Skip to content

vllm.model_executor.layers.pooler

PoolingMetadata module-attribute 

PoolingMetadata = Union[PoolingMetadata, PoolingMetadata]

AllPool

Bases: SimplePooler

Source code in vllm/model_executor/layers/pooler.py 
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
class AllPool(SimplePooler):

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

        if isinstance(hidden_states, list):
            for req_state, prompt_len in zip(hidden_states, prompt_lens):
                assert prompt_len == req_state.shape[0], \
                    "partial prefill not supported with ALL pooling"
            return hidden_states

        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

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 
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
def extract_states(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[list[torch.Tensor], torch.Tensor]:
    prompt_lens = self.get_prompt_lens(hidden_states, pooling_metadata)

    if isinstance(hidden_states, list):
        for req_state, prompt_len in zip(hidden_states, prompt_lens):
            assert prompt_len == req_state.shape[0], \
                "partial prefill not supported with ALL pooling"
        return hidden_states

    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

CLSPool

Bases: SimplePooler

Source code in vllm/model_executor/layers/pooler.py 
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
class CLSPool(SimplePooler):

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

        if isinstance(hidden_states, list):
            result = []
            for req_state, prompt_len in zip(hidden_states, prompt_lens):
                assert prompt_len == req_state.shape[0], \
                    "partial prefill not supported with CLS pooling"
                result.append(req_state[0])
            return result

        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]

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 
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
def extract_states(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[list[torch.Tensor], torch.Tensor]:
    prompt_lens = self.get_prompt_lens(hidden_states, pooling_metadata)

    if isinstance(hidden_states, list):
        result = []
        for req_state, prompt_len in zip(hidden_states, prompt_lens):
            assert prompt_len == req_state.shape[0], \
                "partial prefill not supported with CLS pooling"
            result.append(req_state[0])
        return result

    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]

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 
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
425
426
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
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,
        classifier: nn.Module,
        pooler: Optional[nn.Module] = None,
    ):
        super().__init__()
        self.classifier = classifier
        self.pooler = pooler

        if config.task == "score":
            self.default_activation_function = \
                get_cross_encoder_activation_function(config.hf_config)
        elif config.task == "classify":
            self.default_activation_function = nn.Sigmoid() \
                if config.hf_config.num_labels == 1 else nn.Softmax()
        else:
            raise NotImplementedError(f"task={config.task!r} is not supported"
                                      " with the classification pooler")

    def get_prompt_lens(
        self,
        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

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

        pooled_data = list[torch.Tensor]()
        if isinstance(hidden_states, list):
            for req_state, prompt_len in zip(hidden_states, prompt_lens):
                assert prompt_len == req_state.shape[0], \
                    "partial prefill not supported with classifier"
            pooled_data = hidden_states
        else:
            offset = 0
            for prompt_len in prompt_lens:
                pooled_data_i = hidden_states[offset:offset + prompt_len]
                offset += prompt_len
                pooled_data.append(pooled_data_i)

        pooled_data_lst = []
        for pooled_data_i in pooled_data:

            if self.pooler is not None:
                final_shape_tensor = self.pooler(pooled_data_i)
            else:
                final_shape_tensor = self.classifier(pooled_data_i)

            pooled_data_lst.append(final_shape_tensor)

        pooled_output = torch.stack(pooled_data_lst)

        if self.pooler is not None:
            # apply classifier once on the full batch if possible
            pooled_output = self.classifier(pooled_output)

        # shape: (batch_size, num_labels)
        scores = self.default_activation_function(pooled_output)

        pooled_outputs = [PoolingSequenceGroupOutput(data) for data in scores]
        return PoolerOutput(outputs=pooled_outputs)

classifier instance-attribute 

classifier = classifier

default_activation_function instance-attribute 

default_activation_function = (
    get_cross_encoder_activation_function(hf_config)
)

pooler instance-attribute 

pooler = pooler

__init__ 

__init__(
    config: ModelConfig,
    classifier: Module,
    pooler: Optional[Module] = None,
)
Source code in vllm/model_executor/layers/pooler.py 
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
def __init__(
    self,
    config: ModelConfig,
    classifier: nn.Module,
    pooler: Optional[nn.Module] = None,
):
    super().__init__()
    self.classifier = classifier
    self.pooler = pooler

    if config.task == "score":
        self.default_activation_function = \
            get_cross_encoder_activation_function(config.hf_config)
    elif config.task == "classify":
        self.default_activation_function = nn.Sigmoid() \
            if config.hf_config.num_labels == 1 else nn.Softmax()
    else:
        raise NotImplementedError(f"task={config.task!r} is not supported"
                                  " with the classification pooler")

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 
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
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
def forward(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> PoolerOutput:
    """Pools sentence pair scores from the hidden_states."""
    prompt_lens = self.get_prompt_lens(hidden_states, pooling_metadata)

    pooled_data = list[torch.Tensor]()
    if isinstance(hidden_states, list):
        for req_state, prompt_len in zip(hidden_states, prompt_lens):
            assert prompt_len == req_state.shape[0], \
                "partial prefill not supported with classifier"
        pooled_data = hidden_states
    else:
        offset = 0
        for prompt_len in prompt_lens:
            pooled_data_i = hidden_states[offset:offset + prompt_len]
            offset += prompt_len
            pooled_data.append(pooled_data_i)

    pooled_data_lst = []
    for pooled_data_i in pooled_data:

        if self.pooler is not None:
            final_shape_tensor = self.pooler(pooled_data_i)
        else:
            final_shape_tensor = self.classifier(pooled_data_i)

        pooled_data_lst.append(final_shape_tensor)

    pooled_output = torch.stack(pooled_data_lst)

    if self.pooler is not None:
        # apply classifier once on the full batch if possible
        pooled_output = self.classifier(pooled_output)

    # shape: (batch_size, num_labels)
    scores = self.default_activation_function(pooled_output)

    pooled_outputs = [PoolingSequenceGroupOutput(data) for data in scores]
    return PoolerOutput(outputs=pooled_outputs)

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 
401
402
403
404
405
406
407
408
409
410
def get_prompt_lens(
    self,
    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

LastPool

Bases: SimplePooler

Source code in vllm/model_executor/layers/pooler.py 
134
135
136
137
138
139
140
141
142
143
144
145
146
147
class LastPool(SimplePooler):

    def extract_states(
        self,
        hidden_states: Union[torch.Tensor, list[torch.Tensor]],
        pooling_metadata: PoolingMetadata,
    ) -> Union[list[torch.Tensor], torch.Tensor]:
        if isinstance(hidden_states, list):
            return [h[-1] for h in hidden_states]

        prompt_lens = self.get_prompt_lens(hidden_states, pooling_metadata)

        last_token_flat_indices = torch.cumsum(prompt_lens, dim=0) - 1
        return hidden_states[last_token_flat_indices]

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 
136
137
138
139
140
141
142
143
144
145
146
147
def extract_states(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[list[torch.Tensor], torch.Tensor]:
    if isinstance(hidden_states, list):
        return [h[-1] for h in hidden_states]

    prompt_lens = self.get_prompt_lens(hidden_states, pooling_metadata)

    last_token_flat_indices = torch.cumsum(prompt_lens, dim=0) - 1
    return hidden_states[last_token_flat_indices]

MeanPool

Bases: SimplePooler

Source code in vllm/model_executor/layers/pooler.py 
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
class MeanPool(SimplePooler):

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

        if isinstance(hidden_states, list):
            result = []
            for req_state, prompt_len in zip(hidden_states, prompt_lens):
                assert prompt_len == req_state.shape[0], \
                    "partial prefill not supported with mean pooling"
                result.append(torch.mean(req_state, dim=0,
                                         dtype=torch.float32))
            return result

        # 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)

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 
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
def extract_states(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[list[torch.Tensor], torch.Tensor]:
    prompt_lens = self.get_prompt_lens(hidden_states, pooling_metadata)

    if isinstance(hidden_states, list):
        result = []
        for req_state, prompt_len in zip(hidden_states, prompt_lens):
            assert prompt_len == req_state.shape[0], \
                "partial prefill not supported with mean pooling"
            result.append(torch.mean(req_state, dim=0,
                                     dtype=torch.float32))
        return result

    # 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)

Pooler

Bases: Module

Source code in vllm/model_executor/layers/pooler.py 
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
class Pooler(nn.Module):

    @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,
    ) -> SimplePooler:
        return SimplePooler.from_pooling_type(
            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,
        )

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,
) -> SimplePooler
Source code in vllm/model_executor/layers/pooler.py 
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
@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,
) -> SimplePooler:
    return SimplePooler.from_pooling_type(
        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,
    )

PoolerHead

Bases: Module

Source code in vllm/model_executor/layers/pooler.py 
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
class PoolerHead(nn.Module):

    def __init__(self, *, normalize: bool, softmax: bool) -> None:
        super().__init__()

        self.normalize = normalize
        self.softmax = softmax

    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)
                ]

        if self.normalize:
            if isinstance(pooled_data, list):
                pooled_data = [
                    F.normalize(data, p=2, dim=-1) for data in pooled_data
                ]
            else:
                pooled_data = F.normalize(pooled_data, p=2, dim=-1)

        if self.softmax:
            if isinstance(pooled_data, list):
                pooled_data = [
                    F.softmax(data, dim=-1)
                    if data.shape[-1] >= 2 else F.sigmoid(data)
                    for data in pooled_data
                ]
            else:
                if pooled_data.shape[-1] >= 2:
                    pooled_data = F.softmax(pooled_data, dim=-1)
                else:
                    pooled_data = F.sigmoid(pooled_data)

        # 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
        return pooled_data

normalize instance-attribute 

normalize = normalize

softmax instance-attribute 

softmax = softmax

__init__ 

__init__(*, normalize: bool, softmax: bool) -> None
Source code in vllm/model_executor/layers/pooler.py 
272
273
274
275
276
def __init__(self, *, normalize: bool, softmax: bool) -> None:
    super().__init__()

    self.normalize = normalize
    self.softmax = softmax

forward 

forward(
    pooled_data: Union[list[Tensor], Tensor],
    pooling_metadata: PoolingMetadata,
)
Source code in vllm/model_executor/layers/pooler.py 
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
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)
            ]

    if self.normalize:
        if isinstance(pooled_data, list):
            pooled_data = [
                F.normalize(data, p=2, dim=-1) for data in pooled_data
            ]
        else:
            pooled_data = F.normalize(pooled_data, p=2, dim=-1)

    if self.softmax:
        if isinstance(pooled_data, list):
            pooled_data = [
                F.softmax(data, dim=-1)
                if data.shape[-1] >= 2 else F.sigmoid(data)
                for data in pooled_data
            ]
        else:
            if pooled_data.shape[-1] >= 2:
                pooled_data = F.softmax(pooled_data, dim=-1)
            else:
                pooled_data = F.sigmoid(pooled_data)

    # 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
    return pooled_data

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

SimplePooler

Bases: Module

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 
 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
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
class SimplePooler(nn.Module):
    """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.
    """

    @staticmethod
    def from_pooling_type(
        pooling_type: PoolingType,
        *,
        normalize: bool,
        softmax: bool,
        step_tag_id: Optional[int] = None,
        returned_token_ids: Optional[list[int]] = None,
    ) -> "SimplePooler":
        if pooling_type == PoolingType.LAST:
            assert step_tag_id is None and returned_token_ids is None
            return LastPool(normalize=normalize, softmax=softmax)
        if pooling_type == PoolingType.ALL:
            assert step_tag_id is None and returned_token_ids is None
            return AllPool(normalize=normalize, softmax=softmax)
        if pooling_type == PoolingType.CLS:
            assert step_tag_id is None and returned_token_ids is None
            return CLSPool(normalize=normalize, softmax=softmax)
        if pooling_type == PoolingType.MEAN:
            assert step_tag_id is None and returned_token_ids is None
            return MeanPool(normalize=normalize, softmax=softmax)
        if pooling_type == PoolingType.STEP:
            return StepPool(normalize=normalize,
                            softmax=softmax,
                            step_tag_id=step_tag_id,
                            returned_token_ids=returned_token_ids)

        assert_never(pooling_type)

    def __init__(self, *, normalize: bool, softmax: bool) -> None:
        super().__init__()

        self.head = PoolerHead(normalize=normalize, softmax=softmax)

    def get_prompt_lens(
        self,
        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

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

    def build_output(self, data: torch.Tensor) -> PoolingSequenceGroupOutput:
        return PoolingSequenceGroupOutput(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)
        pooled_outputs = [self.build_output(data) for data in pooled_data]
        return PoolerOutput(outputs=pooled_outputs)

head instance-attribute 

head = PoolerHead(normalize=normalize, softmax=softmax)

__init__ 

__init__(*, normalize: bool, softmax: bool) -> None
Source code in vllm/model_executor/layers/pooler.py 
75
76
77
78
def __init__(self, *, normalize: bool, softmax: bool) -> None:
    super().__init__()

    self.head = PoolerHead(normalize=normalize, softmax=softmax)

build_output 

build_output(data: Tensor) -> PoolingSequenceGroupOutput
Source code in vllm/model_executor/layers/pooler.py 
98
99
def build_output(self, data: torch.Tensor) -> PoolingSequenceGroupOutput:
    return PoolingSequenceGroupOutput(data)

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 
91
92
93
94
95
96
def extract_states(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[list[torch.Tensor], torch.Tensor]:
    raise NotImplementedError

forward 

forward(
    hidden_states: Union[Tensor, list[Tensor]],
    pooling_metadata: PoolingMetadata,
) -> PoolerOutput
Source code in vllm/model_executor/layers/pooler.py 
101
102
103
104
105
106
107
108
109
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)
    pooled_outputs = [self.build_output(data) for data in pooled_data]
    return PoolerOutput(outputs=pooled_outputs)

from_pooling_type staticmethod 

from_pooling_type(
    pooling_type: PoolingType,
    *,
    normalize: bool,
    softmax: bool,
    step_tag_id: Optional[int] = None,
    returned_token_ids: Optional[list[int]] = None,
) -> SimplePooler
Source code in vllm/model_executor/layers/pooler.py 
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
@staticmethod
def from_pooling_type(
    pooling_type: PoolingType,
    *,
    normalize: bool,
    softmax: bool,
    step_tag_id: Optional[int] = None,
    returned_token_ids: Optional[list[int]] = None,
) -> "SimplePooler":
    if pooling_type == PoolingType.LAST:
        assert step_tag_id is None and returned_token_ids is None
        return LastPool(normalize=normalize, softmax=softmax)
    if pooling_type == PoolingType.ALL:
        assert step_tag_id is None and returned_token_ids is None
        return AllPool(normalize=normalize, softmax=softmax)
    if pooling_type == PoolingType.CLS:
        assert step_tag_id is None and returned_token_ids is None
        return CLSPool(normalize=normalize, softmax=softmax)
    if pooling_type == PoolingType.MEAN:
        assert step_tag_id is None and returned_token_ids is None
        return MeanPool(normalize=normalize, softmax=softmax)
    if pooling_type == PoolingType.STEP:
        return StepPool(normalize=normalize,
                        softmax=softmax,
                        step_tag_id=step_tag_id,
                        returned_token_ids=returned_token_ids)

    assert_never(pooling_type)

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 
80
81
82
83
84
85
86
87
88
89
def get_prompt_lens(
    self,
    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

StepPool

Bases: SimplePooler

Source code in vllm/model_executor/layers/pooler.py 
205
206
207
208
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
234
235
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
264
265
266
267
class StepPool(SimplePooler):

    def __init__(
        self,
        *,
        normalize: bool,
        softmax: bool,
        step_tag_id: Optional[int] = None,
        returned_token_ids: Optional[list[int]] = None,
    ):
        super().__init__(normalize=normalize, softmax=softmax)

        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]:
        prompt_lens = self.get_prompt_lens(hidden_states, pooling_metadata)
        prompt_token_ids = self.get_prompt_token_ids(pooling_metadata)

        pooled_data_lst = list[torch.Tensor]()
        if isinstance(hidden_states, list):
            for req_state, prompt_len in zip(hidden_states, prompt_lens):
                assert prompt_len == req_state.shape[0], \
                    "partial prefill not supported with step pooling"
            pooled_data_lst = hidden_states
        else:
            offset = 0
            for prompt_len in prompt_lens:
                pooled_data_i = hidden_states[offset:offset + prompt_len]
                offset += prompt_len
                pooled_data_lst.append(pooled_data_i)

        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

returned_token_ids instance-attribute 

returned_token_ids = returned_token_ids

step_tag_id instance-attribute 

step_tag_id = step_tag_id

__init__ 

__init__(
    *,
    normalize: bool,
    softmax: bool,
    step_tag_id: Optional[int] = None,
    returned_token_ids: Optional[list[int]] = None,
)
Source code in vllm/model_executor/layers/pooler.py 
207
208
209
210
211
212
213
214
215
216
217
218
def __init__(
    self,
    *,
    normalize: bool,
    softmax: bool,
    step_tag_id: Optional[int] = None,
    returned_token_ids: Optional[list[int]] = None,
):
    super().__init__(normalize=normalize, softmax=softmax)

    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 
234
235
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
264
265
266
267
def extract_states(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[list[torch.Tensor], torch.Tensor]:
    prompt_lens = self.get_prompt_lens(hidden_states, pooling_metadata)
    prompt_token_ids = self.get_prompt_token_ids(pooling_metadata)

    pooled_data_lst = list[torch.Tensor]()
    if isinstance(hidden_states, list):
        for req_state, prompt_len in zip(hidden_states, prompt_lens):
            assert prompt_len == req_state.shape[0], \
                "partial prefill not supported with step pooling"
        pooled_data_lst = hidden_states
    else:
        offset = 0
        for prompt_len in prompt_lens:
            pooled_data_i = hidden_states[offset:offset + prompt_len]
            offset += prompt_len
            pooled_data_lst.append(pooled_data_i)

    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

get_prompt_token_ids 

get_prompt_token_ids(
    pooling_metadata: PoolingMetadata,
) -> list[Tensor]
Source code in vllm/model_executor/layers/pooler.py 
220
221
222
223
224
225
226
227
228
229
230
231
232
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()
    ]