| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import Callable, Set |
| 4 | from typing import TypeAlias |
| 5 | |
| 6 | import torch |
| 7 | |
| 8 | from vllm.config import PoolerConfig, get_current_vllm_config |
| 9 | from vllm.model_executor.layers.pooler import ( |
| 10 | ClassifierFn, |
| 11 | PoolingParamsUpdate, |
| 12 | ProjectorFn, |
| 13 | ) |
| 14 | from vllm.model_executor.layers.pooler.abstract import Pooler |
| 15 | from vllm.model_executor.layers.pooler.activations import ( |
| 16 | PoolerActivation, |
| 17 | PoolerNormalize, |
| 18 | resolve_classifier_act_fn, |
| 19 | ) |
| 20 | from vllm.model_executor.models.adapters import _load_st_projector |
| 21 | from vllm.tasks import POOLING_TASKS, PoolingTask |
| 22 | from vllm.v1.pool.metadata import PoolingMetadata |
| 23 | |
| 24 | from .heads import ( |
| 25 | TokenClassifierPoolerHead, |
| 26 | TokenEmbeddingPoolerHead, |
| 27 | TokenPoolerHead, |
| 28 | TokenPoolerHeadOutputItem, |
| 29 | ) |
| 30 | from .methods import ( |
| 31 | TokenPoolingMethod, |
| 32 | TokenPoolingMethodOutputItem, |
| 33 | get_tok_pooling_method, |
| 34 | ) |
| 35 | |
| 36 | TokenPoolingFn: TypeAlias = Callable[ |
| 37 | [torch.Tensor, PoolingMetadata], |
| 38 | list[TokenPoolingMethodOutputItem], |
| 39 | ] |
| 40 | TokenPoolingHeadFn: TypeAlias = Callable[ |
| 41 | [list[TokenPoolingMethodOutputItem], PoolingMetadata], |
| 42 | list[TokenPoolerHeadOutputItem], |
| 43 | ] |
| 44 | |
| 45 | TokenPoolerOutput: TypeAlias = list[torch.Tensor | None] |
| 46 | |
| 47 | |
| 48 | class TokenPooler(Pooler): |
| 49 | """A layer that pools specific information from hidden states. |
| 50 | |
| 51 | This layer does the following: |
| 52 | 1. Extracts specific tokens or aggregates data based on pooling method. |
| 53 | 2. Postprocesses the output based on pooling head. |
| 54 | 3. Returns structured results as `PoolerOutput`. |
| 55 | """ |
| 56 | |
| 57 | def __init__( |
| 58 | self, |
| 59 | pooling: TokenPoolingMethod | TokenPoolingFn, |
| 60 | head: TokenPoolerHead | TokenPoolingHeadFn | None = None, |
| 61 | ) -> None: |
| 62 | super().__init__() |
| 63 | |
| 64 | self.pooling = pooling |
| 65 | self.head = head |
| 66 | |
| 67 | def extra_repr(self) -> str: |
| 68 | head_name = self.head.__class__.__name__ if self.head is not None else None |
| 69 | return f"pooling={self.pooling.__class__.__name__}, head={head_name}" |
| 70 | |
| 71 | def get_supported_tasks(self) -> Set[PoolingTask]: |
| 72 | tasks = set(POOLING_TASKS) |
| 73 | |
| 74 | if isinstance(self.pooling, TokenPoolingMethod): |
| 75 | tasks &= self.pooling.get_supported_tasks() |
| 76 | if isinstance(self.head, TokenPoolerHead): |
| 77 | tasks &= self.head.get_supported_tasks() |
| 78 | |
| 79 | return tasks |
| 80 | |
| 81 | def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate: |
| 82 | updates = PoolingParamsUpdate() |
| 83 | |
| 84 | if isinstance(self.pooling, TokenPoolingMethod): |
| 85 | updates |= self.pooling.get_pooling_updates(task) |
| 86 | |
| 87 | return updates |
| 88 | |
| 89 | def forward( |
| 90 | self, |
| 91 | hidden_states: torch.Tensor, |
| 92 | pooling_metadata: PoolingMetadata, |
| 93 | ) -> TokenPoolerOutput: |
| 94 | pooled_data = self.pooling(hidden_states, pooling_metadata) |
| 95 | if self.head is not None: |
| 96 | pooled_data = self.head(pooled_data, pooling_metadata) |
| 97 | return pooled_data |
| 98 | |
| 99 | |
| 100 | def pooler_for_token_embed( |
| 101 | pooler_config: PoolerConfig, projector: ProjectorFn | None = None |
| 102 | ) -> TokenPooler: |
| 103 | pooling = get_tok_pooling_method(pooler_config.get_tok_pooling_type()) |
| 104 | |
| 105 | vllm_config = get_current_vllm_config() |
| 106 | model_config = vllm_config.model_config |
| 107 | head = TokenEmbeddingPoolerHead( |
| 108 | head_dtype=model_config.head_dtype, |
| 109 | projector=projector |
| 110 | if projector is not None |
| 111 | else _load_st_projector(model_config), |
| 112 | activation=PoolerNormalize(), |
| 113 | ) |
| 114 | |
| 115 | return TokenPooler(pooling=pooling, head=head) |
| 116 | |
| 117 | |
| 118 | def pooler_for_token_classify( |
| 119 | pooler_config: PoolerConfig, |
| 120 | *, |
| 121 | pooling: TokenPoolingMethod | TokenPoolingFn | None = None, |
| 122 | classifier: ClassifierFn | None = None, |
| 123 | act_fn: PoolerActivation | None = None, |
| 124 | ) -> TokenPooler: |
| 125 | if pooling is None: |
| 126 | pooling = get_tok_pooling_method(pooler_config.get_tok_pooling_type()) |
| 127 | |
| 128 | vllm_config = get_current_vllm_config() |
| 129 | model_config = vllm_config.model_config |
| 130 | if model_config.pooler_config is None: |
| 131 | raise ValueError( |
| 132 | "model_config.pooler_config must be set for token classification pooling" |
| 133 | ) |
| 134 | head = TokenClassifierPoolerHead( |
| 135 | head_dtype=model_config.head_dtype, |
| 136 | classifier=classifier, |
| 137 | logit_mean=model_config.pooler_config.logit_mean, |
| 138 | logit_sigma=model_config.pooler_config.logit_sigma, |
| 139 | activation=resolve_classifier_act_fn( |
| 140 | model_config, static_num_labels=False, act_fn=act_fn |
| 141 | ), |
| 142 | ) |
| 143 | |
| 144 | return TokenPooler(pooling=pooling, head=head) |