Skip to content

vllm.model_executor.models.modernbert

ModernBertAttention

Bases: Module

Source code in vllm/model_executor/models/modernbert.py 
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
class ModernBertAttention(nn.Module):
    def __init__(
        self, config: ModernBertConfig, layer_id: int | None = None, prefix: str = ""
    ):
        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,
            prefix=f"{prefix}.Wqkv",
        )

        if layer_types := getattr(config, "layer_types", None):
            # Transformers v5
            layer_type = layer_types[layer_id]
            rope_parameters = config.rope_parameters[layer_type]
            sliding_window: int | None = None
            if layer_type == "sliding_attention":
                sliding_window = config.local_attention // 2
        else:
            # Transformers v4
            sliding_window = None
            if layer_id % config.global_attn_every_n_layers != 0:
                sliding_window = config.local_attention // 2
                rope_theta = (
                    config.local_rope_theta
                    if config.local_rope_theta is not None
                    else config.global_rope_theta
                )
            else:
                rope_theta = config.global_rope_theta
            rope_parameters = {"rope_type": "default", "rope_theta": rope_theta}

        self.rotary_emb = get_rope(
            head_size=self.head_dim,
            max_position=config.max_position_embeddings,
            rope_parameters=rope_parameters,
            dtype=torch.float16,
        )
        self.attn = EncoderOnlyAttention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            prefix=f"{layer_id}.attn",
            per_layer_sliding_window=sliding_window,
        )
        self.Wo = RowParallelLinear(
            config.hidden_size,
            config.hidden_size,
            bias=config.attention_bias,
            prefix=f"{prefix}.Wo",
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: torch.Tensor,
    ) -> 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,
    prefix=f"{prefix}.Wo",
)

Wqkv instance-attribute 

Wqkv = QKVParallelLinear(
    hidden_size,
    head_dim,
    num_heads,
    bias=attention_bias,
    prefix=f"{prefix}.Wqkv",
)

all_head_size instance-attribute 

all_head_size = head_dim * num_heads

attn instance-attribute 

attn = EncoderOnlyAttention(
    num_heads,
    head_dim,
    scaling,
    prefix=f"{layer_id}.attn",
    per_layer_sliding_window=sliding_window,
)

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

num_heads instance-attribute 

num_heads = num_attention_heads

rotary_emb instance-attribute 

rotary_emb = get_rope(
    head_size=head_dim,
    max_position=max_position_embeddings,
    rope_parameters=rope_parameters,
    dtype=float16,
)

scaling instance-attribute 

scaling = head_dim ** -0.5

__init__ 

__init__(
    config: ModernBertConfig,
    layer_id: int | None = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/modernbert.py 
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
def __init__(
    self, config: ModernBertConfig, layer_id: int | None = None, prefix: str = ""
):
    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,
        prefix=f"{prefix}.Wqkv",
    )

    if layer_types := getattr(config, "layer_types", None):
        # Transformers v5
        layer_type = layer_types[layer_id]
        rope_parameters = config.rope_parameters[layer_type]
        sliding_window: int | None = None
        if layer_type == "sliding_attention":
            sliding_window = config.local_attention // 2
    else:
        # Transformers v4
        sliding_window = None
        if layer_id % config.global_attn_every_n_layers != 0:
            sliding_window = config.local_attention // 2
            rope_theta = (
                config.local_rope_theta
                if config.local_rope_theta is not None
                else config.global_rope_theta
            )
        else:
            rope_theta = config.global_rope_theta
        rope_parameters = {"rope_type": "default", "rope_theta": rope_theta}

    self.rotary_emb = get_rope(
        head_size=self.head_dim,
        max_position=config.max_position_embeddings,
        rope_parameters=rope_parameters,
        dtype=torch.float16,
    )
    self.attn = EncoderOnlyAttention(
        self.num_heads,
        self.head_dim,
        self.scaling,
        prefix=f"{layer_id}.attn",
        per_layer_sliding_window=sliding_window,
    )
    self.Wo = RowParallelLinear(
        config.hidden_size,
        config.hidden_size,
        bias=config.attention_bias,
        prefix=f"{prefix}.Wo",
    )

forward 

forward(
    hidden_states: Tensor, position_ids: Tensor
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
129
130
131
132
133
134
135
136
137
138
139
140
def forward(
    self,
    hidden_states: torch.Tensor,
    position_ids: torch.Tensor,
) -> 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 
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
class ModernBertEmbeddings(nn.Module):
    def __init__(self, config: ModernBertConfig):
        super().__init__()
        self.config = config
        self.tok_embeddings = VocabParallelEmbedding(
            config.vocab_size, config.hidden_size
        )
        eps = (
            getattr(config, "norm_eps", None)
            or getattr(config, "layer_norm_eps", None)
            or 1e-5
        )
        self.norm = nn.LayerNorm(config.hidden_size, eps=eps, bias=config.norm_bias)

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.tok_embeddings(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor:
        if inputs_embeds is not None:
            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=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 
36
37
38
39
40
41
42
43
44
45
46
47
def __init__(self, config: ModernBertConfig):
    super().__init__()
    self.config = config
    self.tok_embeddings = VocabParallelEmbedding(
        config.vocab_size, config.hidden_size
    )
    eps = (
        getattr(config, "norm_eps", None)
        or getattr(config, "layer_norm_eps", None)
        or 1e-5
    )
    self.norm = nn.LayerNorm(config.hidden_size, eps=eps, bias=config.norm_bias)

embed_input_ids 

embed_input_ids(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
49
50
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.tok_embeddings(input_ids)

forward 

forward(
    input_ids: Tensor, inputs_embeds: Tensor | None = None
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
52
53
54
55
56
57
58
59
60
61
62
def forward(
    self,
    input_ids: torch.Tensor,
    inputs_embeds: torch.Tensor | None = None,
) -> torch.Tensor:
    if inputs_embeds is not None:
        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 
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
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,
                    prefix=f"{prefix}.layers.{layer_id}",
                )
                for layer_id in range(config.num_hidden_layers)
            ]
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: torch.Tensor,
    ) -> 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,
                prefix=f"{prefix}.layers.{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 
198
199
200
201
202
203
204
205
206
207
208
209
210
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,
                prefix=f"{prefix}.layers.{layer_id}",
            )
            for layer_id in range(config.num_hidden_layers)
        ]
    )

forward 

forward(
    hidden_states: Tensor, position_ids: Tensor
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
212
213
214
215
216
217
218
219
def forward(
    self,
    hidden_states: torch.Tensor,
    position_ids: torch.Tensor,
) -> torch.Tensor:
    for i, layer in enumerate(self.layers):
        hidden_states = layer(hidden_states, position_ids)
    return hidden_states

ModernBertForSequenceClassification

Bases: Module, SupportsCrossEncoding

Source code in vllm/model_executor/models/modernbert.py 
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
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
@default_pooling_type(seq_pooling_type="CLS")
class ModernBertForSequenceClassification(nn.Module, SupportsCrossEncoding):
    is_pooling_model = True

    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,
            dtype=vllm_config.model_config.head_dtype,
        )

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

        self.pooling = ModernBertPooler(vllm_config.model_config)

        self.pooler = DispatchPooler.for_seq_cls(
            pooler_config,
            pooling=self.pooling,
            classifier=self.classifier,
        )

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)

    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["pooling." + name[len("head") + 1 :]]
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)

    def forward(
        self,
        input_ids: torch.LongTensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor:
        return self.model(
            input_ids=input_ids,
            inputs_embeds=inputs_embeds,
            positions=positions,
        )

classifier instance-attribute 

classifier = Linear(
    hidden_size, num_labels, dtype=head_dtype
)

config instance-attribute 

config = config

is_pooling_model class-attribute instance-attribute 

is_pooling_model = True

model instance-attribute 

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

pooler instance-attribute 

pooler = for_seq_cls(
    pooler_config, pooling=pooling, classifier=classifier
)

pooling instance-attribute 

pooling = ModernBertPooler(model_config)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/modernbert.py 
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
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,
        dtype=vllm_config.model_config.head_dtype,
    )

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

    self.pooling = ModernBertPooler(vllm_config.model_config)

    self.pooler = DispatchPooler.for_seq_cls(
        pooler_config,
        pooling=self.pooling,
        classifier=self.classifier,
    )

embed_input_ids 

embed_input_ids(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
353
354
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.model.embed_input_ids(input_ids)

forward 

forward(
    input_ids: LongTensor | None,
    positions: Tensor,
    intermediate_tensors: IntermediateTensors | None = None,
    inputs_embeds: Tensor | None = None,
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
380
381
382
383
384
385
386
387
388
389
390
391
def forward(
    self,
    input_ids: torch.LongTensor | None,
    positions: torch.Tensor,
    intermediate_tensors: IntermediateTensors | None = None,
    inputs_embeds: torch.Tensor | None = None,
) -> torch.Tensor:
    return self.model(
        input_ids=input_ids,
        inputs_embeds=inputs_embeds,
        positions=positions,
    )

load_weights 

load_weights(weights: Iterable[tuple[str, Tensor]])
Source code in vllm/model_executor/models/modernbert.py 
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
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["pooling." + name[len("head") + 1 :]]
            weight_loader = getattr(param, "weight_loader", default_weight_loader)
            weight_loader(param, loaded_weight)

ModernBertForTokenClassification

Bases: Module

Source code in vllm/model_executor/models/modernbert.py 
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
@attn_type("encoder_only")
@default_pooling_type(tok_pooling_type="ALL")
class ModernBertForTokenClassification(nn.Module):
    is_pooling_model = True

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        self.head_dtype = vllm_config.model_config.head_dtype
        self.num_labels = config.num_labels
        self.model = ModernBertModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "modernbert")
        )
        self.head = ModernBertPredictionHead(config)
        self.classifier = nn.Linear(
            config.hidden_size, config.num_labels, dtype=self.head_dtype
        )

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

        self.pooler = pooler_for_token_classify(pooler_config)

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)

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

    def forward(
        self,
        input_ids: torch.Tensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor:
        hidden_states = self.model(
            input_ids=input_ids,
            positions=positions,
            inputs_embeds=inputs_embeds,
            intermediate_tensors=intermediate_tensors,
        )
        hidden_states = self.head(hidden_states)
        hidden_states = hidden_states.to(self.head_dtype)
        return self.classifier(hidden_states)

classifier instance-attribute 

classifier = Linear(
    hidden_size, num_labels, dtype=head_dtype
)

head instance-attribute 

head = ModernBertPredictionHead(config)

head_dtype instance-attribute 

head_dtype = head_dtype

is_pooling_model class-attribute instance-attribute 

is_pooling_model = True

model instance-attribute 

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

num_labels instance-attribute 

num_labels = num_labels

pooler instance-attribute 

pooler = pooler_for_token_classify(pooler_config)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/modernbert.py 
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    self.head_dtype = vllm_config.model_config.head_dtype
    self.num_labels = config.num_labels
    self.model = ModernBertModel(
        vllm_config=vllm_config, prefix=maybe_prefix(prefix, "modernbert")
    )
    self.head = ModernBertPredictionHead(config)
    self.classifier = nn.Linear(
        config.hidden_size, config.num_labels, dtype=self.head_dtype
    )

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

    self.pooler = pooler_for_token_classify(pooler_config)

embed_input_ids 

embed_input_ids(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
435
436
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.model.embed_input_ids(input_ids)

forward 

forward(
    input_ids: Tensor | None,
    positions: Tensor,
    intermediate_tensors: IntermediateTensors | None = None,
    inputs_embeds: Tensor | None = None,
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
def forward(
    self,
    input_ids: torch.Tensor | None,
    positions: torch.Tensor,
    intermediate_tensors: IntermediateTensors | None = None,
    inputs_embeds: torch.Tensor | None = None,
) -> torch.Tensor:
    hidden_states = self.model(
        input_ids=input_ids,
        positions=positions,
        inputs_embeds=inputs_embeds,
        intermediate_tensors=intermediate_tensors,
    )
    hidden_states = self.head(hidden_states)
    hidden_states = hidden_states.to(self.head_dtype)
    return self.classifier(hidden_states)

load_weights 

load_weights(weights: Iterable[tuple[str, Tensor]])
Source code in vllm/model_executor/models/modernbert.py 
438
439
440
441
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
    loader = AutoWeightsLoader(self, skip_prefixes=["drop"])
    loaded_params = loader.load_weights(weights)
    return loaded_params

ModernBertLayer

Bases: Module

Source code in vllm/model_executor/models/modernbert.py 
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
class ModernBertLayer(nn.Module):
    def __init__(
        self, config: ModernBertConfig, prefix: str = "", layer_id: int | None = 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, prefix=f"{prefix}.attn"
        )
        self.mlp_norm = nn.LayerNorm(
            config.hidden_size, eps=config.norm_eps, bias=config.norm_bias
        )
        self.mlp = ModernBertMLP(config, prefix=f"{prefix}.mlp")

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: torch.Tensor,
    ) -> torch.Tensor:
        attn_outputs = self.attn(
            hidden_states=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,
    prefix=f"{prefix}.attn",
)

attn_norm instance-attribute 

attn_norm = Identity()

config instance-attribute 

config = config

mlp instance-attribute 

mlp = ModernBertMLP(config, prefix=f'{prefix}.mlp')

mlp_norm instance-attribute 

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

__init__ 

__init__(
    config: ModernBertConfig,
    prefix: str = "",
    layer_id: int | None = None,
)
Source code in vllm/model_executor/models/modernbert.py 
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
def __init__(
    self, config: ModernBertConfig, prefix: str = "", layer_id: int | None = 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, prefix=f"{prefix}.attn"
    )
    self.mlp_norm = nn.LayerNorm(
        config.hidden_size, eps=config.norm_eps, bias=config.norm_bias
    )
    self.mlp = ModernBertMLP(config, prefix=f"{prefix}.mlp")

forward 

forward(
    hidden_states: Tensor, position_ids: Tensor
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
183
184
185
186
187
188
189
190
191
192
193
194
def forward(
    self,
    hidden_states: torch.Tensor,
    position_ids: torch.Tensor,
) -> torch.Tensor:
    attn_outputs = self.attn(
        hidden_states=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 
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
class ModernBertMLP(nn.Module):
    def __init__(self, config: ModernBertConfig, prefix: str = ""):
        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,
            prefix=f"{prefix}.Wo",
        )

    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,
    prefix=f"{prefix}.Wo",
)

act instance-attribute 

act = GELU()

config instance-attribute 

config = config

__init__ 

__init__(config: ModernBertConfig, prefix: str = '')
Source code in vllm/model_executor/models/modernbert.py 
144
145
146
147
148
149
150
151
152
153
154
155
156
def __init__(self, config: ModernBertConfig, prefix: str = ""):
    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,
        prefix=f"{prefix}.Wo",
    )

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
158
159
160
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 
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
269
270
271
272
273
274
275
276
277
278
279
280
@support_torch_compile
@default_pooling_type(seq_pooling_type="CLS")
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, prefix=f"{prefix}.encoder_layer"
        )
        self.final_norm = nn.LayerNorm(
            config.hidden_size, eps=config.norm_eps, bias=config.norm_bias
        )

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embeddings.embed_input_ids(input_ids)

    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: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor:
        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=positions,
        )
        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, prefix=f"{prefix}.encoder_layer"
)

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 
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
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, prefix=f"{prefix}.encoder_layer"
    )
    self.final_norm = nn.LayerNorm(
        config.hidden_size, eps=config.norm_eps, bias=config.norm_bias
    )

embed_input_ids 

embed_input_ids(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
245
246
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.embeddings.embed_input_ids(input_ids)

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: IntermediateTensors | None = None,
    inputs_embeds: Tensor | None = None,
) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: IntermediateTensors | None = None,
    inputs_embeds: torch.Tensor | None = None,
) -> torch.Tensor:
    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=positions,
    )
    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 
248
249
250
251
252
253
254
255
256
257
258
259
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: SequencePooler

Source code in vllm/model_executor/models/modernbert.py 
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
class ModernBertPooler(SequencePooler):
    def __init__(self, model_config: ModelConfig):
        pooler_config = model_config.pooler_config
        assert pooler_config is not None

        config: ModernBertConfig = model_config.hf_config
        hf_pooling_type = config.classifier_pooling.upper()
        # vllm_pooling_type = pooler_config.seq_pooling_type
        # Currently we don't have a way to see if the user set the pooling type
        # explicitly or not, so we always use the HF pooling type for now.

        super().__init__(
            pooling=get_seq_pooling_method(hf_pooling_type),
            # We set this dummy to avoid adding parameters to nn.Module too early
            head=nn.Identity(),
        )

        head_dtype = model_config.head_dtype
        self.dense = nn.Linear(
            config.hidden_size,
            config.hidden_size,
            config.classifier_bias,
            dtype=head_dtype,
        )
        self.act = nn.GELU()
        self.norm = nn.LayerNorm(
            config.hidden_size,
            eps=config.norm_eps,
            bias=config.norm_bias,
            dtype=head_dtype,
        )

        # Use lambdas so that weights are not registered under `self.head`
        self.head = EmbeddingPoolerHead(
            head_dtype=head_dtype,
            projector=lambda x: self.dense(x),
            activation=LambdaPoolerActivation(lambda x: self.norm(self.act(x))),
        )

act instance-attribute 

act = GELU()

dense instance-attribute 

dense = Linear(
    hidden_size,
    hidden_size,
    classifier_bias,
    dtype=head_dtype,
)

head instance-attribute 

head = EmbeddingPoolerHead(
    head_dtype=head_dtype,
    projector=lambda x: dense(x),
    activation=LambdaPoolerActivation(
        lambda x: norm(act(x))
    ),
)

norm instance-attribute 

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

__init__ 

__init__(model_config: ModelConfig)
Source code in vllm/model_executor/models/modernbert.py 
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
def __init__(self, model_config: ModelConfig):
    pooler_config = model_config.pooler_config
    assert pooler_config is not None

    config: ModernBertConfig = model_config.hf_config
    hf_pooling_type = config.classifier_pooling.upper()
    # vllm_pooling_type = pooler_config.seq_pooling_type
    # Currently we don't have a way to see if the user set the pooling type
    # explicitly or not, so we always use the HF pooling type for now.

    super().__init__(
        pooling=get_seq_pooling_method(hf_pooling_type),
        # We set this dummy to avoid adding parameters to nn.Module too early
        head=nn.Identity(),
    )

    head_dtype = model_config.head_dtype
    self.dense = nn.Linear(
        config.hidden_size,
        config.hidden_size,
        config.classifier_bias,
        dtype=head_dtype,
    )
    self.act = nn.GELU()
    self.norm = nn.LayerNorm(
        config.hidden_size,
        eps=config.norm_eps,
        bias=config.norm_bias,
        dtype=head_dtype,
    )

    # Use lambdas so that weights are not registered under `self.head`
    self.head = EmbeddingPoolerHead(
        head_dtype=head_dtype,
        projector=lambda x: self.dense(x),
        activation=LambdaPoolerActivation(lambda x: self.norm(self.act(x))),
    )

ModernBertPredictionHead

Bases: Module

Source code in vllm/model_executor/models/modernbert.py 
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
class ModernBertPredictionHead(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.dense = nn.Linear(
            config.hidden_size, config.hidden_size, bias=config.classifier_bias
        )
        self.act = ACT2FN[config.classifier_activation]
        self.norm = nn.LayerNorm(
            config.hidden_size,
            eps=getattr(config, "norm_eps", 1e-5),
            bias=getattr(config, "norm_bias", True),
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        return self.norm(self.act(self.dense(hidden_states)))

act instance-attribute 

act = ACT2FN[classifier_activation]

config instance-attribute 

config = config

dense instance-attribute 

dense = Linear(
    hidden_size, hidden_size, bias=classifier_bias
)

norm instance-attribute 

norm = LayerNorm(
    hidden_size,
    eps=getattr(config, "norm_eps", 1e-05),
    bias=getattr(config, "norm_bias", True),
)

__init__ 

__init__(config)
Source code in vllm/model_executor/models/modernbert.py 
395
396
397
398
399
400
401
402
403
404
405
406
def __init__(self, config):
    super().__init__()
    self.config = config
    self.dense = nn.Linear(
        config.hidden_size, config.hidden_size, bias=config.classifier_bias
    )
    self.act = ACT2FN[config.classifier_activation]
    self.norm = nn.LayerNorm(
        config.hidden_size,
        eps=getattr(config, "norm_eps", 1e-5),
        bias=getattr(config, "norm_bias", True),
    )

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/modernbert.py 
408
409
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    return self.norm(self.act(self.dense(hidden_states)))