Skip to content

vllm.model_executor.layers.pooler

ClassifierFn module-attribute 

ClassifierFn = Callable[[Tensor], Tensor]

PoolingFn module-attribute 

PoolingFn = Callable[
    [Union[Tensor, list[Tensor]], PoolingMetadata],
    Union[Tensor, list[Tensor]],
]

PoolingMetadata module-attribute 

PoolingMetadata = Union[PoolingMetadata, PoolingMetadata]

_T module-attribute 

_T = TypeVar('_T', Tensor, list[Tensor])

AllPool

Bases: PoolingMethod

Source code in vllm/model_executor/layers/pooler.py 
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
class AllPool(PoolingMethod):

    def forward_one(
        self,
        hidden_states: torch.Tensor,
        prompt_len: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        assert prompt_len is None or prompt_len == hidden_states.shape[0], \
            "partial prefill not supported with ALL pooling"

        return hidden_states

    def forward_all(
        self,
        hidden_states: torch.Tensor,
        prompt_lens: torch.Tensor,
    ) -> Union[list[torch.Tensor], torch.Tensor]:
        offset = 0
        pooled_data = list[torch.Tensor]()

        for prompt_len in prompt_lens:
            pooled_data.append(hidden_states[offset:offset + prompt_len])
            offset += prompt_len

        return pooled_data

forward_all 

forward_all(
    hidden_states: Tensor, prompt_lens: Tensor
) -> Union[list[Tensor], Tensor]
Source code in vllm/model_executor/layers/pooler.py 
221
222
223
224
225
226
227
228
229
230
231
232
233
def forward_all(
    self,
    hidden_states: torch.Tensor,
    prompt_lens: torch.Tensor,
) -> Union[list[torch.Tensor], torch.Tensor]:
    offset = 0
    pooled_data = list[torch.Tensor]()

    for prompt_len in prompt_lens:
        pooled_data.append(hidden_states[offset:offset + prompt_len])
        offset += prompt_len

    return pooled_data

forward_one 

forward_one(
    hidden_states: Tensor,
    prompt_len: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
211
212
213
214
215
216
217
218
219
def forward_one(
    self,
    hidden_states: torch.Tensor,
    prompt_len: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    assert prompt_len is None or prompt_len == hidden_states.shape[0], \
        "partial prefill not supported with ALL pooling"

    return hidden_states

BasePooler

Bases: Module

Source code in vllm/model_executor/layers/pooler.py 
107
108
109
110
111
112
113
114
115
class BasePooler(nn.Module):

    @abstractmethod
    def forward(
        self,
        hidden_states: Union[torch.Tensor, list[torch.Tensor]],
        pooling_metadata: PoolingMetadata,
    ) -> PoolerOutput:
        raise NotImplementedError

forward abstractmethod 

forward(
    hidden_states: Union[Tensor, list[Tensor]],
    pooling_metadata: PoolingMetadata,
) -> PoolerOutput
Source code in vllm/model_executor/layers/pooler.py 
109
110
111
112
113
114
115
@abstractmethod
def forward(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> PoolerOutput:
    raise NotImplementedError

BasePoolerActivation

Bases: Module, ABC

Source code in vllm/model_executor/layers/pooler.py 
269
270
271
272
273
274
275
276
277
class BasePoolerActivation(nn.Module, ABC):

    @abstractmethod
    def forward(self, pooled_data: _T) -> _T:
        # shape:
        # classify (& score) -> (batch_size, num_classes)
        # embed -> (batch_size, embedding_dim) or list(embedding_dim)
        #          (batch_size, dimensions) or list(dimensions) if using MRL
        raise NotImplementedError

forward abstractmethod 

forward(pooled_data: _T) -> _T
Source code in vllm/model_executor/layers/pooler.py 
271
272
273
274
275
276
277
@abstractmethod
def forward(self, pooled_data: _T) -> _T:
    # shape:
    # classify (& score) -> (batch_size, num_classes)
    # embed -> (batch_size, embedding_dim) or list(embedding_dim)
    #          (batch_size, dimensions) or list(dimensions) if using MRL
    raise NotImplementedError

CLSPool

Bases: PoolingMethod

Source code in vllm/model_executor/layers/pooler.py 
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
class CLSPool(PoolingMethod):

    def forward_one(
        self,
        hidden_states: torch.Tensor,
        prompt_len: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        assert prompt_len is None or prompt_len == hidden_states.shape[0], \
            "partial prefill not supported with CLS pooling"

        return hidden_states[0]

    def forward_all(
        self,
        hidden_states: torch.Tensor,
        prompt_lens: torch.Tensor,
    ) -> Union[list[torch.Tensor], torch.Tensor]:
        first_token_flat_indices = torch.zeros_like(prompt_lens)
        first_token_flat_indices[1:] += torch.cumsum(prompt_lens, dim=0)[:-1]
        return hidden_states[first_token_flat_indices]

forward_all 

forward_all(
    hidden_states: Tensor, prompt_lens: Tensor
) -> Union[list[Tensor], Tensor]
Source code in vllm/model_executor/layers/pooler.py 
181
182
183
184
185
186
187
188
def forward_all(
    self,
    hidden_states: torch.Tensor,
    prompt_lens: torch.Tensor,
) -> Union[list[torch.Tensor], torch.Tensor]:
    first_token_flat_indices = torch.zeros_like(prompt_lens)
    first_token_flat_indices[1:] += torch.cumsum(prompt_lens, dim=0)[:-1]
    return hidden_states[first_token_flat_indices]

forward_one 

forward_one(
    hidden_states: Tensor,
    prompt_len: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
171
172
173
174
175
176
177
178
179
def forward_one(
    self,
    hidden_states: torch.Tensor,
    prompt_len: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    assert prompt_len is None or prompt_len == hidden_states.shape[0], \
        "partial prefill not supported with CLS pooling"

    return hidden_states[0]

ClassifierPooler

Bases: Module

A pooling layer for classification tasks.

This layer does the following: 1. Applies a classification layer to the hidden states. 2. Optionally applies a pooler layer. 3. Applies an activation function to the output. In the case of classification models it is either sigmoid or softmax. In the case of scoring models, the same behavior is configuration dependent, as in the sentence-transformers library.

Source code in vllm/model_executor/layers/pooler.py 
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
class ClassifierPooler(nn.Module):
    """A pooling layer for classification tasks.

    This layer does the following:
    1. Applies a classification layer to the hidden states.
    2. Optionally applies a pooler layer.
    3. Applies an activation function to the output. In the case of
       classification models it is either sigmoid or softmax. In the
       case of scoring models, the same behavior is configuration
       dependent, as in the sentence-transformers library.
    """

    def __init__(
        self,
        config: ModelConfig,
        pooling: PoolingFn,
        classifier: ClassifierFn,
        act_fn: Optional[PoolerActivation] = None,
    ) -> None:
        super().__init__()

        self.pooling = pooling
        self.classifier = classifier

        self.classification_act_fn = get_classification_activation_function(
            config.hf_config) if act_fn is None else act_fn
        self.cross_encoder_act_fn = get_cross_encoder_activation_function(
            config.hf_config) if act_fn is None else act_fn

    def _get_act_fn(self, use_cross_encoder: bool):
        return (self.cross_encoder_act_fn
                if use_cross_encoder else self.classification_act_fn)

    def forward(
        self,
        hidden_states: Union[torch.Tensor, list[torch.Tensor]],
        pooling_metadata: PoolingMetadata,
    ) -> PoolerOutput:
        """Pools sentence pair scores from the hidden_states."""
        pooled_data = self.pooling(hidden_states, pooling_metadata)

        # apply classifier once on the full batch if possible
        if isinstance(pooled_data, torch.Tensor):
            pooled_output = self.classifier(pooled_data)
        elif len({data.shape for data in pooled_data}) <= 1:
            pooled_output = self.classifier(torch.stack(pooled_data))
        else:
            pooled_output = [self.classifier(data) for data in pooled_data]

        if isinstance(pooling_metadata, V0PoolingMetadata):
            use_cross_encoder_list = [
                pooling_param.use_cross_encoder
                for _, pooling_param in pooling_metadata.seq_groups
            ]
        else:
            use_cross_encoder_list = [
                pooling_param.use_cross_encoder
                for pooling_param in pooling_metadata.pooling_params
            ]

        # shape of scores: (batch_size, num_labels)
        if all(use_cross_encoder == use_cross_encoder_list[0]
               for use_cross_encoder in use_cross_encoder_list):
            act_fn = self._get_act_fn(use_cross_encoder_list[0])
            scores = act_fn(pooled_output)
        else:
            scores = torch.stack([
                self._get_act_fn(use_cross_encoder)(vecs)
                for use_cross_encoder, vecs in zip(use_cross_encoder_list,
                                                   pooled_output)
            ])

        return build_output(scores)

classification_act_fn instance-attribute 

classification_act_fn = (
    get_classification_activation_function(hf_config)
    if act_fn is None
    else act_fn
)

classifier instance-attribute 

classifier = classifier

cross_encoder_act_fn instance-attribute 

cross_encoder_act_fn = (
    get_cross_encoder_activation_function(hf_config)
    if act_fn is None
    else act_fn
)

pooling instance-attribute 

pooling = pooling

__init__ 

__init__(
    config: ModelConfig,
    pooling: PoolingFn,
    classifier: ClassifierFn,
    act_fn: Optional[PoolerActivation] = None,
) -> None
Source code in vllm/model_executor/layers/pooler.py 
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
def __init__(
    self,
    config: ModelConfig,
    pooling: PoolingFn,
    classifier: ClassifierFn,
    act_fn: Optional[PoolerActivation] = None,
) -> None:
    super().__init__()

    self.pooling = pooling
    self.classifier = classifier

    self.classification_act_fn = get_classification_activation_function(
        config.hf_config) if act_fn is None else act_fn
    self.cross_encoder_act_fn = get_cross_encoder_activation_function(
        config.hf_config) if act_fn is None else act_fn

_get_act_fn 

_get_act_fn(use_cross_encoder: bool)
Source code in vllm/model_executor/layers/pooler.py 
617
618
619
def _get_act_fn(self, use_cross_encoder: bool):
    return (self.cross_encoder_act_fn
            if use_cross_encoder else self.classification_act_fn)

forward 

forward(
    hidden_states: Union[Tensor, list[Tensor]],
    pooling_metadata: PoolingMetadata,
) -> PoolerOutput

Pools sentence pair scores from the hidden_states.

Source code in vllm/model_executor/layers/pooler.py 
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
def forward(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> PoolerOutput:
    """Pools sentence pair scores from the hidden_states."""
    pooled_data = self.pooling(hidden_states, pooling_metadata)

    # apply classifier once on the full batch if possible
    if isinstance(pooled_data, torch.Tensor):
        pooled_output = self.classifier(pooled_data)
    elif len({data.shape for data in pooled_data}) <= 1:
        pooled_output = self.classifier(torch.stack(pooled_data))
    else:
        pooled_output = [self.classifier(data) for data in pooled_data]

    if isinstance(pooling_metadata, V0PoolingMetadata):
        use_cross_encoder_list = [
            pooling_param.use_cross_encoder
            for _, pooling_param in pooling_metadata.seq_groups
        ]
    else:
        use_cross_encoder_list = [
            pooling_param.use_cross_encoder
            for pooling_param in pooling_metadata.pooling_params
        ]

    # shape of scores: (batch_size, num_labels)
    if all(use_cross_encoder == use_cross_encoder_list[0]
           for use_cross_encoder in use_cross_encoder_list):
        act_fn = self._get_act_fn(use_cross_encoder_list[0])
        scores = act_fn(pooled_output)
    else:
        scores = torch.stack([
            self._get_act_fn(use_cross_encoder)(vecs)
            for use_cross_encoder, vecs in zip(use_cross_encoder_list,
                                               pooled_output)
        ])

    return build_output(scores)

LambdaPoolerActivation

Bases: PoolerActivation

Source code in vllm/model_executor/layers/pooler.py 
335
336
337
338
339
340
341
342
343
class LambdaPoolerActivation(PoolerActivation):

    def __init__(self, fn: Callable[[torch.Tensor], torch.Tensor]):
        super().__init__()

        self.fn = fn

    def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
        return self.fn(pooled_data)

fn instance-attribute 

fn = fn

__init__ 

__init__(fn: Callable[[Tensor], Tensor])
Source code in vllm/model_executor/layers/pooler.py 
337
338
339
340
def __init__(self, fn: Callable[[torch.Tensor], torch.Tensor]):
    super().__init__()

    self.fn = fn

forward_chunk 

forward_chunk(pooled_data: Tensor) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
342
343
def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
    return self.fn(pooled_data)

LastPool

Bases: PoolingMethod

Source code in vllm/model_executor/layers/pooler.py 
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
class LastPool(PoolingMethod):

    def forward_one(
        self,
        hidden_states: torch.Tensor,
        prompt_len: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        return hidden_states[-1]

    def forward_all(
        self,
        hidden_states: torch.Tensor,
        prompt_lens: torch.Tensor,
    ) -> Union[list[torch.Tensor], torch.Tensor]:
        last_token_flat_indices = torch.cumsum(prompt_lens, dim=0) - 1
        return hidden_states[last_token_flat_indices]

forward_all 

forward_all(
    hidden_states: Tensor, prompt_lens: Tensor
) -> Union[list[Tensor], Tensor]
Source code in vllm/model_executor/layers/pooler.py 
200
201
202
203
204
205
206
def forward_all(
    self,
    hidden_states: torch.Tensor,
    prompt_lens: torch.Tensor,
) -> Union[list[torch.Tensor], torch.Tensor]:
    last_token_flat_indices = torch.cumsum(prompt_lens, dim=0) - 1
    return hidden_states[last_token_flat_indices]

forward_one 

forward_one(
    hidden_states: Tensor,
    prompt_len: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
193
194
195
196
197
198
def forward_one(
    self,
    hidden_states: torch.Tensor,
    prompt_len: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    return hidden_states[-1]

MeanPool

Bases: PoolingMethod

Source code in vllm/model_executor/layers/pooler.py 
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
class MeanPool(PoolingMethod):

    def forward_one(
        self,
        hidden_states: torch.Tensor,
        prompt_len: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        assert prompt_len is None or prompt_len == hidden_states.shape[0], \
            "partial prefill not supported with MEAN pooling"

        return hidden_states.mean(dim=0, dtype=torch.float32)

    def forward_all(
        self,
        hidden_states: torch.Tensor,
        prompt_lens: torch.Tensor,
    ) -> Union[list[torch.Tensor], torch.Tensor]:
        # Use float32 for torch.cumsum in MeanPool,
        # otherwise precision will be lost significantly.
        cumsum = torch.cumsum(hidden_states, dim=0, dtype=torch.float32)

        start_indices = torch.cat([
            torch.tensor([0], device=hidden_states.device),
            torch.cumsum(prompt_lens[:-1], dim=0)
        ])
        end_indices = torch.cumsum(prompt_lens, dim=0)
        return (cumsum[end_indices - 1] - cumsum[start_indices] +
                hidden_states[start_indices]) / prompt_lens.unsqueeze(1)

forward_all 

forward_all(
    hidden_states: Tensor, prompt_lens: Tensor
) -> Union[list[Tensor], Tensor]
Source code in vllm/model_executor/layers/pooler.py 
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
def forward_all(
    self,
    hidden_states: torch.Tensor,
    prompt_lens: torch.Tensor,
) -> Union[list[torch.Tensor], torch.Tensor]:
    # Use float32 for torch.cumsum in MeanPool,
    # otherwise precision will be lost significantly.
    cumsum = torch.cumsum(hidden_states, dim=0, dtype=torch.float32)

    start_indices = torch.cat([
        torch.tensor([0], device=hidden_states.device),
        torch.cumsum(prompt_lens[:-1], dim=0)
    ])
    end_indices = torch.cumsum(prompt_lens, dim=0)
    return (cumsum[end_indices - 1] - cumsum[start_indices] +
            hidden_states[start_indices]) / prompt_lens.unsqueeze(1)

forward_one 

forward_one(
    hidden_states: Tensor,
    prompt_len: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
238
239
240
241
242
243
244
245
246
def forward_one(
    self,
    hidden_states: torch.Tensor,
    prompt_len: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    assert prompt_len is None or prompt_len == hidden_states.shape[0], \
        "partial prefill not supported with MEAN pooling"

    return hidden_states.mean(dim=0, dtype=torch.float32)

Pooler

Bases: Module

Source code in vllm/model_executor/layers/pooler.py 
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
class Pooler(nn.Module):

    @staticmethod
    def from_config_with_defaults(
        pooler_config: PoolerConfig,
        pooling_type: PoolingType,
        normalize: bool,
        softmax: bool,
        step_tag_id: Optional[int] = None,
        returned_token_ids: Optional[list[int]] = None,
    ) -> BasePooler:
        resolved_config = ResolvedPoolingConfig.from_config_with_defaults(
            pooler_config=pooler_config,
            pooling_type=pooling_type,
            normalize=normalize,
            softmax=softmax,
            step_tag_id=step_tag_id,
            returned_token_ids=returned_token_ids,
        )

        if pooling_type == PoolingType.STEP:
            return StepPooler.from_config(resolved_config)

        return SimplePooler.from_config(resolved_config)

from_config_with_defaults staticmethod 

from_config_with_defaults(
    pooler_config: PoolerConfig,
    pooling_type: PoolingType,
    normalize: bool,
    softmax: bool,
    step_tag_id: Optional[int] = None,
    returned_token_ids: Optional[list[int]] = None,
) -> BasePooler
Source code in vllm/model_executor/layers/pooler.py 
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
@staticmethod
def from_config_with_defaults(
    pooler_config: PoolerConfig,
    pooling_type: PoolingType,
    normalize: bool,
    softmax: bool,
    step_tag_id: Optional[int] = None,
    returned_token_ids: Optional[list[int]] = None,
) -> BasePooler:
    resolved_config = ResolvedPoolingConfig.from_config_with_defaults(
        pooler_config=pooler_config,
        pooling_type=pooling_type,
        normalize=normalize,
        softmax=softmax,
        step_tag_id=step_tag_id,
        returned_token_ids=returned_token_ids,
    )

    if pooling_type == PoolingType.STEP:
        return StepPooler.from_config(resolved_config)

    return SimplePooler.from_config(resolved_config)

PoolerActivation

Bases: BasePoolerActivation

Source code in vllm/model_executor/layers/pooler.py 
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
class PoolerActivation(BasePoolerActivation):

    @staticmethod
    def wraps(module: nn.Module):
        if isinstance(module, nn.Identity):
            return PoolerIdentity()
        if isinstance(module, (nn.Sigmoid, nn.Softmax)):
            return PoolerClassify()

        return LambdaPoolerActivation(module)

    @abstractmethod
    def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
        raise NotImplementedError

    def forward(self, pooled_data: _T) -> _T:
        if isinstance(pooled_data, list):
            return [self.forward_chunk(data) for data in pooled_data]

        return self.forward_chunk(pooled_data)

forward 

forward(pooled_data: _T) -> _T
Source code in vllm/model_executor/layers/pooler.py 
295
296
297
298
299
def forward(self, pooled_data: _T) -> _T:
    if isinstance(pooled_data, list):
        return [self.forward_chunk(data) for data in pooled_data]

    return self.forward_chunk(pooled_data)

forward_chunk abstractmethod 

forward_chunk(pooled_data: Tensor) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
291
292
293
@abstractmethod
def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
    raise NotImplementedError

wraps staticmethod 

wraps(module: Module)
Source code in vllm/model_executor/layers/pooler.py 
282
283
284
285
286
287
288
289
@staticmethod
def wraps(module: nn.Module):
    if isinstance(module, nn.Identity):
        return PoolerIdentity()
    if isinstance(module, (nn.Sigmoid, nn.Softmax)):
        return PoolerClassify()

    return LambdaPoolerActivation(module)

PoolerClassify

Bases: PoolerActivation

Source code in vllm/model_executor/layers/pooler.py 
315
316
317
318
319
320
321
322
class PoolerClassify(PoolerActivation):

    def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
        num_labels = pooled_data.shape[-1]
        if num_labels < 2:
            return F.sigmoid(pooled_data.float()).to(pooled_data.dtype)

        return F.softmax(pooled_data.float(), dim=-1).to(pooled_data.dtype)

forward_chunk 

forward_chunk(pooled_data: Tensor) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
317
318
319
320
321
322
def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
    num_labels = pooled_data.shape[-1]
    if num_labels < 2:
        return F.sigmoid(pooled_data.float()).to(pooled_data.dtype)

    return F.softmax(pooled_data.float(), dim=-1).to(pooled_data.dtype)

PoolerHead

Bases: Module

Source code in vllm/model_executor/layers/pooler.py 
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
class PoolerHead(nn.Module):

    @classmethod
    def from_config_with_defaults(
        cls,
        pooler_config: PoolerConfig,
        pooling_type: PoolingType,
        normalize: bool,
        softmax: bool,
    ) -> "PoolerHead":
        resolved_config = ResolvedPoolingConfig.from_config_with_defaults(
            pooler_config=pooler_config,
            pooling_type=pooling_type,
            normalize=normalize,
            softmax=softmax,
            step_tag_id=None,
            returned_token_ids=None,
        )

        return cls.from_config(resolved_config)

    @classmethod
    def from_config(cls, pooler_config: ResolvedPoolingConfig) -> "PoolerHead":
        if pooler_config.normalize and pooler_config.softmax:
            raise ValueError("`normalize=True` and `softmax=True` should not "
                             "be set together")

        activation: PoolerActivation
        if pooler_config.normalize:
            activation = PoolerNormalize()
        elif pooler_config.softmax:
            activation = PoolerClassify()
        else:
            activation = PoolerIdentity()

        return cls(activation)

    def __init__(self, activation: PoolerActivation) -> None:
        super().__init__()

        self.activation = activation

    def forward(self, pooled_data: Union[list[torch.Tensor], torch.Tensor],
                pooling_metadata: PoolingMetadata):

        # Using float32 in PoolerHead
        if isinstance(pooled_data, list):
            for i in range(len(pooled_data)):
                pooled_data[i] = pooled_data[i].to(torch.float32)
        else:
            pooled_data = pooled_data.to(torch.float32)

        # for matryoshka representation
        if isinstance(pooling_metadata, V0PoolingMetadata):
            dimensions_list = [
                pooling_param.dimensions
                for _, pooling_param in pooling_metadata.seq_groups
            ]
        else:
            assert isinstance(pooled_data, list)
            dimensions_list = [
                pooling_param.dimensions
                for pooling_param in pooling_metadata.pooling_params
            ]
        if any(d is not None for d in dimensions_list):
            # change the output dimension
            assert len(pooled_data) == len(dimensions_list)
            if len(set(dimensions_list)) == 1 and not isinstance(
                    pooled_data, list):
                # if all dimensions are the same
                d = dimensions_list[0]
                pooled_data = pooled_data[..., :d]
            else:
                pooled_data = [
                    vecs if d is None else vecs[..., :d]
                    for vecs, d in zip(pooled_data, dimensions_list)
                ]

        return self.activation(pooled_data)

activation instance-attribute 

activation = activation

__init__ 

__init__(activation: PoolerActivation) -> None
Source code in vllm/model_executor/layers/pooler.py 
383
384
385
386
def __init__(self, activation: PoolerActivation) -> None:
    super().__init__()

    self.activation = activation

forward 

forward(
    pooled_data: Union[list[Tensor], Tensor],
    pooling_metadata: PoolingMetadata,
)
Source code in vllm/model_executor/layers/pooler.py 
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
def forward(self, pooled_data: Union[list[torch.Tensor], torch.Tensor],
            pooling_metadata: PoolingMetadata):

    # Using float32 in PoolerHead
    if isinstance(pooled_data, list):
        for i in range(len(pooled_data)):
            pooled_data[i] = pooled_data[i].to(torch.float32)
    else:
        pooled_data = pooled_data.to(torch.float32)

    # for matryoshka representation
    if isinstance(pooling_metadata, V0PoolingMetadata):
        dimensions_list = [
            pooling_param.dimensions
            for _, pooling_param in pooling_metadata.seq_groups
        ]
    else:
        assert isinstance(pooled_data, list)
        dimensions_list = [
            pooling_param.dimensions
            for pooling_param in pooling_metadata.pooling_params
        ]
    if any(d is not None for d in dimensions_list):
        # change the output dimension
        assert len(pooled_data) == len(dimensions_list)
        if len(set(dimensions_list)) == 1 and not isinstance(
                pooled_data, list):
            # if all dimensions are the same
            d = dimensions_list[0]
            pooled_data = pooled_data[..., :d]
        else:
            pooled_data = [
                vecs if d is None else vecs[..., :d]
                for vecs, d in zip(pooled_data, dimensions_list)
            ]

    return self.activation(pooled_data)

from_config classmethod 

from_config(
    pooler_config: ResolvedPoolingConfig,
) -> PoolerHead
Source code in vllm/model_executor/layers/pooler.py 
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
@classmethod
def from_config(cls, pooler_config: ResolvedPoolingConfig) -> "PoolerHead":
    if pooler_config.normalize and pooler_config.softmax:
        raise ValueError("`normalize=True` and `softmax=True` should not "
                         "be set together")

    activation: PoolerActivation
    if pooler_config.normalize:
        activation = PoolerNormalize()
    elif pooler_config.softmax:
        activation = PoolerClassify()
    else:
        activation = PoolerIdentity()

    return cls(activation)

from_config_with_defaults classmethod 

from_config_with_defaults(
    pooler_config: PoolerConfig,
    pooling_type: PoolingType,
    normalize: bool,
    softmax: bool,
) -> PoolerHead
Source code in vllm/model_executor/layers/pooler.py 
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
@classmethod
def from_config_with_defaults(
    cls,
    pooler_config: PoolerConfig,
    pooling_type: PoolingType,
    normalize: bool,
    softmax: bool,
) -> "PoolerHead":
    resolved_config = ResolvedPoolingConfig.from_config_with_defaults(
        pooler_config=pooler_config,
        pooling_type=pooling_type,
        normalize=normalize,
        softmax=softmax,
        step_tag_id=None,
        returned_token_ids=None,
    )

    return cls.from_config(resolved_config)

PoolerIdentity

Bases: PoolerActivation

Source code in vllm/model_executor/layers/pooler.py 
302
303
304
305
class PoolerIdentity(PoolerActivation):

    def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
        return pooled_data

forward_chunk 

forward_chunk(pooled_data: Tensor) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
304
305
def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
    return pooled_data

PoolerNormalize

Bases: PoolerActivation

Source code in vllm/model_executor/layers/pooler.py 
308
309
310
311
312
class PoolerNormalize(PoolerActivation):

    def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
        x = F.normalize(pooled_data.float(), p=2, dim=-1)
        return x.to(pooled_data.dtype)

forward_chunk 

forward_chunk(pooled_data: Tensor) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
310
311
312
def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
    x = F.normalize(pooled_data.float(), p=2, dim=-1)
    return x.to(pooled_data.dtype)

PoolerScore

Bases: PoolerActivation

Source code in vllm/model_executor/layers/pooler.py 
325
326
327
328
329
330
331
332
class PoolerScore(PoolerActivation):

    def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
        num_labels = pooled_data.shape[-1]
        if num_labels < 2:
            return F.sigmoid(pooled_data.float()).to(pooled_data.dtype)

        return pooled_data

forward_chunk 

forward_chunk(pooled_data: Tensor) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
327
328
329
330
331
332
def forward_chunk(self, pooled_data: torch.Tensor) -> torch.Tensor:
    num_labels = pooled_data.shape[-1]
    if num_labels < 2:
        return F.sigmoid(pooled_data.float()).to(pooled_data.dtype)

    return pooled_data

PoolingMethod

Bases: Module, ABC

Source code in vllm/model_executor/layers/pooler.py 
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
class PoolingMethod(nn.Module, ABC):

    @staticmethod
    def from_pooling_type(pooling_type: PoolingType) -> "PoolingMethod":
        if pooling_type == PoolingType.LAST:
            return LastPool()
        if pooling_type == PoolingType.ALL:
            return AllPool()
        if pooling_type == PoolingType.CLS:
            return CLSPool()
        if pooling_type == PoolingType.MEAN:
            return MeanPool()

        raise NotImplementedError(f"Unsupported method: {pooling_type}")

    @abstractmethod
    def forward_one(
        self,
        hidden_states: torch.Tensor,
        prompt_len: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        """
        Note:
            `prompt_len=None` means `prompt_len=len(hidden_states)`.
        """
        raise NotImplementedError

    @abstractmethod
    def forward_all(
        self,
        hidden_states: torch.Tensor,
        prompt_lens: torch.Tensor,
    ) -> Union[list[torch.Tensor], torch.Tensor]:
        raise NotImplementedError

    def forward(
        self,
        hidden_states: Union[torch.Tensor, list[torch.Tensor]],
        pooling_metadata: PoolingMetadata,
    ) -> Union[list[torch.Tensor], torch.Tensor]:
        prompt_lens = get_prompt_lens(hidden_states, pooling_metadata)

        if isinstance(hidden_states, list):
            return [
                self.forward_one(h, prompt_len)
                for h, prompt_len in zip(hidden_states, prompt_lens)
            ]

        return self.forward_all(hidden_states, prompt_lens)

forward 

forward(
    hidden_states: Union[Tensor, list[Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[list[Tensor], Tensor]
Source code in vllm/model_executor/layers/pooler.py 
153
154
155
156
157
158
159
160
161
162
163
164
165
166
def forward(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[list[torch.Tensor], torch.Tensor]:
    prompt_lens = get_prompt_lens(hidden_states, pooling_metadata)

    if isinstance(hidden_states, list):
        return [
            self.forward_one(h, prompt_len)
            for h, prompt_len in zip(hidden_states, prompt_lens)
        ]

    return self.forward_all(hidden_states, prompt_lens)

forward_all abstractmethod 

forward_all(
    hidden_states: Tensor, prompt_lens: Tensor
) -> Union[list[Tensor], Tensor]
Source code in vllm/model_executor/layers/pooler.py 
145
146
147
148
149
150
151
@abstractmethod
def forward_all(
    self,
    hidden_states: torch.Tensor,
    prompt_lens: torch.Tensor,
) -> Union[list[torch.Tensor], torch.Tensor]:
    raise NotImplementedError

forward_one abstractmethod 

forward_one(
    hidden_states: Tensor,
    prompt_len: Optional[Tensor] = None,
) -> Tensor
Note

prompt_len=None means prompt_len=len(hidden_states).

Source code in vllm/model_executor/layers/pooler.py 
133
134
135
136
137
138
139
140
141
142
143
@abstractmethod
def forward_one(
    self,
    hidden_states: torch.Tensor,
    prompt_len: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    """
    Note:
        `prompt_len=None` means `prompt_len=len(hidden_states)`.
    """
    raise NotImplementedError

from_pooling_type staticmethod 

from_pooling_type(
    pooling_type: PoolingType,
) -> PoolingMethod
Source code in vllm/model_executor/layers/pooler.py 
120
121
122
123
124
125
126
127
128
129
130
131
@staticmethod
def from_pooling_type(pooling_type: PoolingType) -> "PoolingMethod":
    if pooling_type == PoolingType.LAST:
        return LastPool()
    if pooling_type == PoolingType.ALL:
        return AllPool()
    if pooling_type == PoolingType.CLS:
        return CLSPool()
    if pooling_type == PoolingType.MEAN:
        return MeanPool()

    raise NotImplementedError(f"Unsupported method: {pooling_type}")

PoolingType

Bases: IntEnum

Enumeration for different types of pooling methods.

Source code in vllm/model_executor/layers/pooler.py 
24
25
26
27
28
29
30
class PoolingType(IntEnum):
    """Enumeration for different types of pooling methods."""
    LAST = 0
    ALL = 1
    CLS = 2
    STEP = 3
    MEAN = 4

ALL class-attribute instance-attribute 

ALL = 1

CLS class-attribute instance-attribute 

CLS = 2

LAST class-attribute instance-attribute 

LAST = 0

MEAN class-attribute instance-attribute 

MEAN = 4

STEP class-attribute instance-attribute 

STEP = 3

ResolvedPoolingConfig dataclass

Source code in vllm/model_executor/layers/pooler.py 
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
@dataclass(frozen=True)
class ResolvedPoolingConfig:
    pooling_type: PoolingType

    normalize: bool
    softmax: bool
    step_tag_id: Optional[int]
    returned_token_ids: Optional[list[int]]

    @classmethod
    def from_config_with_defaults(
        cls,
        pooler_config: PoolerConfig,
        pooling_type: PoolingType,
        normalize: bool,
        softmax: bool,
        step_tag_id: Optional[int] = None,
        returned_token_ids: Optional[list[int]] = None,
    ) -> "ResolvedPoolingConfig":
        return cls(
            pooling_type=PoolingType[pooler_config.pooling_type]
            if pooler_config.pooling_type is not None else pooling_type,
            normalize=pooler_config.normalize
            if pooler_config.normalize is not None else normalize,
            softmax=pooler_config.softmax
            if pooler_config.softmax is not None else softmax,
            step_tag_id=pooler_config.step_tag_id
            if pooler_config.step_tag_id is not None else step_tag_id,
            returned_token_ids=pooler_config.returned_token_ids
            if pooler_config.returned_token_ids is not None else
            returned_token_ids,
        )

normalize instance-attribute 

normalize: bool

pooling_type instance-attribute 

pooling_type: PoolingType

returned_token_ids instance-attribute 

returned_token_ids: Optional[list[int]]

softmax instance-attribute 

softmax: bool

step_tag_id instance-attribute 

step_tag_id: Optional[int]

__init__ 

__init__(
    pooling_type: PoolingType,
    normalize: bool,
    softmax: bool,
    step_tag_id: Optional[int],
    returned_token_ids: Optional[list[int]],
) -> None

from_config_with_defaults classmethod 

from_config_with_defaults(
    pooler_config: PoolerConfig,
    pooling_type: PoolingType,
    normalize: bool,
    softmax: bool,
    step_tag_id: Optional[int] = None,
    returned_token_ids: Optional[list[int]] = None,
) -> ResolvedPoolingConfig
Source code in vllm/model_executor/layers/pooler.py 
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
@classmethod
def from_config_with_defaults(
    cls,
    pooler_config: PoolerConfig,
    pooling_type: PoolingType,
    normalize: bool,
    softmax: bool,
    step_tag_id: Optional[int] = None,
    returned_token_ids: Optional[list[int]] = None,
) -> "ResolvedPoolingConfig":
    return cls(
        pooling_type=PoolingType[pooler_config.pooling_type]
        if pooler_config.pooling_type is not None else pooling_type,
        normalize=pooler_config.normalize
        if pooler_config.normalize is not None else normalize,
        softmax=pooler_config.softmax
        if pooler_config.softmax is not None else softmax,
        step_tag_id=pooler_config.step_tag_id
        if pooler_config.step_tag_id is not None else step_tag_id,
        returned_token_ids=pooler_config.returned_token_ids
        if pooler_config.returned_token_ids is not None else
        returned_token_ids,
    )

SimplePooler

Bases: BasePooler

A layer that pools specific information from hidden states.

This layer does the following: 1. Extracts specific tokens or aggregates data based on pooling method. 2. Normalizes output if specified. 3. Returns structured results as PoolerOutput.

Attributes:

Name Type Description
pooling_type

The type of pooling to use.
normalize

Whether to normalize the pooled data.
Source code in vllm/model_executor/layers/pooler.py 
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
class SimplePooler(BasePooler):
    """A layer that pools specific information from hidden states.

    This layer does the following:
    1. Extracts specific tokens or aggregates data based on pooling method.
    2. Normalizes output if specified.
    3. Returns structured results as `PoolerOutput`.

    Attributes:
        pooling_type: The type of pooling to use.
        normalize: Whether to normalize the pooled data.
    """

    @classmethod
    def from_config_with_defaults(
        cls,
        pooler_config: PoolerConfig,
        pooling_type: PoolingType,
        normalize: bool,
        softmax: bool,
    ) -> "SimplePooler":
        resolved_config = ResolvedPoolingConfig.from_config_with_defaults(
            pooler_config=pooler_config,
            pooling_type=pooling_type,
            normalize=normalize,
            softmax=softmax,
        )
        assert resolved_config.pooling_type != PoolingType.STEP

        return cls.from_config(resolved_config)

    @classmethod
    def from_config(
        cls,
        pooler_config: ResolvedPoolingConfig,
    ) -> "SimplePooler":
        pooling = PoolingMethod.from_pooling_type(pooler_config.pooling_type)
        head = PoolerHead.from_config(pooler_config)

        return cls(pooling, head)

    def __init__(self, pooling: PoolingMethod, head: PoolerHead) -> None:
        super().__init__()

        self.pooling = pooling
        self.head = head

    def forward(
        self,
        hidden_states: Union[torch.Tensor, list[torch.Tensor]],
        pooling_metadata: PoolingMetadata,
    ) -> PoolerOutput:
        pooled_data = self.pooling(hidden_states, pooling_metadata)
        pooled_data = self.head(pooled_data, pooling_metadata)
        return build_output(pooled_data)

head instance-attribute 

head = head

pooling instance-attribute 

pooling = pooling

__init__ 

__init__(pooling: PoolingMethod, head: PoolerHead) -> None
Source code in vllm/model_executor/layers/pooler.py 
468
469
470
471
472
def __init__(self, pooling: PoolingMethod, head: PoolerHead) -> None:
    super().__init__()

    self.pooling = pooling
    self.head = head

forward 

forward(
    hidden_states: Union[Tensor, list[Tensor]],
    pooling_metadata: PoolingMetadata,
) -> PoolerOutput
Source code in vllm/model_executor/layers/pooler.py 
474
475
476
477
478
479
480
481
def forward(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> PoolerOutput:
    pooled_data = self.pooling(hidden_states, pooling_metadata)
    pooled_data = self.head(pooled_data, pooling_metadata)
    return build_output(pooled_data)

from_config classmethod 

from_config(
    pooler_config: ResolvedPoolingConfig,
) -> SimplePooler
Source code in vllm/model_executor/layers/pooler.py 
458
459
460
461
462
463
464
465
466
@classmethod
def from_config(
    cls,
    pooler_config: ResolvedPoolingConfig,
) -> "SimplePooler":
    pooling = PoolingMethod.from_pooling_type(pooler_config.pooling_type)
    head = PoolerHead.from_config(pooler_config)

    return cls(pooling, head)

from_config_with_defaults classmethod 

from_config_with_defaults(
    pooler_config: PoolerConfig,
    pooling_type: PoolingType,
    normalize: bool,
    softmax: bool,
) -> SimplePooler
Source code in vllm/model_executor/layers/pooler.py 
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
@classmethod
def from_config_with_defaults(
    cls,
    pooler_config: PoolerConfig,
    pooling_type: PoolingType,
    normalize: bool,
    softmax: bool,
) -> "SimplePooler":
    resolved_config = ResolvedPoolingConfig.from_config_with_defaults(
        pooler_config=pooler_config,
        pooling_type=pooling_type,
        normalize=normalize,
        softmax=softmax,
    )
    assert resolved_config.pooling_type != PoolingType.STEP

    return cls.from_config(resolved_config)

StepPooler

Bases: BasePooler

Source code in vllm/model_executor/layers/pooler.py 
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
class StepPooler(BasePooler):

    @classmethod
    def from_config(cls, pooler_config: ResolvedPoolingConfig) -> "StepPooler":
        assert pooler_config.pooling_type == PoolingType.STEP

        return cls(
            PoolerHead.from_config(pooler_config),
            step_tag_id=pooler_config.step_tag_id,
            returned_token_ids=pooler_config.returned_token_ids,
        )

    def __init__(
        self,
        head: PoolerHead,
        *,
        step_tag_id: Optional[int] = None,
        returned_token_ids: Optional[list[int]] = None,
    ) -> None:
        super().__init__()

        self.pooling = AllPool()
        self.head = head
        self.step_tag_id = step_tag_id
        self.returned_token_ids = returned_token_ids

    def get_prompt_token_ids(
        self,
        pooling_metadata: PoolingMetadata,
    ) -> list[torch.Tensor]:
        if isinstance(pooling_metadata, V1PoolingMetadata):
            return [
                pooling_metadata.prompt_token_ids[i, :num]
                for i, num in enumerate(pooling_metadata.prompt_lens)
            ]
        return [
            torch.tensor(seq_data_i.prompt_token_ids)
            for seq_data_i in pooling_metadata.seq_data.values()
        ]

    def extract_states(
        self,
        hidden_states: Union[torch.Tensor, list[torch.Tensor]],
        pooling_metadata: PoolingMetadata,
    ) -> Union[list[torch.Tensor], torch.Tensor]:
        pooled_data_lst = self.pooling(hidden_states, pooling_metadata)
        prompt_token_ids = self.get_prompt_token_ids(pooling_metadata)

        pooled_data = list[torch.Tensor]()
        returned_token_ids = self.returned_token_ids
        step_tag_id = self.step_tag_id

        for data, token_id in zip(pooled_data_lst, prompt_token_ids):
            if returned_token_ids is not None and len(returned_token_ids) > 0:
                data = data[:, returned_token_ids]

            if step_tag_id is not None:
                data = data[token_id == step_tag_id]
            pooled_data.append(data)

        return pooled_data

    def forward(
        self,
        hidden_states: Union[torch.Tensor, list[torch.Tensor]],
        pooling_metadata: PoolingMetadata,
    ) -> PoolerOutput:
        pooled_data = self.extract_states(hidden_states, pooling_metadata)
        pooled_data = self.head(pooled_data, pooling_metadata)
        return build_output(pooled_data)

head instance-attribute 

head = head

pooling instance-attribute 

pooling = AllPool()

returned_token_ids instance-attribute 

returned_token_ids = returned_token_ids

step_tag_id instance-attribute 

step_tag_id = step_tag_id

__init__ 

__init__(
    head: PoolerHead,
    *,
    step_tag_id: Optional[int] = None,
    returned_token_ids: Optional[list[int]] = None,
) -> None
Source code in vllm/model_executor/layers/pooler.py 
496
497
498
499
500
501
502
503
504
505
506
507
508
def __init__(
    self,
    head: PoolerHead,
    *,
    step_tag_id: Optional[int] = None,
    returned_token_ids: Optional[list[int]] = None,
) -> None:
    super().__init__()

    self.pooling = AllPool()
    self.head = head
    self.step_tag_id = step_tag_id
    self.returned_token_ids = returned_token_ids

extract_states 

extract_states(
    hidden_states: Union[Tensor, list[Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[list[Tensor], Tensor]
Source code in vllm/model_executor/layers/pooler.py 
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
def extract_states(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[list[torch.Tensor], torch.Tensor]:
    pooled_data_lst = self.pooling(hidden_states, pooling_metadata)
    prompt_token_ids = self.get_prompt_token_ids(pooling_metadata)

    pooled_data = list[torch.Tensor]()
    returned_token_ids = self.returned_token_ids
    step_tag_id = self.step_tag_id

    for data, token_id in zip(pooled_data_lst, prompt_token_ids):
        if returned_token_ids is not None and len(returned_token_ids) > 0:
            data = data[:, returned_token_ids]

        if step_tag_id is not None:
            data = data[token_id == step_tag_id]
        pooled_data.append(data)

    return pooled_data

forward 

forward(
    hidden_states: Union[Tensor, list[Tensor]],
    pooling_metadata: PoolingMetadata,
) -> PoolerOutput
Source code in vllm/model_executor/layers/pooler.py 
546
547
548
549
550
551
552
553
def forward(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> PoolerOutput:
    pooled_data = self.extract_states(hidden_states, pooling_metadata)
    pooled_data = self.head(pooled_data, pooling_metadata)
    return build_output(pooled_data)

from_config classmethod 

from_config(
    pooler_config: ResolvedPoolingConfig,
) -> StepPooler
Source code in vllm/model_executor/layers/pooler.py 
486
487
488
489
490
491
492
493
494
@classmethod
def from_config(cls, pooler_config: ResolvedPoolingConfig) -> "StepPooler":
    assert pooler_config.pooling_type == PoolingType.STEP

    return cls(
        PoolerHead.from_config(pooler_config),
        step_tag_id=pooler_config.step_tag_id,
        returned_token_ids=pooler_config.returned_token_ids,
    )

get_prompt_token_ids 

get_prompt_token_ids(
    pooling_metadata: PoolingMetadata,
) -> list[Tensor]
Source code in vllm/model_executor/layers/pooler.py 
510
511
512
513
514
515
516
517
518
519
520
521
522
def get_prompt_token_ids(
    self,
    pooling_metadata: PoolingMetadata,
) -> list[torch.Tensor]:
    if isinstance(pooling_metadata, V1PoolingMetadata):
        return [
            pooling_metadata.prompt_token_ids[i, :num]
            for i, num in enumerate(pooling_metadata.prompt_lens)
        ]
    return [
        torch.tensor(seq_data_i.prompt_token_ids)
        for seq_data_i in pooling_metadata.seq_data.values()
    ]

build_output 

build_output(all_data: Tensor) -> PoolerOutput
Source code in vllm/model_executor/layers/pooler.py 
102
103
104
def build_output(all_data: torch.Tensor) -> PoolerOutput:
    all_outputs = [PoolingSequenceGroupOutput(data) for data in all_data]
    return PoolerOutput(outputs=all_outputs)

get_classification_activation_function 

get_classification_activation_function(
    config: PretrainedConfig,
)
Source code in vllm/model_executor/layers/pooler.py 
79
80
def get_classification_activation_function(config: PretrainedConfig):
    return PoolerClassify()

get_cross_encoder_activation_function 

get_cross_encoder_activation_function(
    config: PretrainedConfig,
)
Source code in vllm/model_executor/layers/pooler.py 
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
def get_cross_encoder_activation_function(config: PretrainedConfig):
    function_name: Optional[str] = None
    if (hasattr(config, "sentence_transformers")
            and "activation_fn" in config.sentence_transformers):
        function_name = config.sentence_transformers["activation_fn"]
    elif (hasattr(config, "sbert_ce_default_activation_function")
          and config.sbert_ce_default_activation_function is not None):
        function_name = config.sbert_ce_default_activation_function

    if function_name is not None:
        assert function_name.startswith("torch.nn.modules."), (
            "Loading of activation functions is restricted to "
            "torch.nn.modules for security reasons")
        fn = resolve_obj_by_qualname(function_name)()
        return PoolerActivation.wraps(fn)

    return PoolerScore()

get_prompt_lens 

get_prompt_lens(
    hidden_states: Union[Tensor, list[Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Tensor
Source code in vllm/model_executor/layers/pooler.py 
67
68
69
70
71
72
73
74
75
76
def get_prompt_lens(
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> torch.Tensor:
    if isinstance(pooling_metadata, V1PoolingMetadata):
        return pooling_metadata.prompt_lens

    assert isinstance(hidden_states, torch.Tensor)
    return PoolingTensors.from_pooling_metadata(
        pooling_metadata, hidden_states.device).prompt_lens