Skip to content

vllm.model_executor.models.modernbert

ModernBertAttention

Bases: Module

Source code in vllm/model_executor/models/modernbert.py 
 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
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
class ModernBertAttention(nn.Module):

    def __init__(self,
                 config: ModernBertConfig,
                 layer_id: Optional[int] = None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.layer_id = layer_id
        self.deterministic_flash_attn = config.deterministic_flash_attn
        self.num_heads = config.num_attention_heads
        assert self.num_heads % tp_size == 0
        self.head_dim = config.hidden_size // config.num_attention_heads
        self.all_head_size = self.head_dim * self.num_heads
        self.scaling = self.head_dim**-0.5
        self.Wqkv = QKVParallelLinear(
            config.hidden_size,
            self.head_dim,
            self.num_heads,
            bias=config.attention_bias,
        )

        if layer_id % config.global_attn_every_n_layers != 0:
            self.local_attention = (config.local_attention // 2,
                                    config.local_attention // 2)
        else:
            self.local_attention = (-1, -1)

        rope_theta = config.global_rope_theta
        if self.local_attention != (
                -1, -1) and config.local_rope_theta is not None:
            rope_theta = config.local_rope_theta
        self.rotary_emb = ModernBertRotaryEmbedding(config=config,
                                                    head_size=self.head_dim,
                                                    dim=self.head_dim,
                                                    base=rope_theta)
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
                              prefix=f"{layer_id}.attn",
                              attn_type=AttentionType.ENCODER_ONLY)
        self.Wo = RowParallelLinear(config.hidden_size,
                                    config.hidden_size,
                                    bias=config.attention_bias)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: Optional[torch.LongTensor] = None,
    ) -> torch.Tensor:
        qkv, _ = self.Wqkv(hidden_states)
        q, k, v = qkv.split([self.all_head_size] * 3, dim=-1)
        q, k = self.rotary_emb(position_ids, q, k)
        attn_outputs = self.attn(q, k, v)
        hidden_states = attn_outputs
        hidden_states, _ = self.Wo(hidden_states)
        return hidden_states

Wo instance-attribute 

Wo = RowParallelLinear(
    hidden_size, hidden_size, bias=attention_bias
)

Wqkv instance-attribute 

Wqkv = QKVParallelLinear(
    hidden_size, head_dim, num_heads, bias=attention_bias
)

all_head_size instance-attribute 

all_head_size = head_dim * num_heads

attn instance-attribute 

attn = Attention(
    num_heads,
    head_dim,
    scaling,
    prefix=f"{layer_id}.attn",
    attn_type=ENCODER_ONLY,
)

config instance-attribute 

config = config

deterministic_flash_attn instance-attribute 

deterministic_flash_attn = deterministic_flash_attn

head_dim instance-attribute 

head_dim = hidden_size // num_attention_heads

hidden_size instance-attribute 

hidden_size = hidden_size

layer_id instance-attribute 

layer_id = layer_id

local_attention instance-attribute 

local_attention = (
    local_attention // 2,
    local_attention // 2,
)

num_heads instance-attribute 

num_heads = num_attention_heads

rotary_emb instance-attribute 

rotary_emb = ModernBertRotaryEmbedding(
    config=config,
    head_size=head_dim,
    dim=head_dim,
    base=rope_theta,
)

scaling instance-attribute 

scaling = head_dim ** -0.5

__init__ 

__init__(
    config: ModernBertConfig, layer_id: Optional[int] = None
)
Source code in vllm/model_executor/models/modernbert.py 
 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
110
111
112
def __init__(self,
             config: ModernBertConfig,
             layer_id: Optional[int] = None):
    super().__init__()
    self.config = config
    self.hidden_size = config.hidden_size
    tp_size = get_tensor_model_parallel_world_size()
    self.layer_id = layer_id
    self.deterministic_flash_attn = config.deterministic_flash_attn
    self.num_heads = config.num_attention_heads
    assert self.num_heads % tp_size == 0
    self.head_dim = config.hidden_size // config.num_attention_heads
    self.all_head_size = self.head_dim * self.num_heads
    self.scaling = self.head_dim**-0.5
    self.Wqkv = QKVParallelLinear(
        config.hidden_size,
        self.head_dim,
        self.num_heads,
        bias=config.attention_bias,
    )

    if layer_id % config.global_attn_every_n_layers != 0:
        self.local_attention = (config.local_attention // 2,
                                config.local_attention // 2)
    else:
        self.local_attention = (-1, -1)

    rope_theta = config.global_rope_theta
    if self.local_attention != (
            -1, -1) and config.local_rope_theta is not None:
        rope_theta = config.local_rope_theta
    self.rotary_emb = ModernBertRotaryEmbedding(config=config,
                                                head_size=self.head_dim,
                                                dim=self.head_dim,
                                                base=rope_theta)
    self.attn = Attention(self.num_heads,
                          self.head_dim,
                          self.scaling,
                          prefix=f"{layer_id}.attn",
                          attn_type=AttentionType.ENCODER_ONLY)
    self.Wo = RowParallelLinear(config.hidden_size,
                                config.hidden_size,
                                bias=config.attention_bias)

forward 

forward(
    hidden_states: Tensor,
    position_ids: Optional[LongTensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
114
115
116
117
118
119
120
121
122
123
124
125
def forward(
    self,
    hidden_states: torch.Tensor,
    position_ids: Optional[torch.LongTensor] = None,
) -> torch.Tensor:
    qkv, _ = self.Wqkv(hidden_states)
    q, k, v = qkv.split([self.all_head_size] * 3, dim=-1)
    q, k = self.rotary_emb(position_ids, q, k)
    attn_outputs = self.attn(q, k, v)
    hidden_states = attn_outputs
    hidden_states, _ = self.Wo(hidden_states)
    return hidden_states

ModernBertEmbeddings

Bases: Module

Source code in vllm/model_executor/models/modernbert.py 
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
class ModernBertEmbeddings(nn.Module):

    def __init__(self, config: ModernBertConfig):

        super().__init__()
        self.config = config
        self.tok_embeddings = VocabParallelEmbedding(config.vocab_size,
                                                     config.hidden_size)
        self.norm = nn.LayerNorm(config.hidden_size,
                                 eps=config.layer_norm_eps,
                                 bias=config.norm_bias)

    def forward(
        self,
        input_ids: torch.Tensor,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        if inputs_embeds:
            return self.norm(inputs_embeds)
        else:
            inputs_embeds = self.tok_embeddings(input_ids)
            embeddings = self.norm(inputs_embeds)
            return embeddings

config instance-attribute 

config = config

norm instance-attribute 

norm = LayerNorm(
    hidden_size, eps=layer_norm_eps, bias=norm_bias
)

tok_embeddings instance-attribute 

tok_embeddings = VocabParallelEmbedding(
    vocab_size, hidden_size
)

__init__ 

__init__(config: ModernBertConfig)
Source code in vllm/model_executor/models/modernbert.py 
31
32
33
34
35
36
37
38
39
def __init__(self, config: ModernBertConfig):

    super().__init__()
    self.config = config
    self.tok_embeddings = VocabParallelEmbedding(config.vocab_size,
                                                 config.hidden_size)
    self.norm = nn.LayerNorm(config.hidden_size,
                             eps=config.layer_norm_eps,
                             bias=config.norm_bias)

forward 

forward(
    input_ids: Tensor,
    inputs_embeds: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
41
42
43
44
45
46
47
48
49
50
51
def forward(
    self,
    input_ids: torch.Tensor,
    inputs_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    if inputs_embeds:
        return self.norm(inputs_embeds)
    else:
        inputs_embeds = self.tok_embeddings(input_ids)
        embeddings = self.norm(inputs_embeds)
        return embeddings

ModernBertEncoderLayer

Bases: Module

Source code in vllm/model_executor/models/modernbert.py 
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
class ModernBertEncoderLayer(nn.Module):

    def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        self.layers = nn.ModuleList([
            ModernBertLayer(config=config, layer_id=layer_id)
            for layer_id in range(config.num_hidden_layers)
        ])

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: Optional[torch.LongTensor] = None,
    ) -> torch.Tensor:
        for i, layer in enumerate(self.layers):
            hidden_states = layer(hidden_states, position_ids)
        return hidden_states

layers instance-attribute 

layers = ModuleList(
    [
        ModernBertLayer(config=config, layer_id=layer_id)
        for layer_id in range(num_hidden_layers)
    ]
)

__init__ 

__init__(vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/modernbert.py 
181
182
183
184
185
186
187
def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    self.layers = nn.ModuleList([
        ModernBertLayer(config=config, layer_id=layer_id)
        for layer_id in range(config.num_hidden_layers)
    ])

forward 

forward(
    hidden_states: Tensor,
    position_ids: Optional[LongTensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
189
190
191
192
193
194
195
196
def forward(
    self,
    hidden_states: torch.Tensor,
    position_ids: Optional[torch.LongTensor] = None,
) -> torch.Tensor:
    for i, layer in enumerate(self.layers):
        hidden_states = layer(hidden_states, position_ids)
    return hidden_states

ModernBertForSequenceClassification

Bases: Module, SupportsV0Only, SupportsCrossEncoding

Source code in vllm/model_executor/models/modernbert.py 
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
340
341
342
class ModernBertForSequenceClassification(nn.Module, SupportsV0Only,
                                          SupportsCrossEncoding):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        self.config = config
        self.model = ModernBertModel(vllm_config=vllm_config,
                                     prefix=maybe_prefix(prefix, "modernbert"))
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
        self._pooler = ClassifierPooler(
            vllm_config.model_config,
            pooling=ModernBertPooler(config),
            classifier=self.classifier,
        )

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

        self_weights = []

        def weight_filter():
            for name, weight in weights:
                if name.startswith("model."):
                    yield name[len("model."):], weight
                else:
                    self_weights.append((name, weight))

        self.model.load_weights(weight_filter())

        params_dict = dict(self.named_parameters())

        for name, loaded_weight in self_weights:
            if name.startswith("classifier"):
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
            if name.startswith("head"):
                param = params_dict["_pooler.pooler." + name[len("head") + 1:]]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)

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

_pooler instance-attribute 

_pooler = ClassifierPooler(
    model_config,
    pooling=ModernBertPooler(config),
    classifier=classifier,
)

classifier instance-attribute 

classifier = Linear(hidden_size, num_labels)

config instance-attribute 

config = config

model instance-attribute 

model = ModernBertModel(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "modernbert"),
)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/modernbert.py 
284
285
286
287
288
289
290
291
292
293
294
295
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    self.config = config
    self.model = ModernBertModel(vllm_config=vllm_config,
                                 prefix=maybe_prefix(prefix, "modernbert"))
    self.classifier = nn.Linear(config.hidden_size, config.num_labels)
    self._pooler = ClassifierPooler(
        vllm_config.model_config,
        pooling=ModernBertPooler(config),
        classifier=self.classifier,
    )

forward 

forward(
    input_ids: Optional[LongTensor],
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
331
332
333
334
335
336
337
338
339
340
341
342
def forward(
    self,
    input_ids: Optional[torch.LongTensor],
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    return self.model(
        input_ids=input_ids,
        inputs_embeds=inputs_embeds,
        position_ids=positions,
    )

load_weights 

load_weights(weights: Iterable[tuple[str, Tensor]])
Source code in vllm/model_executor/models/modernbert.py 
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
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):

    self_weights = []

    def weight_filter():
        for name, weight in weights:
            if name.startswith("model."):
                yield name[len("model."):], weight
            else:
                self_weights.append((name, weight))

    self.model.load_weights(weight_filter())

    params_dict = dict(self.named_parameters())

    for name, loaded_weight in self_weights:
        if name.startswith("classifier"):
            param = params_dict[name]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, loaded_weight)
        if name.startswith("head"):
            param = params_dict["_pooler.pooler." + name[len("head") + 1:]]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, loaded_weight)

pooler 

pooler(
    hidden_states: Tensor, pooling_metadata: PoolingMetadata
) -> Optional[PoolerOutput]
Source code in vllm/model_executor/models/modernbert.py 
324
325
326
327
328
329
def pooler(
    self,
    hidden_states: torch.Tensor,
    pooling_metadata: PoolingMetadata,
) -> Optional[PoolerOutput]:
    return self._pooler(hidden_states, pooling_metadata)

ModernBertLayer

Bases: Module

Source code in vllm/model_executor/models/modernbert.py 
146
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
class ModernBertLayer(nn.Module):

    def __init__(self,
                 config: ModernBertConfig,
                 prefix: str = "",
                 layer_id: Optional[int] = None):
        super().__init__()
        self.config = config
        if layer_id == 0:
            self.attn_norm = nn.Identity()
        else:
            self.attn_norm = nn.LayerNorm(config.hidden_size,
                                          eps=config.norm_eps,
                                          bias=config.norm_bias)
        self.attn = ModernBertAttention(config=config, layer_id=layer_id)
        self.mlp_norm = nn.LayerNorm(config.hidden_size,
                                     eps=config.norm_eps,
                                     bias=config.norm_bias)
        self.mlp = ModernBertMLP(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: Optional[torch.LongTensor] = None,
    ):
        attn_outputs = self.attn(self.attn_norm(hidden_states),
                                 position_ids=position_ids)
        hidden_states = hidden_states + attn_outputs
        mlp_output = self.mlp(self.mlp_norm(hidden_states))
        hidden_states = hidden_states + mlp_output
        return hidden_states

attn instance-attribute 

attn = ModernBertAttention(config=config, layer_id=layer_id)

attn_norm instance-attribute 

attn_norm = Identity()

config instance-attribute 

config = config

mlp instance-attribute 

mlp = ModernBertMLP(config)

mlp_norm instance-attribute 

mlp_norm = LayerNorm(
    hidden_size, eps=norm_eps, bias=norm_bias
)

__init__ 

__init__(
    config: ModernBertConfig,
    prefix: str = "",
    layer_id: Optional[int] = None,
)
Source code in vllm/model_executor/models/modernbert.py 
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
def __init__(self,
             config: ModernBertConfig,
             prefix: str = "",
             layer_id: Optional[int] = None):
    super().__init__()
    self.config = config
    if layer_id == 0:
        self.attn_norm = nn.Identity()
    else:
        self.attn_norm = nn.LayerNorm(config.hidden_size,
                                      eps=config.norm_eps,
                                      bias=config.norm_bias)
    self.attn = ModernBertAttention(config=config, layer_id=layer_id)
    self.mlp_norm = nn.LayerNorm(config.hidden_size,
                                 eps=config.norm_eps,
                                 bias=config.norm_bias)
    self.mlp = ModernBertMLP(config)

forward 

forward(
    hidden_states: Tensor,
    position_ids: Optional[LongTensor] = None,
)
Source code in vllm/model_executor/models/modernbert.py 
166
167
168
169
170
171
172
173
174
175
176
def forward(
    self,
    hidden_states: torch.Tensor,
    position_ids: Optional[torch.LongTensor] = None,
):
    attn_outputs = self.attn(self.attn_norm(hidden_states),
                             position_ids=position_ids)
    hidden_states = hidden_states + attn_outputs
    mlp_output = self.mlp(self.mlp_norm(hidden_states))
    hidden_states = hidden_states + mlp_output
    return hidden_states

ModernBertMLP

Bases: Module

Source code in vllm/model_executor/models/modernbert.py 
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
class ModernBertMLP(nn.Module):

    def __init__(self, config: ModernBertConfig):
        super().__init__()
        self.config = config
        self.Wi = nn.Linear(config.hidden_size,
                            int(config.intermediate_size) * 2,
                            bias=config.mlp_bias)
        self.act = nn.GELU()
        self.Wo = RowParallelLinear(config.intermediate_size,
                                    config.hidden_size,
                                    bias=config.mlp_bias)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        input, gate = self.Wi(hidden_states).chunk(2, dim=-1)
        return self.Wo(self.act(input) * gate)[0]

Wi instance-attribute 

Wi = Linear(
    hidden_size, int(intermediate_size) * 2, bias=mlp_bias
)

Wo instance-attribute 

Wo = RowParallelLinear(
    intermediate_size, hidden_size, bias=mlp_bias
)

act instance-attribute 

act = GELU()

config instance-attribute 

config = config

__init__ 

__init__(config: ModernBertConfig)
Source code in vllm/model_executor/models/modernbert.py 
130
131
132
133
134
135
136
137
138
139
def __init__(self, config: ModernBertConfig):
    super().__init__()
    self.config = config
    self.Wi = nn.Linear(config.hidden_size,
                        int(config.intermediate_size) * 2,
                        bias=config.mlp_bias)
    self.act = nn.GELU()
    self.Wo = RowParallelLinear(config.intermediate_size,
                                config.hidden_size,
                                bias=config.mlp_bias)

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
141
142
143
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    input, gate = self.Wi(hidden_states).chunk(2, dim=-1)
    return self.Wo(self.act(input) * gate)[0]

ModernBertModel

Bases: Module

Source code in vllm/model_executor/models/modernbert.py 
199
200
201
202
203
204
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
@support_torch_compile
class ModernBertModel(nn.Module):
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={"layers.": "encoder_layer.layers."})

    def __init__(
        self,
        vllm_config: VllmConfig,
        prefix: str = "",
    ):
        super().__init__()
        config = vllm_config.model_config.hf_config
        self.config = config
        self.embeddings = ModernBertEmbeddings(config)
        self.encoder_layer = ModernBertEncoderLayer(vllm_config)
        self.final_norm = nn.LayerNorm(config.hidden_size,
                                       eps=config.norm_eps,
                                       bias=config.norm_bias)

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        weights = self.hf_to_vllm_mapper.apply(weights)
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if name.endswith(".bias") and name not in params_dict:
                continue
            param = params_dict[name]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        positions: Optional[torch.Tensor] = None,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
    ) -> torch.Tensor:
        position_ids = positions if positions is not None else position_ids
        if inputs_embeds is not None:
            hidden_states = inputs_embeds
        else:
            hidden_states = self.embeddings(input_ids=input_ids,
                                            inputs_embeds=inputs_embeds)

        outputs = self.encoder_layer(
            hidden_states=hidden_states,
            position_ids=position_ids,
        )
        norm_outputs = self.final_norm(outputs)
        return norm_outputs

config instance-attribute 

config = config

embeddings instance-attribute 

embeddings = ModernBertEmbeddings(config)

encoder_layer instance-attribute 

encoder_layer = ModernBertEncoderLayer(vllm_config)

final_norm instance-attribute 

final_norm = LayerNorm(
    hidden_size, eps=norm_eps, bias=norm_bias
)

hf_to_vllm_mapper class-attribute instance-attribute 

hf_to_vllm_mapper = WeightsMapper(
    orig_to_new_prefix={"layers.": "encoder_layer.layers."}
)

__init__ 

__init__(vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/modernbert.py 
204
205
206
207
208
209
210
211
212
213
214
215
216
def __init__(
    self,
    vllm_config: VllmConfig,
    prefix: str = "",
):
    super().__init__()
    config = vllm_config.model_config.hf_config
    self.config = config
    self.embeddings = ModernBertEmbeddings(config)
    self.encoder_layer = ModernBertEncoderLayer(vllm_config)
    self.final_norm = nn.LayerNorm(config.hidden_size,
                                   eps=config.norm_eps,
                                   bias=config.norm_bias)

forward 

forward(
    input_ids: Optional[LongTensor] = None,
    positions: Optional[Tensor] = None,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    position_ids: Optional[LongTensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
def forward(
    self,
    input_ids: Optional[torch.LongTensor] = None,
    positions: Optional[torch.Tensor] = None,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.LongTensor] = None,
) -> torch.Tensor:
    position_ids = positions if positions is not None else position_ids
    if inputs_embeds is not None:
        hidden_states = inputs_embeds
    else:
        hidden_states = self.embeddings(input_ids=input_ids,
                                        inputs_embeds=inputs_embeds)

    outputs = self.encoder_layer(
        hidden_states=hidden_states,
        position_ids=position_ids,
    )
    norm_outputs = self.final_norm(outputs)
    return norm_outputs

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/modernbert.py 
218
219
220
221
222
223
224
225
226
227
228
229
230
231
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    weights = self.hf_to_vllm_mapper.apply(weights)
    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if name.endswith(".bias") and name not in params_dict:
            continue
        param = params_dict[name]
        weight_loader = getattr(param, "weight_loader",
                                default_weight_loader)
        weight_loader(param, loaded_weight)
        loaded_params.add(name)
    return loaded_params

ModernBertPooler

Bases: BasePooler

Source code in vllm/model_executor/models/modernbert.py 
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
class ModernBertPooler(BasePooler):

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

        pooling_type = PoolingType[config.classifier_pooling.upper()]
        self.pooling = PoolingMethod.from_pooling_type(pooling_type)
        self.dense = nn.Linear(config.hidden_size, config.hidden_size,
                               config.classifier_bias)
        self.pooling_type = config.classifier_pooling
        self.act = nn.GELU()
        self.norm = nn.LayerNorm(config.hidden_size,
                                 eps=config.norm_eps,
                                 bias=config.norm_bias)

    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)
        pooled_output = self.norm(self.act(self.dense(pooled_output)))
        return pooled_output

act instance-attribute 

act = GELU()

dense instance-attribute 

dense = Linear(hidden_size, hidden_size, classifier_bias)

norm instance-attribute 

norm = LayerNorm(hidden_size, eps=norm_eps, bias=norm_bias)

pooling instance-attribute 

pooling = from_pooling_type(pooling_type)

pooling_type instance-attribute 

pooling_type = classifier_pooling

__init__ 

__init__(config: ModernBertConfig)
Source code in vllm/model_executor/models/modernbert.py 
258
259
260
261
262
263
264
265
266
267
268
269
def __init__(self, config: ModernBertConfig):
    super().__init__()

    pooling_type = PoolingType[config.classifier_pooling.upper()]
    self.pooling = PoolingMethod.from_pooling_type(pooling_type)
    self.dense = nn.Linear(config.hidden_size, config.hidden_size,
                           config.classifier_bias)
    self.pooling_type = config.classifier_pooling
    self.act = nn.GELU()
    self.norm = nn.LayerNorm(config.hidden_size,
                             eps=config.norm_eps,
                             bias=config.norm_bias)

forward 

forward(
    hidden_states: Union[Tensor, list[Tensor]],
    pooling_metadata: PoolingMetadata,
) -> Union[Tensor, list[Tensor]]
Source code in vllm/model_executor/models/modernbert.py 
271
272
273
274
275
276
277
278
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)
    pooled_output = self.norm(self.act(self.dense(pooled_output)))
    return pooled_output

ModernBertRotaryEmbedding

Bases: RotaryEmbedding

Source code in vllm/model_executor/models/modernbert.py 
54
55
56
57
58
59
60
61
62
63
64
65
class ModernBertRotaryEmbedding(RotaryEmbedding):

    def __init__(self, config: ModernBertConfig, head_size: int, dim: int,
                 base: float):
        super().__init__(
            head_size=head_size,
            rotary_dim=dim,
            max_position_embeddings=config.max_position_embeddings,
            base=base,
            is_neox_style=True,
            dtype=torch.float16)
        self.config = config

config instance-attribute 

config = config

__init__ 

__init__(
    config: ModernBertConfig,
    head_size: int,
    dim: int,
    base: float,
)
Source code in vllm/model_executor/models/modernbert.py 
56
57
58
59
60
61
62
63
64
65
def __init__(self, config: ModernBertConfig, head_size: int, dim: int,
             base: float):
    super().__init__(
        head_size=head_size,
        rotary_dim=dim,
        max_position_embeddings=config.max_position_embeddings,
        base=base,
        is_neox_style=True,
        dtype=torch.float16)
    self.config = config