| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import Iterable, Iterator |
| 4 | from itertools import chain |
| 5 | from typing import TYPE_CHECKING |
| 6 | |
| 7 | from vllm.v1.sample.logits_processor.interface import ( |
| 8 | AddedRequest, |
| 9 | BatchUpdate, |
| 10 | MovedRequest, |
| 11 | RemovedRequest, |
| 12 | ) |
| 13 | |
| 14 | if TYPE_CHECKING: |
| 15 | from vllm.v1.sample.logits_processor.interface import LogitsProcessor |
| 16 | |
| 17 | |
| 18 | class BatchUpdateBuilder: |
| 19 | """Helps track persistent batch state changes and build |
| 20 | a batch update data structure for logitsprocs |
| 21 | Assumptions: |
| 22 | * All information about requests removed from persistent batch |
| 23 | during a step is aggregated in self._removed through calls to |
| 24 | self.removed_append() at the beginning of a step. This must happen |
| 25 | before the first time that self.removed, self.pop_removed() |
| 26 | or self.peek_removed() are invoked in a given step |
| 27 | * After the first time that self.removed, self.pop_removed() |
| 28 | or self.peek_removed() are read in a step, no new removals |
| 29 | are registered using self.removed_append() |
| 30 | * Elements of self._removed are never directly modified, added or |
| 31 | removed (i.e. modification is only via self.removed_append() and |
| 32 | self.pop_removed()) |
| 33 | Guarantees under above assumptions: |
| 34 | * self.removed is always sorted in descending order |
| 35 | * self.pop_removed() and self.peek_removed() both return |
| 36 | the lowest removed request index in the current step |
| 37 | """ |
| 38 | |
| 39 | _removed: list[RemovedRequest] |
| 40 | _is_removed_sorted: bool |
| 41 | added: list[AddedRequest] |
| 42 | moved: list[MovedRequest] |
| 43 | |
| 44 | def __init__( |
| 45 | self, |
| 46 | removed: list[RemovedRequest] | None = None, |
| 47 | added: list[AddedRequest] | None = None, |
| 48 | moved: list[MovedRequest] | None = None, |
| 49 | ) -> None: |
| 50 | self._removed = removed or [] |
| 51 | self.added = added or [] |
| 52 | self.moved = moved or [] |
| 53 | self._is_removed_sorted = False |
| 54 | |
| 55 | # Used to track changes in the pooling case |
| 56 | # where we don't populate the added list. |
| 57 | self.batch_changed = False |
| 58 | |
| 59 | def _ensure_removed_sorted(self) -> None: |
| 60 | """Sort removed request indices in |
| 61 | descending order. |
| 62 | Idempotent after first call in a |
| 63 | given step, until reset. |
| 64 | """ |
| 65 | if not self._is_removed_sorted: |
| 66 | self._removed.sort(reverse=True) |
| 67 | self._is_removed_sorted = True |
| 68 | |
| 69 | @property |
| 70 | def removed(self) -> list[RemovedRequest]: |
| 71 | """Removed request indices sorted in |
| 72 | descending order""" |
| 73 | self._ensure_removed_sorted() |
| 74 | return self._removed |
| 75 | |
| 76 | def removed_append(self, index: int) -> None: |
| 77 | """Register the removal of a request from the persistent batch. |
| 78 | |
| 79 | Must not be called after the first time self.removed, |
| 80 | self.pop_removed() or self.peek_removed() are invoked. |
| 81 | |
| 82 | Args: |
| 83 | index: request index |
| 84 | |
| 85 | """ |
| 86 | if self._is_removed_sorted: |
| 87 | raise RuntimeError( |
| 88 | "Cannot register new removed request after self.removed has been read." |
| 89 | ) |
| 90 | self._removed.append(index) |
| 91 | self.batch_changed = True |
| 92 | |
| 93 | def has_removed(self) -> bool: |
| 94 | return bool(self._removed) |
| 95 | |
| 96 | def peek_removed(self) -> int | None: |
| 97 | """Return lowest removed request index.""" |
| 98 | if self.has_removed(): |
| 99 | self._ensure_removed_sorted() |
| 100 | return self._removed[-1] |
| 101 | return None |
| 102 | |
| 103 | def pop_removed(self) -> int | None: |
| 104 | """Pop lowest removed request index.""" |
| 105 | if self.has_removed(): |
| 106 | self._ensure_removed_sorted() |
| 107 | return self._removed.pop() |
| 108 | return None |
| 109 | |
| 110 | def reset(self) -> bool: |
| 111 | """Returns True if there were any changes to the batch.""" |
| 112 | self._is_removed_sorted = False |
| 113 | self._removed.clear() |
| 114 | self.added.clear() |
| 115 | self.moved.clear() |
| 116 | batch_changed = self.batch_changed |
| 117 | self.batch_changed = False |
| 118 | return batch_changed |
| 119 | |
| 120 | def get_and_reset(self, batch_size: int) -> BatchUpdate | None: |
| 121 | """Generate a logitsprocs batch update data structure and reset |
| 122 | internal batch update builder state. |
| 123 | |
| 124 | Args: |
| 125 | batch_size: current persistent batch size |
| 126 | |
| 127 | Returns: |
| 128 | Frozen logitsprocs batch update instance; `None` if no updates |
| 129 | |
| 130 | """ |
| 131 | # Reset removal-sorting logic |
| 132 | self._is_removed_sorted = False |
| 133 | self.batch_changed = False |
| 134 | if not any((self._removed, self.moved, self.added)): |
| 135 | # No update; short-circuit |
| 136 | return None |
| 137 | # Build batch state update |
| 138 | batch_update = BatchUpdate( |
| 139 | batch_size=batch_size, |
| 140 | removed=self._removed, |
| 141 | moved=self.moved, |
| 142 | added=self.added, |
| 143 | ) |
| 144 | self._removed = [] |
| 145 | self.moved = [] |
| 146 | self.added = [] |
| 147 | return batch_update |
| 148 | |
| 149 | |
| 150 | class LogitsProcessors: |
| 151 | """Encapsulates initialized logitsproc objects.""" |
| 152 | |
| 153 | def __init__(self, logitsprocs: Iterable["LogitsProcessor"] | None = None) -> None: |
| 154 | self.argmax_invariant: list[LogitsProcessor] = [] |
| 155 | self.non_argmax_invariant: list[LogitsProcessor] = [] |
| 156 | if logitsprocs: |
| 157 | for logitproc in logitsprocs: |
| 158 | ( |
| 159 | self.argmax_invariant |
| 160 | if logitproc.is_argmax_invariant() |
| 161 | else self.non_argmax_invariant |
| 162 | ).append(logitproc) |
| 163 | |
| 164 | @property |
| 165 | def all(self) -> Iterator["LogitsProcessor"]: |
| 166 | """Iterator over all logits processors.""" |
| 167 | return chain(self.argmax_invariant, self.non_argmax_invariant) |