Skip to content

vllm.model_executor.models.starcoder2

PyTorch Starcoder2 model.

Starcoder2Attention

Bases: Module

Source code in vllm/model_executor/models/starcoder2.py 
 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
class Starcoder2Attention(nn.Module):

    def __init__(self,
                 config: Starcoder2Config,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.config = config

        self.hidden_size = config.hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = config.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 = config.num_key_value_heads
        if self.total_num_kv_heads >= tp_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_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_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = self.hidden_size // self.total_num_heads
        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.rope_theta = config.rope_theta
        self.max_position_embeddings = config.max_position_embeddings
        self.use_bias = config.use_bias

        self.qkv_proj = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=self.use_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            self.hidden_size,
            bias=self.use_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=self.max_position_embeddings,
            base=int(self.rope_theta),
            is_neox_style=True,
        )
        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,
        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.o_proj(attn_output)
        return output

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",
)

config instance-attribute 

config = config

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

max_position_embeddings instance-attribute 

max_position_embeddings = max_position_embeddings

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)

o_proj instance-attribute 

o_proj = RowParallelLinear(
    total_num_heads * head_dim,
    hidden_size,
    bias=use_bias,
    quant_config=quant_config,
    prefix=f"{prefix}.o_proj",
)

q_size instance-attribute 

q_size = num_heads * head_dim

qkv_proj instance-attribute 

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

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_embeddings,
    base=int(rope_theta),
    is_neox_style=True,
)

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 = num_key_value_heads

use_bias instance-attribute 

use_bias = use_bias

__init__ 

__init__(
    config: Starcoder2Config,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/starcoder2.py 
 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
def __init__(self,
             config: Starcoder2Config,
             cache_config: Optional[CacheConfig] = None,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()
    self.config = config

    self.hidden_size = config.hidden_size
    tp_size = get_tensor_model_parallel_world_size()
    self.total_num_heads = config.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 = config.num_key_value_heads
    if self.total_num_kv_heads >= tp_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_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_size % self.total_num_kv_heads == 0
    self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
    self.head_dim = self.hidden_size // self.total_num_heads
    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.rope_theta = config.rope_theta
    self.max_position_embeddings = config.max_position_embeddings
    self.use_bias = config.use_bias

    self.qkv_proj = QKVParallelLinear(
        self.hidden_size,
        self.head_dim,
        self.total_num_heads,
        self.total_num_kv_heads,
        bias=self.use_bias,
        quant_config=quant_config,
        prefix=f"{prefix}.qkv_proj",
    )
    self.o_proj = RowParallelLinear(
        self.total_num_heads * self.head_dim,
        self.hidden_size,
        bias=self.use_bias,
        quant_config=quant_config,
        prefix=f"{prefix}.o_proj",
    )
    self.rotary_emb = get_rope(
        self.head_dim,
        rotary_dim=self.head_dim,
        max_position=self.max_position_embeddings,
        base=int(self.rope_theta),
        is_neox_style=True,
    )
    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(positions: Tensor, hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/starcoder2.py 
118
119
120
121
122
123
124
125
126
127
128
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.o_proj(attn_output)
    return output

Starcoder2DecoderLayer

Bases: Module

Source code in vllm/model_executor/models/starcoder2.py 
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
class Starcoder2DecoderLayer(nn.Module):

    def __init__(self,
                 config: Starcoder2Config,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = Starcoder2Attention(config,
                                             cache_config,
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.self_attn")
        self.mlp = Starcoder2MLP(config,
                                 quant_config=quant_config,
                                 prefix=f"{prefix}.mlp")
        self.input_layernorm = nn.LayerNorm(config.hidden_size,
                                            eps=config.norm_epsilon)
        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
                                                     eps=config.norm_epsilon)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        # Self Attention
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states

hidden_size instance-attribute 

hidden_size = hidden_size

input_layernorm instance-attribute 

input_layernorm = LayerNorm(hidden_size, eps=norm_epsilon)

mlp instance-attribute 

mlp = Starcoder2MLP(
    config,
    quant_config=quant_config,
    prefix=f"{prefix}.mlp",
)

post_attention_layernorm instance-attribute 

post_attention_layernorm = LayerNorm(
    hidden_size, eps=norm_epsilon
)

self_attn instance-attribute 

self_attn = Starcoder2Attention(
    config,
    cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.self_attn",
)

__init__ 

__init__(
    config: Starcoder2Config,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/starcoder2.py 
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
def __init__(self,
             config: Starcoder2Config,
             cache_config: Optional[CacheConfig] = None,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()
    self.hidden_size = config.hidden_size
    self.self_attn = Starcoder2Attention(config,
                                         cache_config,
                                         quant_config=quant_config,
                                         prefix=f"{prefix}.self_attn")
    self.mlp = Starcoder2MLP(config,
                             quant_config=quant_config,
                             prefix=f"{prefix}.mlp")
    self.input_layernorm = nn.LayerNorm(config.hidden_size,
                                        eps=config.norm_epsilon)
    self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
                                                 eps=config.norm_epsilon)

forward 

forward(positions: Tensor, hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/starcoder2.py 
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    # Self Attention
    residual = hidden_states
    hidden_states = self.input_layernorm(hidden_states)
    hidden_states = self.self_attn(
        positions=positions,
        hidden_states=hidden_states,
    )
    hidden_states = residual + hidden_states

    # Fully Connected
    residual = hidden_states
    hidden_states = self.post_attention_layernorm(hidden_states)
    hidden_states = self.mlp(hidden_states)
    hidden_states = residual + hidden_states

    return hidden_states

Starcoder2ForCausalLM

Bases: Module, SupportsPP

Source code in vllm/model_executor/models/starcoder2.py 
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
class Starcoder2ForCausalLM(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
        self.model = Starcoder2Model(vllm_config=vllm_config,
                                     prefix=maybe_prefix(prefix, "model"))
        self.vocab_size = config.vocab_size
        self.unpadded_vocab_size = config.vocab_size
        if config.tie_word_embeddings:
            self.lm_head = self.model.embed_tokens
        else:
            self.unpadded_vocab_size = config.vocab_size
            self.lm_head = ParallelLMHead(
                self.unpadded_vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
                padding_size=DEFAULT_VOCAB_PADDING_SIZE,
                quant_config=quant_config,
                prefix=f"{prefix}.lm_head",
            )
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.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.model(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,
            # Models trained using ColossalAI may include these tensors in
            # the checkpoint. Skip them.
            skip_prefixes=(["lm_head.weight"]
                           if self.config.tie_word_embeddings else None),
        )
        return loader.load_weights(weights)

config instance-attribute 

config = config

lm_head instance-attribute 

lm_head = embed_tokens

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
)

model instance-attribute 

model = Starcoder2Model(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "model"),
)

unpadded_vocab_size instance-attribute 

unpadded_vocab_size = vocab_size

vocab_size instance-attribute 

vocab_size = vocab_size

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/starcoder2.py 
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
def __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
    self.model = Starcoder2Model(vllm_config=vllm_config,
                                 prefix=maybe_prefix(prefix, "model"))
    self.vocab_size = config.vocab_size
    self.unpadded_vocab_size = config.vocab_size
    if config.tie_word_embeddings:
        self.lm_head = self.model.embed_tokens
    else:
        self.unpadded_vocab_size = config.vocab_size
        self.lm_head = ParallelLMHead(
            self.unpadded_vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            padding_size=DEFAULT_VOCAB_PADDING_SIZE,
            quant_config=quant_config,
            prefix=f"{prefix}.lm_head",
        )
    self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                            config.vocab_size)
    self.make_empty_intermediate_tensors = (
        self.model.make_empty_intermediate_tensors)

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]
Source code in vllm/model_executor/models/starcoder2.py 
338
339
340
341
342
343
344
345
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/starcoder2.py 
327
328
329
330
331
332
333
334
335
336
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.model(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/starcoder2.py 
324
325
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.model.get_input_embeddings(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/starcoder2.py 
347
348
349
350
351
352
353
354
355
356
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(
        self,
        # Models trained using ColossalAI may include these tensors in
        # the checkpoint. Skip them.
        skip_prefixes=(["lm_head.weight"]
                       if self.config.tie_word_embeddings else None),
    )
    return loader.load_weights(weights)

Starcoder2MLP

Bases: Module

Source code in vllm/model_executor/models/starcoder2.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
class Starcoder2MLP(nn.Module):

    def __init__(self,
                 config: Starcoder2Config,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.c_fc = ColumnParallelLinear(
            config.hidden_size,
            config.intermediate_size,
            bias=config.use_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.c_fc",
        )
        self.c_proj = RowParallelLinear(
            config.intermediate_size,
            config.hidden_size,
            bias=config.use_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.c_proj",
        )
        self.act = get_act_fn(config.hidden_act)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.c_fc(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.c_proj(hidden_states)
        return hidden_states

act instance-attribute 

act = get_act_fn(hidden_act)

c_fc instance-attribute 

c_fc = ColumnParallelLinear(
    hidden_size,
    intermediate_size,
    bias=use_bias,
    quant_config=quant_config,
    prefix=f"{prefix}.c_fc",
)

c_proj instance-attribute 

c_proj = RowParallelLinear(
    intermediate_size,
    hidden_size,
    bias=use_bias,
    quant_config=quant_config,
    prefix=f"{prefix}.c_proj",
)

__init__ 

__init__(
    config: Starcoder2Config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/starcoder2.py 
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
def __init__(self,
             config: Starcoder2Config,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()
    self.c_fc = ColumnParallelLinear(
        config.hidden_size,
        config.intermediate_size,
        bias=config.use_bias,
        quant_config=quant_config,
        prefix=f"{prefix}.c_fc",
    )
    self.c_proj = RowParallelLinear(
        config.intermediate_size,
        config.hidden_size,
        bias=config.use_bias,
        quant_config=quant_config,
        prefix=f"{prefix}.c_proj",
    )
    self.act = get_act_fn(config.hidden_act)

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/starcoder2.py 
154
155
156
157
158
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    hidden_states, _ = self.c_fc(hidden_states)
    hidden_states = self.act(hidden_states)
    hidden_states, _ = self.c_proj(hidden_states)
    return hidden_states

Starcoder2Model

Bases: Module

Source code in vllm/model_executor/models/starcoder2.py 
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
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
287
288
289
290
291
292
293
@support_torch_compile
class Starcoder2Model(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.config = config
        self.vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=f"{prefix}.embed_tokens")
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: Starcoder2DecoderLayer(
                config, cache_config, quant_config=quant_config, prefix=prefix
            ),
            prefix=f"{prefix}.layers",
        )
        self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_epsilon)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(["hidden_states"],
                                                    config.hidden_size))

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: 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 layer in self.layers[self.start_layer:self.end_layer]:
            hidden_states = layer(positions, hidden_states)
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({"hidden_states": hidden_states})
        hidden_states = self.norm(hidden_states)
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]

        params_dict = dict(self.named_parameters(remove_duplicate=False))
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            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)
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    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

config instance-attribute 

config = config

embed_tokens instance-attribute 

embed_tokens = VocabParallelEmbedding(
    vocab_size,
    hidden_size,
    quant_config=quant_config,
    prefix=f"{prefix}.embed_tokens",
)

make_empty_intermediate_tensors instance-attribute 

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

norm instance-attribute 

norm = LayerNorm(hidden_size, eps=norm_epsilon)

vocab_size instance-attribute 

vocab_size = vocab_size

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/starcoder2.py 
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
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.config = config
    self.vocab_size = config.vocab_size

    self.embed_tokens = VocabParallelEmbedding(
        config.vocab_size,
        config.hidden_size,
        quant_config=quant_config,
        prefix=f"{prefix}.embed_tokens")
    self.start_layer, self.end_layer, self.layers = make_layers(
        config.num_hidden_layers,
        lambda prefix: Starcoder2DecoderLayer(
            config, cache_config, quant_config=quant_config, prefix=prefix
        ),
        prefix=f"{prefix}.layers",
    )
    self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_epsilon)
    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(["hidden_states"],
                                                config.hidden_size))

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[IntermediateTensors],
    inputs_embeds: Optional[Tensor] = None,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/starcoder2.py 
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
def forward(
    self,
    input_ids: torch.Tensor,
    positions: 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 layer in self.layers[self.start_layer:self.end_layer]:
        hidden_states = layer(positions, hidden_states)
    if not get_pp_group().is_last_rank:
        return IntermediateTensors({"hidden_states": hidden_states})
    hidden_states = self.norm(hidden_states)
    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/starcoder2.py 
235
236
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.embed_tokens(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/starcoder2.py 
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
287
288
289
290
291
292
293
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    stacked_params_mapping = [
        # (param_name, shard_name, shard_id)
        ("qkv_proj", "q_proj", "q"),
        ("qkv_proj", "k_proj", "k"),
        ("qkv_proj", "v_proj", "v"),
    ]

    params_dict = dict(self.named_parameters(remove_duplicate=False))
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        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)
            if is_pp_missing_parameter(name, self):
                continue
            param = params_dict[name]
            weight_loader = param.weight_loader
            weight_loader(param, loaded_weight, shard_id)
            break
        else:
            name = maybe_remap_kv_scale_name(name, params_dict)
            if name is None:
                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