Skip to content

vllm.model_executor.models.bert_with_rope

BertWithRope

Bases: Module, SupportsV0Only, SupportsQuant

Source code in vllm/model_executor/models/bert_with_rope.py 
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
class BertWithRope(nn.Module, SupportsV0Only, SupportsQuant):
    hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={"model.": ""})

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        self.vllm_config = vllm_config
        self.config = self.config_verify(vllm_config)
        self.embeddings = BertWithRopeEmbedding(self.config)
        self.encoder = BertWithRopeEncoder(
            vllm_config=vllm_config,
            bias=getattr(self.config, "bias", True),
            rotary_kwargs=self.config.rotary_kwargs,
            prefix=f"{prefix}.encoder")

    def config_verify(self, vllm_config):
        raise NotImplementedError

    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:
        if inputs_embeds is not None:
            hidden_states = inputs_embeds
        else:
            hidden_states = self.embeddings(input_ids=input_ids,
                                            token_type_ids=token_type_ids)
        return self.encoder(positions, hidden_states)

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

        if self.config.hidden_act in ["silu", "geglu"]:
            stacked_params_mapping = [
                # (param_name, shard_name, shard_id)
                ("gate_up_proj", "gate_proj", 0),
                ("gate_up_proj", "up_proj", 1),
            ]
        else:
            stacked_params_mapping = []

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "pooler" in name:
                continue
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                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

config instance-attribute 

config = config_verify(vllm_config)

embeddings instance-attribute 

embeddings = BertWithRopeEmbedding(config)

encoder instance-attribute 

encoder = BertWithRopeEncoder(
    vllm_config=vllm_config,
    bias=getattr(config, "bias", True),
    rotary_kwargs=rotary_kwargs,
    prefix=f"{prefix}.encoder",
)

hf_to_vllm_mapper class-attribute instance-attribute 

hf_to_vllm_mapper = WeightsMapper(
    orig_to_new_prefix={"model.": ""}
)

vllm_config instance-attribute 

vllm_config = vllm_config

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/bert_with_rope.py 
403
404
405
406
407
408
409
410
411
412
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    self.vllm_config = vllm_config
    self.config = self.config_verify(vllm_config)
    self.embeddings = BertWithRopeEmbedding(self.config)
    self.encoder = BertWithRopeEncoder(
        vllm_config=vllm_config,
        bias=getattr(self.config, "bias", True),
        rotary_kwargs=self.config.rotary_kwargs,
        prefix=f"{prefix}.encoder")

config_verify 

config_verify(vllm_config)
Source code in vllm/model_executor/models/bert_with_rope.py 
414
415
def config_verify(self, vllm_config):
    raise NotImplementedError

forward 

forward(
    input_ids: Optional[Tensor],
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    token_type_ids: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/bert_with_rope.py 
417
418
419
420
421
422
423
424
425
426
427
428
429
430
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:
    if inputs_embeds is not None:
        hidden_states = inputs_embeds
    else:
        hidden_states = self.embeddings(input_ids=input_ids,
                                        token_type_ids=token_type_ids)
    return self.encoder(positions, hidden_states)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/bert_with_rope.py 
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
459
460
461
462
463
464
465
466
467
468
469
470
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    weights = self.hf_to_vllm_mapper.apply(weights)

    if self.config.hidden_act in ["silu", "geglu"]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
    else:
        stacked_params_mapping = []

    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if "pooler" in name:
            continue
        for (param_name, weight_name, shard_id) in stacked_params_mapping:
            if weight_name not in name:
                continue
            name = name.replace(weight_name, param_name)
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue
            param = params_dict[name]
            weight_loader = param.weight_loader
            weight_loader(param, loaded_weight, shard_id)
            break
        else:
            # Skip loading extra bias for GPTQ models.
            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

BertWithRopeAttention

Bases: Module

Source code in vllm/model_executor/models/bert_with_rope.py 
 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 BertWithRopeAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_attention_heads: int,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = True,
        rotary_kwargs: Optional[dict] = None,
        prefix: str = "",
    ):
        super().__init__()

        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()

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

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

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

        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5

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

        self.rotary_emb = get_rope(**rotary_kwargs)

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

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

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

attn instance-attribute 

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

head_dim instance-attribute 

head_dim = hidden_size // total_num_heads

hidden_size instance-attribute 

hidden_size = hidden_size

kv_size instance-attribute 

kv_size = num_kv_heads * head_dim

num_heads instance-attribute 

num_heads = total_num_heads // tp_size

num_kv_heads instance-attribute 

num_kv_heads = max(1, total_num_kv_heads // tp_size)

out_proj instance-attribute 

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

q_size instance-attribute 

q_size = num_heads * head_dim

qkv_proj instance-attribute 

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

rotary_emb instance-attribute 

rotary_emb = get_rope(**rotary_kwargs)

scaling instance-attribute 

scaling = head_dim ** -0.5

total_num_heads instance-attribute 

total_num_heads = num_attention_heads

total_num_kv_heads instance-attribute 

total_num_kv_heads = total_num_heads

__init__ 

__init__(
    hidden_size: int,
    num_attention_heads: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    bias: bool = True,
    rotary_kwargs: Optional[dict] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert_with_rope.py 
 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
def __init__(
    self,
    hidden_size: int,
    num_attention_heads: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    bias: bool = True,
    rotary_kwargs: Optional[dict] = None,
    prefix: str = "",
):
    super().__init__()

    self.hidden_size = hidden_size
    tp_size = get_tensor_model_parallel_world_size()

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

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

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

    self.q_size = self.num_heads * self.head_dim
    self.kv_size = self.num_kv_heads * self.head_dim
    self.scaling = self.head_dim**-0.5

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

    self.rotary_emb = get_rope(**rotary_kwargs)

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

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

forward 

forward(positions: Tensor, hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/bert_with_rope.py 
130
131
132
133
134
135
136
137
138
139
140
def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    qkv, _ = self.qkv_proj(hidden_states)
    q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
    q, k = self.rotary_emb(positions, q, k)
    attn_output = self.attn(q, k, v)
    output, _ = self.out_proj(attn_output)
    return output

BertWithRopeBlock

Bases: Module

Source code in vllm/model_executor/models/bert_with_rope.py 
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
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
class BertWithRopeBlock(nn.Module):

    def __init__(self,
                 config: PretrainedConfig,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 moe: bool = False,
                 bias: bool = True,
                 rotary_kwargs: Optional[dict] = None,
                 prefix: str = ""):
        super().__init__()
        self.attn = BertWithRopeAttention(
            hidden_size=config.hidden_size,
            num_attention_heads=config.num_attention_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            bias=bias,
            rotary_kwargs=rotary_kwargs,
            prefix=f"{prefix}.attention")

        if moe:
            self.mlp = NomicMoELayer(config=config, )
        else:
            if config.hidden_act in ["silu", "geglu"]:
                self.mlp = BertWithRopeGatedMLP(
                    hidden_size=config.hidden_size,
                    intermediate_size=config.intermediate_size,
                    hidden_act=config.hidden_act,
                    bias=bias,
                    quant_config=quant_config,
                    prefix=f"{prefix}.mlp")
            else:
                self.mlp = BertWithRopeMLP(
                    hidden_size=config.hidden_size,
                    intermediate_size=config.intermediate_size,
                    hidden_act=config.hidden_act,
                    bias=bias,
                    quant_config=quant_config,
                    prefix=f"{prefix}.mlp")

        self.attn_ln = nn.LayerNorm(config.hidden_size,
                                    eps=config.layer_norm_eps)
        self.mlp_ln = nn.LayerNorm(config.hidden_size,
                                   eps=config.layer_norm_eps)

    def forward(self, positions: torch.Tensor, hidden_states: torch.Tensor):
        attn_output = self.attn(positions, hidden_states)
        hidden_states = self.attn_ln(hidden_states + attn_output)
        mlp_out = self.mlp(hidden_states)
        hidden_states = self.mlp_ln(hidden_states + mlp_out)
        return hidden_states

attn instance-attribute 

attn = BertWithRopeAttention(
    hidden_size=hidden_size,
    num_attention_heads=num_attention_heads,
    cache_config=cache_config,
    quant_config=quant_config,
    bias=bias,
    rotary_kwargs=rotary_kwargs,
    prefix=f"{prefix}.attention",
)

attn_ln instance-attribute 

attn_ln = LayerNorm(hidden_size, eps=layer_norm_eps)

mlp instance-attribute 

mlp = NomicMoELayer(config=config)

mlp_ln instance-attribute 

mlp_ln = LayerNorm(hidden_size, eps=layer_norm_eps)

__init__ 

__init__(
    config: PretrainedConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    moe: bool = False,
    bias: bool = True,
    rotary_kwargs: Optional[dict] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert_with_rope.py 
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
343
344
345
346
347
348
349
350
351
352
353
354
355
356
def __init__(self,
             config: PretrainedConfig,
             cache_config: Optional[CacheConfig] = None,
             quant_config: Optional[QuantizationConfig] = None,
             moe: bool = False,
             bias: bool = True,
             rotary_kwargs: Optional[dict] = None,
             prefix: str = ""):
    super().__init__()
    self.attn = BertWithRopeAttention(
        hidden_size=config.hidden_size,
        num_attention_heads=config.num_attention_heads,
        cache_config=cache_config,
        quant_config=quant_config,
        bias=bias,
        rotary_kwargs=rotary_kwargs,
        prefix=f"{prefix}.attention")

    if moe:
        self.mlp = NomicMoELayer(config=config, )
    else:
        if config.hidden_act in ["silu", "geglu"]:
            self.mlp = BertWithRopeGatedMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                bias=bias,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp")
        else:
            self.mlp = BertWithRopeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                bias=bias,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp")

    self.attn_ln = nn.LayerNorm(config.hidden_size,
                                eps=config.layer_norm_eps)
    self.mlp_ln = nn.LayerNorm(config.hidden_size,
                               eps=config.layer_norm_eps)

forward 

forward(positions: Tensor, hidden_states: Tensor)
Source code in vllm/model_executor/models/bert_with_rope.py 
358
359
360
361
362
363
def forward(self, positions: torch.Tensor, hidden_states: torch.Tensor):
    attn_output = self.attn(positions, hidden_states)
    hidden_states = self.attn_ln(hidden_states + attn_output)
    mlp_out = self.mlp(hidden_states)
    hidden_states = self.mlp_ln(hidden_states + mlp_out)
    return hidden_states

BertWithRopeEmbedding

Bases: Module

Source code in vllm/model_executor/models/bert_with_rope.py 
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
class BertWithRopeEmbedding(nn.Module):

    def __init__(self, config: PretrainedConfig):

        super().__init__()
        if config.position_embedding_type not in ["rope", "rotary"]:
            raise ValueError("Only 'rotary'('rope') position_embedding_type" +
                             " is supported")

        self.word_embeddings = VocabParallelEmbedding(config.vocab_size,
                                                      config.hidden_size)
        if config.type_vocab_size > 0:
            self.token_type_embeddings = VocabParallelEmbedding(
                config.type_vocab_size, config.hidden_size)
        else:
            self.token_type_embeddings = None

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

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

        embeddings = inputs_embeds
        if self.token_type_embeddings is not None:
            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 += token_type_embeddings

        embeddings = self.LayerNorm(embeddings)
        return embeddings

LayerNorm instance-attribute 

LayerNorm = LayerNorm(hidden_size, eps=layer_norm_eps)

token_type_embeddings instance-attribute 

token_type_embeddings = VocabParallelEmbedding(
    type_vocab_size, hidden_size
)

word_embeddings instance-attribute 

word_embeddings = VocabParallelEmbedding(
    vocab_size, hidden_size
)

__init__ 

__init__(config: PretrainedConfig)
Source code in vllm/model_executor/models/bert_with_rope.py 
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
def __init__(self, config: PretrainedConfig):

    super().__init__()
    if config.position_embedding_type not in ["rope", "rotary"]:
        raise ValueError("Only 'rotary'('rope') position_embedding_type" +
                         " is supported")

    self.word_embeddings = VocabParallelEmbedding(config.vocab_size,
                                                  config.hidden_size)
    if config.type_vocab_size > 0:
        self.token_type_embeddings = VocabParallelEmbedding(
            config.type_vocab_size, config.hidden_size)
    else:
        self.token_type_embeddings = None

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

forward 

forward(
    input_ids: Tensor,
    token_type_ids: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/bert_with_rope.py 
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
def forward(
    self,
    input_ids: torch.Tensor,
    token_type_ids: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    input_shape = input_ids.size()
    inputs_embeds = self.word_embeddings(input_ids)

    embeddings = inputs_embeds
    if self.token_type_embeddings is not None:
        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 += token_type_embeddings

    embeddings = self.LayerNorm(embeddings)
    return embeddings

BertWithRopeEncoder

Bases: Module

Source code in vllm/model_executor/models/bert_with_rope.py 
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
@support_torch_compile
class BertWithRopeEncoder(nn.Module):

    def __init__(self,
                 vllm_config: VllmConfig,
                 bias: bool = True,
                 rotary_kwargs: Optional[dict] = None,
                 prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        every_n = getattr(config, "moe_every_n_layers", 0)
        self.layers = nn.ModuleList([
            BertWithRopeBlock(config=config,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              bias=bias,
                              moe=every_n > 0 and (layer_idx % every_n == 1),
                              rotary_kwargs=rotary_kwargs,
                              prefix=f"{prefix}.layer.{layer_idx}")
            for layer_idx in range(config.num_hidden_layers)
        ])

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

layers instance-attribute 

layers = ModuleList(
    [
        BertWithRopeBlock(
            config=config,
            cache_config=cache_config,
            quant_config=quant_config,
            bias=bias,
            moe=every_n > 0 and layer_idx % every_n == 1,
            rotary_kwargs=rotary_kwargs,
            prefix=f"{prefix}.layer.{layer_idx}",
        )
        for layer_idx in range(num_hidden_layers)
    ]
)

__init__ 

__init__(
    vllm_config: VllmConfig,
    bias: bool = True,
    rotary_kwargs: Optional[dict] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert_with_rope.py 
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
def __init__(self,
             vllm_config: VllmConfig,
             bias: bool = True,
             rotary_kwargs: Optional[dict] = None,
             prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    cache_config = vllm_config.cache_config
    quant_config = vllm_config.quant_config
    every_n = getattr(config, "moe_every_n_layers", 0)
    self.layers = nn.ModuleList([
        BertWithRopeBlock(config=config,
                          cache_config=cache_config,
                          quant_config=quant_config,
                          bias=bias,
                          moe=every_n > 0 and (layer_idx % every_n == 1),
                          rotary_kwargs=rotary_kwargs,
                          prefix=f"{prefix}.layer.{layer_idx}")
        for layer_idx in range(config.num_hidden_layers)
    ])

forward 

forward(positions: Tensor, hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/bert_with_rope.py 
390
391
392
393
394
395
396
397
def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    for layer in self.layers:
        hidden_states = layer(positions, hidden_states)
    return hidden_states

BertWithRopeGatedMLP

Bases: Module

Source code in vllm/model_executor/models/bert_with_rope.py 
143
144
145
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
class BertWithRopeGatedMLP(nn.Module):

    def __init__(self,
                 hidden_size: int,
                 intermediate_size: int,
                 hidden_act: str,
                 bias: bool = True,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.act_fn = get_act_and_mul_fn(hidden_act)
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(input_size=intermediate_size,
                                           output_size=hidden_size,
                                           bias=bias,
                                           quant_config=quant_config,
                                           prefix=f"{prefix}.down_proj")

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        gate_up, _ = self.gate_up_proj(hidden_states)
        hidden_states = self.act_fn(gate_up)
        hidden_states, _ = self.down_proj(hidden_states)
        return hidden_states

act_fn instance-attribute 

act_fn = get_act_and_mul_fn(hidden_act)

down_proj instance-attribute 

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

gate_up_proj instance-attribute 

gate_up_proj = MergedColumnParallelLinear(
    hidden_size,
    [intermediate_size] * 2,
    bias=bias,
    quant_config=quant_config,
    prefix=f"{prefix}.gate_up_proj",
)

__init__ 

__init__(
    hidden_size: int,
    intermediate_size: int,
    hidden_act: str,
    bias: bool = True,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert_with_rope.py 
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
def __init__(self,
             hidden_size: int,
             intermediate_size: int,
             hidden_act: str,
             bias: bool = True,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()
    self.act_fn = get_act_and_mul_fn(hidden_act)
    self.gate_up_proj = MergedColumnParallelLinear(
        hidden_size,
        [intermediate_size] * 2,
        bias=bias,
        quant_config=quant_config,
        prefix=f"{prefix}.gate_up_proj",
    )
    self.down_proj = RowParallelLinear(input_size=intermediate_size,
                                       output_size=hidden_size,
                                       bias=bias,
                                       quant_config=quant_config,
                                       prefix=f"{prefix}.down_proj")

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/bert_with_rope.py 
167
168
169
170
171
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    gate_up, _ = self.gate_up_proj(hidden_states)
    hidden_states = self.act_fn(gate_up)
    hidden_states, _ = self.down_proj(hidden_states)
    return hidden_states

BertWithRopeMLP

Bases: Module

Source code in vllm/model_executor/models/bert_with_rope.py 
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
class BertWithRopeMLP(nn.Module):

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

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.up_proj(hidden_states)
        hidden_states = self.act_fn(hidden_states)
        hidden_states, _ = self.down_proj(hidden_states)
        return hidden_states

act_fn instance-attribute 

act_fn = get_act_fn(hidden_act)

down_proj instance-attribute 

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

up_proj instance-attribute 

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

__init__ 

__init__(
    hidden_size: int,
    intermediate_size: int,
    hidden_act: str,
    bias: bool = True,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/bert_with_rope.py 
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
def __init__(self,
             hidden_size: int,
             intermediate_size: int,
             hidden_act: str,
             bias: bool = True,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()
    self.act_fn = get_act_fn(hidden_act)
    self.up_proj = ColumnParallelLinear(input_size=hidden_size,
                                        output_size=intermediate_size,
                                        bias=bias,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.up_proj")
    self.down_proj = RowParallelLinear(input_size=intermediate_size,
                                       output_size=hidden_size,
                                       bias=bias,
                                       quant_config=quant_config,
                                       prefix=f"{prefix}.down_proj")

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/bert_with_rope.py 
196
197
198
199
200
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    hidden_states, _ = self.up_proj(hidden_states)
    hidden_states = self.act_fn(hidden_states)
    hidden_states, _ = self.down_proj(hidden_states)
    return hidden_states

GteNewModel

Bases: BertWithRope

Source code in vllm/model_executor/models/bert_with_rope.py 
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
class GteNewModel(BertWithRope):
    # for https://huggingface.co/Alibaba-NLP/new-impl

    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_substr={
            "new.": "",
            "layer": "layers",
            "attention.qkv_proj": "attn.qkv_proj",
            "attention.o_proj": "attn.out_proj",
        })

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

        # GteNewModel only gate_up_proj does not have bias.
        # Hack method learned from vllm/model_executor/models/glm.py
        for layer in self.encoder.layers:
            layer.mlp.gate_up_proj.bias = None
            layer.mlp.gate_up_proj.skip_bias_add = True

    def config_verify(self, vllm_config):
        config = vllm_config.model_config.hf_config

        assert config.__class__.__name__ == "NewConfig"
        assert config.hidden_act == "gelu"

        config.hidden_act = "geglu"

        head_dim = config.hidden_size // config.num_attention_heads
        config.rotary_kwargs = {
            "head_size": head_dim,
            "rotary_dim": getattr(config, "rotary_emb_dim", head_dim),
            "max_position": config.max_position_embeddings,
            "base": config.rope_theta,
            "rope_scaling": getattr(config, "rope_scaling", None)
        }
        return config

    def split_up_gate_proj(self, weights: Iterable[tuple[str, torch.Tensor]]):
        n = "mlp.up_gate_proj"
        for name, weight in weights:
            if n in name:
                up, gate = weight.chunk(2, dim=0)
                yield name.replace(n, "mlp.up_proj"), up
                yield name.replace(n, "mlp.gate_proj"), gate
            else:
                yield name, weight

    def ignore_unnecessary_layers(self,
                                  weights: Iterable[tuple[str, torch.Tensor]]):
        for name, weight in weights:
            if name.startswith("classifier"):
                continue
            yield name, weight

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        weights = self.ignore_unnecessary_layers(weights)
        weights = self.split_up_gate_proj(weights)
        return super().load_weights(weights)

hf_to_vllm_mapper class-attribute instance-attribute 

hf_to_vllm_mapper = WeightsMapper(
    orig_to_new_substr={
        "new.": "",
        "layer": "layers",
        "attention.qkv_proj": "attn.qkv_proj",
        "attention.o_proj": "attn.out_proj",
    }
)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/bert_with_rope.py 
544
545
546
547
548
549
550
551
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__(vllm_config=vllm_config, prefix=prefix)

    # GteNewModel only gate_up_proj does not have bias.
    # Hack method learned from vllm/model_executor/models/glm.py
    for layer in self.encoder.layers:
        layer.mlp.gate_up_proj.bias = None
        layer.mlp.gate_up_proj.skip_bias_add = True

config_verify 

config_verify(vllm_config)
Source code in vllm/model_executor/models/bert_with_rope.py 
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
def config_verify(self, vllm_config):
    config = vllm_config.model_config.hf_config

    assert config.__class__.__name__ == "NewConfig"
    assert config.hidden_act == "gelu"

    config.hidden_act = "geglu"

    head_dim = config.hidden_size // config.num_attention_heads
    config.rotary_kwargs = {
        "head_size": head_dim,
        "rotary_dim": getattr(config, "rotary_emb_dim", head_dim),
        "max_position": config.max_position_embeddings,
        "base": config.rope_theta,
        "rope_scaling": getattr(config, "rope_scaling", None)
    }
    return config

ignore_unnecessary_layers 

ignore_unnecessary_layers(
    weights: Iterable[tuple[str, Tensor]],
)
Source code in vllm/model_executor/models/bert_with_rope.py 
581
582
583
584
585
586
def ignore_unnecessary_layers(self,
                              weights: Iterable[tuple[str, torch.Tensor]]):
    for name, weight in weights:
        if name.startswith("classifier"):
            continue
        yield name, weight

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/bert_with_rope.py 
588
589
590
591
592
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    weights = self.ignore_unnecessary_layers(weights)
    weights = self.split_up_gate_proj(weights)
    return super().load_weights(weights)

split_up_gate_proj 

split_up_gate_proj(weights: Iterable[tuple[str, Tensor]])
Source code in vllm/model_executor/models/bert_with_rope.py 
571
572
573
574
575
576
577
578
579
def split_up_gate_proj(self, weights: Iterable[tuple[str, torch.Tensor]]):
    n = "mlp.up_gate_proj"
    for name, weight in weights:
        if n in name:
            up, gate = weight.chunk(2, dim=0)
            yield name.replace(n, "mlp.up_proj"), up
            yield name.replace(n, "mlp.gate_proj"), gate
        else:
            yield name, weight

JinaRobertaModel

Bases: BertWithRope

Source code in vllm/model_executor/models/bert_with_rope.py 
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
class JinaRobertaModel(BertWithRope):
    # for https://huggingface.co/jinaai/jina-embeddings-v3

    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_substr={
            "emb_ln": "embeddings.LayerNorm",
            "mixer.Wqkv": "attn.qkv_proj",
            "mixer.out_proj": "attn.out_proj",
            "norm1": "attn_ln",
            "mlp.fc1.": "mlp.up_proj.",
            "mlp.fc2": "mlp.down_proj",
            "norm2": "mlp_ln",
        })

    def config_verify(self, vllm_config):
        config = vllm_config.model_config.hf_config

        assert config.__class__.__name__ == "XLMRobertaFlashConfig"

        head_dim = config.hidden_size // config.num_attention_heads
        config.rotary_kwargs = {
            "head_size": head_dim,
            "rotary_dim": getattr(config, "rotary_emb_dim", head_dim),
            "max_position": config.max_position_embeddings,
            "base": getattr(config, "rope_theta", config.rotary_emb_base),
            "rope_scaling": getattr(config, "rope_scaling", None)
        }
        return config

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

    @torch.inference_mode()
    def jina_merge_lora_weights(self, weights: Iterable[tuple[str,
                                                              torch.Tensor]]):
        # use for jina-embeddings-v3
        # Merge Lora weights into a single weight tensor.
        # This is a temporary solution until we have a better way to handle

        scaling = self.config.lora_alpha / self.config.lora_rank
        device = self.vllm_config.device_config.device

        weights = {name: weight for name, weight in weights}

        o = ".original"
        a = ".0.lora_A"
        b = ".0.lora_B"

        # text-matching
        i = -1

        for name in list(weights.keys()):
            if o in name:
                dtype = weights[name].dtype
                shape = weights[name].shape
                weight_name = name[:-len(o)]

                if "embeddings" in weight_name:
                    B = weights[weight_name + a][i].to(device).float()
                    A = weights[weight_name + b][i].to(device).float()
                else:
                    B = weights[weight_name + b][i].to(device).float()
                    A = weights[weight_name + a][i].to(device).float()

                weight = (weights[weight_name + o].to(device) +
                          torch.matmul(B, A).view(shape) * scaling)
                weight = weight.cpu().to(dtype)

                weights[weight_name.replace(".parametrizations", "")] = weight

                del weights[weight_name + o], weights[weight_name +
                                                      a], weights[weight_name +
                                                                  b]

        return [(name, weight) for name, weight in weights.items()]

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        weights = self.jina_merge_lora_weights(weights)
        return super().load_weights(weights)

hf_to_vllm_mapper class-attribute instance-attribute 

hf_to_vllm_mapper = WeightsMapper(
    orig_to_new_substr={
        "emb_ln": "embeddings.LayerNorm",
        "mixer.Wqkv": "attn.qkv_proj",
        "mixer.out_proj": "attn.out_proj",
        "norm1": "attn_ln",
        "mlp.fc1.": "mlp.up_proj.",
        "mlp.fc2": "mlp.down_proj",
        "norm2": "mlp_ln",
    }
)

config_verify 

config_verify(vllm_config)
Source code in vllm/model_executor/models/bert_with_rope.py 
638
639
640
641
642
643
644
645
646
647
648
649
650
651
def config_verify(self, vllm_config):
    config = vllm_config.model_config.hf_config

    assert config.__class__.__name__ == "XLMRobertaFlashConfig"

    head_dim = config.hidden_size // config.num_attention_heads
    config.rotary_kwargs = {
        "head_size": head_dim,
        "rotary_dim": getattr(config, "rotary_emb_dim", head_dim),
        "max_position": config.max_position_embeddings,
        "base": getattr(config, "rope_theta", config.rotary_emb_base),
        "rope_scaling": getattr(config, "rope_scaling", None)
    }
    return config

forward 

forward(
    input_ids: Tensor,
    position_ids: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    token_type_ids: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/bert_with_rope.py 
653
654
655
656
657
658
659
660
661
662
663
664
665
def forward(
    self,
    input_ids: torch.Tensor,
    position_ids: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
    token_type_ids: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    return super().forward(input_ids=input_ids,
                           positions=position_ids,
                           intermediate_tensors=intermediate_tensors,
                           inputs_embeds=inputs_embeds,
                           token_type_ids=token_type_ids)

jina_merge_lora_weights 

jina_merge_lora_weights(
    weights: Iterable[tuple[str, Tensor]],
)
Source code in vllm/model_executor/models/bert_with_rope.py 
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
@torch.inference_mode()
def jina_merge_lora_weights(self, weights: Iterable[tuple[str,
                                                          torch.Tensor]]):
    # use for jina-embeddings-v3
    # Merge Lora weights into a single weight tensor.
    # This is a temporary solution until we have a better way to handle

    scaling = self.config.lora_alpha / self.config.lora_rank
    device = self.vllm_config.device_config.device

    weights = {name: weight for name, weight in weights}

    o = ".original"
    a = ".0.lora_A"
    b = ".0.lora_B"

    # text-matching
    i = -1

    for name in list(weights.keys()):
        if o in name:
            dtype = weights[name].dtype
            shape = weights[name].shape
            weight_name = name[:-len(o)]

            if "embeddings" in weight_name:
                B = weights[weight_name + a][i].to(device).float()
                A = weights[weight_name + b][i].to(device).float()
            else:
                B = weights[weight_name + b][i].to(device).float()
                A = weights[weight_name + a][i].to(device).float()

            weight = (weights[weight_name + o].to(device) +
                      torch.matmul(B, A).view(shape) * scaling)
            weight = weight.cpu().to(dtype)

            weights[weight_name.replace(".parametrizations", "")] = weight

            del weights[weight_name + o], weights[weight_name +
                                                  a], weights[weight_name +
                                                              b]

    return [(name, weight) for name, weight in weights.items()]

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/bert_with_rope.py 
711
712
713
714
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    weights = self.jina_merge_lora_weights(weights)
    return super().load_weights(weights)

NomicBertModel

Bases: BertWithRope

Source code in vllm/model_executor/models/bert_with_rope.py 
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
class NomicBertModel(BertWithRope):
    # for https://huggingface.co/nomic-ai/nomic-bert-2048

    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_substr={
            "emb_ln": "embeddings.LayerNorm",
            "attn.Wqkv": "attn.qkv_proj",
            "norm1": "attn_ln",
            "mlp.fc1.": "mlp.up_proj.",
            "mlp.fc11": "mlp.up_proj",
            "mlp.fc12": "mlp.gate_proj",
            "mlp.fc2": "mlp.down_proj",
            "norm2": "mlp_ln",
        })

    def config_verify(self, vllm_config):
        config = vllm_config.model_config.hf_config

        assert config.__class__.__name__ == "NomicBertConfig"
        assert config.activation_function in ["swiglu", "gelu"]
        config.position_embedding_type = getattr(config,
                                                 "position_embedding_type",
                                                 "rope")

        if config.activation_function == "swiglu":
            config.hidden_act = "silu"
        else:
            config.hidden_act = config.activation_function

        assert (config.mlp_fc1_bias == config.mlp_fc2_bias ==
                config.qkv_proj_bias)
        config.bias = config.qkv_proj_bias

        assert config.rotary_emb_scale_base is None
        assert not config.rotary_emb_interleaved

        config.layer_norm_eps = config.layer_norm_epsilon
        config.intermediate_size = config.n_inner
        config.hidden_size = config.n_embd
        config.num_hidden_layers = config.n_layer

        head_dim = config.hidden_size // config.num_attention_heads
        rotary_emb_dim = head_dim * config.rotary_emb_fraction
        config.rotary_kwargs = {
            "head_size": head_dim,
            "rotary_dim": rotary_emb_dim,
            "max_position": config.max_trained_positions,
            "base": getattr(config, "rope_theta", config.rotary_emb_base),
            "rope_scaling": getattr(config, "rope_scaling", None)
        }

        # we ignore config.rotary_scaling_factor so that for datasets shorter
        # than max_trained_positions 2048, the results are consistent
        # with SentenceTransformer.
        # The context extension uses vllm style rope_theta and rope_scaling.
        # See #17785

        return config

hf_to_vllm_mapper class-attribute instance-attribute 

hf_to_vllm_mapper = WeightsMapper(
    orig_to_new_substr={
        "emb_ln": "embeddings.LayerNorm",
        "attn.Wqkv": "attn.qkv_proj",
        "norm1": "attn_ln",
        "mlp.fc1.": "mlp.up_proj.",
        "mlp.fc11": "mlp.up_proj",
        "mlp.fc12": "mlp.gate_proj",
        "mlp.fc2": "mlp.down_proj",
        "norm2": "mlp_ln",
    }
)

config_verify 

config_verify(vllm_config)
Source code in vllm/model_executor/models/bert_with_rope.py 
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
def config_verify(self, vllm_config):
    config = vllm_config.model_config.hf_config

    assert config.__class__.__name__ == "NomicBertConfig"
    assert config.activation_function in ["swiglu", "gelu"]
    config.position_embedding_type = getattr(config,
                                             "position_embedding_type",
                                             "rope")

    if config.activation_function == "swiglu":
        config.hidden_act = "silu"
    else:
        config.hidden_act = config.activation_function

    assert (config.mlp_fc1_bias == config.mlp_fc2_bias ==
            config.qkv_proj_bias)
    config.bias = config.qkv_proj_bias

    assert config.rotary_emb_scale_base is None
    assert not config.rotary_emb_interleaved

    config.layer_norm_eps = config.layer_norm_epsilon
    config.intermediate_size = config.n_inner
    config.hidden_size = config.n_embd
    config.num_hidden_layers = config.n_layer

    head_dim = config.hidden_size // config.num_attention_heads
    rotary_emb_dim = head_dim * config.rotary_emb_fraction
    config.rotary_kwargs = {
        "head_size": head_dim,
        "rotary_dim": rotary_emb_dim,
        "max_position": config.max_trained_positions,
        "base": getattr(config, "rope_theta", config.rotary_emb_base),
        "rope_scaling": getattr(config, "rope_scaling", None)
    }

    # we ignore config.rotary_scaling_factor so that for datasets shorter
    # than max_trained_positions 2048, the results are consistent
    # with SentenceTransformer.
    # The context extension uses vllm style rope_theta and rope_scaling.
    # See #17785

    return config

NomicExpertMLP

Bases: Module

Source code in vllm/model_executor/models/bert_with_rope.py 
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
class NomicExpertMLP(nn.Module):

    def __init__(self, hidden_size: int, ffn_hidden_size: int,
                 moe_num_experts: int, ffn_act_fn: str):
        super().__init__()
        self.hidden_size = hidden_size
        self.ffn_hidden_size = ffn_hidden_size
        self.moe_num_experts = moe_num_experts

        self.w1 = nn.Parameter(
            torch.empty(moe_num_experts * ffn_hidden_size, hidden_size))
        self.w2 = nn.Parameter(
            torch.empty(moe_num_experts * ffn_hidden_size, hidden_size))
        self.activation_fn = get_act_fn(ffn_act_fn)

    def forward(self, x: torch.Tensor, expert_idx: int) -> torch.Tensor:
        expert_w1 = self.w1.view(self.moe_num_experts, self.ffn_hidden_size,
                                 self.hidden_size)[expert_idx]
        expert_w2 = self.w2.view(self.moe_num_experts, self.ffn_hidden_size,
                                 self.hidden_size)[expert_idx]

        x1 = x.matmul(expert_w1.t())
        act_out = self.activation_fn(x1)
        x2 = act_out.matmul(expert_w2)
        return x2

activation_fn instance-attribute 

activation_fn = get_act_fn(ffn_act_fn)

ffn_hidden_size instance-attribute 

ffn_hidden_size = ffn_hidden_size

hidden_size instance-attribute 

hidden_size = hidden_size

moe_num_experts instance-attribute 

moe_num_experts = moe_num_experts

w1 instance-attribute 

w1 = Parameter(
    empty(moe_num_experts * ffn_hidden_size, hidden_size)
)

w2 instance-attribute 

w2 = Parameter(
    empty(moe_num_experts * ffn_hidden_size, hidden_size)
)

__init__ 

__init__(
    hidden_size: int,
    ffn_hidden_size: int,
    moe_num_experts: int,
    ffn_act_fn: str,
)
Source code in vllm/model_executor/models/bert_with_rope.py 
223
224
225
226
227
228
229
230
231
232
233
234
def __init__(self, hidden_size: int, ffn_hidden_size: int,
             moe_num_experts: int, ffn_act_fn: str):
    super().__init__()
    self.hidden_size = hidden_size
    self.ffn_hidden_size = ffn_hidden_size
    self.moe_num_experts = moe_num_experts

    self.w1 = nn.Parameter(
        torch.empty(moe_num_experts * ffn_hidden_size, hidden_size))
    self.w2 = nn.Parameter(
        torch.empty(moe_num_experts * ffn_hidden_size, hidden_size))
    self.activation_fn = get_act_fn(ffn_act_fn)

forward 

forward(x: Tensor, expert_idx: int) -> Tensor
Source code in vllm/model_executor/models/bert_with_rope.py 
236
237
238
239
240
241
242
243
244
245
def forward(self, x: torch.Tensor, expert_idx: int) -> torch.Tensor:
    expert_w1 = self.w1.view(self.moe_num_experts, self.ffn_hidden_size,
                             self.hidden_size)[expert_idx]
    expert_w2 = self.w2.view(self.moe_num_experts, self.ffn_hidden_size,
                             self.hidden_size)[expert_idx]

    x1 = x.matmul(expert_w1.t())
    act_out = self.activation_fn(x1)
    x2 = act_out.matmul(expert_w2)
    return x2

NomicExperts

Bases: Module

Source code in vllm/model_executor/models/bert_with_rope.py 
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
281
282
283
284
285
286
class NomicExperts(nn.Module):

    def __init__(self, config, hidden_size: int, ffn_hidden_size: int,
                 moe_num_experts: int):
        super().__init__()
        self.moe_num_experts = moe_num_experts

        self.mlp = NomicExpertMLP(hidden_size=config.n_embd,
                                  ffn_hidden_size=config.n_inner,
                                  moe_num_experts=moe_num_experts,
                                  ffn_act_fn=config.hidden_act)
        self.bias = nn.Parameter(torch.zeros(config.n_embd))

    def forward(self, x: torch.Tensor, weights: torch.Tensor,
                top_weights: torch.Tensor,
                top_experts: torch.LongTensor) -> torch.Tensor:
        q_len, hidden_size = x.shape
        x = x.view(-1, hidden_size)
        out = torch.zeros_like(x)

        expert_mask = nn.functional.one_hot(
            top_experts, num_classes=self.moe_num_experts).permute(2, 1, 0)
        for expert_idx in range(0, self.moe_num_experts):
            topk_idx, token_idx = torch.where(expert_mask[expert_idx])
            if token_idx.shape[0] == 0:
                continue

            token_list = token_idx.tolist()
            topk_list = topk_idx.tolist()

            expert_tokens = x[None, token_list].reshape(-1, hidden_size)
            expert_out = self.mlp(
                expert_tokens, expert_idx) * top_weights[token_list, topk_list,
                                                         None]

            out.index_add_(0, token_idx, expert_out)

        out = out.reshape(q_len, hidden_size)
        return out + self.bias

bias instance-attribute 

bias = Parameter(zeros(n_embd))

mlp instance-attribute 

mlp = NomicExpertMLP(
    hidden_size=n_embd,
    ffn_hidden_size=n_inner,
    moe_num_experts=moe_num_experts,
    ffn_act_fn=hidden_act,
)

moe_num_experts instance-attribute 

moe_num_experts = moe_num_experts

__init__ 

__init__(
    config,
    hidden_size: int,
    ffn_hidden_size: int,
    moe_num_experts: int,
)
Source code in vllm/model_executor/models/bert_with_rope.py 
250
251
252
253
254
255
256
257
258
259
def __init__(self, config, hidden_size: int, ffn_hidden_size: int,
             moe_num_experts: int):
    super().__init__()
    self.moe_num_experts = moe_num_experts

    self.mlp = NomicExpertMLP(hidden_size=config.n_embd,
                              ffn_hidden_size=config.n_inner,
                              moe_num_experts=moe_num_experts,
                              ffn_act_fn=config.hidden_act)
    self.bias = nn.Parameter(torch.zeros(config.n_embd))

forward 

forward(
    x: Tensor,
    weights: Tensor,
    top_weights: Tensor,
    top_experts: LongTensor,
) -> Tensor
Source code in vllm/model_executor/models/bert_with_rope.py 
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
def forward(self, x: torch.Tensor, weights: torch.Tensor,
            top_weights: torch.Tensor,
            top_experts: torch.LongTensor) -> torch.Tensor:
    q_len, hidden_size = x.shape
    x = x.view(-1, hidden_size)
    out = torch.zeros_like(x)

    expert_mask = nn.functional.one_hot(
        top_experts, num_classes=self.moe_num_experts).permute(2, 1, 0)
    for expert_idx in range(0, self.moe_num_experts):
        topk_idx, token_idx = torch.where(expert_mask[expert_idx])
        if token_idx.shape[0] == 0:
            continue

        token_list = token_idx.tolist()
        topk_list = topk_idx.tolist()

        expert_tokens = x[None, token_list].reshape(-1, hidden_size)
        expert_out = self.mlp(
            expert_tokens, expert_idx) * top_weights[token_list, topk_list,
                                                     None]

        out.index_add_(0, token_idx, expert_out)

    out = out.reshape(q_len, hidden_size)
    return out + self.bias

NomicMoELayer

Bases: Module

Source code in vllm/model_executor/models/bert_with_rope.py 
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
class NomicMoELayer(nn.Module):

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

        self.router = NomicRouter(
            config.n_embd,
            moe_num_experts=config.num_experts,
            moe_top_k=config.moe_top_k,
        )

        self.experts = NomicExperts(
            config,
            hidden_size=config.n_embd,
            ffn_hidden_size=config.n_inner,
            moe_num_experts=config.num_experts,
        )

    def forward(self, x: torch.Tensor):
        weights, top_weights, top_experts = self.router(x)
        out = self.experts(x, weights, top_weights, top_experts)
        return out

experts instance-attribute 

experts = NomicExperts(
    config,
    hidden_size=n_embd,
    ffn_hidden_size=n_inner,
    moe_num_experts=num_experts,
)

router instance-attribute 

router = NomicRouter(
    n_embd, moe_num_experts=num_experts, moe_top_k=moe_top_k
)

__init__ 

__init__(config: PretrainedConfig)
Source code in vllm/model_executor/models/bert_with_rope.py 
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
def __init__(self, config: PretrainedConfig):
    super().__init__()

    self.router = NomicRouter(
        config.n_embd,
        moe_num_experts=config.num_experts,
        moe_top_k=config.moe_top_k,
    )

    self.experts = NomicExperts(
        config,
        hidden_size=config.n_embd,
        ffn_hidden_size=config.n_inner,
        moe_num_experts=config.num_experts,
    )

forward 

forward(x: Tensor)
Source code in vllm/model_executor/models/bert_with_rope.py 
307
308
309
310
def forward(self, x: torch.Tensor):
    weights, top_weights, top_experts = self.router(x)
    out = self.experts(x, weights, top_weights, top_experts)
    return out

NomicRouter

Bases: Module

Source code in vllm/model_executor/models/bert_with_rope.py 
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
class NomicRouter(nn.Module):

    def __init__(self, hidden_size: int, moe_num_experts: int, moe_top_k: int):
        super().__init__()
        self.moe_top_k = moe_top_k
        self.layer = ReplicatedLinear(hidden_size, moe_num_experts, bias=False)

    def forward(
        self, x: torch.Tensor
    ) -> tuple[torch.Tensor, torch.Tensor, torch.LongTensor]:
        weights = self.layer(x.view(-1, x.shape[-1]))[0].softmax(
            dim=-1, dtype=torch.float32)
        top_weights, top_experts = torch.topk(weights, self.moe_top_k, dim=-1)
        weights = weights.to(x.dtype)
        top_weights = top_weights.to(x.dtype)
        return weights, top_weights, top_experts  # type: ignore

layer instance-attribute 

layer = ReplicatedLinear(
    hidden_size, moe_num_experts, bias=False
)

moe_top_k instance-attribute 

moe_top_k = moe_top_k

__init__ 

__init__(
    hidden_size: int, moe_num_experts: int, moe_top_k: int
)
Source code in vllm/model_executor/models/bert_with_rope.py 
205
206
207
208
def __init__(self, hidden_size: int, moe_num_experts: int, moe_top_k: int):
    super().__init__()
    self.moe_top_k = moe_top_k
    self.layer = ReplicatedLinear(hidden_size, moe_num_experts, bias=False)

forward 

forward(x: Tensor) -> tuple[Tensor, Tensor, LongTensor]
Source code in vllm/model_executor/models/bert_with_rope.py 
210
211
212
213
214
215
216
217
218
def forward(
    self, x: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor, torch.LongTensor]:
    weights = self.layer(x.view(-1, x.shape[-1]))[0].softmax(
        dim=-1, dtype=torch.float32)
    top_weights, top_experts = torch.topk(weights, self.moe_top_k, dim=-1)
    weights = weights.to(x.dtype)
    top_weights = top_weights.to(x.dtype)
    return weights, top_weights, top_experts  # type: ignore

SnowflakeGteNewModel

Bases: GteNewModel

Source code in vllm/model_executor/models/bert_with_rope.py 
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
class SnowflakeGteNewModel(GteNewModel):
    # for Snowflake/snowflake-arctic-embed-m-v2.0

    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_substr={
            "layer": "layers",
            "attention.qkv_proj": "attn.qkv_proj",
            "attention.o_proj": "attn.out_proj",
        })

    def config_verify(self, vllm_config):
        config = vllm_config.model_config.hf_config

        assert config.__class__.__name__ == "GteConfig"
        assert config.hidden_act == "gelu"

        config.hidden_act = "geglu"

        head_dim = config.hidden_size // config.num_attention_heads
        config.rotary_kwargs = {
            "head_size": head_dim,
            "rotary_dim": getattr(config, "rotary_emb_dim", head_dim),
            "max_position": config.max_position_embeddings,
            "base": config.rope_theta,
            "rope_scaling": getattr(config, "rope_scaling", None)
        }
        return config

hf_to_vllm_mapper class-attribute instance-attribute 

hf_to_vllm_mapper = WeightsMapper(
    orig_to_new_substr={
        "layer": "layers",
        "attention.qkv_proj": "attn.qkv_proj",
        "attention.o_proj": "attn.out_proj",
    }
)

config_verify 

config_verify(vllm_config)
Source code in vllm/model_executor/models/bert_with_rope.py 
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
def config_verify(self, vllm_config):
    config = vllm_config.model_config.hf_config

    assert config.__class__.__name__ == "GteConfig"
    assert config.hidden_act == "gelu"

    config.hidden_act = "geglu"

    head_dim = config.hidden_size // config.num_attention_heads
    config.rotary_kwargs = {
        "head_size": head_dim,
        "rotary_dim": getattr(config, "rotary_emb_dim", head_dim),
        "max_position": config.max_position_embeddings,
        "base": config.rope_theta,
        "rope_scaling": getattr(config, "rope_scaling", None)
    }
    return config