Skip to content

vllm.model_executor.models.minicpm_eagle

Inference-only EagleMiniCPM model compatible with HuggingFace weights.

EagleMiniCPMDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/minicpm_eagle.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
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
class EagleMiniCPMDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.cache_config = cache_config
        self.quant_config = quant_config
        self.hidden_size = config.hidden_size
        self.rope_theta = getattr(config, "rope_theta", 10000)
        self.rope_scaling = getattr(config, "rope_scaling", None)
        self.max_position_embeddings = getattr(config,
                                               "max_position_embeddings", 8192)
        self.prefix = prefix
        self._init_attn_block()
        self._init_ffn_block()

    def _init_attn_block(self):
        self.input_layernorm = RMSNorm(self.config.hidden_size,
                                       eps=self.config.rms_norm_eps)
        self.self_attn = EagleMiniCPMAttention(
            hidden_size=self.hidden_size,
            num_heads=self.config.num_attention_heads,
            num_kv_heads=self.config.num_key_value_heads,
            rope_theta=self.rope_theta,
            rope_scaling=self.rope_scaling,
            max_position_embeddings=self.max_position_embeddings,
            cache_config=self.cache_config,
            quant_config=self.quant_config,
            prefix=f"{self.prefix}.self_attn",
        )

    def _init_ffn_block(self):
        self.post_attention_layernorm = RMSNorm(self.config.hidden_size,
                                                eps=self.config.rms_norm_eps)
        self.num_experts = getattr(self.config, "num_experts", 0)
        if self.num_experts == 0:
            self.mlp = EagleMiniCPMMLP(
                hidden_size=self.hidden_size,
                intermediate_size=self.config.intermediate_size,
                hidden_act=self.config.hidden_act,
                hidden_act_param=getattr(self.config, "hidden_act_param", 0.),
                quant_config=self.quant_config,
            )
        else:
            self.mlp = EagleMiniCPMMoE(
                num_experts=self.config.num_experts,
                top_k=self.config.num_experts_per_tok,
                hidden_size=self.config.hidden_size,
                intermediate_size=self.config.intermediate_size)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> tuple[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 * \
            (self.config.scale_depth / math.sqrt(self.config.mup_denominator))

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states * \
            (self.config.scale_depth / math.sqrt(self.config.mup_denominator))

        return hidden_states, None

cache_config instance-attribute 

cache_config = cache_config

config instance-attribute 

config = config

hidden_size instance-attribute 

hidden_size = hidden_size

max_position_embeddings instance-attribute 

max_position_embeddings = getattr(
    config, "max_position_embeddings", 8192
)

prefix instance-attribute 

prefix = prefix

quant_config instance-attribute 

quant_config = quant_config

rope_scaling instance-attribute 

rope_scaling = getattr(config, 'rope_scaling', None)

rope_theta instance-attribute 

rope_theta = getattr(config, 'rope_theta', 10000)

__init__ 

__init__(
    config: PretrainedConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/minicpm_eagle.py 
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
def __init__(
    self,
    config: PretrainedConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.config = config
    self.cache_config = cache_config
    self.quant_config = quant_config
    self.hidden_size = config.hidden_size
    self.rope_theta = getattr(config, "rope_theta", 10000)
    self.rope_scaling = getattr(config, "rope_scaling", None)
    self.max_position_embeddings = getattr(config,
                                           "max_position_embeddings", 8192)
    self.prefix = prefix
    self._init_attn_block()
    self._init_ffn_block()

_init_attn_block 

_init_attn_block()
Source code in vllm/model_executor/models/minicpm_eagle.py 
75
76
77
78
79
80
81
82
83
84
85
86
87
88
def _init_attn_block(self):
    self.input_layernorm = RMSNorm(self.config.hidden_size,
                                   eps=self.config.rms_norm_eps)
    self.self_attn = EagleMiniCPMAttention(
        hidden_size=self.hidden_size,
        num_heads=self.config.num_attention_heads,
        num_kv_heads=self.config.num_key_value_heads,
        rope_theta=self.rope_theta,
        rope_scaling=self.rope_scaling,
        max_position_embeddings=self.max_position_embeddings,
        cache_config=self.cache_config,
        quant_config=self.quant_config,
        prefix=f"{self.prefix}.self_attn",
    )

_init_ffn_block 

_init_ffn_block()
Source code in vllm/model_executor/models/minicpm_eagle.py 
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
def _init_ffn_block(self):
    self.post_attention_layernorm = RMSNorm(self.config.hidden_size,
                                            eps=self.config.rms_norm_eps)
    self.num_experts = getattr(self.config, "num_experts", 0)
    if self.num_experts == 0:
        self.mlp = EagleMiniCPMMLP(
            hidden_size=self.hidden_size,
            intermediate_size=self.config.intermediate_size,
            hidden_act=self.config.hidden_act,
            hidden_act_param=getattr(self.config, "hidden_act_param", 0.),
            quant_config=self.quant_config,
        )
    else:
        self.mlp = EagleMiniCPMMoE(
            num_experts=self.config.num_experts,
            top_k=self.config.num_experts_per_tok,
            hidden_size=self.config.hidden_size,
            intermediate_size=self.config.intermediate_size)

forward 

forward(
    positions: Tensor,
    hidden_states: Tensor,
    residual: Optional[Tensor],
) -> tuple[Tensor, Tensor]
Source code in vllm/model_executor/models/minicpm_eagle.py 
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
) -> tuple[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 * \
        (self.config.scale_depth / math.sqrt(self.config.mup_denominator))

    # Fully Connected
    residual = hidden_states
    hidden_states = self.post_attention_layernorm(hidden_states)
    hidden_states = self.mlp(hidden_states)
    hidden_states = residual + hidden_states * \
        (self.config.scale_depth / math.sqrt(self.config.mup_denominator))

    return hidden_states, None

EagleMiniCPMForCausalLM

Bases: Module, SupportsLoRA, SupportsPP

Source code in vllm/model_executor/models/minicpm_eagle.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
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
class EagleMiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    # LoRA specific attributes
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }
    embedding_padding_modules = ["lm_head"]

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.speculative_config.draft_model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.prefix = prefix
        self.vllm_config = vllm_config
        self.config = config
        self.lora_config = lora_config
        self.cache_config = cache_config
        self.quant_config = quant_config

        target_layer_num = vllm_config.model_config.get_num_layers(
            vllm_config.parallel_config)

        self.model = self._init_model(vllm_config=vllm_config,
                                      prefix=maybe_prefix(prefix, "model"),
                                      start_layer=target_layer_num)

        unpadded_vocab_size = config.vocab_size
        if lora_config:
            unpadded_vocab_size += lora_config.lora_extra_vocab_size
        self.lm_head = ParallelLMHead(
            unpadded_vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            padding_size=DEFAULT_VOCAB_PADDING_SIZE
            # We need bigger padding if using lora for kernel
            # compatibility
            if not lora_config else lora_config.lora_vocab_padding_size,
            quant_config=quant_config,
        )
        if config.tie_word_embeddings:
            self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
        self.scale_width = self.config.hidden_size / self.config.dim_model_base

        self.logits_processor = LogitsProcessor(unpadded_vocab_size,
                                                config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

    def _init_model(self,
                    *,
                    vllm_config: VllmConfig,
                    prefix: str = "",
                    start_layer: int = 0):
        return EagleMiniCPMModel(vllm_config=vllm_config,
                                 prefix=prefix,
                                 start_layer=start_layer)

    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,
        hidden_states: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        hidden_states, hidden_states2 = self.model(input_ids, positions,
                                                   hidden_states)
        hidden_states = hidden_states / self.scale_width
        hidden_states2 = hidden_states2 / self.scale_width
        return hidden_states, hidden_states2

    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,
            skip_prefixes=(["lm_head."]
                           if self.config.tie_word_embeddings else None),
        )
        return loader.load_weights(weights)

cache_config instance-attribute 

cache_config = cache_config

config instance-attribute 

config = config

embedding_modules class-attribute instance-attribute 

embedding_modules = {
    "embed_tokens": "input_embeddings",
    "lm_head": "output_embeddings",
}

embedding_padding_modules class-attribute instance-attribute 

embedding_padding_modules = ['lm_head']

lm_head instance-attribute 

lm_head = ParallelLMHead(
    unpadded_vocab_size,
    hidden_size,
    org_num_embeddings=vocab_size,
    padding_size=DEFAULT_VOCAB_PADDING_SIZE
    if not lora_config
    else lora_vocab_padding_size,
    quant_config=quant_config,
)

logits_processor instance-attribute 

logits_processor = LogitsProcessor(
    unpadded_vocab_size, vocab_size
)

lora_config instance-attribute 

lora_config = lora_config

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

model instance-attribute 

model = _init_model(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "model"),
    start_layer=target_layer_num,
)

packed_modules_mapping class-attribute instance-attribute 

packed_modules_mapping = {
    "qkv_proj": ["q_proj", "k_proj", "v_proj"],
    "gate_up_proj": ["gate_proj", "up_proj"],
}

prefix instance-attribute 

prefix = prefix

quant_config instance-attribute 

quant_config = quant_config

scale_width instance-attribute 

scale_width = hidden_size / dim_model_base

vllm_config instance-attribute 

vllm_config = vllm_config

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/minicpm_eagle.py 
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
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.speculative_config.draft_model_config.hf_config
    cache_config = vllm_config.cache_config
    quant_config = vllm_config.quant_config
    lora_config = vllm_config.lora_config

    self.prefix = prefix
    self.vllm_config = vllm_config
    self.config = config
    self.lora_config = lora_config
    self.cache_config = cache_config
    self.quant_config = quant_config

    target_layer_num = vllm_config.model_config.get_num_layers(
        vllm_config.parallel_config)

    self.model = self._init_model(vllm_config=vllm_config,
                                  prefix=maybe_prefix(prefix, "model"),
                                  start_layer=target_layer_num)

    unpadded_vocab_size = config.vocab_size
    if lora_config:
        unpadded_vocab_size += lora_config.lora_extra_vocab_size
    self.lm_head = ParallelLMHead(
        unpadded_vocab_size,
        config.hidden_size,
        org_num_embeddings=config.vocab_size,
        padding_size=DEFAULT_VOCAB_PADDING_SIZE
        # We need bigger padding if using lora for kernel
        # compatibility
        if not lora_config else lora_config.lora_vocab_padding_size,
        quant_config=quant_config,
    )
    if config.tie_word_embeddings:
        self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
    self.scale_width = self.config.hidden_size / self.config.dim_model_base

    self.logits_processor = LogitsProcessor(unpadded_vocab_size,
                                            config.vocab_size)
    self.make_empty_intermediate_tensors = (
        self.model.make_empty_intermediate_tensors)

_init_model 

_init_model(
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    start_layer: int = 0,
)
Source code in vllm/model_executor/models/minicpm_eagle.py 
351
352
353
354
355
356
357
358
def _init_model(self,
                *,
                vllm_config: VllmConfig,
                prefix: str = "",
                start_layer: int = 0):
    return EagleMiniCPMModel(vllm_config=vllm_config,
                             prefix=prefix,
                             start_layer=start_layer)

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]
Source code in vllm/model_executor/models/minicpm_eagle.py 
375
376
377
378
379
380
381
382
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,
    hidden_states: Tensor,
) -> tuple[Tensor, Tensor]
Source code in vllm/model_executor/models/minicpm_eagle.py 
363
364
365
366
367
368
369
370
371
372
373
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
    hidden_states, hidden_states2 = self.model(input_ids, positions,
                                               hidden_states)
    hidden_states = hidden_states / self.scale_width
    hidden_states2 = hidden_states2 / self.scale_width
    return hidden_states, hidden_states2

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/minicpm_eagle.py 
360
361
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/minicpm_eagle.py 
384
385
386
387
388
389
390
391
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(
        self,
        skip_prefixes=(["lm_head."]
                       if self.config.tie_word_embeddings else None),
    )
    return loader.load_weights(weights)

EagleMiniCPMModel

Bases: Module

Source code in vllm/model_executor/models/minicpm_eagle.py 
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
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
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
283
284
285
@support_torch_compile
class EagleMiniCPMModel(nn.Module):

    def __init__(self,
                 *,
                 vllm_config: VllmConfig,
                 prefix: str = "",
                 start_layer: int = 0):
        super().__init__()

        config = vllm_config.speculative_config.draft_model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
        self.cache_config = cache_config
        self.quant_config = quant_config
        lora_vocab = (lora_config.lora_extra_vocab_size *
                      (lora_config.max_loras or 1)) if lora_config else 0
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size
        self.fc = torch.nn.Linear(self.config.hidden_size * 2,
                                  self.config.hidden_size,
                                  bias=False)
        self.input_norm1 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.input_norm2 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.embed_tokens = VocabParallelEmbedding(
            self.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
        )
        self.num_experts = getattr(self.config, "num_experts", 0)
        self._init_layers(prefix, config, cache_config, quant_config,
                          start_layer)
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], self.config.hidden_size))

    def _init_layers(
        self,
        prefix: str,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig],
        quant_config: Optional[QuantizationConfig],
        start_layer: int,
    ):
        self.eagle_layers = nn.ModuleList([
            EagleMiniCPMDecoderLayer(
                config,
                cache_config,
                quant_config,
                f"{prefix}.eagle_layers.{i + start_layer}",
            ) for i in range(self.config.num_hidden_layers)
        ])

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        embedding = self.embed_tokens(input_ids)
        return embedding * self.config.scale_emb

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        input_embeds = self.get_input_embeddings(input_ids)
        input_embeds = self.input_norm1(input_embeds)
        hidden_states = self.input_norm2(hidden_states)

        hidden_states = self.fc(
            torch.cat((input_embeds, hidden_states), dim=-1))
        residual = None
        for layer in self.eagle_layers:
            hidden_states, residual = layer(
                positions,
                hidden_states,
                residual,
            )

        return hidden_states, 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"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
        expert_params_mapping = [
            # (param_name, weight_name, expert_id)
            ("ws" if weight_name in ["w1", "w3"] else "w2s",
             f"experts.{expert_id}.{weight_name}.weight", expert_id)
            for expert_id in range(self.num_experts)
            for weight_name in ["w1", "w2", "w3"]
        ]
        params_dict = dict(self.named_parameters())

        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
            if ("rotary_emb.cos_cached" in name
                    or "rotary_emb.sin_cached" in name):
                # Models trained using ColossalAI may include these tensors in
                # the checkpoint. Skip them.
                continue
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                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:
                for param_name, weight_name, expert_id 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,
                                  expert_id=expert_id)
                    break
                else:
                    # 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

cache_config instance-attribute 

cache_config = cache_config

config instance-attribute 

config = config

embed_tokens instance-attribute 

embed_tokens = VocabParallelEmbedding(
    vocab_size, hidden_size, org_num_embeddings=vocab_size
)

fc instance-attribute 

fc = Linear(hidden_size * 2, hidden_size, bias=False)

input_norm1 instance-attribute 

input_norm1 = RMSNorm(hidden_size, eps=rms_norm_eps)

input_norm2 instance-attribute 

input_norm2 = RMSNorm(hidden_size, eps=rms_norm_eps)

make_empty_intermediate_tensors instance-attribute 

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

norm instance-attribute 

norm = RMSNorm(hidden_size, eps=rms_norm_eps)

num_experts instance-attribute 

num_experts = getattr(config, 'num_experts', 0)

org_vocab_size instance-attribute 

org_vocab_size = vocab_size

quant_config instance-attribute 

quant_config = quant_config

vocab_size instance-attribute 

vocab_size = vocab_size + lora_vocab

__init__ 

__init__(
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    start_layer: int = 0,
)
Source code in vllm/model_executor/models/minicpm_eagle.py 
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
166
167
168
169
170
171
172
173
def __init__(self,
             *,
             vllm_config: VllmConfig,
             prefix: str = "",
             start_layer: int = 0):
    super().__init__()

    config = vllm_config.speculative_config.draft_model_config.hf_config
    cache_config = vllm_config.cache_config
    quant_config = vllm_config.quant_config
    lora_config = vllm_config.lora_config

    self.config = config
    self.cache_config = cache_config
    self.quant_config = quant_config
    lora_vocab = (lora_config.lora_extra_vocab_size *
                  (lora_config.max_loras or 1)) if lora_config else 0
    self.vocab_size = config.vocab_size + lora_vocab
    self.org_vocab_size = config.vocab_size
    self.fc = torch.nn.Linear(self.config.hidden_size * 2,
                              self.config.hidden_size,
                              bias=False)
    self.input_norm1 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
    self.input_norm2 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
    self.embed_tokens = VocabParallelEmbedding(
        self.vocab_size,
        config.hidden_size,
        org_num_embeddings=config.vocab_size,
    )
    self.num_experts = getattr(self.config, "num_experts", 0)
    self._init_layers(prefix, config, cache_config, quant_config,
                      start_layer)
    self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], self.config.hidden_size))

_init_layers 

_init_layers(
    prefix: str,
    config: PretrainedConfig,
    cache_config: Optional[CacheConfig],
    quant_config: Optional[QuantizationConfig],
    start_layer: int,
)
Source code in vllm/model_executor/models/minicpm_eagle.py 
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
def _init_layers(
    self,
    prefix: str,
    config: PretrainedConfig,
    cache_config: Optional[CacheConfig],
    quant_config: Optional[QuantizationConfig],
    start_layer: int,
):
    self.eagle_layers = nn.ModuleList([
        EagleMiniCPMDecoderLayer(
            config,
            cache_config,
            quant_config,
            f"{prefix}.eagle_layers.{i + start_layer}",
        ) for i in range(self.config.num_hidden_layers)
    ])

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    hidden_states: Tensor,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/minicpm_eagle.py 
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> Union[torch.Tensor, IntermediateTensors]:
    input_embeds = self.get_input_embeddings(input_ids)
    input_embeds = self.input_norm1(input_embeds)
    hidden_states = self.input_norm2(hidden_states)

    hidden_states = self.fc(
        torch.cat((input_embeds, hidden_states), dim=-1))
    residual = None
    for layer in self.eagle_layers:
        hidden_states, residual = layer(
            positions,
            hidden_states,
            residual,
        )

    return hidden_states, hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/minicpm_eagle.py 
192
193
194
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    embedding = self.embed_tokens(input_ids)
    return embedding * self.config.scale_emb

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/minicpm_eagle.py 
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
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"),
        ("gate_up_proj", "gate_proj", 0),
        ("gate_up_proj", "up_proj", 1),
    ]
    expert_params_mapping = [
        # (param_name, weight_name, expert_id)
        ("ws" if weight_name in ["w1", "w3"] else "w2s",
         f"experts.{expert_id}.{weight_name}.weight", expert_id)
        for expert_id in range(self.num_experts)
        for weight_name in ["w1", "w2", "w3"]
    ]
    params_dict = dict(self.named_parameters())

    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if "rotary_emb.inv_freq" in name:
            continue
        if ("rotary_emb.cos_cached" in name
                or "rotary_emb.sin_cached" in name):
            # Models trained using ColossalAI may include these tensors in
            # the checkpoint. Skip them.
            continue
        for (param_name, weight_name, shard_id) in stacked_params_mapping:
            if weight_name not in name:
                continue
            name = name.replace(weight_name, param_name)
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue
            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:
            for param_name, weight_name, expert_id 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,
                              expert_id=expert_id)
                break
            else:
                # 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