                for input_id in list(cached_encoder_input_ids):
+                    mm_positions = request.mm_positions[input_id]
+                    start_pos = mm_positions["offset"]
+                    num_tokens = mm_positions["length"]
+                    if start_pos + num_tokens <= request.num_computed_tokens:
+                        # The encoder output is already processed and stored
+                        # in the decoder's KV cache.
+                        self.encoder_cache_manager.free_encoder_input(
+                            request, input_id)
+
+            # Add newly generated spec token ids to the request.
+            if spec_token_ids is not None:
+                request.spec_token_ids = spec_token_ids[req_index]
+
+            stopped = False
+            new_logprobs = None
+            new_token_ids = generated_token_ids
+
+            # Append generated tokens and check for stop. Note that if
+            # a request is still being prefilled, we expect the model runner
+            # to return empty token ids for the request.
+            for num_new, output_token_id in enumerate(new_token_ids, 1):
+                request.append_output_token_ids(output_token_id)
+
+                # Check for stop and update request state.
+                # This must be called before we make the EngineCoreOutput.
+                stopped = check_stop(request, self.max_model_len)
+                if stopped:
+                    self._free_request(request)
+                    del new_token_ids[num_new:]  # Trim new tokens if needed.
+                    break
+
+            # Extract sample logprobs if needed.
+            if request.sampling_params.logprobs is not None and logprobs:
+                # NOTE: once we support N tokens per step (spec decode),
+                # the outer lists can be of length > 1.
+                new_logprobs = logprobs.slice(req_index, req_index + 1)
+
+            if new_token_ids and request.use_structured_output:
+                # NOTE: structured_output_request
+                # should not be None if use_structured_output, we have
+                # check above, so safe to ignore type warning
+                request.structured_output_request.grammar.accept_tokens(  # type: ignore[union-attr]
+                    req_id, new_token_ids)
+
+            # Get prompt logprobs for this request.
+            prompt_logprobs_tensors = prompt_logprobs_dict.get(req_id)
+            if new_token_ids:
+                # Add EngineCoreOutput for this Request.
+                outputs.append(
+                    EngineCoreOutput(
+                        request_id=req_id,
+                        new_token_ids=new_token_ids,
+                        finish_reason=request.get_finished_reason(),
+                        new_logprobs=new_logprobs,
+                        new_prompt_logprobs_tensors=prompt_logprobs_tensors,
+                        stop_reason=request.stop_reason,
+                        events=request.take_events()))
+            else:
+                # Invariant: EngineCore returns no partial prefill outputs.
+                assert not prompt_logprobs_tensors
+
+            self.scheduled_req_ids.remove(req_id)
+            if not stopped:
+                new_running.append(request)
+
+        self.running = new_running
+        engine_core_outputs = EngineCoreOutputs(
+            outputs=outputs,
+            scheduler_stats=self.make_stats(spec_decoding_stats),
+        )
+        if self.include_finished_set:
+            #TODO currently sending duplicates here, improve this
+            engine_core_outputs.finished_requests = (
+                scheduler_output.finished_req_ids | self.finished_req_ids)
+
+        return engine_core_outputs
+
+    def add_request(self, request: Request) -> None:
+        self.waiting.append(request)
+        self.requests[request.request_id] = request
+        if self.log_stats:
+            request.record_event(EngineCoreEventType.QUEUED)
+
+    def finish_requests(
+        self,
+        request_ids: Union[str, Iterable[str]],
+        finished_status: RequestStatus,
+    ) -> None:
+        """Handles the finish signal from outside the scheduler.
+
+        For example, the API server can abort a request when the client
+        disconnects.
+        """
+        assert RequestStatus.is_finished(finished_status)
+        if isinstance(request_ids, str):
+            request_ids = (request_ids, )
+        else:
+            request_ids = set(request_ids)
+
+        for req_id in request_ids:
+            request = self.requests.get(req_id)
+            if request is None:
+                # Invalid request ID.
+                continue
+
+            if request.status == RequestStatus.RUNNING:
+                self.running.remove(request)
+                self.scheduled_req_ids.discard(request.request_id)
+            else:
+                self.waiting.remove(request)
+            request.status = finished_status
+            self._free_request(request)
+
+    def _free_request(self, request: Request) -> None:
+        assert request.is_finished()
+        self.kv_cache_manager.free(request)
+        self.kv_cache_manager.free_block_hashes(request)
+        self.encoder_cache_manager.free(request)
+        self._cached_reqs_data.pop(request.request_id, None)
+        del self.requests[request.request_id]
+        self.finished_req_ids.add(request.request_id)
+
+    def get_num_unfinished_requests(self) -> int:
+        return len(self.waiting) + len(self.running)
+
+    def has_finished_requests(self) -> bool:
+        return len(self.finished_req_ids) > 0
+
+    def get_num_unscheduled_requests(self) -> int:
+        """Number of requests that are not being processed by the executor."""
+        return self.get_num_unfinished_requests() - len(self.scheduled_req_ids)
+
+    def reset_prefix_cache(self) -> bool:
+        return self.kv_cache_manager.reset_prefix_cache()
+
+    def make_stats(
+        self,
+        spec_decoding_stats: Optional[SpecDecodingStats] = None,
+    ) -> Optional[SchedulerStats]:
+        if not self.log_stats:
+            return None
+        return SchedulerStats(
+            num_running_reqs=len(self.running),
+            num_waiting_reqs=len(self.waiting),
+            gpu_cache_usage=self.kv_cache_manager.usage,
+            prefix_cache_stats=self.kv_cache_manager.make_prefix_cache_stats(),
+            spec_decoding_stats=spec_decoding_stats,
+        )
+
+    def make_spec_decoding_stats(
+        self,
+        spec_decoding_stats: Optional[SpecDecodingStats],
+        num_draft_tokens: int,
+        num_accepted_tokens: int,
+    ) -> Optional[SpecDecodingStats]:
+        if not self.log_stats:
+            return None
+        if spec_decoding_stats is None:
+            spec_decoding_stats = SpecDecodingStats()
+        spec_decoding_stats.observe(num_draft_tokens=num_draft_tokens,
+                                    num_accepted_tokens=num_accepted_tokens)
+        return spec_decoding_stats

vllm/v1/core/sched/utils.py

@@ -0,0 +1,22 @@
+# SPDX-License-Identifier: Apache-2.0
+from vllm.v1.request import Request, RequestStatus
+
+
+def check_stop(request: Request, max_model_len: int) -> bool:
+    if (request.num_tokens >= max_model_len
+            or request.num_output_tokens >= request.max_tokens):
+        request.status = RequestStatus.FINISHED_LENGTH_CAPPED
+        return True
+
+    sampling_params = request.sampling_params
+    last_token_id = request.output_token_ids[-1]
+    if (not sampling_params.ignore_eos
+            and last_token_id == request.eos_token_id):
+        request.status = RequestStatus.FINISHED_STOPPED
+        return True
+
+    if last_token_id in (sampling_params.stop_token_ids or ()):
+        request.status = RequestStatus.FINISHED_STOPPED
+        request.stop_reason = last_token_id
+        return True
+    return False