vllm-cpu 0.8.5.post2__cp310-cp310-manylinux_2_17_x86_64.whl

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

vllm/v1/core/specialized_manager.py

@@ -0,0 +1,161 @@
+# SPDX-License-Identifier: Apache-2.0
+from abc import ABC, abstractmethod
+
+from vllm.utils import cdiv
+from vllm.v1.core.block_pool import BlockPool
+from vllm.v1.core.kv_cache_utils import BlockHashType, KVCacheBlock
+from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheSpec,
+                                        SlidingWindowSpec)
+
+
+class SpecializedManager(ABC):
+    """
+    An abstract base class for specialized managers that handle the kv
+    cache management logic of different attention layers.
+    """
+
+    def __init__(
+        self,
+        kv_cache_spec: KVCacheSpec,
+        block_pool: BlockPool,
+    ) -> None:
+        """
+        Initializes the SpecializedManager.
+        Args:
+            kv_cache_spec: The kv_cache_spec for this manager.
+            block_pool: The block pool.
+        """
+
+        self.block_size = kv_cache_spec.block_size
+        self.kv_cache_spec = kv_cache_spec
+        self.block_pool = block_pool
+
+    @abstractmethod
+    def find_longest_cache_hit(
+            self, block_hashes: list[BlockHashType]) -> list[KVCacheBlock]:
+        """
+        Get the longest cache hit prefix of the blocks. If no cache hit is 
+        found, return an empty list.
+
+        Args:
+            block_hashes: The block hashes of the request.
+        Returns:
+            A list of cached blocks with skipped blocks replaced by null block.
+            For example, sliding window manager should return a list like
+            [NULL, NULL, KVCacheBlock(7), KVCacheBlock(8)] for block size 4 and 
+            sliding window 8. 
+        """
+
+        raise NotImplementedError
+
+    @abstractmethod
+    def remove_skipped_blocks(self, blocks: list[KVCacheBlock],
+                              num_computed_tokens: int) -> list[KVCacheBlock]:
+        """
+        Remove the blocks that are no longer needed from `blocks`. The removed 
+        blocks should be replaced by null_block. Return the removed blocks in 
+        eviction order, where the first returned block should be evicted first.
+        Don't free the removed blocks in this function.
+
+        Args:
+            blocks: The list of blocks to be updated.
+            num_computed_tokens: The number of tokens that have been computed.
+        Returns:
+            The removed blocks in eviction order.
+        """
+        raise NotImplementedError
+
+
+class FullAttentionManager(SpecializedManager):
+
+    def find_longest_cache_hit(
+            self, block_hashes: list[BlockHashType]) -> list[KVCacheBlock]:
+        computed_blocks: list[KVCacheBlock] = []
+        for block_hash in block_hashes:
+            # block_hashes is a chain of block hashes. If a block hash is not
+            # in the cached_block_hash_to_id, the following block hashes are
+            # not computed yet for sure.
+            if cached_block := self.block_pool.get_cached_block(block_hash):
+                computed_blocks.append(cached_block)
+            else:
+                break
+        return computed_blocks
+
+    def remove_skipped_blocks(self, blocks: list[KVCacheBlock],
+                              num_computed_tokens: int) -> list[KVCacheBlock]:
+        # No need to remove blocks for full attention.
+        return []
+
+
+class SlidingWindowManager(SpecializedManager):
+
+    def __init__(self, kv_cache_spec: SlidingWindowSpec,
+                 block_pool: BlockPool):
+        super().__init__(kv_cache_spec, block_pool)
+        self.sliding_window = kv_cache_spec.sliding_window
+        # The number of contiguous blocks needed for prefix cache hit.
+        # -1 since the input token itself is also included in the window
+        self.sliding_window_contiguous_blocks = cdiv(
+            (kv_cache_spec.sliding_window - 1), self.block_size)
+        self._null_block = block_pool.null_block
+
+    def find_longest_cache_hit(
+            self, block_hashes: list[BlockHashType]) -> list[KVCacheBlock]:
+        # TODO: reduce i by sliding_window_contiguous_blocks when cache miss, to
+        # optimize the time complexity from O(len(block_hashes)) to
+        # O(len(block_hashes) / sliding_window_contiguous_blocks +
+        # sliding_window_contiguous_blocks),
+        # which is good for low cache hit rate scenarios.
+        computed_blocks = [self._null_block] * len(block_hashes)
+        num_contiguous_blocks = 0
+
+        # Search from right to left and early stop when a match is found.
+        for i in range(len(block_hashes) - 1, -1, -1):
+            if cached_block := self.block_pool.get_cached_block(
+                    block_hashes[i]):
+                computed_blocks[i] = cached_block
+                num_contiguous_blocks += 1
+                if (num_contiguous_blocks
+                        >= self.sliding_window_contiguous_blocks):
+                    # Trim the trailing blocks.
+                    # E.g., [NULL, NULL, 8, 3, NULL, 9] -> [NULL, NULL, 8, 3]
+                    # when sliding_window_contiguous_blocks=2.
+                    del computed_blocks[i + num_contiguous_blocks:]
+                    return computed_blocks
+            else:
+                num_contiguous_blocks = 0
+        # The first `num_contiguous_blocks` is a cache hit even if
+        # `num_contiguous_blocks < sliding_window_contiguous_blocks`.
+        del computed_blocks[num_contiguous_blocks:]
+        return computed_blocks
+
+    def remove_skipped_blocks(self, blocks: list[KVCacheBlock],
+                              num_computed_tokens: int) -> list[KVCacheBlock]:
+        # Remove the blocks that are no longer be in the sliding window and
+        # skipped during the attention computation.
+        last_useful_token = num_computed_tokens - self.sliding_window + 1
+        last_useful_block = last_useful_token // self.block_size
+
+        removed_blocks: list[KVCacheBlock] = []
+        for i in range(last_useful_block - 1, -1, -1):
+            if blocks[i] == self._null_block:
+                # If the block is already a null block, the blocks before it
+                # should also have been set to null blocks by the previous calls
+                # to this function.
+                break
+            removed_blocks.append(blocks[i])
+            blocks[i] = self._null_block
+        return removed_blocks
+
+
+spec_manager_map: dict[type[KVCacheSpec], type[SpecializedManager]] = {
+    FullAttentionSpec: FullAttentionManager,
+    SlidingWindowSpec: SlidingWindowManager,
+}
+
+
+def get_specialized_manager(kv_cache_spec: KVCacheSpec,
+                            block_pool: BlockPool) -> SpecializedManager:
+    manager_class = spec_manager_map[type(kv_cache_spec)]
+    manager = manager_class(kv_cache_spec, block_pool)
+    return manager

vllm/v1/engine/__init__.py

@@ -0,0 +1,166 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import enum
+import time
+from collections.abc import Sequence
+from typing import Any, Optional, Union
+
+import msgspec
+
+from vllm.lora.request import LoRARequest
+from vllm.multimodal import MultiModalKwargs
+from vllm.multimodal.inputs import PlaceholderRange
+from vllm.sampling_params import SamplingParams
+from vllm.v1.metrics.stats import SchedulerStats
+from vllm.v1.outputs import LogprobsLists, LogprobsTensors
+
+# These are possible values of RequestOutput.finish_reason,
+# so form part of the external API.
+FINISH_REASON_STRINGS = ("stop", "length", "abort")
+
+
+class FinishReason(enum.IntEnum):
+    """
+    Reason a request finished - stop, length, or abort.
+
+    Int rather than Str for more compact serialization.
+
+    stop - a stop string was emitted
+    length - max_tokens was consumed, or max_model_len was reached
+    abort - aborted for another reason
+
+    """
+    STOP = 0
+    LENGTH = 1
+    ABORT = 2
+
+    def __str__(self):
+        return FINISH_REASON_STRINGS[self.value]
+
+
+class EngineCoreRequest(
+        msgspec.Struct,
+        array_like=True,  # type: ignore[call-arg]
+        omit_defaults=True,  # type: ignore[call-arg]
+        gc=False):  # type: ignore[call-arg]
+
+    # NOTE: prompt and prompt_token_ids should be DecoderOnlyInput,
+    # but this object is currently not playing well with msgspec
+    # due to circular imports and typing we have in data.py
+
+    request_id: str
+    prompt_token_ids: list[int]
+    mm_inputs: Optional[Sequence[Optional[MultiModalKwargs]]]
+    mm_hashes: Optional[list[str]]
+    mm_placeholders: Optional[list[PlaceholderRange]]
+    sampling_params: SamplingParams
+    eos_token_id: Optional[int]
+    arrival_time: float
+    lora_request: Optional[LoRARequest]
+
+    # Used in DP case to indicate which wave of requests this is expected to
+    # belong to, to cover a race condition where the request is sent before
+    # a wave finished notification is received.
+    current_wave: int = 0
+
+
+class EngineCoreEventType(enum.IntEnum):
+    """The type of engine core request event."""
+    QUEUED = 1
+    SCHEDULED = 2
+    PREEMPTED = 3
+
+
+class EngineCoreEvent(msgspec.Struct):
+    """A timestamped engine core event associated with a request.
+
+    The timestamp is a monotonic timestamps and is used for by the engine
+    frontend to calculate intervals between engine core events. These
+    timestamps should not be compared with timestamps from other processes.
+    """
+    type: EngineCoreEventType
+    timestamp: float
+
+    @classmethod
+    def new_event(cls,
+                  event_type: EngineCoreEventType,
+                  timestamp: Optional[float] = None) -> "EngineCoreEvent":
+        timestamp = time.monotonic() if timestamp is None else timestamp
+        return cls(event_type, timestamp)
+
+
+class EngineCoreOutput(
+        msgspec.Struct,
+        array_like=True,  # type: ignore[call-arg]
+        omit_defaults=True,  # type: ignore[call-arg]
+        gc=False):  # type: ignore[call-arg]
+
+    request_id: str
+    new_token_ids: list[int]
+
+    new_logprobs: Optional[LogprobsLists] = None
+    new_prompt_logprobs_tensors: Optional[LogprobsTensors] = None
+
+    finish_reason: Optional[FinishReason] = None
+    stop_reason: Union[int, str, None] = None
+    events: Optional[list[EngineCoreEvent]] = None
+
+    @property
+    def finished(self) -> bool:
+        return self.finish_reason is not None
+
+
+class UtilityOutput(
+        msgspec.Struct,
+        array_like=True,  # type: ignore[call-arg]
+        gc=False):  # type: ignore[call-arg]
+
+    call_id: int
+
+    # Non-None implies the call failed, result should be None.
+    failure_message: Optional[str] = None
+    result: Any = None
+
+
+class EngineCoreOutputs(
+        msgspec.Struct,
+        array_like=True,  # type: ignore[call-arg]
+        omit_defaults=True,  # type: ignore[call-arg]
+        gc=False):  # type: ignore[call-arg]
+
+    #NOTE(Nick): We could consider ways to make this more compact,
+    # e.g. columnwise layout
+
+    engine_index: int = 0
+
+    # [num_reqs]
+    outputs: list[EngineCoreOutput] = []
+    scheduler_stats: Optional[SchedulerStats] = None
+    timestamp: float = 0.0
+
+    utility_output: Optional[UtilityOutput] = None
+    finished_requests: Optional[set[str]] = None
+
+    # In DP case, used to signal that the current wave of requests
+    # has finished and the engines are paused.
+    wave_complete: Optional[int] = None
+    # In DP case, used to signal that a request was received for an
+    # "old" wave, so the next wave needs to be started in other engines.
+    start_wave: Optional[int] = None
+
+    def __post_init__(self):
+        if self.timestamp == 0.0:
+            self.timestamp = time.monotonic()
+
+
+class EngineCoreRequestType(enum.Enum):
+    """
+    Request types defined as hex byte strings, so it can be sent over sockets
+    without separate encoding step.
+    """
+    ADD = b'\x00'
+    ABORT = b'\x01'
+    START_DP_WAVE = b'\x02'
+    UTILITY = b'\x03'
+    # Sentinel used within EngineCoreProc.
+    EXECUTOR_FAILED = b'\x04'