Skip to content

vllm.model_executor.models.dbrx

DbrxAttention

Bases: Module

Source code in vllm/model_executor/models/dbrx.py 
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
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 DbrxAttention(nn.Module):

    def __init__(
        self,
        config: DbrxConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.d_model = config.d_model
        self.total_num_heads = config.n_heads
        self.head_dim = self.d_model // self.total_num_heads
        self.total_num_kv_heads = config.attn_config.kv_n_heads
        self.clip_qkv = config.attn_config.clip_qkv
        self.rope_theta = config.attn_config.rope_theta
        self.max_position = config.max_seq_len

        # pylint: disable=invalid-name
        self.Wqkv = QKVParallelLinear(
            self.d_model,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
        )
        self.out_proj = RowParallelLinear(
            self.d_model,
            self.d_model,
            bias=False,
            quant_config=quant_config,
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=self.max_position,
            base=int(self.rope_theta),
            is_neox_style=True,
        )

        tp_world_size = get_tensor_model_parallel_world_size()
        self.tp_size = tp_world_size
        assert self.total_num_heads % tp_world_size == 0
        self.num_heads = self.total_num_heads // tp_world_size
        if self.total_num_kv_heads >= tp_world_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_world_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_world_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_world_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.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
                              num_kv_heads=self.num_kv_heads,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.Wqkv(hidden_states)
        if self.clip_qkv is not None:
            qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(position_ids, q, k)
        attn_output = self.attn(q, k, v)
        hidden_states, _ = self.out_proj(attn_output)
        return hidden_states

Wqkv instance-attribute 

Wqkv = QKVParallelLinear(
    d_model,
    head_dim,
    total_num_heads,
    total_num_kv_heads,
    bias=False,
    quant_config=quant_config,
)

attn instance-attribute 

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

clip_qkv instance-attribute 

clip_qkv = clip_qkv

d_model instance-attribute 

d_model = d_model

head_dim instance-attribute 

head_dim = d_model // total_num_heads

kv_size instance-attribute 

kv_size = num_kv_heads * head_dim

max_position instance-attribute 

max_position = max_seq_len

num_heads instance-attribute 

num_heads = total_num_heads // tp_world_size

num_kv_heads instance-attribute 

num_kv_heads = max(1, total_num_kv_heads // tp_world_size)

out_proj instance-attribute 

out_proj = RowParallelLinear(
    d_model, d_model, bias=False, quant_config=quant_config
)

q_size instance-attribute 

q_size = num_heads * head_dim

rope_theta instance-attribute 

rope_theta = rope_theta

rotary_emb instance-attribute 

rotary_emb = get_rope(
    head_dim,
    rotary_dim=head_dim,
    max_position=max_position,
    base=int(rope_theta),
    is_neox_style=True,
)

scaling instance-attribute 

scaling = head_dim ** -0.5

total_num_heads instance-attribute 

total_num_heads = n_heads

total_num_kv_heads instance-attribute 

total_num_kv_heads = kv_n_heads

tp_size instance-attribute 

tp_size = tp_world_size

__init__ 

__init__(
    config: DbrxConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/dbrx.py 
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
def __init__(
    self,
    config: DbrxConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.d_model = config.d_model
    self.total_num_heads = config.n_heads
    self.head_dim = self.d_model // self.total_num_heads
    self.total_num_kv_heads = config.attn_config.kv_n_heads
    self.clip_qkv = config.attn_config.clip_qkv
    self.rope_theta = config.attn_config.rope_theta
    self.max_position = config.max_seq_len

    # pylint: disable=invalid-name
    self.Wqkv = QKVParallelLinear(
        self.d_model,
        self.head_dim,
        self.total_num_heads,
        self.total_num_kv_heads,
        bias=False,
        quant_config=quant_config,
    )
    self.out_proj = RowParallelLinear(
        self.d_model,
        self.d_model,
        bias=False,
        quant_config=quant_config,
    )
    self.rotary_emb = get_rope(
        self.head_dim,
        rotary_dim=self.head_dim,
        max_position=self.max_position,
        base=int(self.rope_theta),
        is_neox_style=True,
    )

    tp_world_size = get_tensor_model_parallel_world_size()
    self.tp_size = tp_world_size
    assert self.total_num_heads % tp_world_size == 0
    self.num_heads = self.total_num_heads // tp_world_size
    if self.total_num_kv_heads >= tp_world_size:
        # Number of KV heads is greater than TP size, so we partition
        # the KV heads across multiple tensor parallel GPUs.
        assert self.total_num_kv_heads % tp_world_size == 0
    else:
        # Number of KV heads is less than TP size, so we replicate
        # the KV heads across multiple tensor parallel GPUs.
        assert tp_world_size % self.total_num_kv_heads == 0
    self.num_kv_heads = max(1, self.total_num_kv_heads // tp_world_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.attn = Attention(self.num_heads,
                          self.head_dim,
                          self.scaling,
                          num_kv_heads=self.num_kv_heads,
                          cache_config=cache_config,
                          quant_config=quant_config,
                          prefix=f"{prefix}.attn")

forward 

forward(
    position_ids: Tensor, hidden_states: Tensor
) -> Tensor
Source code in vllm/model_executor/models/dbrx.py 
233
234
235
236
237
238
239
240
241
242
243
244
245
def forward(
    self,
    position_ids: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    qkv, _ = self.Wqkv(hidden_states)
    if self.clip_qkv is not None:
        qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
    q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
    q, k = self.rotary_emb(position_ids, q, k)
    attn_output = self.attn(q, k, v)
    hidden_states, _ = self.out_proj(attn_output)
    return hidden_states

DbrxBlock

Bases: Module

Source code in vllm/model_executor/models/dbrx.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
class DbrxBlock(nn.Module):

    def __init__(
        self,
        config: DbrxConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.norm_attn_norm = DbrxFusedNormAttention(
            config,
            cache_config,
            quant_config,
            prefix=f"{prefix}.norm_attn_norm")
        self.ffn = DbrxMoE(config, quant_config, prefix=f"{prefix}.ffn")

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        hidden_states, residual = self.norm_attn_norm(
            position_ids=position_ids,
            hidden_states=hidden_states,
        )
        hidden_states = self.ffn(hidden_states)
        hidden_states = hidden_states + residual
        return hidden_states

ffn instance-attribute 

ffn = DbrxMoE(config, quant_config, prefix=f'{prefix}.ffn')

norm_attn_norm instance-attribute 

norm_attn_norm = DbrxFusedNormAttention(
    config,
    cache_config,
    quant_config,
    prefix=f"{prefix}.norm_attn_norm",
)

__init__ 

__init__(
    config: DbrxConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/dbrx.py 
285
286
287
288
289
290
291
292
293
294
295
296
297
298
def __init__(
    self,
    config: DbrxConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.norm_attn_norm = DbrxFusedNormAttention(
        config,
        cache_config,
        quant_config,
        prefix=f"{prefix}.norm_attn_norm")
    self.ffn = DbrxMoE(config, quant_config, prefix=f"{prefix}.ffn")

forward 

forward(
    position_ids: Tensor, hidden_states: Tensor
) -> Tensor
Source code in vllm/model_executor/models/dbrx.py 
300
301
302
303
304
305
306
307
308
309
310
311
def forward(
    self,
    position_ids: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    hidden_states, residual = self.norm_attn_norm(
        position_ids=position_ids,
        hidden_states=hidden_states,
    )
    hidden_states = self.ffn(hidden_states)
    hidden_states = hidden_states + residual
    return hidden_states

DbrxExperts

Bases: FusedMoE

Source code in vllm/model_executor/models/dbrx.py 
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 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
class DbrxExperts(FusedMoE):

    def __init__(
        self,
        config: DbrxConfig,
        quant_config: Optional[QuantizationConfig] = None,
        params_dtype: Optional[torch.dtype] = None,
        prefix: str = "",
    ):
        super().__init__(
            num_experts=config.ffn_config.moe_num_experts,
            top_k=config.ffn_config.moe_top_k,
            hidden_size=config.d_model,
            intermediate_size=config.ffn_config.ffn_hidden_size,
            params_dtype=params_dtype,
            reduce_results=True,
            renormalize=True,
            quant_config=quant_config,
            tp_size=get_tensor_model_parallel_world_size(),
            prefix=prefix,
        )
        self.config = config
        self.d_model = config.d_model
        self.intermediate_size = (self.config.ffn_config.ffn_hidden_size //
                                  self.tp_size)

    # Define custom weight loader for dbrx model
    def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
                      weight_name: str, param_name: str):
        tp_rank = get_tensor_model_parallel_rank()
        param_data = param.data
        shard_size = self.intermediate_size
        shard = slice(tp_rank * shard_size, (tp_rank + 1) * shard_size)
        # DBRX uses GLU for each experts.
        # GLU has 3 linear layers: w1, v1 and w2.
        if weight_name.endswith("w1"):
            if param_name.endswith("weight"):
                loaded_weight = torch.reshape(
                    loaded_weight,
                    [-1, self.intermediate_size * self.tp_size, self.d_model],
                )
                param_data[:, 0:shard_size, :] = loaded_weight[:, shard, :]
            elif param_name.endswith("weight_scale"):
                param_data[:, 0] = loaded_weight
            else:
                param_data = loaded_weight
        if weight_name.endswith("v1"):
            if param_name.endswith("weight"):
                loaded_weight = torch.reshape(
                    loaded_weight,
                    [-1, self.intermediate_size * self.tp_size, self.d_model],
                )
                param_data[:, shard_size:2 *
                           shard_size, :] = loaded_weight[:, shard, :]
            elif param_name.endswith("weight_scale"):
                param_data[:, 1] = loaded_weight
            else:
                param_data[:] = loaded_weight
        if weight_name.endswith("w2"):
            if param_name.endswith("weight"):
                loaded_weight = torch.reshape(
                    loaded_weight,
                    [-1, self.intermediate_size * self.tp_size, self.d_model],
                ).transpose(1, 2)
                param_data[:] = loaded_weight[:, :, shard]
            else:
                param_data[:] = loaded_weight

config instance-attribute 

config = config

d_model instance-attribute 

d_model = d_model

intermediate_size instance-attribute 

intermediate_size = ffn_hidden_size // tp_size

__init__ 

__init__(
    config: DbrxConfig,
    quant_config: Optional[QuantizationConfig] = None,
    params_dtype: Optional[dtype] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/dbrx.py 
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
def __init__(
    self,
    config: DbrxConfig,
    quant_config: Optional[QuantizationConfig] = None,
    params_dtype: Optional[torch.dtype] = None,
    prefix: str = "",
):
    super().__init__(
        num_experts=config.ffn_config.moe_num_experts,
        top_k=config.ffn_config.moe_top_k,
        hidden_size=config.d_model,
        intermediate_size=config.ffn_config.ffn_hidden_size,
        params_dtype=params_dtype,
        reduce_results=True,
        renormalize=True,
        quant_config=quant_config,
        tp_size=get_tensor_model_parallel_world_size(),
        prefix=prefix,
    )
    self.config = config
    self.d_model = config.d_model
    self.intermediate_size = (self.config.ffn_config.ffn_hidden_size //
                              self.tp_size)

weight_loader 

weight_loader(
    param: Parameter,
    loaded_weight: Tensor,
    weight_name: str,
    param_name: str,
)
Source code in vllm/model_executor/models/dbrx.py 
 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 weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
                  weight_name: str, param_name: str):
    tp_rank = get_tensor_model_parallel_rank()
    param_data = param.data
    shard_size = self.intermediate_size
    shard = slice(tp_rank * shard_size, (tp_rank + 1) * shard_size)
    # DBRX uses GLU for each experts.
    # GLU has 3 linear layers: w1, v1 and w2.
    if weight_name.endswith("w1"):
        if param_name.endswith("weight"):
            loaded_weight = torch.reshape(
                loaded_weight,
                [-1, self.intermediate_size * self.tp_size, self.d_model],
            )
            param_data[:, 0:shard_size, :] = loaded_weight[:, shard, :]
        elif param_name.endswith("weight_scale"):
            param_data[:, 0] = loaded_weight
        else:
            param_data = loaded_weight
    if weight_name.endswith("v1"):
        if param_name.endswith("weight"):
            loaded_weight = torch.reshape(
                loaded_weight,
                [-1, self.intermediate_size * self.tp_size, self.d_model],
            )
            param_data[:, shard_size:2 *
                       shard_size, :] = loaded_weight[:, shard, :]
        elif param_name.endswith("weight_scale"):
            param_data[:, 1] = loaded_weight
        else:
            param_data[:] = loaded_weight
    if weight_name.endswith("w2"):
        if param_name.endswith("weight"):
            loaded_weight = torch.reshape(
                loaded_weight,
                [-1, self.intermediate_size * self.tp_size, self.d_model],
            ).transpose(1, 2)
            param_data[:] = loaded_weight[:, :, shard]
        else:
            param_data[:] = loaded_weight

DbrxForCausalLM

Bases: Module, SupportsPP

Source code in vllm/model_executor/models/dbrx.py 
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
471
472
class DbrxForCausalLM(nn.Module, SupportsPP):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        self.config = config
        if config.tie_word_embeddings:
            raise ValueError(
                "tie_word_embeddings is not supported for Dbrx models.")
        self.quant_config = quant_config
        self.unpadded_vocab_size = config.vocab_size
        self.transformer = DbrxModel(vllm_config=vllm_config,
                                     prefix=maybe_prefix(
                                         prefix, "transformer"))
        self.lm_head = ParallelLMHead(
            config.vocab_size,
            config.d_model,
            org_num_embeddings=config.vocab_size,
            padding_size=DEFAULT_VOCAB_PADDING_SIZE,
            quant_config=quant_config,
        )
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.transformer.make_empty_intermediate_tensors)

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.transformer.get_input_embeddings(input_ids)

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

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states,
                                       sampling_metadata)
        return logits

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

config instance-attribute 

config = config

lm_head instance-attribute 

lm_head = ParallelLMHead(
    vocab_size,
    d_model,
    org_num_embeddings=vocab_size,
    padding_size=DEFAULT_VOCAB_PADDING_SIZE,
    quant_config=quant_config,
)

logits_processor instance-attribute 

logits_processor = LogitsProcessor(
    unpadded_vocab_size, vocab_size
)

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

quant_config instance-attribute 

quant_config = quant_config

transformer instance-attribute 

transformer = DbrxModel(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "transformer"),
)

unpadded_vocab_size instance-attribute 

unpadded_vocab_size = vocab_size

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/dbrx.py 
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    quant_config = vllm_config.quant_config
    self.config = config
    if config.tie_word_embeddings:
        raise ValueError(
            "tie_word_embeddings is not supported for Dbrx models.")
    self.quant_config = quant_config
    self.unpadded_vocab_size = config.vocab_size
    self.transformer = DbrxModel(vllm_config=vllm_config,
                                 prefix=maybe_prefix(
                                     prefix, "transformer"))
    self.lm_head = ParallelLMHead(
        config.vocab_size,
        config.d_model,
        org_num_embeddings=config.vocab_size,
        padding_size=DEFAULT_VOCAB_PADDING_SIZE,
        quant_config=quant_config,
    )
    self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                            config.vocab_size)
    self.make_empty_intermediate_tensors = (
        self.transformer.make_empty_intermediate_tensors)

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]
Source code in vllm/model_executor/models/dbrx.py 
460
461
462
463
464
465
466
467
def compute_logits(
    self,
    hidden_states: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[torch.Tensor]:
    logits = self.logits_processor(self.lm_head, hidden_states,
                                   sampling_metadata)
    return logits

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/dbrx.py 
449
450
451
452
453
454
455
456
457
458
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
    hidden_states = self.transformer(input_ids, positions,
                                     intermediate_tensors, inputs_embeds)
    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/dbrx.py 
446
447
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.transformer.get_input_embeddings(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/dbrx.py 
469
470
471
472
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(self)
    return loader.load_weights(weights)

DbrxFusedNormAttention

Bases: Module

Source code in vllm/model_executor/models/dbrx.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
class DbrxFusedNormAttention(nn.Module):

    def __init__(
        self,
        config: DbrxConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.d_model = config.d_model
        self.attn = DbrxAttention(config,
                                  cache_config,
                                  quant_config,
                                  prefix=f"{prefix}.attn")
        self.norm_1 = nn.LayerNorm(self.d_model)
        self.norm_2 = nn.LayerNorm(self.d_model)

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.norm_1(hidden_states)
        x = self.attn(
            position_ids=position_ids,
            hidden_states=hidden_states,
        )
        hidden_states = residual + x
        residual = hidden_states
        hidden_states = self.norm_2(hidden_states)
        return hidden_states, residual

attn instance-attribute 

attn = DbrxAttention(
    config,
    cache_config,
    quant_config,
    prefix=f"{prefix}.attn",
)

d_model instance-attribute 

d_model = d_model

norm_1 instance-attribute 

norm_1 = LayerNorm(d_model)

norm_2 instance-attribute 

norm_2 = LayerNorm(d_model)

__init__ 

__init__(
    config: DbrxConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/dbrx.py 
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
def __init__(
    self,
    config: DbrxConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.d_model = config.d_model
    self.attn = DbrxAttention(config,
                              cache_config,
                              quant_config,
                              prefix=f"{prefix}.attn")
    self.norm_1 = nn.LayerNorm(self.d_model)
    self.norm_2 = nn.LayerNorm(self.d_model)

forward 

forward(
    position_ids: Tensor, hidden_states: Tensor
) -> Tensor
Source code in vllm/model_executor/models/dbrx.py 
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
def forward(
    self,
    position_ids: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    residual = hidden_states
    hidden_states = self.norm_1(hidden_states)
    x = self.attn(
        position_ids=position_ids,
        hidden_states=hidden_states,
    )
    hidden_states = residual + x
    residual = hidden_states
    hidden_states = self.norm_2(hidden_states)
    return hidden_states, residual

DbrxMoE

Bases: Module

A tensor-parallel MoE implementation for DBRX.

Each expert's weights are sharded across all ranks and a fused MoE kernel is used for the forward pass, and finally we reduce the outputs across ranks.

Source code in vllm/model_executor/models/dbrx.py 
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
class DbrxMoE(nn.Module):
    """A tensor-parallel MoE implementation for DBRX.

    Each expert's weights are sharded across all ranks and a fused MoE
    kernel is used for the forward pass, and finally we reduce the outputs
    across ranks.
    """

    def __init__(
        self,
        config: DbrxConfig,
        quant_config: Optional[QuantizationConfig] = None,
        params_dtype: Optional[torch.dtype] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.d_model = config.d_model
        if params_dtype is None:
            params_dtype = torch.get_default_dtype()
        self.params_dtype = params_dtype

        self.router = DbrxRouter(config, self.params_dtype)

        self.experts = DbrxExperts(config=config,
                                   quant_config=quant_config,
                                   params_dtype=self.params_dtype,
                                   prefix=f"{prefix}.experts")

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        orig_shape = hidden_states.shape
        hidden_states = hidden_states.view(-1, self.d_model)
        # router_logits: (num_tokens, n_experts)
        router_logits = self.router(hidden_states)
        final_hidden_states = self.experts(hidden_states, router_logits)
        return final_hidden_states.view(orig_shape)

d_model instance-attribute 

d_model = d_model

experts instance-attribute 

experts = DbrxExperts(
    config=config,
    quant_config=quant_config,
    params_dtype=params_dtype,
    prefix=f"{prefix}.experts",
)

params_dtype instance-attribute 

params_dtype = params_dtype

router instance-attribute 

router = DbrxRouter(config, params_dtype)

__init__ 

__init__(
    config: DbrxConfig,
    quant_config: Optional[QuantizationConfig] = None,
    params_dtype: Optional[dtype] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/dbrx.py 
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
def __init__(
    self,
    config: DbrxConfig,
    quant_config: Optional[QuantizationConfig] = None,
    params_dtype: Optional[torch.dtype] = None,
    prefix: str = "",
):
    super().__init__()
    self.d_model = config.d_model
    if params_dtype is None:
        params_dtype = torch.get_default_dtype()
    self.params_dtype = params_dtype

    self.router = DbrxRouter(config, self.params_dtype)

    self.experts = DbrxExperts(config=config,
                               quant_config=quant_config,
                               params_dtype=self.params_dtype,
                               prefix=f"{prefix}.experts")

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/dbrx.py 
159
160
161
162
163
164
165
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    orig_shape = hidden_states.shape
    hidden_states = hidden_states.view(-1, self.d_model)
    # router_logits: (num_tokens, n_experts)
    router_logits = self.router(hidden_states)
    final_hidden_states = self.experts(hidden_states, router_logits)
    return final_hidden_states.view(orig_shape)

DbrxModel

Bases: Module

Source code in vllm/model_executor/models/dbrx.py 
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
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
class DbrxModel(nn.Module):

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

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

        self.quant_config = quant_config
        self.wte = VocabParallelEmbedding(
            config.vocab_size,
            config.d_model,
        )
        self.start_layer, self.end_layer, self.blocks = make_layers(
            config.n_layers,
            lambda prefix: DbrxBlock(
                config, cache_config, quant_config, prefix=prefix),
            prefix=f"{prefix}.blocks",
        )
        self.norm_f = nn.LayerNorm(config.d_model, eps=1e-5)
        for module in self.modules():
            if hasattr(module, "bias") and isinstance(module.bias,
                                                      nn.Parameter):
                # Remove the bias term in Linear and LayerNorm.
                module.register_parameter("bias", None)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(["hidden_states"],
                                                    config.d_model))

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.wte(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors],
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
        else:
            assert intermediate_tensors
            hidden_states = intermediate_tensors["hidden_states"]
        for block in self.blocks[self.start_layer:self.end_layer]:
            hidden_states = block(position_ids, hidden_states)
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({"hidden_states": hidden_states})
        hidden_states = self.norm_f(hidden_states)
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        expert_params_mapping = [(
            "w13" if weight_name in ["w1", "v1"] else "w2",
            f"mlp.{weight_name}",
        ) for weight_name in ["w1", "v1", "w2"]]
        params_dict = dict(self.named_parameters(remove_duplicate=False))
        loaded_params: set[str] = set()

        for name, loaded_weight in weights:
            if (self.quant_config is not None and
                (scale_name := self.quant_config.get_cache_scale(name))):
                # Loading kv cache quantization scales
                param = params_dict[scale_name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                                 loaded_weight[0])
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue

            if name.endswith(("w1", "w2", "v1")):
                name = name + "_weight"
            for param_name, weight_name in expert_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, weight_name, name)
                break

            else:
                if is_pp_missing_parameter(name, self):
                    continue
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    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

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors_factory(
        ["hidden_states"], d_model
    )
)

norm_f instance-attribute 

norm_f = LayerNorm(d_model, eps=1e-05)

quant_config instance-attribute 

quant_config = quant_config

wte instance-attribute 

wte = VocabParallelEmbedding(vocab_size, d_model)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/dbrx.py 
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
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()

    config = vllm_config.model_config.hf_config
    cache_config = vllm_config.cache_config
    quant_config = vllm_config.quant_config

    self.quant_config = quant_config
    self.wte = VocabParallelEmbedding(
        config.vocab_size,
        config.d_model,
    )
    self.start_layer, self.end_layer, self.blocks = make_layers(
        config.n_layers,
        lambda prefix: DbrxBlock(
            config, cache_config, quant_config, prefix=prefix),
        prefix=f"{prefix}.blocks",
    )
    self.norm_f = nn.LayerNorm(config.d_model, eps=1e-5)
    for module in self.modules():
        if hasattr(module, "bias") and isinstance(module.bias,
                                                  nn.Parameter):
            # Remove the bias term in Linear and LayerNorm.
            module.register_parameter("bias", None)
    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(["hidden_states"],
                                                config.d_model))

forward 

forward(
    input_ids: Tensor,
    position_ids: Tensor,
    intermediate_tensors: Optional[IntermediateTensors],
    inputs_embeds: Optional[Tensor] = None,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/dbrx.py 
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
def forward(
    self,
    input_ids: torch.Tensor,
    position_ids: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors],
    inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
    if get_pp_group().is_first_rank:
        if inputs_embeds is not None:
            hidden_states = inputs_embeds
        else:
            hidden_states = self.get_input_embeddings(input_ids)
    else:
        assert intermediate_tensors
        hidden_states = intermediate_tensors["hidden_states"]
    for block in self.blocks[self.start_layer:self.end_layer]:
        hidden_states = block(position_ids, hidden_states)
    if not get_pp_group().is_last_rank:
        return IntermediateTensors({"hidden_states": hidden_states})
    hidden_states = self.norm_f(hidden_states)
    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/dbrx.py 
344
345
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.wte(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/dbrx.py 
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    expert_params_mapping = [(
        "w13" if weight_name in ["w1", "v1"] else "w2",
        f"mlp.{weight_name}",
    ) for weight_name in ["w1", "v1", "w2"]]
    params_dict = dict(self.named_parameters(remove_duplicate=False))
    loaded_params: set[str] = set()

    for name, loaded_weight in weights:
        if (self.quant_config is not None and
            (scale_name := self.quant_config.get_cache_scale(name))):
            # Loading kv cache quantization scales
            param = params_dict[scale_name]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                             loaded_weight[0])
            weight_loader(param, loaded_weight)
            loaded_params.add(scale_name)
            continue

        if name.endswith(("w1", "w2", "v1")):
            name = name + "_weight"
        for param_name, weight_name in expert_params_mapping:
            if weight_name not in name:
                continue
            name = name.replace(weight_name, param_name)
            if is_pp_missing_parameter(name, self):
                continue
            param = params_dict[name]
            weight_loader = param.weight_loader
            weight_loader(param, loaded_weight, weight_name, name)
            break

        else:
            if is_pp_missing_parameter(name, self):
                continue
            # Remapping the name of FP8 kv-scale.
            name = maybe_remap_kv_scale_name(name, params_dict)
            if name is None:
                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

DbrxRouter

Bases: Module

A Router implementation for DBRX that returns logits for each expert per token.

Source code in vllm/model_executor/models/dbrx.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
class DbrxRouter(nn.Module):
    """A Router implementation for DBRX that returns logits for each expert
    per token.
    """

    def __init__(
        self,
        config: DbrxConfig,
        params_dtype: Optional[torch.dtype] = None,
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.num_total_experts = config.ffn_config.moe_num_experts
        self.d_model = config.d_model
        self.layer = ReplicatedLinear(
            self.d_model,
            self.num_total_experts,
            bias=False,
            params_dtype=params_dtype,
            quant_config=None,
        )

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

d_model instance-attribute 

d_model = d_model

layer instance-attribute 

layer = ReplicatedLinear(
    d_model,
    num_total_experts,
    bias=False,
    params_dtype=params_dtype,
    quant_config=None,
)

num_total_experts instance-attribute 

num_total_experts = moe_num_experts

tp_size instance-attribute 

tp_size = get_tensor_model_parallel_world_size()

__init__ 

__init__(
    config: DbrxConfig, params_dtype: Optional[dtype] = None
)
Source code in vllm/model_executor/models/dbrx.py 
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
def __init__(
    self,
    config: DbrxConfig,
    params_dtype: Optional[torch.dtype] = None,
):
    super().__init__()
    self.tp_size = get_tensor_model_parallel_world_size()
    self.num_total_experts = config.ffn_config.moe_num_experts
    self.d_model = config.d_model
    self.layer = ReplicatedLinear(
        self.d_model,
        self.num_total_experts,
        bias=False,
        params_dtype=params_dtype,
        quant_config=None,
    )

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/dbrx.py 
57
58
59
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    router_logits, _ = self.layer(hidden_states)
    return router_logits