| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from typing import ( |
| 4 | TYPE_CHECKING, |
| 5 | Any, |
| 6 | ClassVar, |
| 7 | Literal, |
| 8 | Protocol, |
| 9 | overload, |
| 10 | runtime_checkable, |
| 11 | ) |
| 12 | |
| 13 | import torch |
| 14 | import torch.nn as nn |
| 15 | from typing_extensions import TypeIs, TypeVar |
| 16 | |
| 17 | from vllm.logger import init_logger |
| 18 | from vllm.tasks import ScoreType |
| 19 | from vllm.utils.func_utils import supports_kw |
| 20 | |
| 21 | if TYPE_CHECKING: |
| 22 | from vllm.config import VllmConfig |
| 23 | from vllm.config.model import AttnTypeStr |
| 24 | from vllm.config.pooler import SequencePoolingType, TokenPoolingType |
| 25 | from vllm.model_executor.layers.pooler import Pooler |
| 26 | else: |
| 27 | VllmConfig = Any |
| 28 | Pooler = Any |
| 29 | SequencePoolingType = Any |
| 30 | TokenPoolingType = Any |
| 31 | AttnTypeStr = Any |
| 32 | |
| 33 | logger = init_logger(__name__) |
| 34 | |
| 35 | # The type of hidden states |
| 36 | # Currently, T = torch.Tensor for all models except for Medusa |
| 37 | # which has T = list[torch.Tensor] |
| 38 | T = TypeVar("T", default=torch.Tensor) |
| 39 | T_co = TypeVar("T_co", default=torch.Tensor, covariant=True) |
| 40 | |
| 41 | # NOTE: Unlike those in `interfaces.py`, we don't define `ClassVar` tags |
| 42 | # for the base interfaces to avoid breaking OOT registration for existing models |
| 43 | # that don't inherit from the base interface classes |
| 44 | |
| 45 | |
| 46 | @runtime_checkable |
| 47 | class VllmModel(Protocol[T_co]): |
| 48 | """The interface required for all models in vLLM.""" |
| 49 | |
| 50 | def __init__(self, vllm_config: VllmConfig, prefix: str = "") -> None: ... |
| 51 | |
| 52 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 53 | """Apply token embeddings to `input_ids`.""" |
| 54 | ... |
| 55 | |
| 56 | def forward(self, input_ids: torch.Tensor, positions: torch.Tensor) -> T_co: ... |
| 57 | |
| 58 | |
| 59 | def _check_vllm_model_init(model: type[object] | object) -> bool: |
| 60 | model_init = getattr(model, "__init__") # noqa: B009 |
| 61 | return supports_kw(model_init, "vllm_config") |
| 62 | |
| 63 | |
| 64 | def _check_vllm_model_embed_input_ids(model: type[object] | object) -> bool: |
| 65 | model_embed_input_ids = getattr(model, "embed_input_ids", None) |
| 66 | if not callable(model_embed_input_ids): |
| 67 | logger.warning( |
| 68 | "The model (%s) is missing the `embed_input_ids` method.", |
| 69 | model, |
| 70 | ) |
| 71 | return False |
| 72 | |
| 73 | return True |
| 74 | |
| 75 | |
| 76 | def _check_vllm_model_forward(model: type[object] | object) -> bool: |
| 77 | model_forward = getattr(model, "forward", None) |
| 78 | if not callable(model_forward): |
| 79 | return False |
| 80 | |
| 81 | vllm_kws = ("input_ids", "positions") |
| 82 | missing_kws = tuple(kw for kw in vllm_kws if not supports_kw(model_forward, kw)) |
| 83 | |
| 84 | if missing_kws and (isinstance(model, type) and issubclass(model, nn.Module)): |
| 85 | logger.warning( |
| 86 | "The model (%s) is missing " |
| 87 | "vLLM-specific keywords from its `forward` method: %s", |
| 88 | model, |
| 89 | missing_kws, |
| 90 | ) |
| 91 | |
| 92 | return len(missing_kws) == 0 |
| 93 | |
| 94 | |
| 95 | @overload |
| 96 | def is_vllm_model(model: type[object]) -> TypeIs[type[VllmModel]]: ... |
| 97 | |
| 98 | |
| 99 | @overload |
| 100 | def is_vllm_model(model: object) -> TypeIs[VllmModel]: ... |
| 101 | |
| 102 | |
| 103 | def is_vllm_model( |
| 104 | model: type[object] | object, |
| 105 | ) -> TypeIs[type[VllmModel]] | TypeIs[VllmModel]: |
| 106 | return ( |
| 107 | _check_vllm_model_init(model) |
| 108 | and _check_vllm_model_embed_input_ids(model) |
| 109 | and _check_vllm_model_forward(model) |
| 110 | ) |
| 111 | |
| 112 | |
| 113 | @runtime_checkable |
| 114 | class VllmModelForTextGeneration(VllmModel[T], Protocol[T]): |
| 115 | """The interface required for all generative models in vLLM.""" |
| 116 | |
| 117 | def compute_logits( |
| 118 | self, |
| 119 | hidden_states: T, |
| 120 | ) -> T | None: |
| 121 | """Return `None` if TP rank > 0.""" |
| 122 | ... |
| 123 | |
| 124 | |
| 125 | @overload |
| 126 | def is_text_generation_model( |
| 127 | model: type[object], |
| 128 | ) -> TypeIs[type[VllmModelForTextGeneration]]: ... |
| 129 | |
| 130 | |
| 131 | @overload |
| 132 | def is_text_generation_model(model: object) -> TypeIs[VllmModelForTextGeneration]: ... |
| 133 | |
| 134 | |
| 135 | def is_text_generation_model( |
| 136 | model: type[object] | object, |
| 137 | ) -> TypeIs[type[VllmModelForTextGeneration]] | TypeIs[VllmModelForTextGeneration]: |
| 138 | if not is_vllm_model(model): |
| 139 | return False |
| 140 | |
| 141 | if isinstance(model, type): |
| 142 | return isinstance(model, VllmModelForTextGeneration) |
| 143 | |
| 144 | return isinstance(model, VllmModelForTextGeneration) |
| 145 | |
| 146 | |
| 147 | @runtime_checkable |
| 148 | class VllmModelForPooling(VllmModel[T_co], Protocol[T_co]): |
| 149 | """The interface required for all pooling models in vLLM.""" |
| 150 | |
| 151 | is_pooling_model: ClassVar[Literal[True]] = True |
| 152 | """ |
| 153 | A flag that indicates this model supports pooling. |
| 154 | |
| 155 | Note: |
| 156 | There is no need to redefine this flag if this class is in the |
| 157 | MRO of your model class. |
| 158 | """ |
| 159 | |
| 160 | default_seq_pooling_type: ClassVar[SequencePoolingType] = "LAST" |
| 161 | """ |
| 162 | Indicates the [vllm.config.pooler.PoolerConfig.seq_pooling_type][] |
| 163 | to use by default. |
| 164 | |
| 165 | You can use the |
| 166 | [vllm.model_executor.models.interfaces_base.default_pooling_type][] |
| 167 | decorator to conveniently set this field. |
| 168 | """ |
| 169 | |
| 170 | default_tok_pooling_type: ClassVar[TokenPoolingType] = "ALL" |
| 171 | """ |
| 172 | Indicates the [vllm.config.pooler.PoolerConfig.tok_pooling_type][] |
| 173 | to use by default. |
| 174 | |
| 175 | You can use the |
| 176 | [vllm.model_executor.models.interfaces_base.default_pooling_type][] |
| 177 | decorator to conveniently set this field. |
| 178 | """ |
| 179 | |
| 180 | attn_type: ClassVar[AttnTypeStr] = "decoder" |
| 181 | """ |
| 182 | Indicates the |
| 183 | [vllm.config.model.ModelConfig.attn_type][] |
| 184 | to use by default. |
| 185 | |
| 186 | You can use the |
| 187 | [vllm.model_executor.models.interfaces_base.attn_type][] |
| 188 | decorator to conveniently set this field. |
| 189 | """ |
| 190 | |
| 191 | score_type: ClassVar[ScoreType] = "bi-encoder" |
| 192 | """ |
| 193 | Indicates the |
| 194 | [vllm.config.model.ModelConfig.score_type][] |
| 195 | to use by default. |
| 196 | |
| 197 | Scoring API handles score/rerank for:\n |
| 198 | - "classify" task (score_type: cross-encoder models)\n |
| 199 | - "embed" task (score_type: bi-encoder models)\n |
| 200 | - "token_embed" task (score_type: late interaction models)\n |
| 201 | |
| 202 | score_type defaults to bi-encoder, then the Score API uses the "embed" task.\n |
| 203 | If you set score_type to cross-encoder via |
| 204 | [vllm.model_executor.models.interfaces.SupportsCrossEncoding][], |
| 205 | then the Score API uses the "score" task.\n |
| 206 | If you set score_type to late-interaction via |
| 207 | [vllm.model_executor.models.interfaces.SupportsLateInteraction][], |
| 208 | then the Score API uses the "token_embed" task.\n |
| 209 | """ |
| 210 | |
| 211 | pooler: Pooler |
| 212 | """The pooler is only called on TP rank 0.""" |
| 213 | |
| 214 | |
| 215 | @overload |
| 216 | def is_pooling_model(model: type[object]) -> TypeIs[type[VllmModelForPooling]]: ... |
| 217 | |
| 218 | |
| 219 | @overload |
| 220 | def is_pooling_model(model: object) -> TypeIs[VllmModelForPooling]: ... |
| 221 | |
| 222 | |
| 223 | def is_pooling_model( |
| 224 | model: type[object] | object, |
| 225 | ) -> TypeIs[type[VllmModelForPooling]] | TypeIs[VllmModelForPooling]: |
| 226 | if not is_vllm_model(model): |
| 227 | return False |
| 228 | |
| 229 | return getattr(model, "is_pooling_model", False) |
| 230 | |
| 231 | |
| 232 | _T = TypeVar("_T", bound=type[nn.Module]) |
| 233 | |
| 234 | |
| 235 | def default_pooling_type( |
| 236 | *, |
| 237 | seq_pooling_type: SequencePoolingType = "LAST", |
| 238 | tok_pooling_type: TokenPoolingType = "ALL", |
| 239 | ): |
| 240 | """Decorator to set `VllmModelForPooling.default_*_pooling_type`.""" |
| 241 | |
| 242 | def func(model: _T) -> _T: |
| 243 | model.default_seq_pooling_type = seq_pooling_type # type: ignore |
| 244 | model.default_tok_pooling_type = tok_pooling_type # type: ignore |
| 245 | return model |
| 246 | |
| 247 | return func |
| 248 | |
| 249 | |
| 250 | def get_default_seq_pooling_type( |
| 251 | model: type[object] | object, |
| 252 | ) -> SequencePoolingType: |
| 253 | return getattr(model, "default_seq_pooling_type", "LAST") |
| 254 | |
| 255 | |
| 256 | def get_default_tok_pooling_type( |
| 257 | model: type[object] | object, |
| 258 | ) -> TokenPoolingType: |
| 259 | return getattr(model, "default_tok_pooling_type", "ALL") |
| 260 | |
| 261 | |
| 262 | def attn_type(attn_type: AttnTypeStr): |
| 263 | """Decorator to set `VllmModelForPooling.attn_type`.""" |
| 264 | |
| 265 | def func(model: _T) -> _T: |
| 266 | model.attn_type = attn_type # type: ignore |
| 267 | return model |
| 268 | |
| 269 | return func |
| 270 | |
| 271 | |
| 272 | def get_attn_type(model: type[object] | object) -> AttnTypeStr: |
| 273 | return getattr(model, "attn_type", "decoder") |
| 274 | |
| 275 | |
| 276 | def get_score_type(model: type[object] | object) -> ScoreType: |
| 277 | score_types: set[ScoreType] = set() |
| 278 | model_cls = model if isinstance(model, type) else type(model) |
| 279 | for m in model_cls.__mro__: |
| 280 | score_type: ScoreType = getattr(m, "score_type", "bi-encoder") |
| 281 | if score_type != "bi-encoder": |
| 282 | score_types.add(score_type) |
| 283 | assert len(score_types) < 2 |
| 284 | return "bi-encoder" if not score_types else next(iter(score_types)) |