|
| 1 | +import enum |
| 2 | +from dataclasses import dataclass |
| 3 | +from typing import Dict, List, Set |
| 4 | + |
| 5 | + |
| 6 | +class RequsetStatus(enum.Enum): |
| 7 | + """The status of Sentences""" |
| 8 | + |
| 9 | + WAITING = enum.auto() |
| 10 | + RUNNING = enum.auto() |
| 11 | + ABORTED = enum.auto() |
| 12 | + OVERLENGTH = enum.auto() |
| 13 | + COMPLETED = enum.auto() |
| 14 | + LENGTH_CAPPED = enum.auto() |
| 15 | + |
| 16 | + @staticmethod |
| 17 | + def is_finished(status: "RequsetStatus") -> bool: |
| 18 | + return status in [ |
| 19 | + RequsetStatus.OVERLENGTH, |
| 20 | + RequsetStatus.COMPLETED, |
| 21 | + RequsetStatus.LENGTH_CAPPED, |
| 22 | + ] |
| 23 | + |
| 24 | + @staticmethod |
| 25 | + def is_running(status: "RequsetStatus") -> bool: |
| 26 | + return status == RequsetStatus.RUNNING |
| 27 | + |
| 28 | + @staticmethod |
| 29 | + def is_waiting(status: "RequsetStatus") -> bool: |
| 30 | + return status == RequsetStatus.WAITING |
| 31 | + |
| 32 | + |
| 33 | +class Sequence: |
| 34 | + """Store information of input sequence. |
| 35 | +
|
| 36 | + Args: |
| 37 | + request_id: The ID of input sequence. |
| 38 | + prompt: The prompt of input sequence. |
| 39 | + token_id: The tokens ID of input sequence. |
| 40 | + block_size: The block size of input sequence. |
| 41 | + sample_params: The sample_params of input sequence. |
| 42 | + block_table_index: The index of input sequence in block_table. |
| 43 | + """ |
| 44 | + |
| 45 | + def __init__( |
| 46 | + self, |
| 47 | + request_id: int, |
| 48 | + prompt: str, |
| 49 | + token_id: List[int], |
| 50 | + block_size: int, |
| 51 | + sample_params, # SampleParams needs to be imported later. |
| 52 | + block_table_index: int, |
| 53 | + ): |
| 54 | + self.request_id = request_id |
| 55 | + self.prompt = prompt |
| 56 | + self.input_token_id = token_id |
| 57 | + self.blokc_size = block_size |
| 58 | + self.sample_params = sample_params |
| 59 | + self.output_token_id = [] |
| 60 | + self.status = RequsetStatus.WAITING |
| 61 | + self.block_table_index = block_table_index |
| 62 | + |
| 63 | + def get_sentence_len(self) -> None: |
| 64 | + """ |
| 65 | + Get length of current sentence. |
| 66 | + """ |
| 67 | + return len(self.input_token_id) + len(self.output_token_id) |
| 68 | + |
| 69 | + def get_input_len(self) -> None: |
| 70 | + """ |
| 71 | + Get length of input sentence. |
| 72 | + """ |
| 73 | + return len(self.input_token_id) |
| 74 | + |
| 75 | + def get_output_len(self) -> None: |
| 76 | + """ |
| 77 | + Get output length of current sentence. |
| 78 | + """ |
| 79 | + return len(self.output_token_id) |
| 80 | + |
| 81 | + def check_finish(self) -> bool: |
| 82 | + """ |
| 83 | + Check whether inference is over. |
| 84 | + """ |
| 85 | + return RequsetStatus.is_finished(self.status) |
| 86 | + |
| 87 | + def __repr__(self) -> str: |
| 88 | + return ( |
| 89 | + f"Request ID(request_id={self.request_id}, " |
| 90 | + f"prompt={self.prompt}, " |
| 91 | + f"status={self.status.name}, " |
| 92 | + f"sample_params={self.sample_params}, " |
| 93 | + f"logical block number={len(self._logical_blocks)}" |
| 94 | + ) |
| 95 | + |
| 96 | + |
| 97 | +@dataclass |
| 98 | +class BatchHandler: |
| 99 | + """ |
| 100 | + Information to be passed and used for a batch of sequences. |
| 101 | + """ |
| 102 | + |
| 103 | + sequences_set: Set[Sequence] |
| 104 | + block_table: Dict[int, int] |
| 105 | + |
| 106 | + @classmethod |
| 107 | + def init_batch(cls, seqs: List[Sequence]) -> "BatchHandler": |
| 108 | + """ |
| 109 | + Initializes inference batches by input sentence list. |
| 110 | +
|
| 111 | + Args: |
| 112 | + seqs (List[Sequence]): List of input sequence. |
| 113 | + """ |
| 114 | + sequences_set = set() |
| 115 | + block_table = {} |
| 116 | + for seq in seqs: |
| 117 | + if seq in sequences_set: |
| 118 | + print("The sequence is already in sequences_set.") |
| 119 | + assert ( |
| 120 | + seq.request_id in block_table |
| 121 | + ), "The sequence has been added to sequences_set, but it has not been added to block_table." |
| 122 | + continue |
| 123 | + assert ( |
| 124 | + seq.request_id not in block_table |
| 125 | + ), "The sequence has not been added to sequences_set, but it is already in block_table." |
| 126 | + |
| 127 | + sequences_set.add(seq) |
| 128 | + block_table[seq.request_id] = seq.block_table_index |
| 129 | + |
| 130 | + return cls(sequences_set=sequences_set, block_table=block_table) |
| 131 | + |
| 132 | + def clear_batch(self) -> None: |
| 133 | + """ |
| 134 | + Clear sequence set and block table. |
| 135 | + """ |
| 136 | + for seq in self.sequences_set: |
| 137 | + if not seq.check_finish(): |
| 138 | + seq.status = RequsetStatus.ABORTED |
| 139 | + self.sequences_set.clear() |
| 140 | + self.block_table.clear() |
| 141 | + |
| 142 | + def fliter_batch(self) -> None: |
| 143 | + """ |
| 144 | + Remove completed sentences from a batch. |
| 145 | + """ |
| 146 | + for seq in self.sequences_set: |
| 147 | + if seq.check_finish(): |
| 148 | + self.sequences_set.reomve(seq) |
| 149 | + del self.block_table[seq.request_id] |
| 150 | + |
| 151 | + def add_seqs(self, seqs: List[Sequence]) -> None: |
| 152 | + """ |
| 153 | + Add new sequence to batch |
| 154 | +
|
| 155 | + Args: |
| 156 | + seqs (List[Sequence]): The list of new sequences. |
| 157 | + """ |
| 158 | + for seq in seqs: |
| 159 | + if seq in self.sequences_set: |
| 160 | + print("The sequence is already in sequences_set.") |
| 161 | + assert ( |
| 162 | + seq.request_id in self.block_table |
| 163 | + ), "The sequence has been added to sequences_set, but it has not been added to block_table." |
| 164 | + continue |
| 165 | + assert ( |
| 166 | + seq.request_id not in self.block_table |
| 167 | + ), "The sequence has not been added to sequences_set, but it is already in block_table." |
| 168 | + self.sequences_set.add(seq) |
| 169 | + self.block_table[seq.request_id] = seq.block_table_index |
0 commit comments