Skip to content

vllm.model_executor.models.bert

BertAttention

Bases: Module

Source code in vllm/model_executor/models/bert.py 
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
class BertAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_attention_heads: int,
        layer_norm_eps: float,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()

        self.self = BertSelfAttention(hidden_size=hidden_size,
                                      num_attention_heads=num_attention_heads,
                                      cache_config=cache_config,
                                      quant_config=quant_config,
                                      prefix=f"{prefix}.output")

        self.output = BertSelfOutput(hidden_size=hidden_size,
                                     layer_norm_eps=layer_norm_eps,
                                     quant_config=quant_config,
                                     prefix=f"{prefix}.output")

    def forward(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        self_output = self.self(hidden_states)
        return self.output(self_output, hidden_states)

output instance-attribute 

output = BertSelfOutput(
    hidden_size=hidden_size,
    layer_norm_eps=layer_norm_eps,
    quant_config=quant_config,
    prefix=f"{prefix}.output",
)

self instance-attribute 

self = BertSelfAttention(
    hidden_size=hidden_size,
    num_attention_heads=num_attention_heads,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.output",
)

__init__ 

__init__(
    hidden_size: int,
    num_attention_heads: int,
    layer_norm_eps: float,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert.py 
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
def __init__(
    self,
    hidden_size: int,
    num_attention_heads: int,
    layer_norm_eps: float,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()

    self.self = BertSelfAttention(hidden_size=hidden_size,
                                  num_attention_heads=num_attention_heads,
                                  cache_config=cache_config,
                                  quant_config=quant_config,
                                  prefix=f"{prefix}.output")

    self.output = BertSelfOutput(hidden_size=hidden_size,
                                 layer_norm_eps=layer_norm_eps,
                                 quant_config=quant_config,
                                 prefix=f"{prefix}.output")

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/bert.py 
208
209
210
211
212
213
def forward(
    self,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    self_output = self.self(hidden_states)
    return self.output(self_output, hidden_states)

BertEmbedding

Bases: Module

Source code in vllm/model_executor/models/bert.py 
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
class BertEmbedding(nn.Module):

    def __init__(self, config: BertConfig):

        super().__init__()
        self.size = config.hidden_size
        self.word_embeddings = VocabParallelEmbedding(config.vocab_size,
                                                      config.hidden_size)
        self.position_embeddings = VocabParallelEmbedding(
            config.max_position_embeddings, config.hidden_size)
        self.token_type_embeddings = VocabParallelEmbedding(
            config.type_vocab_size, config.hidden_size)
        self.LayerNorm = nn.LayerNorm(config.hidden_size,
                                      eps=config.layer_norm_eps)

        self.register_buffer(
            "position_ids",
            torch.arange(config.max_position_embeddings).unsqueeze(0),
        )
        self.position_embedding_type = config.position_embedding_type
        if self.position_embedding_type != "absolute":
            raise ValueError("Only 'absolute' position_embedding_type" +
                             " is supported")

    def forward(
        self,
        input_ids: torch.Tensor,
        seq_lens: torch.Tensor,
        position_ids: torch.Tensor,
        token_type_ids: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        input_shape = input_ids.size()

        # Input embeddings.
        inputs_embeds = self.word_embeddings(input_ids)

        # Position embeddings.
        position_embeddings = self.position_embeddings(position_ids)

        if token_type_ids is None:
            token_type_ids = torch.zeros(input_shape,
                                         dtype=torch.long,
                                         device=inputs_embeds.device)

        token_type_embeddings = self.token_type_embeddings(token_type_ids)

        embeddings = inputs_embeds + token_type_embeddings + position_embeddings
        embeddings = self.LayerNorm(embeddings)
        return embeddings

LayerNorm instance-attribute 

LayerNorm = LayerNorm(hidden_size, eps=layer_norm_eps)

position_embedding_type instance-attribute 

position_embedding_type = position_embedding_type

position_embeddings instance-attribute 

position_embeddings = VocabParallelEmbedding(
    max_position_embeddings, hidden_size
)

size instance-attribute 

size = hidden_size

token_type_embeddings instance-attribute 

token_type_embeddings = VocabParallelEmbedding(
    type_vocab_size, hidden_size
)

word_embeddings instance-attribute 

word_embeddings = VocabParallelEmbedding(
    vocab_size, hidden_size
)

__init__ 

__init__(config: BertConfig)
Source code in vllm/model_executor/models/bert.py 
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
def __init__(self, config: BertConfig):

    super().__init__()
    self.size = config.hidden_size
    self.word_embeddings = VocabParallelEmbedding(config.vocab_size,
                                                  config.hidden_size)
    self.position_embeddings = VocabParallelEmbedding(
        config.max_position_embeddings, config.hidden_size)
    self.token_type_embeddings = VocabParallelEmbedding(
        config.type_vocab_size, config.hidden_size)
    self.LayerNorm = nn.LayerNorm(config.hidden_size,
                                  eps=config.layer_norm_eps)

    self.register_buffer(
        "position_ids",
        torch.arange(config.max_position_embeddings).unsqueeze(0),
    )
    self.position_embedding_type = config.position_embedding_type
    if self.position_embedding_type != "absolute":
        raise ValueError("Only 'absolute' position_embedding_type" +
                         " is supported")

forward 

forward(
    input_ids: Tensor,
    seq_lens: Tensor,
    position_ids: Tensor,
    token_type_ids: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/bert.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    seq_lens: torch.Tensor,
    position_ids: torch.Tensor,
    token_type_ids: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    input_shape = input_ids.size()

    # Input embeddings.
    inputs_embeds = self.word_embeddings(input_ids)

    # Position embeddings.
    position_embeddings = self.position_embeddings(position_ids)

    if token_type_ids is None:
        token_type_ids = torch.zeros(input_shape,
                                     dtype=torch.long,
                                     device=inputs_embeds.device)

    token_type_embeddings = self.token_type_embeddings(token_type_ids)

    embeddings = inputs_embeds + token_type_embeddings + position_embeddings
    embeddings = self.LayerNorm(embeddings)
    return embeddings

BertEmbeddingModel

Bases: Module, SupportsV0Only, SupportsQuant

A model that uses Bert to provide embedding functionalities.

This class encapsulates the BertModel and provides an interface for embedding operations and customized pooling functions.

Attributes:

Name Type Description
model

An instance of BertModel used for forward operations.
_pooler

An instance of Pooler used for pooling operations.
Source code in vllm/model_executor/models/bert.py 
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
482
483
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
class BertEmbeddingModel(nn.Module, SupportsV0Only, SupportsQuant):
    """A model that uses Bert to provide embedding functionalities.

    This class encapsulates the BertModel and provides an interface for
    embedding operations and customized pooling functions.

    Attributes:
        model: An instance of BertModel used for forward operations.
        _pooler: An instance of Pooler used for pooling operations.
    """

    is_pooling_model = True

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        pooler_config = vllm_config.model_config.pooler_config
        assert pooler_config is not None

        self.model = self._build_model(vllm_config=vllm_config,
                                       prefix=maybe_prefix(prefix, "model"))
        self.pooler = self._build_pooler(pooler_config)

    def forward(
        self,
        input_ids: Optional[torch.Tensor],
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        return self.model(input_ids=input_ids,
                          position_ids=positions,
                          inputs_embeds=inputs_embeds,
                          intermediate_tensors=intermediate_tensors)

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
        weights_list = list(weights)

        has_model_prefix = any(
            name.startswith("model.") for name, _ in weights_list)
        if not has_model_prefix:
            mapper = WeightsMapper(orig_to_new_prefix={"": "model."})

        loader = AutoWeightsLoader(self, skip_prefixes=["lm_head."])
        return loader.load_weights(weights_list, mapper=mapper)

    def _build_model(self,
                     vllm_config: VllmConfig,
                     prefix: str = "") -> BertModel:
        return BertModel(vllm_config=vllm_config,
                         prefix=prefix,
                         embedding_class=BertEmbedding)

    def _build_pooler(self, pooler_config: PoolerConfig) -> Pooler:
        return DispatchPooler({
            "encode":
            Pooler.for_encode(pooler_config),
            "embed":
            Pooler.for_embed(
                pooler_config,
                default_pooling_type=PoolingType.CLS,
            ),
        })

is_pooling_model class-attribute instance-attribute 

is_pooling_model = True

model instance-attribute 

model = _build_model(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "model"),
)

pooler instance-attribute 

pooler = _build_pooler(pooler_config)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/bert.py 
463
464
465
466
467
468
469
470
471
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()

    pooler_config = vllm_config.model_config.pooler_config
    assert pooler_config is not None

    self.model = self._build_model(vllm_config=vllm_config,
                                   prefix=maybe_prefix(prefix, "model"))
    self.pooler = self._build_pooler(pooler_config)

_build_model 

_build_model(
    vllm_config: VllmConfig, prefix: str = ""
) -> BertModel
Source code in vllm/model_executor/models/bert.py 
496
497
498
499
500
501
def _build_model(self,
                 vllm_config: VllmConfig,
                 prefix: str = "") -> BertModel:
    return BertModel(vllm_config=vllm_config,
                     prefix=prefix,
                     embedding_class=BertEmbedding)

_build_pooler 

_build_pooler(pooler_config: PoolerConfig) -> Pooler
Source code in vllm/model_executor/models/bert.py 
503
504
505
506
507
508
509
510
511
512
def _build_pooler(self, pooler_config: PoolerConfig) -> Pooler:
    return DispatchPooler({
        "encode":
        Pooler.for_encode(pooler_config),
        "embed":
        Pooler.for_embed(
            pooler_config,
            default_pooling_type=PoolingType.CLS,
        ),
    })

forward 

forward(
    input_ids: Optional[Tensor],
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/bert.py 
473
474
475
476
477
478
479
480
481
482
483
def forward(
    self,
    input_ids: Optional[torch.Tensor],
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    return self.model(input_ids=input_ids,
                      position_ids=positions,
                      inputs_embeds=inputs_embeds,
                      intermediate_tensors=intermediate_tensors)

load_weights 

load_weights(weights: Iterable[tuple[str, Tensor]])
Source code in vllm/model_executor/models/bert.py 
485
486
487
488
489
490
491
492
493
494
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
    weights_list = list(weights)

    has_model_prefix = any(
        name.startswith("model.") for name, _ in weights_list)
    if not has_model_prefix:
        mapper = WeightsMapper(orig_to_new_prefix={"": "model."})

    loader = AutoWeightsLoader(self, skip_prefixes=["lm_head."])
    return loader.load_weights(weights_list, mapper=mapper)

BertEncoder

Bases: Module

Source code in vllm/model_executor/models/bert.py 
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
@support_torch_compile
class BertEncoder(nn.Module):

    def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        self.layer = nn.ModuleList([
            BertLayer(config=config,
                      cache_config=cache_config,
                      quant_config=quant_config,
                      prefix=f"{prefix}.layer.{layer_idx}")
            for layer_idx in range(config.num_hidden_layers)
        ])

    def forward(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        for layer in self.layer:
            hidden_states = layer(hidden_states)
        return hidden_states

layer instance-attribute 

layer = ModuleList(
    [
        (
            BertLayer(
                config=config,
                cache_config=cache_config,
                quant_config=quant_config,
                prefix=f"{prefix}.layer.{layer_idx}",
            )
        )
        for layer_idx in (range(num_hidden_layers))
    ]
)

__init__ 

__init__(vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/bert.py 
125
126
127
128
129
130
131
132
133
134
135
136
def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    cache_config = vllm_config.cache_config
    quant_config = vllm_config.quant_config
    self.layer = nn.ModuleList([
        BertLayer(config=config,
                  cache_config=cache_config,
                  quant_config=quant_config,
                  prefix=f"{prefix}.layer.{layer_idx}")
        for layer_idx in range(config.num_hidden_layers)
    ])

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/bert.py 
138
139
140
141
142
143
144
def forward(
    self,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    for layer in self.layer:
        hidden_states = layer(hidden_states)
    return hidden_states

BertForSequenceClassification

Bases: Module, SupportsV0Only, SupportsCrossEncoding, SupportsQuant

A model that uses Bert to provide embedding functionalities.

This class encapsulates the BertModel and provides an interface for embedding operations and customized pooling functions.

Attributes:

Name Type Description
model

An instance of BertModel used for forward operations.
_pooler

An instance of Pooler used for pooling operations.
Source code in vllm/model_executor/models/bert.py 
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
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
class BertForSequenceClassification(nn.Module, SupportsV0Only,
                                    SupportsCrossEncoding, SupportsQuant):
    """A model that uses Bert to provide embedding functionalities.

   This class encapsulates the BertModel and provides an interface for
   embedding operations and customized pooling functions.

   Attributes:
       model: An instance of BertModel used for forward operations.
       _pooler: An instance of Pooler used for pooling operations.
   """

    is_pooling_model = True

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config

        self.num_labels = config.num_labels
        self.bert = BertPoolingModel(vllm_config=vllm_config,
                                     prefix=maybe_prefix(prefix, "bert"),
                                     embedding_class=BertEmbedding)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

        pooler_config = vllm_config.model_config.pooler_config
        assert pooler_config is not None

        self.pooler = DispatchPooler({
            "encode":
            Pooler.for_encode(pooler_config),
            "classify":
            ClassifierPooler(
                pooling=self.bert.pooler,
                classifier=self.classifier,
                act_fn=ClassifierPooler.act_fn_for_seq_cls(
                    vllm_config.model_config),
            ),
            "score":
            ClassifierPooler(
                pooling=self.bert.pooler,
                classifier=self.classifier,
                act_fn=ClassifierPooler.act_fn_for_cross_encoder(
                    vllm_config.model_config),
            ),
        })

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
        loader = AutoWeightsLoader(self)
        loaded_params = loader.load_weights(weights)
        return loaded_params

    def forward(
        self,
        input_ids: Optional[torch.Tensor],
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        token_type_ids: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        return self.bert(input_ids=input_ids,
                         position_ids=positions,
                         inputs_embeds=inputs_embeds,
                         intermediate_tensors=intermediate_tensors,
                         token_type_ids=token_type_ids)

bert instance-attribute 

bert = BertPoolingModel(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "bert"),
    embedding_class=BertEmbedding,
)

classifier instance-attribute 

classifier = Linear(hidden_size, num_labels)

is_pooling_model class-attribute instance-attribute 

is_pooling_model = True

num_labels instance-attribute 

num_labels = num_labels

pooler instance-attribute 

pooler = DispatchPooler(
    {
        "encode": for_encode(pooler_config),
        "classify": ClassifierPooler(
            pooling=pooler,
            classifier=classifier,
            act_fn=act_fn_for_seq_cls(model_config),
        ),
        "score": ClassifierPooler(
            pooling=pooler,
            classifier=classifier,
            act_fn=act_fn_for_cross_encoder(model_config),
        ),
    }
)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/bert.py 
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
554
555
556
557
558
559
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config

    self.num_labels = config.num_labels
    self.bert = BertPoolingModel(vllm_config=vllm_config,
                                 prefix=maybe_prefix(prefix, "bert"),
                                 embedding_class=BertEmbedding)
    self.classifier = nn.Linear(config.hidden_size, config.num_labels)

    pooler_config = vllm_config.model_config.pooler_config
    assert pooler_config is not None

    self.pooler = DispatchPooler({
        "encode":
        Pooler.for_encode(pooler_config),
        "classify":
        ClassifierPooler(
            pooling=self.bert.pooler,
            classifier=self.classifier,
            act_fn=ClassifierPooler.act_fn_for_seq_cls(
                vllm_config.model_config),
        ),
        "score":
        ClassifierPooler(
            pooling=self.bert.pooler,
            classifier=self.classifier,
            act_fn=ClassifierPooler.act_fn_for_cross_encoder(
                vllm_config.model_config),
        ),
    })

forward 

forward(
    input_ids: Optional[Tensor],
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    token_type_ids: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/bert.py 
566
567
568
569
570
571
572
573
574
575
576
577
578
def forward(
    self,
    input_ids: Optional[torch.Tensor],
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
    token_type_ids: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    return self.bert(input_ids=input_ids,
                     position_ids=positions,
                     inputs_embeds=inputs_embeds,
                     intermediate_tensors=intermediate_tensors,
                     token_type_ids=token_type_ids)

load_weights 

load_weights(weights: Iterable[tuple[str, Tensor]])
Source code in vllm/model_executor/models/bert.py 
561
562
563
564
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
    loader = AutoWeightsLoader(self)
    loaded_params = loader.load_weights(weights)
    return loaded_params

BertIntermediate

Bases: Module

Source code in vllm/model_executor/models/bert.py 
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
class BertIntermediate(nn.Module):

    def __init__(self,
                 hidden_size: int,
                 intermediate_size: int,
                 hidden_act: str,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.dense = ColumnParallelLinear(input_size=hidden_size,
                                          output_size=intermediate_size,
                                          bias=True,
                                          quant_config=quant_config,
                                          prefix=f"{prefix}.dense")
        self.intermediate_act_fn = get_act_fn(hidden_act)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.dense(hidden_states)
        hidden_states = self.intermediate_act_fn(hidden_states)
        return hidden_states

dense instance-attribute 

dense = ColumnParallelLinear(
    input_size=hidden_size,
    output_size=intermediate_size,
    bias=True,
    quant_config=quant_config,
    prefix=f"{prefix}.dense",
)

intermediate_act_fn instance-attribute 

intermediate_act_fn = get_act_fn(hidden_act)

__init__ 

__init__(
    hidden_size: int,
    intermediate_size: int,
    hidden_act: str,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert.py 
295
296
297
298
299
300
301
302
303
304
305
306
307
def __init__(self,
             hidden_size: int,
             intermediate_size: int,
             hidden_act: str,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()
    self.dense = ColumnParallelLinear(input_size=hidden_size,
                                      output_size=intermediate_size,
                                      bias=True,
                                      quant_config=quant_config,
                                      prefix=f"{prefix}.dense")
    self.intermediate_act_fn = get_act_fn(hidden_act)

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/bert.py 
309
310
311
312
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    hidden_states, _ = self.dense(hidden_states)
    hidden_states = self.intermediate_act_fn(hidden_states)
    return hidden_states

BertLayer

Bases: Module

Source code in vllm/model_executor/models/bert.py 
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
class BertLayer(nn.Module):

    def __init__(self,
                 config: BertConfig,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()

        self.attention = BertAttention(
            hidden_size=config.hidden_size,
            num_attention_heads=config.num_attention_heads,
            layer_norm_eps=config.layer_norm_eps,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attention")

        self.intermediate = BertIntermediate(
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            prefix=f"{prefix}.intermediate")

        self.output = BertOutput(hidden_size=config.hidden_size,
                                 intermediate_size=config.intermediate_size,
                                 layer_norm_eps=config.layer_norm_eps,
                                 quant_config=quant_config,
                                 prefix=f"{prefix}.output")

    def forward(self, hidden_states: torch.Tensor):
        attn_output = self.attention(hidden_states)
        intermediate_output = self.intermediate(attn_output)
        output = self.output(intermediate_output, attn_output)
        return output

attention instance-attribute 

attention = BertAttention(
    hidden_size=hidden_size,
    num_attention_heads=num_attention_heads,
    layer_norm_eps=layer_norm_eps,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.attention",
)

intermediate instance-attribute 

intermediate = BertIntermediate(
    hidden_size=hidden_size,
    intermediate_size=intermediate_size,
    hidden_act=hidden_act,
    quant_config=quant_config,
    prefix=f"{prefix}.intermediate",
)

output instance-attribute 

output = BertOutput(
    hidden_size=hidden_size,
    intermediate_size=intermediate_size,
    layer_norm_eps=layer_norm_eps,
    quant_config=quant_config,
    prefix=f"{prefix}.output",
)

__init__ 

__init__(
    config: BertConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert.py 
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
def __init__(self,
             config: BertConfig,
             cache_config: Optional[CacheConfig] = None,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()

    self.attention = BertAttention(
        hidden_size=config.hidden_size,
        num_attention_heads=config.num_attention_heads,
        layer_norm_eps=config.layer_norm_eps,
        cache_config=cache_config,
        quant_config=quant_config,
        prefix=f"{prefix}.attention")

    self.intermediate = BertIntermediate(
        hidden_size=config.hidden_size,
        intermediate_size=config.intermediate_size,
        hidden_act=config.hidden_act,
        quant_config=quant_config,
        prefix=f"{prefix}.intermediate")

    self.output = BertOutput(hidden_size=config.hidden_size,
                             intermediate_size=config.intermediate_size,
                             layer_norm_eps=config.layer_norm_eps,
                             quant_config=quant_config,
                             prefix=f"{prefix}.output")

forward 

forward(hidden_states: Tensor)
Source code in vllm/model_executor/models/bert.py 
177
178
179
180
181
def forward(self, hidden_states: torch.Tensor):
    attn_output = self.attention(hidden_states)
    intermediate_output = self.intermediate(attn_output)
    output = self.output(intermediate_output, attn_output)
    return output

BertModel

Bases: Module, SupportsQuant

Source code in vllm/model_executor/models/bert.py 
340
341
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
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
class BertModel(nn.Module, SupportsQuant):

    is_pooling_model = True

    packed_modules_mapping = {"qkv_proj": ["query", "key", "value"]}

    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        embedding_class: type[nn.Module] = BertEmbedding,
    ) -> None:
        super().__init__()

        config = vllm_config.model_config.hf_config
        self.embeddings = embedding_class(config)
        self.encoder = BertEncoder(vllm_config=vllm_config,
                                   prefix=f"{prefix}.encoder")

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        token_type_ids: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        if inputs_embeds is not None:
            hidden_states = inputs_embeds
        else:
            attn_metadata = get_forward_context().attn_metadata
            assert hasattr(attn_metadata, "seq_lens_tensor")
            hidden_states = self.embeddings(
                input_ids=input_ids,
                seq_lens=attn_metadata.seq_lens_tensor,
                position_ids=position_ids,
                token_type_ids=token_type_ids)
        return self.encoder(hidden_states)

    def _load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "query", "q"),
            ("qkv_proj", "key", "k"),
            ("qkv_proj", "value", "v"),
        ]

        loaded_stacked_params = []
        other_weights = []
        params_dict = dict(self.named_parameters())
        for name, loaded_weight in weights:
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue

                name = name.replace(weight_name, param_name)
                if name not in params_dict:
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                loaded_stacked_params.append(name)
                break
            else:
                if name in params_dict:
                    other_weights.append((name, loaded_weight))

        return other_weights, loaded_stacked_params

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        other_weights, loaded_stacked_params = self._load_weights(weights)

        loader = AutoWeightsLoader(self, skip_prefixes=["pooler."])
        loaded_params = loader.load_weights(other_weights)
        loaded_params.update(loaded_stacked_params)
        return loaded_params

embeddings instance-attribute 

embeddings = embedding_class(config)

encoder instance-attribute 

encoder = BertEncoder(
    vllm_config=vllm_config, prefix=f"{prefix}.encoder"
)

is_pooling_model class-attribute instance-attribute 

is_pooling_model = True

packed_modules_mapping class-attribute instance-attribute 

packed_modules_mapping = {
    "qkv_proj": ["query", "key", "value"]
}

__init__ 

__init__(
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    embedding_class: type[Module] = BertEmbedding,
) -> None
Source code in vllm/model_executor/models/bert.py 
346
347
348
349
350
351
352
353
354
355
356
357
358
def __init__(
    self,
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    embedding_class: type[nn.Module] = BertEmbedding,
) -> None:
    super().__init__()

    config = vllm_config.model_config.hf_config
    self.embeddings = embedding_class(config)
    self.encoder = BertEncoder(vllm_config=vllm_config,
                               prefix=f"{prefix}.encoder")

_load_weights 

_load_weights(weights: Iterable[tuple[str, Tensor]])
Source code in vllm/model_executor/models/bert.py 
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
def _load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
    stacked_params_mapping = [
        # (param_name, shard_name, shard_id)
        ("qkv_proj", "query", "q"),
        ("qkv_proj", "key", "k"),
        ("qkv_proj", "value", "v"),
    ]

    loaded_stacked_params = []
    other_weights = []
    params_dict = dict(self.named_parameters())
    for name, loaded_weight in weights:
        for (param_name, weight_name, shard_id) in stacked_params_mapping:
            if weight_name not in name:
                continue

            name = name.replace(weight_name, param_name)
            if name not in params_dict:
                continue
            param = params_dict[name]
            weight_loader = param.weight_loader
            weight_loader(param, loaded_weight, shard_id)
            loaded_stacked_params.append(name)
            break
        else:
            if name in params_dict:
                other_weights.append((name, loaded_weight))

    return other_weights, loaded_stacked_params

forward 

forward(
    input_ids: Tensor,
    position_ids: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    token_type_ids: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/bert.py 
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
def forward(
    self,
    input_ids: torch.Tensor,
    position_ids: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
    token_type_ids: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    if inputs_embeds is not None:
        hidden_states = inputs_embeds
    else:
        attn_metadata = get_forward_context().attn_metadata
        assert hasattr(attn_metadata, "seq_lens_tensor")
        hidden_states = self.embeddings(
            input_ids=input_ids,
            seq_lens=attn_metadata.seq_lens_tensor,
            position_ids=position_ids,
            token_type_ids=token_type_ids)
    return self.encoder(hidden_states)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/bert.py 
410
411
412
413
414
415
416
417
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    other_weights, loaded_stacked_params = self._load_weights(weights)

    loader = AutoWeightsLoader(self, skip_prefixes=["pooler."])
    loaded_params = loader.load_weights(other_weights)
    loaded_params.update(loaded_stacked_params)
    return loaded_params

BertOutput

Bases: Module

Source code in vllm/model_executor/models/bert.py 
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
class BertOutput(nn.Module):

    def __init__(self,
                 hidden_size: int,
                 intermediate_size: int,
                 layer_norm_eps: float,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()

        self.dense = RowParallelLinear(input_size=intermediate_size,
                                       output_size=hidden_size,
                                       bias=True,
                                       quant_config=quant_config,
                                       prefix=f"{prefix}.dense")

        self.LayerNorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)

    def forward(self, hidden_states: torch.Tensor,
                input_tensor: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.dense(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states

LayerNorm instance-attribute 

LayerNorm = LayerNorm(hidden_size, eps=layer_norm_eps)

dense instance-attribute 

dense = RowParallelLinear(
    input_size=intermediate_size,
    output_size=hidden_size,
    bias=True,
    quant_config=quant_config,
    prefix=f"{prefix}.dense",
)

__init__ 

__init__(
    hidden_size: int,
    intermediate_size: int,
    layer_norm_eps: float,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert.py 
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
def __init__(self,
             hidden_size: int,
             intermediate_size: int,
             layer_norm_eps: float,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()

    self.dense = RowParallelLinear(input_size=intermediate_size,
                                   output_size=hidden_size,
                                   bias=True,
                                   quant_config=quant_config,
                                   prefix=f"{prefix}.dense")

    self.LayerNorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)

forward 

forward(
    hidden_states: Tensor, input_tensor: Tensor
) -> Tensor
Source code in vllm/model_executor/models/bert.py 
333
334
335
336
337
def forward(self, hidden_states: torch.Tensor,
            input_tensor: torch.Tensor) -> torch.Tensor:
    hidden_states, _ = self.dense(hidden_states)
    hidden_states = self.LayerNorm(hidden_states + input_tensor)
    return hidden_states

BertPooler

Bases: Pooler

Source code in vllm/model_executor/models/bert.py 
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
class BertPooler(Pooler):

    def __init__(self, config: BertConfig):
        super().__init__()

        self.pooling = PoolingMethod.from_pooling_type(PoolingType.CLS)
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = nn.Tanh()

    def get_supported_tasks(self) -> Set[PoolingTask]:
        return self.pooling.get_supported_tasks()

    def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate:
        return self.pooling.get_pooling_updates(task)

    def _head(self, pooled_output: torch.Tensor):
        pooled_output = self.dense(pooled_output)
        pooled_output = self.activation(pooled_output)
        return pooled_output

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

        if isinstance(pooled_output, list):
            pooled_output = [self._head(output) for output in pooled_output]
        else:
            pooled_output = self._head(pooled_output)

        return pooled_output

activation instance-attribute 

activation = Tanh()

dense instance-attribute 

dense = Linear(hidden_size, hidden_size)

pooling instance-attribute 

pooling = from_pooling_type(CLS)

__init__ 

__init__(config: BertConfig)
Source code in vllm/model_executor/models/bert.py 
89
90
91
92
93
94
def __init__(self, config: BertConfig):
    super().__init__()

    self.pooling = PoolingMethod.from_pooling_type(PoolingType.CLS)
    self.dense = nn.Linear(config.hidden_size, config.hidden_size)
    self.activation = nn.Tanh()

_head 

_head(pooled_output: Tensor)
Source code in vllm/model_executor/models/bert.py 
102
103
104
105
def _head(self, pooled_output: torch.Tensor):
    pooled_output = self.dense(pooled_output)
    pooled_output = self.activation(pooled_output)
    return pooled_output

forward 

forward(
    hidden_states: Union[Tensor, list[Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[Tensor, list[Tensor]]
Source code in vllm/model_executor/models/bert.py 
107
108
109
110
111
112
113
114
115
116
117
118
119
def forward(
    self,
    hidden_states: Union[torch.Tensor, list[torch.Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[torch.Tensor, list[torch.Tensor]]:
    pooled_output = self.pooling(hidden_states, pooling_metadata)

    if isinstance(pooled_output, list):
        pooled_output = [self._head(output) for output in pooled_output]
    else:
        pooled_output = self._head(pooled_output)

    return pooled_output

get_pooling_updates 

get_pooling_updates(
    task: PoolingTask,
) -> PoolingParamsUpdate
Source code in vllm/model_executor/models/bert.py 
 99
100
def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate:
    return self.pooling.get_pooling_updates(task)

get_supported_tasks 

get_supported_tasks() -> Set[PoolingTask]
Source code in vllm/model_executor/models/bert.py 
96
97
def get_supported_tasks(self) -> Set[PoolingTask]:
    return self.pooling.get_supported_tasks()

BertPoolingModel

Bases: BertModel

Source code in vllm/model_executor/models/bert.py 
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
class BertPoolingModel(BertModel):

    is_pooling_model = True

    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        embedding_class: type[nn.Module] = BertEmbedding,
    ) -> None:
        super().__init__(
            vllm_config=vllm_config,
            prefix=prefix,
            embedding_class=embedding_class,
        )

        config = vllm_config.model_config.hf_config
        self.pooler = BertPooler(config)

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        other_weights, loaded_stacked_params = self._load_weights(weights)

        loader = AutoWeightsLoader(self)
        loaded_params = loader.load_weights(other_weights)
        loaded_params.update(loaded_stacked_params)
        return loaded_params

is_pooling_model class-attribute instance-attribute 

is_pooling_model = True

pooler instance-attribute 

pooler = BertPooler(config)

__init__ 

__init__(
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    embedding_class: type[Module] = BertEmbedding,
) -> None
Source code in vllm/model_executor/models/bert.py 
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
def __init__(
    self,
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    embedding_class: type[nn.Module] = BertEmbedding,
) -> None:
    super().__init__(
        vllm_config=vllm_config,
        prefix=prefix,
        embedding_class=embedding_class,
    )

    config = vllm_config.model_config.hf_config
    self.pooler = BertPooler(config)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/bert.py 
440
441
442
443
444
445
446
447
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    other_weights, loaded_stacked_params = self._load_weights(weights)

    loader = AutoWeightsLoader(self)
    loaded_params = loader.load_weights(other_weights)
    loaded_params.update(loaded_stacked_params)
    return loaded_params

BertSelfAttention

Bases: Module

Source code in vllm/model_executor/models/bert.py 
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
268
class BertSelfAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_attention_heads: int,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()

        self.total_num_heads = num_attention_heads
        assert self.total_num_heads % tp_size == 0

        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = self.total_num_heads
        self.head_dim = self.hidden_size // self.total_num_heads
        assert self.head_dim * self.total_num_heads == self.hidden_size

        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)

        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.qkv_proj = QKVParallelLinear(
            hidden_size=self.hidden_size,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=True,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj")

        self.attn = Attention(num_heads=self.num_heads,
                              head_size=self.head_dim,
                              scale=self.scaling,
                              num_kv_heads=self.num_kv_heads,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn",
                              attn_type=AttentionType.ENCODER_ONLY)

    def forward(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        output = self.attn(q, k, v)
        return output

attn instance-attribute 

attn = Attention(
    num_heads=num_heads,
    head_size=head_dim,
    scale=scaling,
    num_kv_heads=num_kv_heads,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.attn",
    attn_type=ENCODER_ONLY,
)

head_dim instance-attribute 

head_dim = hidden_size // total_num_heads

hidden_size instance-attribute 

hidden_size = hidden_size

kv_size instance-attribute 

kv_size = num_kv_heads * head_dim

num_heads instance-attribute 

num_heads = total_num_heads // tp_size

num_kv_heads instance-attribute 

num_kv_heads = max(1, total_num_kv_heads // tp_size)

q_size instance-attribute 

q_size = num_heads * head_dim

qkv_proj instance-attribute 

qkv_proj = QKVParallelLinear(
    hidden_size=hidden_size,
    head_size=head_dim,
    total_num_heads=total_num_heads,
    total_num_kv_heads=total_num_kv_heads,
    bias=True,
    quant_config=quant_config,
    prefix=f"{prefix}.qkv_proj",
)

scaling instance-attribute 

scaling = head_dim ** -0.5

total_num_heads instance-attribute 

total_num_heads = num_attention_heads

total_num_kv_heads instance-attribute 

total_num_kv_heads = total_num_heads

__init__ 

__init__(
    hidden_size: int,
    num_attention_heads: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert.py 
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
def __init__(
    self,
    hidden_size: int,
    num_attention_heads: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.hidden_size = hidden_size
    tp_size = get_tensor_model_parallel_world_size()

    self.total_num_heads = num_attention_heads
    assert self.total_num_heads % tp_size == 0

    self.num_heads = self.total_num_heads // tp_size
    self.total_num_kv_heads = self.total_num_heads
    self.head_dim = self.hidden_size // self.total_num_heads
    assert self.head_dim * self.total_num_heads == self.hidden_size

    self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)

    self.q_size = self.num_heads * self.head_dim
    self.kv_size = self.num_kv_heads * self.head_dim
    self.scaling = self.head_dim**-0.5
    self.qkv_proj = QKVParallelLinear(
        hidden_size=self.hidden_size,
        head_size=self.head_dim,
        total_num_heads=self.total_num_heads,
        total_num_kv_heads=self.total_num_kv_heads,
        bias=True,
        quant_config=quant_config,
        prefix=f"{prefix}.qkv_proj")

    self.attn = Attention(num_heads=self.num_heads,
                          head_size=self.head_dim,
                          scale=self.scaling,
                          num_kv_heads=self.num_kv_heads,
                          cache_config=cache_config,
                          quant_config=quant_config,
                          prefix=f"{prefix}.attn",
                          attn_type=AttentionType.ENCODER_ONLY)

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/bert.py 
261
262
263
264
265
266
267
268
def forward(
    self,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    qkv, _ = self.qkv_proj(hidden_states)
    q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
    output = self.attn(q, k, v)
    return output

BertSelfOutput

Bases: Module

Source code in vllm/model_executor/models/bert.py 
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
class BertSelfOutput(nn.Module):

    def __init__(self,
                 hidden_size: int,
                 layer_norm_eps: float,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.dense = RowParallelLinear(input_size=hidden_size,
                                       output_size=hidden_size,
                                       bias=True,
                                       quant_config=quant_config,
                                       prefix=f"{prefix}.dense")
        self.LayerNorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)

    def forward(self, hidden_states: torch.Tensor,
                input_tensor: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.dense(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states

LayerNorm instance-attribute 

LayerNorm = LayerNorm(hidden_size, eps=layer_norm_eps)

dense instance-attribute 

dense = RowParallelLinear(
    input_size=hidden_size,
    output_size=hidden_size,
    bias=True,
    quant_config=quant_config,
    prefix=f"{prefix}.dense",
)

__init__ 

__init__(
    hidden_size: int,
    layer_norm_eps: float,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert.py 
273
274
275
276
277
278
279
280
281
282
283
284
def __init__(self,
             hidden_size: int,
             layer_norm_eps: float,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()
    self.dense = RowParallelLinear(input_size=hidden_size,
                                   output_size=hidden_size,
                                   bias=True,
                                   quant_config=quant_config,
                                   prefix=f"{prefix}.dense")
    self.LayerNorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)

forward 

forward(
    hidden_states: Tensor, input_tensor: Tensor
) -> Tensor
Source code in vllm/model_executor/models/bert.py 
286
287
288
289
290
def forward(self, hidden_states: torch.Tensor,
            input_tensor: torch.Tensor) -> torch.Tensor:
    hidden_states, _ = self.dense(hidden_states)
    hidden_states = self.LayerNorm(hidden_states + input_tensor)
    return hidden_states