| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import dataclasses |
| 4 | from collections.abc import Mapping, Set |
| 5 | from itertools import groupby |
| 6 | |
| 7 | import torch |
| 8 | |
| 9 | from vllm.config import PoolerConfig |
| 10 | from vllm.model_executor.layers.pooler import PoolingParamsUpdate |
| 11 | from vllm.tasks import PoolingTask |
| 12 | from vllm.v1.pool.metadata import PoolingMetadata |
| 13 | |
| 14 | from .abstract import Pooler, PoolerOutput |
| 15 | from .common import ClassifierFn |
| 16 | from .seqwise import ( |
| 17 | SequencePoolingFn, |
| 18 | SequencePoolingMethod, |
| 19 | pooler_for_classify, |
| 20 | pooler_for_embed, |
| 21 | ) |
| 22 | from .tokwise import pooler_for_token_classify, pooler_for_token_embed |
| 23 | |
| 24 | |
| 25 | class DispatchPooler(Pooler): |
| 26 | """Dispatches calls to a sub-pooler based on the pooling task.""" |
| 27 | |
| 28 | @classmethod |
| 29 | def for_embedding(cls, pooler_config: PoolerConfig): |
| 30 | return cls( |
| 31 | { |
| 32 | "token_embed": pooler_for_token_embed(pooler_config), |
| 33 | "embed": pooler_for_embed(pooler_config), |
| 34 | }, |
| 35 | ) |
| 36 | |
| 37 | @classmethod |
| 38 | def for_seq_cls( |
| 39 | cls, |
| 40 | pooler_config: PoolerConfig, |
| 41 | *, |
| 42 | pooling: SequencePoolingMethod | SequencePoolingFn | None = None, |
| 43 | classifier: ClassifierFn | None = None, |
| 44 | ): |
| 45 | return cls( |
| 46 | { |
| 47 | "token_classify": pooler_for_token_classify( |
| 48 | pooler_config, |
| 49 | classifier=classifier, |
| 50 | ), |
| 51 | "classify": pooler_for_classify( |
| 52 | pooler_config, |
| 53 | pooling=pooling, |
| 54 | classifier=classifier, |
| 55 | ), |
| 56 | } |
| 57 | ) |
| 58 | |
| 59 | def __init__(self, poolers_by_task: Mapping[PoolingTask, Pooler]) -> None: |
| 60 | super().__init__() |
| 61 | |
| 62 | for task, pooler in poolers_by_task.items(): |
| 63 | if task not in pooler.get_supported_tasks(): |
| 64 | raise ValueError( |
| 65 | f"{pooler=} does not support {task=}. " |
| 66 | f"Supported tasks: {pooler.get_supported_tasks()}" |
| 67 | ) |
| 68 | |
| 69 | self.poolers_by_task = poolers_by_task |
| 70 | |
| 71 | def get_supported_tasks(self) -> Set[PoolingTask]: |
| 72 | return set(self.poolers_by_task) |
| 73 | |
| 74 | def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate: |
| 75 | return self.poolers_by_task[task].get_pooling_updates(task) |
| 76 | |
| 77 | def replace_classifier( |
| 78 | self, |
| 79 | new_classifier: ClassifierFn, |
| 80 | ) -> None: |
| 81 | """Replaces the classifier to the LoRA-wrapped version.""" |
| 82 | for task, pooler in self.poolers_by_task.items(): |
| 83 | if task not in {"classify"}: |
| 84 | # Now LoRA only supports classification tasks, |
| 85 | # so we only replace the classifier for "classify" task. |
| 86 | continue |
| 87 | head = getattr(pooler, "head", None) |
| 88 | if head is not None and hasattr(head, "classifier"): |
| 89 | head.classifier = new_classifier |
| 90 | |
| 91 | def forward( |
| 92 | self, |
| 93 | hidden_states: torch.Tensor, |
| 94 | pooling_metadata: PoolingMetadata, |
| 95 | ) -> PoolerOutput: |
| 96 | poolers_by_task = self.poolers_by_task |
| 97 | cursor = pooling_metadata.pooling_cursor |
| 98 | |
| 99 | outputs = list[torch.Tensor | None]() |
| 100 | offset = 0 |
| 101 | token_offset = 0 |
| 102 | for task, group in groupby(pooling_metadata.tasks): |
| 103 | if not (pooler := poolers_by_task.get(task)): |
| 104 | raise ValueError( |
| 105 | f"Unsupported task: {task!r} " |
| 106 | f"Supported tasks: {self.get_supported_tasks()}" |
| 107 | ) |
| 108 | |
| 109 | num_items = len(list(group)) |
| 110 | group_metadata = pooling_metadata[offset : offset + num_items] |
| 111 | if cursor is None: |
| 112 | group_hidden_states = hidden_states |
| 113 | else: |
| 114 | # Slice out this group's tokens so sub-poolers see only their |
| 115 | # portion of the batch. Token offset is computed from the CPU |
| 116 | # `num_scheduled_tokens_cpu` to avoid a GPU->CPU sync. |
| 117 | group_cursor = group_metadata.pooling_cursor |
| 118 | assert group_cursor is not None |
| 119 | num_group_tokens = int(group_cursor.num_scheduled_tokens_cpu.sum()) |
| 120 | group_hidden_states = hidden_states[ |
| 121 | token_offset : token_offset + num_group_tokens |
| 122 | ] |
| 123 | if token_offset: |
| 124 | # Shift first/last indices to be relative to the slice |
| 125 | # so seqwise poolers (which index `hidden_states` directly) |
| 126 | # remain correct. |
| 127 | pooling_cursor = dataclasses.replace( |
| 128 | group_cursor, |
| 129 | first_token_indices_gpu=( |
| 130 | group_cursor.first_token_indices_gpu - token_offset |
| 131 | ), |
| 132 | last_token_indices_gpu=( |
| 133 | group_cursor.last_token_indices_gpu - token_offset |
| 134 | ), |
| 135 | ) |
| 136 | group_metadata = dataclasses.replace( |
| 137 | group_metadata, pooling_cursor=pooling_cursor |
| 138 | ) |
| 139 | token_offset += num_group_tokens |
| 140 | |
| 141 | group_output: PoolerOutput = pooler(group_hidden_states, group_metadata) |
| 142 | |
| 143 | outputs.extend(group_output) |
| 144 | offset += num_items |
| 145 | |
| 146 | return outputs |
| 147 | |
| 148 | def extra_repr(self) -> str: |
| 149 | s = f"supported_task={self.get_supported_tasks()}" |
| 150 | return s |
| 151 | |
| 152 | |
| 153 | class IdentityPooler(Pooler): |
| 154 | def get_supported_tasks(self) -> Set[PoolingTask]: |
| 155 | return {"plugin"} |
| 156 | |
| 157 | def forward( |
| 158 | self, |
| 159 | hidden_states: torch.Tensor, |
| 160 | pooling_metadata: PoolingMetadata, |
| 161 | ) -> PoolerOutput: |
| 162 | return hidden_states |
| 163 | |
| 164 | |
| 165 | class BOSEOSFilter(Pooler): |
| 166 | """Filters the BOS and EOS token results from outputs.""" |
| 167 | |
| 168 | def __init__( |
| 169 | self, |
| 170 | pooler: Pooler, |
| 171 | bos_token_id: int = -1, # -1 disables the filtering |
| 172 | eos_token_id: int = -1, |
| 173 | ) -> None: |
| 174 | super().__init__() |
| 175 | |
| 176 | self.pooler = pooler |
| 177 | self.bos_token_id = bos_token_id |
| 178 | self.eos_token_id = eos_token_id |
| 179 | |
| 180 | def extra_repr(self) -> str: |
| 181 | return f"bos_token_id={self.bos_token_id}, eos_token_id={self.eos_token_id}" |
| 182 | |
| 183 | def get_supported_tasks(self) -> Set[PoolingTask]: |
| 184 | return self.pooler.get_supported_tasks() |
| 185 | |
| 186 | def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate: |
| 187 | return PoolingParamsUpdate(requires_token_ids=True) |
| 188 | |
| 189 | def forward( |
| 190 | self, |
| 191 | hidden_states: torch.Tensor | list[torch.Tensor], |
| 192 | pooling_metadata: PoolingMetadata, |
| 193 | ) -> PoolerOutput: |
| 194 | pooled_outputs = self.pooler(hidden_states, pooling_metadata) |
| 195 | assert isinstance(pooled_outputs, list) |
| 196 | prompt_token_ids = pooling_metadata.get_prompt_token_ids_cpu() |
| 197 | |
| 198 | for i, (prompt_len, token_ids) in enumerate( |
| 199 | zip(pooling_metadata.prompt_lens, prompt_token_ids) |
| 200 | ): |
| 201 | pooled_data = pooled_outputs[i] |
| 202 | assert ( |
| 203 | isinstance(pooled_data, torch.Tensor) |
| 204 | and pooled_data.shape[0] == prompt_len |
| 205 | ) |
| 206 | if int(token_ids[0]) == self.bos_token_id: |
| 207 | pooled_data = pooled_data[1:] |
| 208 | if int(token_ids[-1]) == self.eos_token_id: |
| 209 | pooled_data = pooled_data[:-1] |
| 210 | pooled_outputs[i] = pooled_data.squeeze(-1) |
| 211 | |
| 212 | return pooled_outputs |
| 213 | |
| 214 | |
| 215 | class BgeM3Pooler(Pooler): |
| 216 | def __init__(self, token_classify_pooler: Pooler, embed_pooler: Pooler) -> None: |
| 217 | super().__init__() |
| 218 | self.token_classify_pooler = token_classify_pooler |
| 219 | self.embed_pooler = embed_pooler |
| 220 | |
| 221 | def forward( |
| 222 | self, hidden_states: torch.Tensor, pooling_metadata: PoolingMetadata |
| 223 | ) -> PoolerOutput: |
| 224 | embed_outputs = self.embed_pooler(hidden_states, pooling_metadata) |
| 225 | token_classify_outputs = self.token_classify_pooler( |
| 226 | hidden_states, pooling_metadata |
| 227 | ) |
| 228 | pooler_outputs: list[torch.Tensor] = [] |
| 229 | for embed_output, token_classify_output in zip( |
| 230 | embed_outputs, token_classify_outputs |
| 231 | ): |
| 232 | pooler_outputs.append( |
| 233 | torch.cat( |
| 234 | [embed_output.view(-1), token_classify_output.view(-1)], dim=-1 |
| 235 | ) |
| 236 | ) |
| 237 | |
| 238 | return pooler_outputs |
| 239 | |
| 240 | def get_supported_tasks(self) -> Set[PoolingTask]: |
| 241 | return {"embed&token_classify"} |
| 242 | |
| 243 | def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate: |
| 244 | return self.embed_pooler.get_pooling_updates( |
| 245 | "embed" |
| 246 | ) | self.token_classify_pooler.get_pooling_updates("token_classify") |
| 247 | |
| 248 | def extra_repr(self) -> str: |
| 249 | s = f"supported_task={self.get_supported_tasks()}" |
| 250 | return s |
| 251 | |
| 252 | |
| 253 | __all__ = ["BOSEOSFilter", "DispatchPooler", "IdentityPooler", "BgeM3Pooler"] |