minicpm.py 23.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

ywfang's avatar
ywfang committed
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# Copyright 2023 The vLLM team.
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only MiniCPM model compatible with HuggingFace weights."""
26

ywfang's avatar
ywfang committed
27
import math
28
from collections.abc import Iterable
29
from itertools import islice
30
from typing import Any
ywfang's avatar
ywfang committed
31
32
33

import torch
from torch import nn
34
from transformers import PretrainedConfig
ywfang's avatar
ywfang committed
35

36
from vllm.attention.layer import Attention
37
from vllm.compilation.decorators import support_torch_compile
38
from vllm.config import CacheConfig, VllmConfig
39
40
41
42
43
44
from vllm.distributed import (
    get_pp_group,
    get_tensor_model_parallel_rank,
    get_tensor_model_parallel_world_size,
    tensor_model_parallel_all_reduce,
)
45
from vllm.model_executor.layers.activation import FatreluAndMul, SiluAndMul
46
from vllm.model_executor.layers.fused_moe import fused_experts, fused_topk
ywfang's avatar
ywfang committed
47
from vllm.model_executor.layers.layernorm import RMSNorm
48
49
50
51
52
53
from vllm.model_executor.layers.linear import (
    MergedColumnParallelLinear,
    QKVParallelLinear,
    ReplicatedLinear,
    RowParallelLinear,
)
ywfang's avatar
ywfang committed
54
from vllm.model_executor.layers.logits_processor import LogitsProcessor
55
from vllm.model_executor.layers.quantization import QuantizationConfig
ywfang's avatar
ywfang committed
56
57
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
58
59
60
    ParallelLMHead,
    VocabParallelEmbedding,
)
61
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
ywfang's avatar
ywfang committed
62
from vllm.model_executor.utils import set_weight_attrs
63
from vllm.platforms import current_platform
64
from vllm.sequence import IntermediateTensors
ywfang's avatar
ywfang committed
65

66
from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
67
68
69
70
71
72
73
from .utils import (
    AutoWeightsLoader,
    is_pp_missing_parameter,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
74

ywfang's avatar
ywfang committed
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90

class MiniCPMMoE(nn.Module):
    """A tensor-parallel MoE implementation that shards each expert
    across all ranks.

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

    def __init__(
        self,
        num_experts: int,
        top_k: int,
        hidden_size: int,
        intermediate_size: int,
91
92
        params_dtype: torch.dtype | None = None,
        tp_size: int | None = None,
ywfang's avatar
ywfang committed
93
94
95
96
97
98
99
100
101
102
103
104
    ):
        super().__init__()
        self.tp_size = tp_size or get_tensor_model_parallel_world_size()
        self.num_total_experts = num_experts
        self.top_k = top_k
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size // self.tp_size

        if params_dtype is None:
            params_dtype = torch.get_default_dtype()
        self.params_dtype = params_dtype

105
106
107
108
109
110
111
        self.gate = ReplicatedLinear(
            self.hidden_size,
            self.num_total_experts,
            bias=False,
            params_dtype=self.params_dtype,
            quant_config=None,
        )
ywfang's avatar
ywfang committed
112
113

        self.ws = nn.Parameter(
114
115
116
117
118
119
120
121
            torch.empty(
                self.num_total_experts,
                2 * self.intermediate_size,
                self.hidden_size,
                device=current_platform.device_type,
                dtype=self.params_dtype,
            )
        )
ywfang's avatar
ywfang committed
122
        self.w2s = nn.Parameter(
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
            torch.empty(
                self.num_total_experts,
                self.hidden_size,
                self.intermediate_size,
                device=current_platform.device_type,
                dtype=self.params_dtype,
            )
        )

        set_weight_attrs(
            self.ws,
            {
                "weight_loader": self.weight_loader,
            },
        )
        set_weight_attrs(
            self.w2s,
            {
                "weight_loader": self.weight_loader,
            },
        )

    def weight_loader(
        self,
        param: nn.Parameter,
        loaded_weight: torch.Tensor,
        weight_name: str,
        expert_id: int,
    ):
ywfang's avatar
ywfang committed
152
153
154
155
156
157
158
        tp_rank = get_tensor_model_parallel_rank()
        param_data = param.data
        shard_size = self.intermediate_size
        shard = slice(tp_rank * shard_size, (tp_rank + 1) * shard_size)
        if weight_name.endswith("w1.weight"):
            param_data[expert_id, 0:shard_size, :] = loaded_weight[shard, :]
        if weight_name.endswith("w3.weight"):
159
160
161
            param_data[expert_id, shard_size : 2 * shard_size, :] = loaded_weight[
                shard, :
            ]
ywfang's avatar
ywfang committed
162
163
164
165
166
167
168
169
        if weight_name.endswith("w2.weight"):
            param_data[expert_id, :, :] = loaded_weight[:, shard]

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        num_tokens, hidden_size = hidden_states.shape
        hidden_states = hidden_states.view(-1, self.hidden_size)
        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
170

171
172
173
        topk_weights, topk_ids, _ = fused_topk(
            hidden_states, router_logits, self.top_k, renormalize=True
        )
174

175
176
177
        final_hidden_states = fused_experts(
            hidden_states, self.ws, self.w2s, topk_weights, topk_ids, inplace=True
        )
ywfang's avatar
ywfang committed
178
179

        if self.tp_size > 1:
180
            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
ywfang's avatar
ywfang committed
181
182
183
184
185
186
187
188
189
190

        return final_hidden_states.view(num_tokens, hidden_size)


class MiniCPMMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
191
        hidden_act_param: float,
192
        quant_config: QuantizationConfig | None = None,
193
        prefix: str = "",
ywfang's avatar
ywfang committed
194
195
196
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
197
198
199
200
201
            hidden_size,
            [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
202
203
        )
        self.down_proj = RowParallelLinear(
204
205
206
207
208
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.down_proj",
209
        )
210
211
212
213
214
        if hidden_act == "silu":
            self.act_fn = SiluAndMul()
        elif hidden_act == "fatrelu":
            self.act_fn = FatreluAndMul(threshold=hidden_act_param)
        else:
215
216
217
218
            raise ValueError(
                f"Unsupported activation: {hidden_act}. "
                "Only silu and fatrelu are supported for now."
            )
ywfang's avatar
ywfang committed
219
220
221
222
223
224
225
226
227
228
229
230
231
232

    def forward(self, x):
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class MiniCPMAttention(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
233
        rope_parameters: dict[str, Any] | None = None,
ywfang's avatar
ywfang committed
234
        max_position_embeddings: int = 8192,
235
236
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
237
        prefix: str = "",
ywfang's avatar
ywfang committed
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
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_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 = 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.max_position_embeddings = max_position_embeddings

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
267
            quant_config=quant_config,
268
            prefix=f"{prefix}.qkv_proj",
ywfang's avatar
ywfang committed
269
270
271
272
273
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
274
            quant_config=quant_config,
275
            prefix=f"{prefix}.o_proj",
ywfang's avatar
ywfang committed
276
277
278
279
280
        )

        self.rotary_emb = get_rope(
            self.head_dim,
            max_position=max_position_embeddings,
281
            rope_parameters=rope_parameters,
ywfang's avatar
ywfang committed
282
        )
283

284
285
286
287
288
289
290
291
292
        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",
        )
ywfang's avatar
ywfang committed
293
294
295
296
297
298
299
300
301
302
303
304

    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)
        orig_dtype = q.dtype
        q, k = q.float(), k.float()
        q, k = self.rotary_emb(positions, q, k)
        q, k = q.to(orig_dtype), k.to(orig_dtype)
305
        attn_output = self.attn(q, k, v)
ywfang's avatar
ywfang committed
306
307
308
309
310
311
312
        output, _ = self.o_proj(attn_output)
        return output


class MiniCPMDecoderLayer(nn.Module):
    def __init__(
        self,
313
        config: PretrainedConfig,
314
315
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
316
        prefix: str = "",
ywfang's avatar
ywfang committed
317
318
319
    ) -> None:
        super().__init__()
        self.config = config
ywfang's avatar
ywfang committed
320
321
        self.cache_config = cache_config
        self.quant_config = quant_config
ywfang's avatar
ywfang committed
322
        self.hidden_size = config.hidden_size
323
        self.max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
324
        self.prefix = prefix
ywfang's avatar
ywfang committed
325
326
327
328
        self._init_attn_block()
        self._init_ffn_block()

    def _init_attn_block(self):
329
330
331
        self.input_layernorm = RMSNorm(
            self.config.hidden_size, eps=self.config.rms_norm_eps
        )
ywfang's avatar
ywfang committed
332
333
        self.self_attn = MiniCPMAttention(
            hidden_size=self.hidden_size,
ywfang's avatar
ywfang committed
334
335
            num_heads=self.config.num_attention_heads,
            num_kv_heads=self.config.num_key_value_heads,
336
            rope_parameters=self.config.rope_parameters,
ywfang's avatar
ywfang committed
337
338
339
            max_position_embeddings=self.max_position_embeddings,
            cache_config=self.cache_config,
            quant_config=self.quant_config,
340
            prefix=f"{self.prefix}.self_attn",
ywfang's avatar
ywfang committed
341
        )
ywfang's avatar
ywfang committed
342
343

    def _init_ffn_block(self):
344
345
346
        self.post_attention_layernorm = RMSNorm(
            self.config.hidden_size, eps=self.config.rms_norm_eps
        )
ywfang's avatar
ywfang committed
347
348
349
350
        self.num_experts = getattr(self.config, "num_experts", 0)
        if self.num_experts == 0:
            self.mlp = MiniCPMMLP(
                hidden_size=self.hidden_size,
ywfang's avatar
ywfang committed
351
352
                intermediate_size=self.config.intermediate_size,
                hidden_act=self.config.hidden_act,
353
                hidden_act_param=getattr(self.config, "hidden_act_param", 0.0),
ywfang's avatar
ywfang committed
354
                quant_config=self.quant_config,
ywfang's avatar
ywfang committed
355
356
            )
        else:
ywfang's avatar
ywfang committed
357
358
359
360
            self.mlp = MiniCPMMoE(
                num_experts=self.config.num_experts,
                top_k=self.config.num_experts_per_tok,
                hidden_size=self.config.hidden_size,
361
362
                intermediate_size=self.config.intermediate_size,
            )
ywfang's avatar
ywfang committed
363
364
365
366
367

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
368
        residual: torch.Tensor | None,
369
    ) -> tuple[torch.Tensor, torch.Tensor]:
ywfang's avatar
ywfang committed
370
371
372
373
374
375
376
        # Self Attention
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )
377
378
379
        hidden_states = residual + hidden_states * (
            self.config.scale_depth / math.sqrt(self.config.num_hidden_layers)
        )
ywfang's avatar
ywfang committed
380
381
382
383
384

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
385
386
387
        hidden_states = residual + hidden_states * (
            self.config.scale_depth / math.sqrt(self.config.num_hidden_layers)
        )
ywfang's avatar
ywfang committed
388
389
390
391

        return hidden_states, None


392
@support_torch_compile
ywfang's avatar
ywfang committed
393
class MiniCPMModel(nn.Module):
394
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
ywfang's avatar
ywfang committed
395
        super().__init__()
396
397
398
399
400

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

ywfang's avatar
ywfang committed
401
        self.config = config
ywfang's avatar
ywfang committed
402
403
        self.cache_config = cache_config
        self.quant_config = quant_config
404
405
406

        self.vocab_size = config.vocab_size

ywfang's avatar
ywfang committed
407
408
409
410
        self.embed_tokens = VocabParallelEmbedding(
            self.vocab_size,
            config.hidden_size,
        )
411
        self.num_experts = getattr(self.config, "num_experts", 0)
412
        self._init_layers(prefix, config, cache_config, quant_config)
ywfang's avatar
ywfang committed
413
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
414
415
416

        self.aux_hidden_state_layers = tuple[int, ...]()

417
418
419
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], self.config.hidden_size
        )
ywfang's avatar
ywfang committed
420

421
422
423
424
    def _init_layers(
        self,
        prefix: str,
        config: PretrainedConfig,
425
426
        cache_config: CacheConfig | None,
        quant_config: QuantizationConfig | None,
427
428
429
    ):
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
430
            lambda prefix: MiniCPMDecoderLayer(
431
432
433
434
                config, cache_config, quant_config, prefix=prefix
            ),
            prefix=f"{prefix}.layers",
        )
ywfang's avatar
ywfang committed
435

436
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
ywfang's avatar
ywfang committed
437
438
439
440
441
442
443
        embedding = self.embed_tokens(input_ids)
        return embedding * self.config.scale_emb

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
444
445
446
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]:
447
448
449
450
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
451
                hidden_states = self.embed_input_ids(input_ids)
452
            residual = None
ywfang's avatar
ywfang committed
453
        else:
454
455
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
ywfang's avatar
ywfang committed
456

457
458
        aux_hidden_states = []
        for idx, layer in enumerate(
459
460
            islice(self.layers, self.start_layer, self.end_layer)
        ):
461
462
            if idx in self.aux_hidden_state_layers:
                aux_hidden_states.append(
463
464
                    hidden_states + residual if residual is not None else hidden_states
                )
ywfang's avatar
ywfang committed
465
466
467
468
469
            hidden_states, residual = layer(
                positions,
                hidden_states,
                residual,
            )
470

471
        if not get_pp_group().is_last_rank:
472
473
474
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
475

ywfang's avatar
ywfang committed
476
        hidden_states = self.norm(hidden_states)
477
478
479

        if len(aux_hidden_states) > 0:
            return hidden_states, aux_hidden_states
ywfang's avatar
ywfang committed
480
481
        return hidden_states

482
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
483
484
485
486
487
488
489
490
491
492
        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)
493
494
495
496
497
            (
                "ws" if weight_name in ["w1", "w3"] else "w2s",
                f"experts.{expert_id}.{weight_name}.weight",
                expert_id,
            )
498
499
500
501
            for expert_id in range(self.num_experts)
            for weight_name in ["w1", "w2", "w3"]
        ]
        params_dict = dict(self.named_parameters())
502

503
        loaded_params: set[str] = set()
504
505
506
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
507
            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
508
509
510
                # Models trained using ColossalAI may include these tensors in
                # the checkpoint. Skip them.
                continue
511
            for param_name, weight_name, shard_id in stacked_params_mapping:
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
                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
533
534
535
                    weight_loader(
                        param, loaded_weight, weight_name, expert_id=expert_id
                    )
536
537
538
539
540
541
542
543
                    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]
544
545
546
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
547
548
549
550
                    weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

ywfang's avatar
ywfang committed
551

552
class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
ywfang's avatar
ywfang committed
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
    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",
    }

571
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
ywfang's avatar
ywfang committed
572
        super().__init__()
573
574
575
        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
576

577
        parallel_config = vllm_config.parallel_config
578

579
580
        self.prefix = prefix
        self.vllm_config = vllm_config
ywfang's avatar
ywfang committed
581
        self.config = config
582

ywfang's avatar
ywfang committed
583
584
        self.cache_config = cache_config
        self.quant_config = quant_config
585

586
587
588
        self.model = self._init_model(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
589

590
        self.lm_head = ParallelLMHead(
591
            config.vocab_size,
592
593
            config.hidden_size,
            quant_config=quant_config,
594
            prefix=maybe_prefix(prefix, "lm_head"),
595
596
597
        )
        if config.tie_word_embeddings:
            self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
ywfang's avatar
ywfang committed
598
599
        self.scale_width = self.config.hidden_size / self.config.dim_model_base

600
        self.logits_processor = LogitsProcessor(config.vocab_size)
601
        self.make_empty_intermediate_tensors = (
602
603
            self.model.make_empty_intermediate_tensors
        )
604
605
        if parallel_config.enable_eplb and getattr(config, "num_experts", 0) > 0:
            raise NotImplementedError("EPLB is not supported for MiniCPM yet.")
ywfang's avatar
ywfang committed
606

607
    def _init_model(self, *, vllm_config: VllmConfig, prefix: str = ""):
608
        return MiniCPMModel(vllm_config=vllm_config, prefix=prefix)
ywfang's avatar
ywfang committed
609

610
611
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
612

613
614
615
616
617
618
619
    def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
        self.model.aux_hidden_state_layers = layers

    def get_eagle3_aux_hidden_state_layers(self) -> tuple[int, ...]:
        num_layers = len(self.model.layers)
        return (2, num_layers // 2, num_layers - 3)

ywfang's avatar
ywfang committed
620
621
622
623
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
624
625
626
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]:
627
628
629
        model_output = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
630
631
632
633
634
635
636
637
638
639
640
641
642

        if isinstance(model_output, tuple) and len(model_output) == 2:
            # Aux hidden states are present.
            hidden_states, aux_hidden_states = model_output
            hidden_states = hidden_states / self.scale_width
            return hidden_states, aux_hidden_states
        else:
            # Only hidden states or IntermediateTensors
            if isinstance(model_output, IntermediateTensors):
                return model_output
            else:
                hidden_states = model_output / self.scale_width
                return hidden_states
ywfang's avatar
ywfang committed
643

644
645
646
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
647
    ) -> torch.Tensor | None:
648
        logits = self.logits_processor(self.lm_head, hidden_states)
ywfang's avatar
ywfang committed
649
650
        return logits

651
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
652
653
        loader = AutoWeightsLoader(
            self,
654
            skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
655
656
        )
        return loader.load_weights(weights)