Skip to content

vllm.model_executor.models.roberta

RobertaClassificationHead

Bases: Module

Head for sentence-level classification tasks.

Source code in vllm/model_executor/models/roberta.py 
101
102
103
104
105
106
107
108
109
110
111
112
113
114
class RobertaClassificationHead(nn.Module):
    """Head for sentence-level classification tasks."""

    def __init__(self, config: RobertaConfig):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # CLSPool has already been applied in `pooling`
        x = self.dense(x)
        x = torch.tanh(x)
        x = self.out_proj(x)
        return x

dense instance-attribute 

dense = Linear(hidden_size, hidden_size)

out_proj instance-attribute 

out_proj = Linear(hidden_size, num_labels)

__init__ 

__init__(config: RobertaConfig)
Source code in vllm/model_executor/models/roberta.py 
104
105
106
107
def __init__(self, config: RobertaConfig):
    super().__init__()
    self.dense = nn.Linear(config.hidden_size, config.hidden_size)
    self.out_proj = nn.Linear(config.hidden_size, config.num_labels)

forward 

forward(x: Tensor) -> Tensor
Source code in vllm/model_executor/models/roberta.py 
109
110
111
112
113
114
def forward(self, x: torch.Tensor) -> torch.Tensor:
    # CLSPool has already been applied in `pooling`
    x = self.dense(x)
    x = torch.tanh(x)
    x = self.out_proj(x)
    return x

RobertaEmbedding

Bases: Module

Source code in vllm/model_executor/models/roberta.py 
25
26
27
28
29
30
31
32
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
class RobertaEmbedding(nn.Module):

    def __init__(self, config: RobertaConfig):
        super().__init__()
        self.size = config.hidden_size
        self.word_embeddings = VocabParallelEmbedding(config.vocab_size,
                                                      config.hidden_size)
        self.padding_idx = config.pad_token_id
        self.position_embeddings = nn.Embedding(config.max_position_embeddings,
                                                config.hidden_size,
                                                padding_idx=self.padding_idx)

        self.token_type_embeddings = nn.Embedding(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()
        inputs_embeds = self.word_embeddings(input_ids)

        # Replace position ids because in RoBERTa models
        # they have to start at padding_idx + 1 and ignore
        # existing padding tokens
        # References:
        # - https://github.com/huggingface/transformers/blob/a3d69a8994d673899608a7c17fbf4f953f50474e/src/transformers/models/roberta/modeling_roberta.py#L133
        # - https://github.com/huggingface/transformers/blob/a3d69a8994d673899608a7c17fbf4f953f50474e/src/transformers/models/roberta/modeling_roberta.py#L1669
        pos_list = []
        token_list = []
        offset = 0
        for seq_len in seq_lens:
            pos_list.append(position_ids[offset:offset + seq_len])
            token_list.append(input_ids[offset:offset + seq_len])
            offset += seq_len

        new_pos_list = []
        for positions, tokens in zip(pos_list, token_list):
            # Verify assumption that incoming position are
            # always a sequence from 0 to N.
            expected_pos = torch.arange(positions.size()[0],
                                        dtype=torch.long,
                                        device=inputs_embeds.device)
            assert torch.equal(positions, expected_pos)
            new_pos_list.append(
                create_position_ids_from_input_ids(tokens, self.padding_idx))
        position_ids = torch.cat(new_pos_list)

        # 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)

padding_idx instance-attribute 

padding_idx = pad_token_id

position_embedding_type instance-attribute 

position_embedding_type = position_embedding_type

position_embeddings instance-attribute 

position_embeddings = Embedding(
    max_position_embeddings,
    hidden_size,
    padding_idx=padding_idx,
)

size instance-attribute 

size = hidden_size

token_type_embeddings instance-attribute 

token_type_embeddings = Embedding(
    type_vocab_size, hidden_size
)

word_embeddings instance-attribute 

word_embeddings = VocabParallelEmbedding(
    vocab_size, hidden_size
)

__init__ 

__init__(config: RobertaConfig)
Source code in vllm/model_executor/models/roberta.py 
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
def __init__(self, config: RobertaConfig):
    super().__init__()
    self.size = config.hidden_size
    self.word_embeddings = VocabParallelEmbedding(config.vocab_size,
                                                  config.hidden_size)
    self.padding_idx = config.pad_token_id
    self.position_embeddings = nn.Embedding(config.max_position_embeddings,
                                            config.hidden_size,
                                            padding_idx=self.padding_idx)

    self.token_type_embeddings = nn.Embedding(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/roberta.py 
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
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()
    inputs_embeds = self.word_embeddings(input_ids)

    # Replace position ids because in RoBERTa models
    # they have to start at padding_idx + 1 and ignore
    # existing padding tokens
    # References:
    # - https://github.com/huggingface/transformers/blob/a3d69a8994d673899608a7c17fbf4f953f50474e/src/transformers/models/roberta/modeling_roberta.py#L133
    # - https://github.com/huggingface/transformers/blob/a3d69a8994d673899608a7c17fbf4f953f50474e/src/transformers/models/roberta/modeling_roberta.py#L1669
    pos_list = []
    token_list = []
    offset = 0
    for seq_len in seq_lens:
        pos_list.append(position_ids[offset:offset + seq_len])
        token_list.append(input_ids[offset:offset + seq_len])
        offset += seq_len

    new_pos_list = []
    for positions, tokens in zip(pos_list, token_list):
        # Verify assumption that incoming position are
        # always a sequence from 0 to N.
        expected_pos = torch.arange(positions.size()[0],
                                    dtype=torch.long,
                                    device=inputs_embeds.device)
        assert torch.equal(positions, expected_pos)
        new_pos_list.append(
            create_position_ids_from_input_ids(tokens, self.padding_idx))
    position_ids = torch.cat(new_pos_list)

    # 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

RobertaEmbeddingModel

Bases: BertEmbeddingModel

A model that uses Roberta 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/roberta.py 
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
class RobertaEmbeddingModel(BertEmbeddingModel):
    """A model that uses Roberta 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.
   """

    def _build_model(self,
                     vllm_config: VllmConfig,
                     prefix: str = "") -> Union[BertModel, BertWithRope]:
        if (vllm_config.model_config.hf_config.position_embedding_type ==
                "rotary"):
            return JinaRobertaModel(vllm_config=vllm_config, prefix=prefix)
        else:
            return BertModel(vllm_config=vllm_config,
                             prefix=prefix,
                             embedding_class=RobertaEmbedding)

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
        weights_list = list(weights)
        has_roberta_prefix = any(
            name.startswith("roberta.") for name, _ in weights_list)
        if has_roberta_prefix:
            # For models with the `roberta.` prefix e.g.
            # `FacebookAI/roberta-base`
            mapper = WeightsMapper(orig_to_new_prefix={"roberta.": "model."})
        else:
            # For models without the `roberta.` prefix e.g.
            # `sentence-transformers/stsb-roberta-base-v2`
            mapper = WeightsMapper(orig_to_new_prefix={"": "model."})

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

_build_model 

_build_model(
    vllm_config: VllmConfig, prefix: str = ""
) -> Union[BertModel, BertWithRope]
Source code in vllm/model_executor/models/roberta.py 
128
129
130
131
132
133
134
135
136
137
def _build_model(self,
                 vllm_config: VllmConfig,
                 prefix: str = "") -> Union[BertModel, BertWithRope]:
    if (vllm_config.model_config.hf_config.position_embedding_type ==
            "rotary"):
        return JinaRobertaModel(vllm_config=vllm_config, prefix=prefix)
    else:
        return BertModel(vllm_config=vllm_config,
                         prefix=prefix,
                         embedding_class=RobertaEmbedding)

load_weights 

load_weights(weights: Iterable[tuple[str, Tensor]])
Source code in vllm/model_executor/models/roberta.py 
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
    weights_list = list(weights)
    has_roberta_prefix = any(
        name.startswith("roberta.") for name, _ in weights_list)
    if has_roberta_prefix:
        # For models with the `roberta.` prefix e.g.
        # `FacebookAI/roberta-base`
        mapper = WeightsMapper(orig_to_new_prefix={"roberta.": "model."})
    else:
        # For models without the `roberta.` prefix e.g.
        # `sentence-transformers/stsb-roberta-base-v2`
        mapper = WeightsMapper(orig_to_new_prefix={"": "model."})

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

RobertaForSequenceClassification

Bases: Module, SupportsCrossEncoding, SupportsV0Only

A model that uses Roberta to provide embedding functionalities.

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

Attributes:

Name Type Description
roberta

An instance of BertModel used for forward operations.
_pooler

An instance of Pooler used for pooling operations.
Source code in vllm/model_executor/models/roberta.py 
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
182
183
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
214
215
216
217
218
219
220
class RobertaForSequenceClassification(nn.Module, SupportsCrossEncoding,
                                       SupportsV0Only):
    """A model that uses Roberta to provide embedding functionalities.

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

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

    jina_to_vllm_mapper = WeightsMapper(
        orig_to_new_substr={
            'emb_ln': "embeddings.LayerNorm",
            'layers': "layer",
            'mixer.Wqkv': "attention.self.qkv_proj",
            'mixer.out_proj': "attention.output.dense",
            'norm1': "attention.output.LayerNorm",
            'mlp.fc1': "intermediate.dense",
            'mlp.fc2': "output.dense",
            'norm2': "output.LayerNorm",
        })

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

        self.num_labels = config.num_labels
        self.roberta = BertModel(vllm_config=vllm_config,
                                 prefix=maybe_prefix(prefix, "bert"),
                                 embedding_class=RobertaEmbedding,
                                 add_pooling_layer=False)
        self.classifier = RobertaClassificationHead(config)

        self._pooler = ClassifierPooler(
            vllm_config.model_config,
            pooling=CLSPool(),
            classifier=self.classifier,
        )

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

    def pooler(
        self,
        hidden_states: torch.Tensor,
        pooling_metadata: PoolingMetadata,
    ) -> Optional[PoolerOutput]:
        return self._pooler(hidden_states, pooling_metadata)

    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.roberta(input_ids=input_ids,
                            position_ids=positions,
                            inputs_embeds=inputs_embeds,
                            intermediate_tensors=intermediate_tensors,
                            token_type_ids=token_type_ids)

_pooler instance-attribute 

_pooler = ClassifierPooler(
    model_config, pooling=CLSPool(), classifier=classifier
)

classifier instance-attribute 

classifier = RobertaClassificationHead(config)

jina_to_vllm_mapper class-attribute instance-attribute 

jina_to_vllm_mapper = WeightsMapper(
    orig_to_new_substr={
        "emb_ln": "embeddings.LayerNorm",
        "layers": "layer",
        "mixer.Wqkv": "attention.self.qkv_proj",
        "mixer.out_proj": "attention.output.dense",
        "norm1": "attention.output.LayerNorm",
        "mlp.fc1": "intermediate.dense",
        "mlp.fc2": "output.dense",
        "norm2": "output.LayerNorm",
    }
)

num_labels instance-attribute 

num_labels = num_labels

roberta instance-attribute 

roberta = BertModel(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "bert"),
    embedding_class=RobertaEmbedding,
    add_pooling_layer=False,
)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/roberta.py 
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config

    self.num_labels = config.num_labels
    self.roberta = BertModel(vllm_config=vllm_config,
                             prefix=maybe_prefix(prefix, "bert"),
                             embedding_class=RobertaEmbedding,
                             add_pooling_layer=False)
    self.classifier = RobertaClassificationHead(config)

    self._pooler = ClassifierPooler(
        vllm_config.model_config,
        pooling=CLSPool(),
        classifier=self.classifier,
    )

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/roberta.py 
208
209
210
211
212
213
214
215
216
217
218
219
220
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.roberta(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/roberta.py 
197
198
199
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
    loader = AutoWeightsLoader(self)
    return loader.load_weights(weights, mapper=self.jina_to_vllm_mapper)

pooler 

pooler(
    hidden_states: Tensor, pooling_metadata: PoolingMetadata
) -> Optional[PoolerOutput]
Source code in vllm/model_executor/models/roberta.py 
201
202
203
204
205
206
def pooler(
    self,
    hidden_states: torch.Tensor,
    pooling_metadata: PoolingMetadata,
) -> Optional[PoolerOutput]:
    return self._pooler(hidden_states, pooling_metadata)

create_position_ids_from_input_ids 

create_position_ids_from_input_ids(
    input_ids, padding_idx, past_key_values_length=0
)

Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols are ignored. This is modified from fairseq's utils.make_positions.

Parameters:

Name Type Description Default
x

torch.Tensor x:

 required

Returns: torch.Tensor

Source code in vllm/model_executor/models/roberta.py 
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
def create_position_ids_from_input_ids(input_ids,
                                       padding_idx,
                                       past_key_values_length=0):
    """
    Replace non-padding symbols with their position numbers.
    Position numbers begin at padding_idx+1. Padding symbols
    are ignored. This is modified from fairseq's `utils.make_positions`.

    Args:
        x: torch.Tensor x:

    Returns: torch.Tensor
    """
    # The series of casts and type-conversions here are carefully
    # balanced to both work with ONNX export and XLA.
    mask = input_ids.ne(padding_idx).int()

    incremental_indices = (torch.cumsum(mask, dim=0).type_as(mask) +
                           past_key_values_length) * mask

    return incremental_indices.long() + padding_idx