Skip to content

vllm.model_executor.models.gpt_neox

Inference-only GPT-NeoX model compatible with HuggingFace weights.

GPTNeoXAttention

Bases: Module

Source code in vllm/model_executor/models/gpt_neox.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
class GPTNeoXAttention(nn.Module):

    def __init__(
        self,
        config: GPTNeoXConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.total_num_heads = config.num_attention_heads
        self.hidden_size = config.hidden_size
        self.head_size = self.hidden_size // self.total_num_heads
        self.bias = getattr(config, "attention_bias", True)

        tensor_model_parallel_world_size = (
            get_tensor_model_parallel_world_size())
        assert self.total_num_heads % tensor_model_parallel_world_size == 0
        self.num_heads = (self.total_num_heads //
                          tensor_model_parallel_world_size)

        self.query_key_value = QKVParallelLinear(
            config.hidden_size,
            self.head_size,
            self.total_num_heads,
            bias=self.bias,
            quant_config=quant_config,
        )
        self.dense = RowParallelLinear(
            config.hidden_size,
            config.hidden_size,
            bias=self.bias,
            quant_config=quant_config,
        )
        scaling = self.head_size**-0.5
        rotary_dim = int(self.head_size * config.rotary_pct)
        assert rotary_dim % 2 == 0
        rope_theta = getattr(config, "rope_theta", 10000)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        self.rotary_emb = get_rope(
            self.head_size,
            rotary_dim=rotary_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
        )
        self.attn = Attention(self.num_heads,
                              self.head_size,
                              scaling,
                              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.query_key_value(hidden_states)
        q, k, v = qkv.chunk(chunks=3, dim=-1)
        q, k = self.rotary_emb(position_ids, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.dense(attn_output)
        return output

attn instance-attribute 

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

bias instance-attribute 

bias = getattr(config, 'attention_bias', True)

dense instance-attribute 

dense = RowParallelLinear(
    hidden_size,
    hidden_size,
    bias=bias,
    quant_config=quant_config,
)

head_size instance-attribute 

head_size = hidden_size // total_num_heads

hidden_size instance-attribute 

hidden_size = hidden_size

num_heads instance-attribute 

num_heads = (
    total_num_heads // tensor_model_parallel_world_size
)

query_key_value instance-attribute 

query_key_value = QKVParallelLinear(
    hidden_size,
    head_size,
    total_num_heads,
    bias=bias,
    quant_config=quant_config,
)

rotary_emb instance-attribute 

rotary_emb = get_rope(
    head_size,
    rotary_dim=rotary_dim,
    max_position=max_position_embeddings,
    base=rope_theta,
)

total_num_heads instance-attribute 

total_num_heads = num_attention_heads

__init__ 

__init__(
    config: GPTNeoXConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/gpt_neox.py 
 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
def __init__(
    self,
    config: GPTNeoXConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.total_num_heads = config.num_attention_heads
    self.hidden_size = config.hidden_size
    self.head_size = self.hidden_size // self.total_num_heads
    self.bias = getattr(config, "attention_bias", True)

    tensor_model_parallel_world_size = (
        get_tensor_model_parallel_world_size())
    assert self.total_num_heads % tensor_model_parallel_world_size == 0
    self.num_heads = (self.total_num_heads //
                      tensor_model_parallel_world_size)

    self.query_key_value = QKVParallelLinear(
        config.hidden_size,
        self.head_size,
        self.total_num_heads,
        bias=self.bias,
        quant_config=quant_config,
    )
    self.dense = RowParallelLinear(
        config.hidden_size,
        config.hidden_size,
        bias=self.bias,
        quant_config=quant_config,
    )
    scaling = self.head_size**-0.5
    rotary_dim = int(self.head_size * config.rotary_pct)
    assert rotary_dim % 2 == 0
    rope_theta = getattr(config, "rope_theta", 10000)
    max_position_embeddings = getattr(config, "max_position_embeddings",
                                      8192)
    self.rotary_emb = get_rope(
        self.head_size,
        rotary_dim=rotary_dim,
        max_position=max_position_embeddings,
        base=rope_theta,
    )
    self.attn = Attention(self.num_heads,
                          self.head_size,
                          scaling,
                          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/gpt_neox.py 
104
105
106
107
108
109
110
111
112
113
114
def forward(
    self,
    position_ids: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    qkv, _ = self.query_key_value(hidden_states)
    q, k, v = qkv.chunk(chunks=3, dim=-1)
    q, k = self.rotary_emb(position_ids, q, k)
    attn_output = self.attn(q, k, v)
    output, _ = self.dense(attn_output)
    return output

GPTNeoXForCausalLM

Bases: Module, SupportsPP

Source code in vllm/model_executor/models/gpt_neox.py 
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
class GPTNeoXForCausalLM(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.quant_config = quant_config
        self.gpt_neox = GPTNeoXModel(vllm_config=vllm_config,
                                     prefix=maybe_prefix(prefix, "gpt_neox"))
        self.embed_out = ParallelLMHead(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
        )
        if self.config.tie_word_embeddings:
            self.embed_out.weight = self.gpt_neox.embed_in.weight
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.gpt_neox.make_empty_intermediate_tensors)

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

embed_out instance-attribute 

embed_out = ParallelLMHead(
    vocab_size, hidden_size, quant_config=quant_config
)

gpt_neox instance-attribute 

gpt_neox = GPTNeoXModel(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "gpt_neox"),
)

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

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/gpt_neox.py 
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
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.quant_config = quant_config
    self.gpt_neox = GPTNeoXModel(vllm_config=vllm_config,
                                 prefix=maybe_prefix(prefix, "gpt_neox"))
    self.embed_out = ParallelLMHead(
        config.vocab_size,
        config.hidden_size,
        quant_config=quant_config,
    )
    if self.config.tie_word_embeddings:
        self.embed_out.weight = self.gpt_neox.embed_in.weight
    self.logits_processor = LogitsProcessor(config.vocab_size)
    self.make_empty_intermediate_tensors = (
        self.gpt_neox.make_empty_intermediate_tensors)

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]
Source code in vllm/model_executor/models/gpt_neox.py 
320
321
322
323
324
325
326
327
def compute_logits(
    self,
    hidden_states: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[torch.Tensor]:
    logits = self.logits_processor(self.embed_out, 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/gpt_neox.py 
309
310
311
312
313
314
315
316
317
318
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.gpt_neox(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/gpt_neox.py 
306
307
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.gpt_neox.get_input_embeddings(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/gpt_neox.py 
329
330
331
332
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(self)
    return loader.load_weights(weights)

GPTNeoXLayer

Bases: Module

Source code in vllm/model_executor/models/gpt_neox.py 
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
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
class GPTNeoXLayer(nn.Module):

    def __init__(
        self,
        config: GPTNeoXConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.use_parallel_residual = config.use_parallel_residual
        self.input_layernorm = nn.LayerNorm(config.hidden_size,
                                            eps=config.layer_norm_eps)
        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
                                                     eps=config.layer_norm_eps)
        self.attention = GPTNeoXAttention(config,
                                          cache_config,
                                          quant_config,
                                          prefix=f"{prefix}.attention")
        self.mlp = GPTNeoXMLP(config, quant_config)

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        attn_input = self.input_layernorm(hidden_states)
        attn_output = self.attention(
            position_ids=position_ids,
            hidden_states=attn_input,
        )

        if self.use_parallel_residual:
            # pseudocode:
            # x = x + attn(ln1(x)) + mlp(ln2(x))
            mlp_input = self.post_attention_layernorm(hidden_states)
            mlp_output = self.mlp(mlp_input)
            hidden_states = mlp_output + attn_output + hidden_states
        else:
            # pseudocode:
            # x = x + attn(ln1(x))
            # x = x + mlp(ln2(x))
            attn_output = attn_output + hidden_states
            mlp_input = self.post_attention_layernorm(attn_output)
            mlp_output = self.mlp(mlp_input)
            hidden_states = mlp_output + attn_output
        return hidden_states

attention instance-attribute 

attention = GPTNeoXAttention(
    config,
    cache_config,
    quant_config,
    prefix=f"{prefix}.attention",
)

input_layernorm instance-attribute 

input_layernorm = LayerNorm(hidden_size, eps=layer_norm_eps)

mlp instance-attribute 

mlp = GPTNeoXMLP(config, quant_config)

post_attention_layernorm instance-attribute 

post_attention_layernorm = LayerNorm(
    hidden_size, eps=layer_norm_eps
)

use_parallel_residual instance-attribute 

use_parallel_residual = use_parallel_residual

__init__ 

__init__(
    config: GPTNeoXConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/gpt_neox.py 
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
def __init__(
    self,
    config: GPTNeoXConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.use_parallel_residual = config.use_parallel_residual
    self.input_layernorm = nn.LayerNorm(config.hidden_size,
                                        eps=config.layer_norm_eps)
    self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
                                                 eps=config.layer_norm_eps)
    self.attention = GPTNeoXAttention(config,
                                      cache_config,
                                      quant_config,
                                      prefix=f"{prefix}.attention")
    self.mlp = GPTNeoXMLP(config, quant_config)

forward 

forward(
    position_ids: Tensor, hidden_states: Tensor
) -> Tensor
Source code in vllm/model_executor/models/gpt_neox.py 
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
def forward(
    self,
    position_ids: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    attn_input = self.input_layernorm(hidden_states)
    attn_output = self.attention(
        position_ids=position_ids,
        hidden_states=attn_input,
    )

    if self.use_parallel_residual:
        # pseudocode:
        # x = x + attn(ln1(x)) + mlp(ln2(x))
        mlp_input = self.post_attention_layernorm(hidden_states)
        mlp_output = self.mlp(mlp_input)
        hidden_states = mlp_output + attn_output + hidden_states
    else:
        # pseudocode:
        # x = x + attn(ln1(x))
        # x = x + mlp(ln2(x))
        attn_output = attn_output + hidden_states
        mlp_input = self.post_attention_layernorm(attn_output)
        mlp_output = self.mlp(mlp_input)
        hidden_states = mlp_output + attn_output
    return hidden_states

GPTNeoXMLP

Bases: Module

Source code in vllm/model_executor/models/gpt_neox.py 
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
class GPTNeoXMLP(nn.Module):

    def __init__(
        self,
        config: GPTNeoXConfig,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        self.dense_h_to_4h = ColumnParallelLinear(
            config.hidden_size,
            config.intermediate_size,
            quant_config=quant_config,
        )
        self.dense_4h_to_h = RowParallelLinear(
            config.intermediate_size,
            config.hidden_size,
            quant_config=quant_config,
        )
        self.act = get_act_fn(config.hidden_act)

    def forward(self, hidden_states):
        hidden_states, _ = self.dense_h_to_4h(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.dense_4h_to_h(hidden_states)
        return hidden_states

act instance-attribute 

act = get_act_fn(hidden_act)

dense_4h_to_h instance-attribute 

dense_4h_to_h = RowParallelLinear(
    intermediate_size,
    hidden_size,
    quant_config=quant_config,
)

dense_h_to_4h instance-attribute 

dense_h_to_4h = ColumnParallelLinear(
    hidden_size,
    intermediate_size,
    quant_config=quant_config,
)

__init__ 

__init__(
    config: GPTNeoXConfig,
    quant_config: Optional[QuantizationConfig] = None,
)
Source code in vllm/model_executor/models/gpt_neox.py 
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
def __init__(
    self,
    config: GPTNeoXConfig,
    quant_config: Optional[QuantizationConfig] = None,
):
    super().__init__()
    self.dense_h_to_4h = ColumnParallelLinear(
        config.hidden_size,
        config.intermediate_size,
        quant_config=quant_config,
    )
    self.dense_4h_to_h = RowParallelLinear(
        config.intermediate_size,
        config.hidden_size,
        quant_config=quant_config,
    )
    self.act = get_act_fn(config.hidden_act)

forward 

forward(hidden_states)
Source code in vllm/model_executor/models/gpt_neox.py 
137
138
139
140
141
def forward(self, hidden_states):
    hidden_states, _ = self.dense_h_to_4h(hidden_states)
    hidden_states = self.act(hidden_states)
    hidden_states, _ = self.dense_4h_to_h(hidden_states)
    return hidden_states

GPTNeoXModel

Bases: Module

Source code in vllm/model_executor/models/gpt_neox.py 
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
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
@support_torch_compile
class GPTNeoXModel(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.embed_in = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: GPTNeoXLayer(
                config, cache_config, quant_config, prefix=prefix),
            prefix=f"{prefix}.layers",
        )
        self.final_layer_norm = nn.LayerNorm(config.hidden_size,
                                             eps=config.layer_norm_eps)
        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_in(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:
            hidden_states = intermediate_tensors["hidden_states"]
        for layer in self.layers[self.start_layer:self.end_layer]:
            hidden_states = layer(position_ids, hidden_states)
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({"hidden_states": hidden_states})
        hidden_states = self.final_layer_norm(hidden_states)
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if ("attention.bias" in name or "attention.masked_bias" in name
                    or "rotary_emb.inv_freq" in name):
                continue
            if ("rotary_emb.cos_cached" in name
                    or "rotary_emb.sin_cached" in name):
                # Models trained using OpenRLHF may include
                # these tensors in the checkpoint. Skip them.
                continue
            if is_pp_missing_parameter(name, self):
                continue
            param = params_dict[name]

            if "query_key_value" in name:
                # NOTE: GPT-NeoX's fused QKV's output_dim has the shape of
                # (num_heads * 3 * head_size), while the
                # required shape is (3 * num_heads * head_size).
                # Thus, we need weight conversion.
                output_dim = getattr(param, "output_dim", None)
                num_heads = self.config.num_attention_heads
                if output_dim is not None:
                    loaded_weight_shape = loaded_weight.shape
                    loaded_weight = loaded_weight.view(
                        loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
                        loaded_weight_shape[output_dim + 1:])
                    loaded_weight = loaded_weight.transpose(
                        output_dim, output_dim + 1)
                    loaded_weight = loaded_weight.reshape(loaded_weight_shape)

            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_in instance-attribute 

embed_in = VocabParallelEmbedding(vocab_size, hidden_size)

final_layer_norm instance-attribute 

final_layer_norm = LayerNorm(
    hidden_size, eps=layer_norm_eps
)

make_empty_intermediate_tensors instance-attribute 

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

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/gpt_neox.py 
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
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.embed_in = VocabParallelEmbedding(
        config.vocab_size,
        config.hidden_size,
    )
    self.start_layer, self.end_layer, self.layers = make_layers(
        config.num_hidden_layers,
        lambda prefix: GPTNeoXLayer(
            config, cache_config, quant_config, prefix=prefix),
        prefix=f"{prefix}.layers",
    )
    self.final_layer_norm = nn.LayerNorm(config.hidden_size,
                                         eps=config.layer_norm_eps)
    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(["hidden_states"],
                                                config.hidden_size))

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/gpt_neox.py 
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
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:
        hidden_states = intermediate_tensors["hidden_states"]
    for layer in self.layers[self.start_layer:self.end_layer]:
        hidden_states = layer(position_ids, hidden_states)
    if not get_pp_group().is_last_rank:
        return IntermediateTensors({"hidden_states": hidden_states})
    hidden_states = self.final_layer_norm(hidden_states)
    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/gpt_neox.py 
221
222
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.embed_in(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/gpt_neox.py 
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
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if ("attention.bias" in name or "attention.masked_bias" in name
                or "rotary_emb.inv_freq" in name):
            continue
        if ("rotary_emb.cos_cached" in name
                or "rotary_emb.sin_cached" in name):
            # Models trained using OpenRLHF may include
            # these tensors in the checkpoint. Skip them.
            continue
        if is_pp_missing_parameter(name, self):
            continue
        param = params_dict[name]

        if "query_key_value" in name:
            # NOTE: GPT-NeoX's fused QKV's output_dim has the shape of
            # (num_heads * 3 * head_size), while the
            # required shape is (3 * num_heads * head_size).
            # Thus, we need weight conversion.
            output_dim = getattr(param, "output_dim", None)
            num_heads = self.config.num_attention_heads
            if output_dim is not None:
                loaded_weight_shape = loaded_weight.shape
                loaded_weight = loaded_weight.view(
                    loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
                    loaded_weight_shape[output_dim + 1:])
                loaded_weight = loaded_weight.transpose(
                    output_dim, output_dim + 1)
                loaded_weight = loaded_weight.reshape(loaded_weight_shape)

        weight_loader = getattr(param, "weight_loader",
                                default_weight_loader)
        weight_loader(param, loaded_weight)
        loaded_params.add(name)
    return loaded_params