deepseek.py 17.7 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
# coding=utf-8
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# Copyright 2023 The vLLM team.
# Copyright 2023 DeepSeek-AI 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 Deepseek model."""
24
from typing import Any, Dict, Iterable, List, Optional, Tuple
25
26
27
28
29

import torch
from torch import nn
from transformers import PretrainedConfig

30
from vllm.attention import Attention, AttentionMetadata
31
from vllm.config import CacheConfig
32
33
34
from vllm.distributed import (get_tensor_model_parallel_rank,
                              get_tensor_model_parallel_world_size,
                              tensor_model_parallel_all_reduce)
35
36
37
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import fused_moe
from vllm.model_executor.layers.layernorm import RMSNorm
38
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
39
                                               QKVParallelLinear,
40
                                               ReplicatedLinear,
41
                                               RowParallelLinear)
42
from vllm.model_executor.layers.logits_processor import LogitsProcessor
43
44
from vllm.model_executor.layers.quantization.base_config import (
    QuantizationConfig)
45
from vllm.model_executor.layers.rotary_embedding import get_rope
46
47
from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.layers.vocab_parallel_embedding import (
48
    ParallelLMHead, VocabParallelEmbedding)
49
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
50
from vllm.model_executor.sampling_metadata import SamplingMetadata
51
from vllm.sequence import IntermediateTensors, SamplerOutput
52
53
54
55
56
57
58
59
60


class DeepseekMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
61
        quant_config: Optional[QuantizationConfig] = None,
62
63
64
65
66
67
        reduce_results: bool = True,
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
68
            quant_config=quant_config)
69
70
71
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
72
                                           quant_config=quant_config,
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
                                           reduce_results=reduce_results)
        if hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {hidden_act}. "
                             "Only silu is supported for now.")
        self.act_fn = SiluAndMul()

    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 DeepseekMoE(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
91
        quant_config: Optional[QuantizationConfig] = None,
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
    ):
        super().__init__()
        self.config = config
        self.rank = get_tensor_model_parallel_rank()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.n_routed_experts = config.n_routed_experts
        self.top_k = config.num_experts_per_tok
        if self.tp_size > self.n_routed_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {self.n_routed_experts}.")

        self.experts = nn.ModuleList([
            DeepseekMLP(hidden_size=config.hidden_size,
                        intermediate_size=config.moe_intermediate_size,
                        hidden_act=config.hidden_act,
108
                        quant_config=quant_config,
109
110
111
112
113
114
115
116
                        reduce_results=False)
            for idx in range(self.n_routed_experts)
        ])
        self.pack_params()

        self.gate = ReplicatedLinear(config.hidden_size,
                                     self.n_routed_experts,
                                     bias=False,
117
                                     quant_config=None)
118
119

        if config.n_shared_experts is not None:
120
121
            intermediate_size = (config.moe_intermediate_size *
                                 config.n_shared_experts)
122
123
124
125
            self.shared_experts = DeepseekMLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
126
                quant_config=quant_config,
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
                reduce_results=False,
            )

    def pack_params(self):
        w1 = []
        w2 = []
        for expert in self.experts:
            w1.append(expert.gate_up_proj.weight)
            w2.append(expert.down_proj.weight)
        self.w1 = torch._utils._flatten_dense_tensors(w1)
        w1s = torch._utils._unflatten_dense_tensors(self.w1, w1)
        for data, param in zip(w1s, w1):
            param.data = data
        self.w1 = self.w1.view(len(w1), *w1s[0].shape)

        self.w2 = torch._utils._flatten_dense_tensors(w2)
        w2s = torch._utils._unflatten_dense_tensors(self.w2, w2)
        for data, param in zip(w2s, w2):
            param.data = data

        self.w2 = self.w2.view(len(w2), *w2s[0].shape)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
150
        num_tokens, hidden_dim = hidden_states.shape
151
152
153
        hidden_states = hidden_states.view(-1, hidden_dim)
        if self.config.n_shared_experts is not None:
            shared_output = self.shared_experts(hidden_states)
154
        # router_logits: (num_tokens, n_experts)
155
156
157
158
        router_logits, _ = self.gate(hidden_states)
        final_hidden_states = fused_moe(hidden_states,
                                        self.w1,
                                        self.w2,
159
160
161
                                        router_logits,
                                        self.top_k,
                                        renormalize=self.config.norm_topk_prob,
162
163
164
165
166
167
168
                                        inplace=True)

        if self.config.n_shared_experts is not None:
            final_hidden_states = final_hidden_states + shared_output
        final_hidden_states = tensor_model_parallel_all_reduce(
            final_hidden_states)

169
        return final_hidden_states.view(num_tokens, hidden_dim)
170
171
172
173
174
175
176
177
178
179
180
181


class DeepseekAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[Dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
182
        cache_config: Optional[CacheConfig] = None,
183
        quant_config: Optional[QuantizationConfig] = None,
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
    ) -> 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,
214
            quant_config=quant_config,
215
216
217
218
219
220
        )

        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
221
            quant_config=quant_config,
222
223
224
225
226
227
228
229
230
        )

        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,
        )
231
232
233
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
234
                              num_kv_heads=self.num_kv_heads,
235
236
                              cache_config=cache_config,
                              quant_config=quant_config)
237
238
239
240
241

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
242
243
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
244
245
246
247
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)
248
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
249
250
251
252
253
254
255
256
257
258
        output, _ = self.o_proj(attn_output)
        return output


class DeepseekDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        layer_idx: int,
259
        cache_config: Optional[CacheConfig] = None,
260
        quant_config: Optional[QuantizationConfig] = None,
261
262
263
264
265
266
267
268
269
270
271
272
273
274
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        self.self_attn = DeepseekAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
275
            cache_config=cache_config,
276
            quant_config=quant_config,
277
        )
278
279
280
        if (config.n_routed_experts is not None
                and layer_idx >= config.first_k_dense_replace
                and layer_idx % config.moe_layer_freq == 0):
281
            self.mlp = DeepseekMoE(config=config, quant_config=quant_config)
282
283
284
285
286
        else:
            self.mlp = DeepseekMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
287
                quant_config=quant_config,
288
289
290
291
292
293
294
295
296
297
            )
        self.input_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(config.hidden_size,
                                                eps=config.rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
298
299
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
300
301
302
303
304
305
306
307
308
309
310
311
312
        residual: Optional[torch.Tensor],
    ) -> torch.Tensor:
        # Self Attention
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
            kv_cache=kv_cache,
313
            attn_metadata=attn_metadata,
314
315
316
317
318
319
320
321
322
323
324
        )

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


class DeepseekModel(nn.Module):

325
326
    fall_back_to_pt_during_load = False

327
328
329
    def __init__(
        self,
        config: PretrainedConfig,
330
        cache_config: Optional[CacheConfig] = None,
331
        quant_config: Optional[QuantizationConfig] = None,
332
333
334
335
336
337
338
339
340
341
    ) -> None:
        super().__init__()
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
        self.layers = nn.ModuleList([
342
343
344
345
            DeepseekDecoderLayer(config,
                                 layer_idx,
                                 cache_config,
                                 quant_config=quant_config)
346
347
348
349
350
351
352
353
            for layer_idx in range(config.num_hidden_layers)
        ])
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
354
355
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
356
357
358
359
360
361
    ) -> torch.Tensor:
        hidden_states = self.embed_tokens(input_ids)
        residual = None
        for i in range(len(self.layers)):
            layer = self.layers[i]
            hidden_states, residual = layer(positions, hidden_states,
362
                                            kv_caches[i], attn_metadata,
363
364
365
366
367
368
369
370
371
372
                                            residual)
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


class DeepseekForCausalLM(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
373
        cache_config: Optional[CacheConfig] = None,
374
        quant_config: Optional[QuantizationConfig] = None,
375
376
377
    ) -> None:
        super().__init__()
        self.config = config
378
        self.quant_config = quant_config
379
        self.model = DeepseekModel(config, cache_config, quant_config)
380
381
382
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
                                      quant_config=quant_config)
383
384
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.sampler = Sampler()
385
386
387
388
389

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
390
391
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
392
        intermediate_tensors: Optional[IntermediateTensors] = None,
393
394
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, kv_caches,
395
                                   attn_metadata)
396
397
        return hidden_states

398
399
400
401
402
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
403
        logits = self.logits_processor(self.lm_head, hidden_states,
404
405
406
                                       sampling_metadata)
        return logits

407
408
    def sample(
        self,
409
        logits: Optional[torch.Tensor],
410
411
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
412
        next_tokens = self.sampler(logits, sampling_metadata)
413
414
        return next_tokens

415
    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
416
417
418
419
420
421
422
423
424
425
        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),
        ]

        params_dict = dict(self.named_parameters())
426
        for name, loaded_weight in weights:
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
            if "rotary_emb.inv_freq" in name:
                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
                # Skip experts that are not assigned to this worker.
                if (("mlp.experts." in name or "mlp.shared_experts." in name)
                        and name not in params_dict):
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Skip experts that are not assigned to this worker.
                if (("mlp.experts." in name or "mlp.shared_experts." in name)
                        and name not in params_dict):
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)