Skip to content

vllm.model_executor.models.mpt

MPTAttention

Bases: Module

Source code in vllm/model_executor/models/mpt.py 
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 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 MPTAttention(nn.Module):

    def __init__(
        self,
        config: MPTConfig,
        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.clip_qkv = config.attn_config["clip_qkv"]
        self.qk_ln = config.attn_config["qk_ln"]
        self.alibi_bias_max = config.attn_config["alibi_bias_max"]
        if "kv_n_heads" in config.attn_config:
            self.total_num_kv_heads = config.attn_config['kv_n_heads']
        else:
            self.total_num_kv_heads = self.total_num_heads
        assert not config.attn_config["prefix_lm"]
        assert config.attn_config["alibi"]

        # pylint: disable=invalid-name
        self.Wqkv = QKVParallelLinear(
            self.d_model,
            self.d_model // self.total_num_heads,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=not config.no_bias,
            quant_config=quant_config,
        )
        if self.qk_ln:
            self.q_ln = nn.LayerNorm(self.d_model)
            self.k_ln = nn.LayerNorm(self.d_model)
        self.out_proj = RowParallelLinear(
            self.d_model,
            self.d_model,
            bias=not config.no_bias,
            quant_config=quant_config,
        )

        tp_world_size = get_tensor_model_parallel_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
        # Create the alibi slopes and slice them.
        tp_rank = get_tensor_model_parallel_rank()
        head_start = tp_rank * self.num_heads
        head_end = (tp_rank + 1) * self.num_heads
        alibi_slopes = _get_alibi_slopes(self.total_num_heads,
                                         self.alibi_bias_max)
        alibi_slopes = alibi_slopes[head_start:head_end].tolist()

        self.head_dim = self.d_model // self.total_num_heads
        scaling = self.head_dim**-0.5
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              scaling,
                              alibi_slopes=alibi_slopes,
                              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:
        del position_ids  # unused.
        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)
        if self.qk_ln:
            q = self.q_ln(q)
            k = self.k_ln(k)
        attn_output = self.attn(q, k, v)
        output, _ = self.out_proj(attn_output)
        return output

Wqkv instance-attribute 

Wqkv = QKVParallelLinear(
    d_model,
    d_model // total_num_heads,
    total_num_heads,
    total_num_kv_heads,
    bias=not no_bias,
    quant_config=quant_config,
)

alibi_bias_max instance-attribute 

alibi_bias_max = attn_config['alibi_bias_max']

attn instance-attribute 

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

clip_qkv instance-attribute 

clip_qkv = attn_config['clip_qkv']

d_model instance-attribute 

d_model = d_model

head_dim instance-attribute 

head_dim = d_model // total_num_heads

k_ln instance-attribute 

k_ln = LayerNorm(d_model)

kv_size instance-attribute 

kv_size = num_kv_heads * head_dim

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=not no_bias,
    quant_config=quant_config,
)

q_ln instance-attribute 

q_ln = LayerNorm(d_model)

q_size instance-attribute 

q_size = num_heads * head_dim

qk_ln instance-attribute 

qk_ln = attn_config['qk_ln']

total_num_heads instance-attribute 

total_num_heads = n_heads

total_num_kv_heads instance-attribute 

total_num_kv_heads = attn_config['kv_n_heads']

__init__ 

__init__(
    config: MPTConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/mpt.py 
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 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
def __init__(
    self,
    config: MPTConfig,
    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.clip_qkv = config.attn_config["clip_qkv"]
    self.qk_ln = config.attn_config["qk_ln"]
    self.alibi_bias_max = config.attn_config["alibi_bias_max"]
    if "kv_n_heads" in config.attn_config:
        self.total_num_kv_heads = config.attn_config['kv_n_heads']
    else:
        self.total_num_kv_heads = self.total_num_heads
    assert not config.attn_config["prefix_lm"]
    assert config.attn_config["alibi"]

    # pylint: disable=invalid-name
    self.Wqkv = QKVParallelLinear(
        self.d_model,
        self.d_model // self.total_num_heads,
        self.total_num_heads,
        self.total_num_kv_heads,
        bias=not config.no_bias,
        quant_config=quant_config,
    )
    if self.qk_ln:
        self.q_ln = nn.LayerNorm(self.d_model)
        self.k_ln = nn.LayerNorm(self.d_model)
    self.out_proj = RowParallelLinear(
        self.d_model,
        self.d_model,
        bias=not config.no_bias,
        quant_config=quant_config,
    )

    tp_world_size = get_tensor_model_parallel_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
    # Create the alibi slopes and slice them.
    tp_rank = get_tensor_model_parallel_rank()
    head_start = tp_rank * self.num_heads
    head_end = (tp_rank + 1) * self.num_heads
    alibi_slopes = _get_alibi_slopes(self.total_num_heads,
                                     self.alibi_bias_max)
    alibi_slopes = alibi_slopes[head_start:head_end].tolist()

    self.head_dim = self.d_model // self.total_num_heads
    scaling = self.head_dim**-0.5
    self.attn = Attention(self.num_heads,
                          self.head_dim,
                          scaling,
                          alibi_slopes=alibi_slopes,
                          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/mpt.py 
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
def forward(
    self,
    position_ids: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    del position_ids  # unused.
    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)
    if self.qk_ln:
        q = self.q_ln(q)
        k = self.k_ln(k)
    attn_output = self.attn(q, k, v)
    output, _ = self.out_proj(attn_output)
    return output

MPTBlock

Bases: Module

Source code in vllm/model_executor/models/mpt.py 
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
class MPTBlock(nn.Module):

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

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

attn instance-attribute 

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

ffn instance-attribute 

ffn = MPTMLP(config, quant_config)

norm_1 instance-attribute 

norm_1 = LayerNorm(hidden_size)

norm_2 instance-attribute 

norm_2 = LayerNorm(hidden_size)

__init__ 

__init__(
    config: MPTConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/mpt.py 
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
def __init__(
    self,
    config: MPTConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    hidden_size = config.d_model
    self.norm_1 = nn.LayerNorm(hidden_size)
    self.attn = MPTAttention(config,
                             cache_config,
                             quant_config,
                             prefix=f"{prefix}.attn")
    self.norm_2 = nn.LayerNorm(hidden_size)
    self.ffn = MPTMLP(config, quant_config)

forward 

forward(
    position_ids: Tensor, hidden_states: Tensor
) -> Tensor
Source code in vllm/model_executor/models/mpt.py 
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
def forward(
    self,
    position_ids: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    x = self.norm_1(hidden_states)
    x = self.attn(
        position_ids=position_ids,
        hidden_states=x,
    )
    hidden_states = hidden_states + x
    x = self.norm_2(hidden_states)
    x = self.ffn(x)
    hidden_states = hidden_states + x
    return hidden_states

MPTForCausalLM

Bases: Module, SupportsPP

Source code in vllm/model_executor/models/mpt.py 
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
class MPTForCausalLM(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
        assert config.tie_word_embeddings
        self.quant_config = quant_config

        self.transformer = MPTModel(vllm_config=vllm_config,
                                    prefix=maybe_prefix(prefix, "transformer"))
        self.lm_head = self.transformer.wte
        self.logits_processor = LogitsProcessor(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 = wte

logits_processor instance-attribute 

logits_processor = LogitsProcessor(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 = MPTModel(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "transformer"),
)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/mpt.py 
290
291
292
293
294
295
296
297
298
299
300
301
302
303
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
    assert config.tie_word_embeddings
    self.quant_config = quant_config

    self.transformer = MPTModel(vllm_config=vllm_config,
                                prefix=maybe_prefix(prefix, "transformer"))
    self.lm_head = self.transformer.wte
    self.logits_processor = LogitsProcessor(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/mpt.py 
319
320
321
322
323
324
325
326
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/mpt.py 
308
309
310
311
312
313
314
315
316
317
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/mpt.py 
305
306
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/mpt.py 
328
329
330
331
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(self)
    return loader.load_weights(weights)

MPTMLP

Bases: Module

Source code in vllm/model_executor/models/mpt.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
172
class MPTMLP(nn.Module):

    def __init__(
        self,
        config: MPTConfig,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        hidden_size = config.d_model
        expansion_ratio = config.expansion_ratio
        intermediate_size = expansion_ratio * hidden_size
        self.up_proj = ColumnParallelLinear(
            hidden_size,
            intermediate_size,
            bias=not config.no_bias,
            quant_config=quant_config,
        )
        self.act = get_act_fn("gelu")
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=not config.no_bias,
            quant_config=quant_config,
        )

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

act instance-attribute 

act = get_act_fn('gelu')

down_proj instance-attribute 

down_proj = RowParallelLinear(
    intermediate_size,
    hidden_size,
    bias=not no_bias,
    quant_config=quant_config,
)

up_proj instance-attribute 

up_proj = ColumnParallelLinear(
    hidden_size,
    intermediate_size,
    bias=not no_bias,
    quant_config=quant_config,
)

__init__ 

__init__(
    config: MPTConfig,
    quant_config: Optional[QuantizationConfig] = None,
)
Source code in vllm/model_executor/models/mpt.py 
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
def __init__(
    self,
    config: MPTConfig,
    quant_config: Optional[QuantizationConfig] = None,
):
    super().__init__()
    hidden_size = config.d_model
    expansion_ratio = config.expansion_ratio
    intermediate_size = expansion_ratio * hidden_size
    self.up_proj = ColumnParallelLinear(
        hidden_size,
        intermediate_size,
        bias=not config.no_bias,
        quant_config=quant_config,
    )
    self.act = get_act_fn("gelu")
    self.down_proj = RowParallelLinear(
        intermediate_size,
        hidden_size,
        bias=not config.no_bias,
        quant_config=quant_config,
    )

forward 

forward(x: Tensor) -> Tensor
Source code in vllm/model_executor/models/mpt.py 
168
169
170
171
172
def forward(self, x: torch.Tensor) -> torch.Tensor:
    x, _ = self.up_proj(x)
    x = self.act(x)
    x, _ = self.down_proj(x)
    return x

MPTModel

Bases: Module

Source code in vllm/model_executor/models/mpt.py 
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
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
@support_torch_compile
class MPTModel(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

        assert config.embedding_fraction == 1.0
        assert config.norm_type == "low_precision_layernorm"

        self.wte = VocabParallelEmbedding(
            config.vocab_size,
            config.d_model,
        )
        self.start_layer, self.end_layer, self.blocks = make_layers(
            config.n_layers,
            lambda prefix: MPTBlock(
                config, cache_config, quant_config, prefix=prefix),
            prefix=f"{prefix}.blocks")
        self.norm_f = nn.LayerNorm(config.d_model)
        if config.no_bias:
            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 is not None
            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]:
        params_dict = dict(self.named_parameters(remove_duplicate=False))
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue
            if is_pp_missing_parameter(name, self):
                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)

wte instance-attribute 

wte = VocabParallelEmbedding(vocab_size, d_model)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/mpt.py 
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
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

    assert config.embedding_fraction == 1.0
    assert config.norm_type == "low_precision_layernorm"

    self.wte = VocabParallelEmbedding(
        config.vocab_size,
        config.d_model,
    )
    self.start_layer, self.end_layer, self.blocks = make_layers(
        config.n_layers,
        lambda prefix: MPTBlock(
            config, cache_config, quant_config, prefix=prefix),
        prefix=f"{prefix}.blocks")
    self.norm_f = nn.LayerNorm(config.d_model)
    if config.no_bias:
        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/mpt.py 
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
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 is not None
        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/mpt.py 
244
245
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/mpt.py 
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    params_dict = dict(self.named_parameters(remove_duplicate=False))
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        # Skip loading extra bias for GPTQ models.
        if name.endswith(".bias") and name not in params_dict:
            continue
        if is_pp_missing_parameter(name, self):
            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

_get_alibi_slopes 

_get_alibi_slopes(
    total_num_heads: int, alibi_bias_max: int
) -> Tensor
Source code in vllm/model_executor/models/mpt.py 
36
37
38
39
40
41
42
43
44
45
46
def _get_alibi_slopes(
    total_num_heads: int,
    alibi_bias_max: int,
) -> torch.Tensor:
    next_power_of_2 = 2**math.ceil(math.log2(total_num_heads))
    m = torch.arange(1, next_power_of_2 + 1, dtype=torch.float32)
    m = m.mul(alibi_bias_max / next_power_of_2)
    slopes = 1.0 / torch.pow(2, m)
    if next_power_of_2 != total_num_heads:
        slopes = torch.concat([slopes[1::2], slopes[::2]])[:total_num_heads]
    return slopes