minicpm.py 23.3 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

ywfang's avatar
ywfang committed
3
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."""
import math
26
27
from collections.abc import Iterable
from typing import Any, Optional, Union
ywfang's avatar
ywfang committed
28
29
30

import torch
from torch import nn
31
from transformers import PretrainedConfig
ywfang's avatar
ywfang committed
32

33
from vllm.attention import Attention
34
from vllm.compilation.decorators import support_torch_compile
35
from vllm.config import CacheConfig, VllmConfig
36
from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
37
38
                              get_tensor_model_parallel_world_size,
                              tensor_model_parallel_all_reduce)
39
from vllm.model_executor.layers.activation import FatreluAndMul, SiluAndMul
ywfang's avatar
ywfang committed
40
41
from vllm.model_executor.layers.fused_moe import fused_moe
from vllm.model_executor.layers.layernorm import RMSNorm
42
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
ywfang's avatar
ywfang committed
43
44
45
46
                                               QKVParallelLinear,
                                               ReplicatedLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
47
from vllm.model_executor.layers.quantization import QuantizationConfig
ywfang's avatar
ywfang committed
48
49
50
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
51
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
ywfang's avatar
ywfang committed
52
53
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.model_executor.utils import set_weight_attrs
54
from vllm.platforms import current_platform
55
from vllm.sequence import IntermediateTensors
ywfang's avatar
ywfang committed
56

57
from .interfaces import SupportsLoRA, SupportsPP
58
from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
59
60
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
61

ywfang's avatar
ywfang committed
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

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,
        params_dtype: Optional[torch.dtype] = None,
        tp_size: Optional[int] = None,
    ):
        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

        self.gate = ReplicatedLinear(self.hidden_size,
                                     self.num_total_experts,
                                     bias=False,
                                     params_dtype=self.params_dtype,
96
                                     quant_config=None)
ywfang's avatar
ywfang committed
97
98
99
100
101

        self.ws = nn.Parameter(
            torch.empty(self.num_total_experts,
                        2 * self.intermediate_size,
                        self.hidden_size,
102
                        device=current_platform.device_type,
ywfang's avatar
ywfang committed
103
104
105
106
107
                        dtype=self.params_dtype))
        self.w2s = nn.Parameter(
            torch.empty(self.num_total_experts,
                        self.hidden_size,
                        self.intermediate_size,
108
                        device=current_platform.device_type,
ywfang's avatar
ywfang committed
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
                        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):
        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"):
            param_data[expert_id,
                       shard_size:2 * shard_size, :] = loaded_weight[shard, :]
        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)
        final_hidden_states = fused_moe(hidden_states,
                                        self.ws,
                                        self.w2s,
                                        router_logits,
                                        self.top_k,
                                        renormalize=True,
                                        inplace=True)

        if self.tp_size > 1:
            final_hidden_states = tensor_model_parallel_all_reduce(
                final_hidden_states)

        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,
159
        hidden_act_param: float,
160
        quant_config: Optional[QuantizationConfig] = None,
ywfang's avatar
ywfang committed
161
162
163
164
165
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
166
            quant_config=quant_config)
ywfang's avatar
ywfang committed
167
168
169
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
170
                                           quant_config=quant_config)
171
172
173
174
175
        if hidden_act == "silu":
            self.act_fn = SiluAndMul()
        elif hidden_act == "fatrelu":
            self.act_fn = FatreluAndMul(threshold=hidden_act_param)
        else:
ywfang's avatar
ywfang committed
176
            raise ValueError(f"Unsupported activation: {hidden_act}. "
177
                             "Only silu and fatrelu are supported for now.")
ywfang's avatar
ywfang committed
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193

    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,
        rope_theta: float = 10000,
194
        rope_scaling: Optional[dict[str, Any]] = None,
ywfang's avatar
ywfang committed
195
        max_position_embeddings: int = 8192,
196
        cache_config: Optional[CacheConfig] = None,
197
        quant_config: Optional[QuantizationConfig] = None,
198
        prefix: str = "",
ywfang's avatar
ywfang committed
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
    ) -> 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.rope_theta = rope_theta
        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,
229
            quant_config=quant_config,
ywfang's avatar
ywfang committed
230
231
232
233
234
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
235
            quant_config=quant_config,
ywfang's avatar
ywfang committed
236
237
238
239
240
241
242
243
244
        )

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
        )
245

ywfang's avatar
ywfang committed
246
247
248
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
249
                              num_kv_heads=self.num_kv_heads,
250
                              cache_config=cache_config,
251
252
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")
ywfang's avatar
ywfang committed
253
254
255
256
257
258
259
260
261
262
263
264

    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)
265
        attn_output = self.attn(q, k, v)
ywfang's avatar
ywfang committed
266
267
268
269
270
271
272
273
        output, _ = self.o_proj(attn_output)
        return output


class MiniCPMDecoderLayer(nn.Module):

    def __init__(
        self,
274
        config: PretrainedConfig,
275
        cache_config: Optional[CacheConfig] = None,
276
        quant_config: Optional[QuantizationConfig] = None,
277
        prefix: str = "",
ywfang's avatar
ywfang committed
278
279
280
    ) -> None:
        super().__init__()
        self.config = config
ywfang's avatar
ywfang committed
281
282
        self.cache_config = cache_config
        self.quant_config = quant_config
ywfang's avatar
ywfang committed
283
        self.hidden_size = config.hidden_size
ywfang's avatar
ywfang committed
284
285
286
287
        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)
288
        self.prefix = prefix
ywfang's avatar
ywfang committed
289
290
291
292
293
294
        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)
ywfang's avatar
ywfang committed
295
296
        self.self_attn = MiniCPMAttention(
            hidden_size=self.hidden_size,
ywfang's avatar
ywfang committed
297
298
299
300
301
302
303
            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,
304
            prefix=f"{self.prefix}.self_attn",
ywfang's avatar
ywfang committed
305
        )
ywfang's avatar
ywfang committed
306
307
308
309

    def _init_ffn_block(self):
        self.post_attention_layernorm = RMSNorm(self.config.hidden_size,
                                                eps=self.config.rms_norm_eps)
ywfang's avatar
ywfang committed
310
311
312
313
        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
314
315
                intermediate_size=self.config.intermediate_size,
                hidden_act=self.config.hidden_act,
316
                hidden_act_param=getattr(self.config, "hidden_act_param", 0.),
ywfang's avatar
ywfang committed
317
                quant_config=self.quant_config,
ywfang's avatar
ywfang committed
318
319
            )
        else:
ywfang's avatar
ywfang committed
320
321
322
323
324
            self.mlp = MiniCPMMoE(
                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)
ywfang's avatar
ywfang committed
325
326
327
328
329
330

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
331
    ) -> tuple[torch.Tensor, torch.Tensor]:
ywfang's avatar
ywfang committed
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
        # 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.num_hidden_layers))

        # 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.num_hidden_layers))

        return hidden_states, None


352
@support_torch_compile
ywfang's avatar
ywfang committed
353
354
class MiniCPMModel(nn.Module):

355
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
ywfang's avatar
ywfang committed
356
        super().__init__()
357
358
359
360
361
362

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

ywfang's avatar
ywfang committed
363
        self.config = config
ywfang's avatar
ywfang committed
364
365
        self.cache_config = cache_config
        self.quant_config = quant_config
ywfang's avatar
ywfang committed
366
367
368
369
370
371
372
373
374
        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.embed_tokens = VocabParallelEmbedding(
            self.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
        )
375
        self.num_experts = getattr(self.config, "num_experts", 0)
376
        self._init_layers(prefix, config, cache_config, quant_config)
ywfang's avatar
ywfang committed
377
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
378
379
380
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], self.config.hidden_size))
ywfang's avatar
ywfang committed
381

382
383
384
385
386
387
388
389
390
    def _init_layers(
        self,
        prefix: str,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig],
        quant_config: Optional[QuantizationConfig],
    ):
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
391
392
            lambda prefix: MiniCPMDecoderLayer(
                config, cache_config, quant_config, prefix=prefix),
393
            prefix=f"{prefix}.layers")
ywfang's avatar
ywfang committed
394
395
396
397
398
399
400
401
402

    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,
Alphi's avatar
Alphi committed
403
        intermediate_tensors: Optional[IntermediateTensors] = None,
ywfang's avatar
ywfang committed
404
        inputs_embeds: Optional[torch.Tensor] = None,
405
406
407
408
409
410
411
    ) -> 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)
            residual = None
ywfang's avatar
ywfang committed
412
        else:
413
414
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
ywfang's avatar
ywfang committed
415

416
        for layer in self.layers[self.start_layer:self.end_layer]:
ywfang's avatar
ywfang committed
417
418
419
420
421
            hidden_states, residual = layer(
                positions,
                hidden_states,
                residual,
            )
422
423
424
425
426
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
ywfang's avatar
ywfang committed
427
428
429
        hidden_states = self.norm(hidden_states)
        return hidden_states

430
431
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
        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())
448

449
        loaded_params: set[str] = set()
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
        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

ywfang's avatar
ywfang committed
498

499
class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
ywfang's avatar
ywfang committed
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
    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"]

519
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
ywfang's avatar
ywfang committed
520
        super().__init__()
521
522
523
524
        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
525

526
527
        self.prefix = prefix
        self.vllm_config = vllm_config
ywfang's avatar
ywfang committed
528
        self.config = config
529
        self.lora_config = lora_config
ywfang's avatar
ywfang committed
530
531
        self.cache_config = cache_config
        self.quant_config = quant_config
532

533
534
535
        self.model = self._init_model(vllm_config=vllm_config,
                                      prefix=maybe_prefix(prefix, "model"))

ywfang's avatar
ywfang committed
536
537
538
        unpadded_vocab_size = config.vocab_size
        if lora_config:
            unpadded_vocab_size += lora_config.lora_extra_vocab_size
539
540
541
542
543
544
545
546
547
548
549
550
        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)
ywfang's avatar
ywfang committed
551
552
553
554
        self.scale_width = self.config.hidden_size / self.config.dim_model_base

        self.logits_processor = LogitsProcessor(unpadded_vocab_size,
                                                config.vocab_size)
555
556
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
ywfang's avatar
ywfang committed
557

558
    def _init_model(self, *, vllm_config: VllmConfig, prefix: str = ""):
559
        return MiniCPMModel(vllm_config=vllm_config, prefix=prefix)
ywfang's avatar
ywfang committed
560

561
562
563
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

ywfang's avatar
ywfang committed
564
565
566
567
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
568
        intermediate_tensors: Optional[IntermediateTensors] = None,
569
        inputs_embeds: Optional[torch.Tensor] = None,
570
    ) -> Union[torch.Tensor, IntermediateTensors]:
571
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
572
                                   inputs_embeds) / self.scale_width
ywfang's avatar
ywfang committed
573
574
        return hidden_states

575
576
577
578
579
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
580
        logits = self.logits_processor(self.lm_head, hidden_states,
ywfang's avatar
ywfang committed
581
582
583
                                       sampling_metadata)
        return logits

584
585
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
586
587
588
589
590
591
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=(["lm_head."]
                           if self.config.tie_word_embeddings else None),
        )
        return loader.load_weights(weights)