deepseek_v2.py 35.8 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

wangding zeng's avatar
wangding zeng committed
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
# 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.
24
"""Inference-only DeepseekV2/DeepseekV3 model."""
25
from typing import Any, Dict, Iterable, Optional, Set, Tuple, Union
wangding zeng's avatar
wangding zeng committed
26

27
28
import os
import re
wangding zeng's avatar
wangding zeng committed
29
30
31
32
import torch
from torch import nn
from transformers import PretrainedConfig

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

56
57
from .interfaces import SupportsPP
from .utils import (PPMissingLayer, is_pp_missing_parameter,
58
59
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
60
from vllm import _custom_ops as ops
61

wangding zeng's avatar
wangding zeng committed
62
63
64
65
66
67
68
69
70
71

class DeepseekV2MLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: bool = True,
72
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
73
74
75
76
77
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
78
79
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj")
wangding zeng's avatar
wangding zeng committed
80
81
82
83
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
                                           quant_config=quant_config,
84
85
                                           reduce_results=reduce_results,
                                           prefix=f"{prefix}.down_proj")
wangding zeng's avatar
wangding zeng committed
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
        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 DeepseekV2MoE(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
104
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
105
106
107
108
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.routed_scaling_factor = config.routed_scaling_factor
109
110
111
112
113
114
        self.n_shared_experts = config.n_shared_experts

        if config.hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {config.hidden_act}. "
                             "Only silu is supported for now.")

wangding zeng's avatar
wangding zeng committed
115
        self.gate = ReplicatedLinear(config.hidden_size,
116
                                     config.n_routed_experts,
wangding zeng's avatar
wangding zeng committed
117
                                     bias=False,
118
119
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
        if config.topk_method == "noaux_tc":
            self.gate.e_score_correction_bias = nn.Parameter(
                torch.empty(config.n_routed_experts))
        else:
            self.gate.e_score_correction_bias = None

        self.experts = FusedMoE(
            num_experts=config.n_routed_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            reduce_results=False,
            renormalize=config.norm_topk_prob,
            quant_config=quant_config,
            use_grouped_topk=True,
            num_expert_group=config.n_group,
            topk_group=config.topk_group,
            prefix=f"{prefix}.experts",
            scoring_func=config.scoring_func,
            e_score_correction_bias=self.gate.e_score_correction_bias)

wangding zeng's avatar
wangding zeng committed
141
142
143
144
145
146
147
148
149
        if config.n_shared_experts is not None:
            intermediate_size = (config.moe_intermediate_size *
                                 config.n_shared_experts)
            self.shared_experts = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                reduce_results=False,
150
                prefix=f"{prefix}.shared_experts",
wangding zeng's avatar
wangding zeng committed
151
152
153
154
155
            )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
156
        if self.n_shared_experts is not None:
wangding zeng's avatar
wangding zeng committed
157
158
159
            shared_output = self.shared_experts(hidden_states)
        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
160
161
162
163
164
        if hidden_states.dtype != torch.float16:
            final_hidden_states = self.experts(
                hidden_states=hidden_states,
                router_logits=router_logits) * self.routed_scaling_factor
        else:
165
166
            # Fix FP16 overflow
            # See DeepseekV2DecoderLayer for more details.
167
168
            final_hidden_states = self.experts(hidden_states=hidden_states,
                                               router_logits=router_logits)
169
        if shared_output is not None:
170
171
172
            if hidden_states.dtype != torch.float16:
                final_hidden_states = final_hidden_states + shared_output
            else:
173
174
                # Fix FP16 overflow
                # See DeepseekV2DecoderLayer for more details.
175
176
                final_hidden_states = final_hidden_states + shared_output \
                    * (1. / self.routed_scaling_factor)
177
178
179
        if self.tp_size > 1:
            final_hidden_states = tensor_model_parallel_all_reduce(
                final_hidden_states)
wangding zeng's avatar
wangding zeng committed
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

        return final_hidden_states.view(num_tokens, hidden_dim)


def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
    import math
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


class DeepseekV2Attention(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
        q_lora_rank: int,
        kv_lora_rank: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[Dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
208
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
        self.v_head_dim = v_head_dim
        self.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank
        self.num_heads = num_heads
        tp_size = get_tensor_model_parallel_world_size()
        assert num_heads % tp_size == 0
        self.num_local_heads = num_heads // tp_size
        self.scaling = self.qk_head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        if self.q_lora_rank is not None:
            self.q_a_proj = ReplicatedLinear(self.hidden_size,
                                             self.q_lora_rank,
                                             bias=False,
230
231
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.q_a_proj")
wangding zeng's avatar
wangding zeng committed
232
233
234
235
236
237
            self.q_a_layernorm = RMSNorm(self.q_lora_rank,
                                         eps=config.rms_norm_eps)
            self.q_b_proj = ColumnParallelLinear(q_lora_rank,
                                                 self.num_heads *
                                                 self.qk_head_dim,
                                                 bias=False,
238
239
                                                 quant_config=quant_config,
                                                 prefix=f"{prefix}.q_b_proj")
wangding zeng's avatar
wangding zeng committed
240
241
242
243
244
        else:
            self.q_proj = ColumnParallelLinear(self.hidden_size,
                                               self.num_heads *
                                               self.qk_head_dim,
                                               bias=False,
245
246
                                               quant_config=quant_config,
                                               prefix=f"{prefix}.q_proj")
wangding zeng's avatar
wangding zeng committed
247

248
249
250
251
252
253
        self.kv_a_proj_with_mqa = ReplicatedLinear(
            self.hidden_size,
            self.kv_lora_rank + self.qk_rope_head_dim,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_a_proj_with_mqa")
wangding zeng's avatar
wangding zeng committed
254
255
256
257
258
259
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank,
                                      eps=config.rms_norm_eps)
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
260
261
            quant_config=quant_config,
            prefix=f"{prefix}.kv_b_proj")
wangding zeng's avatar
wangding zeng committed
262
263
264
265
        # O projection.
        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
                                        self.hidden_size,
                                        bias=False,
266
267
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")
268
269
        if rope_scaling:
            rope_scaling["rope_type"] = 'deepseek_yarn'
270

wangding zeng's avatar
wangding zeng committed
271
272
273
274
275
276
277
278
279
280
281
282
283
284
        self.rotary_emb = get_rope(qk_rope_head_dim,
                                   rotary_dim=qk_rope_head_dim,
                                   max_position=max_position_embeddings,
                                   base=rope_theta,
                                   rope_scaling=rope_scaling,
                                   is_neox_style=False)

        if rope_scaling:
            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
            scaling_factor = rope_scaling["factor"]
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

        self.attn = Attention(self.num_local_heads,
285
                              self.qk_head_dim,
wangding zeng's avatar
wangding zeng committed
286
287
288
                              self.scaling,
                              num_kv_heads=self.num_local_heads,
                              cache_config=cache_config,
289
290
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")
wangding zeng's avatar
wangding zeng committed
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

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        if self.q_lora_rank is not None:
            q = self.q_a_proj(hidden_states)[0]
            q = self.q_a_layernorm(q)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads,
                                         self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(-1, self.num_local_heads,
                                                   self.qk_head_dim)
        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim],
                               dim=-1)
        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
        kv_a, _ = latent_cache.split(
            [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        latent_cache = latent_cache.unsqueeze(1)
        kv_a = self.kv_a_layernorm(kv_a.contiguous())
        kv = self.kv_b_proj(kv_a)[0]
        kv = kv.view(-1, self.num_local_heads,
                     self.qk_nope_head_dim + self.v_head_dim)
        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
        k_pe = latent_cache[:, :, self.kv_lora_rank:]
317

wangding zeng's avatar
wangding zeng committed
318
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
319

wangding zeng's avatar
wangding zeng committed
320
321
322
323
        q[..., self.qk_nope_head_dim:] = q_pe
        k = torch.empty_like(q)
        k[..., :self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim:] = k_pe
324
325
326
327
        # padding value to qk_head_dim for alignment
        v = torch.nn.functional.pad(
            v, [0, self.qk_head_dim - self.v_head_dim],
            value=0).view(-1, self.num_local_heads * self.qk_head_dim)
328
        attn_output = self.attn(q, k, v)
wangding zeng's avatar
wangding zeng committed
329
        attn_output = attn_output.view(
330
331
            -1, self.num_local_heads,
            self.qk_head_dim)[..., :self.v_head_dim].reshape(
wangding zeng's avatar
wangding zeng committed
332
333
334
335
336
                -1, self.num_local_heads * self.v_head_dim)
        output, _ = self.o_proj(attn_output)
        return output


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
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
class DeepseekV2MLAAttention(nn.Module):
    """
    Main reference: DeepseekV2 paper, and FlashInfer Implementation
    (https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551).
    
    For more info see MLACommonImpl in: vllm/attention/backends/mla/utils.py
    """

    def __init__(
        self,
        config: PretrainedConfig,
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
        q_lora_rank: Optional[int],
        kv_lora_rank: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[Dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
        self.v_head_dim = v_head_dim

        self.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank

        self.num_heads = num_heads
        tp_size = get_tensor_model_parallel_world_size()
        assert num_heads % tp_size == 0
        self.num_local_heads = num_heads // tp_size

        self.scaling = self.qk_head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        if self.q_lora_rank is not None:
            self.q_a_proj = ReplicatedLinear(self.hidden_size,
                                             self.q_lora_rank,
                                             bias=False,
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.q_a_proj")
            self.q_a_layernorm = RMSNorm(self.q_lora_rank,
                                         eps=config.rms_norm_eps)
            self.q_b_proj = ColumnParallelLinear(q_lora_rank,
                                                 self.num_heads *
                                                 self.qk_head_dim,
                                                 bias=False,
                                                 quant_config=quant_config,
                                                 prefix=f"{prefix}.q_b_proj")
        else:
            self.q_proj = ColumnParallelLinear(self.hidden_size,
                                               self.num_heads *
                                               self.qk_head_dim,
                                               bias=False,
                                               quant_config=quant_config,
                                               prefix=f"{prefix}.q_proj")

        self.kv_a_proj_with_mqa = ReplicatedLinear(
            self.hidden_size,
            self.kv_lora_rank + self.qk_rope_head_dim,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_a_proj_with_mqa")
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank,
                                      eps=config.rms_norm_eps)
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_b_proj")
        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
                                        self.hidden_size,
                                        bias=False,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")

423
424
        if rope_scaling:
            rope_scaling["rope_type"] = 'deepseek_yarn'
425
426
427
428
429
430
431
432
433
434
435
436
        self.rotary_emb = get_rope(qk_rope_head_dim,
                                   rotary_dim=qk_rope_head_dim,
                                   max_position=max_position_embeddings,
                                   base=rope_theta,
                                   rope_scaling=rope_scaling,
                                   is_neox_style=False)
        if rope_scaling:
            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
            scaling_factor = rope_scaling["factor"]
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

437
438
439
440
441
442
        # In the MLA backend, kv_cache includes both k_c and
        # pe (i.e. decoupled position embeddings). In particular,
        # the concat_and_cache_mla op requires
        #     k_c.size(1) + k_pe.size(1) == kv_cache.size(2)
        # i.e.
        #     kv_lora_rank + qk_rope_head_dim == head_size
443
444
        self.mla_attn = Attention(
            num_heads=self.num_local_heads,
445
            head_size=self.kv_lora_rank + self.qk_rope_head_dim,
446
447
448
449
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
            scale=self.scaling,
            num_kv_heads=1,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
            use_mla=True,
            # MLA Args
            q_lora_rank=self.q_lora_rank,
            kv_lora_rank=self.kv_lora_rank,
            qk_nope_head_dim=self.qk_nope_head_dim,
            qk_rope_head_dim=self.qk_rope_head_dim,
            qk_head_dim=self.qk_head_dim,
            v_head_dim=self.v_head_dim,
            rotary_emb=self.rotary_emb,
            q_proj=self.q_proj if self.q_lora_rank is None else self.q_b_proj,
            kv_b_proj=self.kv_b_proj,
            o_proj=self.o_proj,
        )

        self.prefix = prefix
        self.debug_layer_idx = int(self.prefix.split(".")[-2])

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        if self.q_lora_rank is not None:
            ckq = self.q_a_proj(hidden_states)[0]
            hidden_states_or_q_c = self.q_a_layernorm(ckq)
        else:
            hidden_states_or_q_c = hidden_states
        kv_c, k_pe = self.kv_a_proj_with_mqa(hidden_states)[0].split(
            [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())
481
482
483
484
        return self.mla_attn(hidden_states_or_q_c,
                             kv_c_normed,
                             k_pe,
                             output_shape=hidden_states.shape)
485
486


wangding zeng's avatar
wangding zeng committed
487
488
489
490
491
class DeepseekV2DecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
492
        prefix: str,
493
        model_config: ModelConfig,
wangding zeng's avatar
wangding zeng committed
494
495
496
497
498
499
500
501
502
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
    ) -> 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)
503
504
505
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
        layer_idx = int(prefix.split(sep='.')[-1])
506
        self.layer_idx = layer_idx
507
508
509
510
511
        if model_config.use_mla:
            attn_cls = DeepseekV2MLAAttention
        else:
            attn_cls = DeepseekV2Attention
        self.self_attn = attn_cls(
wangding zeng's avatar
wangding zeng committed
512
513
514
515
516
517
518
519
520
521
522
523
524
525
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            qk_nope_head_dim=config.qk_nope_head_dim,
            qk_rope_head_dim=config.qk_rope_head_dim,
            v_head_dim=config.v_head_dim,
            q_lora_rank=config.q_lora_rank
            if hasattr(config, "q_lora_rank") else None,
            kv_lora_rank=config.kv_lora_rank,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
526
            prefix=f"{prefix}.self_attn",
wangding zeng's avatar
wangding zeng committed
527
        )
528

wangding zeng's avatar
wangding zeng committed
529
530
531
        if (config.n_routed_experts is not None
                and layer_idx >= config.first_k_dense_replace
                and layer_idx % config.moe_layer_freq == 0):
532
533
534
535
536
            self.mlp = DeepseekV2MoE(
                config=config,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )
wangding zeng's avatar
wangding zeng committed
537
538
539
540
541
542
        else:
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
543
                prefix=f"{prefix}.mlp",
wangding zeng's avatar
wangding zeng committed
544
545
546
547
548
            )
        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)
549
        self.routed_scaling_factor = config.routed_scaling_factor
wangding zeng's avatar
wangding zeng committed
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        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,
        )

569
570
571
572
        if hidden_states.dtype == torch.float16:
            # Fix FP16 overflow
            # We scale both hidden_states and residual before
            # rmsnorm, and rmsnorm result would not affect by scale.
573
            hidden_states *= 1. / self.routed_scaling_factor
574
575
576
577
578
579
            if self.layer_idx == 0:
                # The residual is shared by all layers, we only scale it on
                # first layer.
                residual *= 1. / self.routed_scaling_factor

        # Fully Connected
wangding zeng's avatar
wangding zeng committed
580
581
582
        hidden_states, residual = self.post_attention_layernorm(
            hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
583
584
585
586
587
588
589
590

        if isinstance(self.mlp,
                      DeepseekV2MLP) and hidden_states.dtype == torch.float16:
            # Fix FP16 overflow
            # Scaling the DeepseekV2MLP output, it is the input of
            # input_layernorm of next decoder layer.
            # The scaling of DeepseekV2MOE output would be done in the forward
            # of DeepseekV2MOE
591
            hidden_states *= 1. / self.routed_scaling_factor
592

wangding zeng's avatar
wangding zeng committed
593
594
595
        return hidden_states, residual


596
@support_torch_compile
wangding zeng's avatar
wangding zeng committed
597
598
599
600
class DeepseekV2Model(nn.Module):

    fall_back_to_pt_during_load = False

601
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
602
        super().__init__()
603
604

        config = vllm_config.model_config.hf_config
605
        model_config = vllm_config.model_config
606
607
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
608
        self.config = config
609

wangding zeng's avatar
wangding zeng committed
610
611
        self.vocab_size = config.vocab_size

612
613
614
615
        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
616
617
                quant_config=quant_config,
                prefix=f"{prefix}.embed_tokens")
618
619
620
621
622
623
624
625
        else:
            self.embed_tokens = PPMissingLayer()

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: DeepseekV2DecoderLayer(
                config,
                prefix,
626
                model_config=model_config,
627
628
629
630
631
632
633
634
635
                cache_config=cache_config,
                quant_config=quant_config,
            ),
            prefix=f"{prefix}.layers")

        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
636
637
638
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))
wangding zeng's avatar
wangding zeng committed
639

640
641
642
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

wangding zeng's avatar
wangding zeng committed
643
644
645
646
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
647
        intermediate_tensors: Optional[IntermediateTensors],
648
        inputs_embeds: Optional[torch.Tensor] = None,
649
    ) -> Union[torch.Tensor, IntermediateTensors]:
650
        if get_pp_group().is_first_rank:
651
652
653
654
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
655
656
657
658
659
660
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

661
        for layer in self.layers[self.start_layer:self.end_layer]:
662
            hidden_states, residual = layer(positions, hidden_states, residual)
663
664
665
666
667
668
669

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })

wangding zeng's avatar
wangding zeng committed
670
671
672
673
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


674
class DeepseekV2ForCausalLM(nn.Module, SupportsPP):
wangding zeng's avatar
wangding zeng committed
675

676
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
677
        super().__init__()
678
679
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
wangding zeng's avatar
wangding zeng committed
680
681
        self.config = config
        self.quant_config = quant_config
682
683
        self.model = DeepseekV2Model(vllm_config=vllm_config,
                                     prefix=maybe_prefix(prefix, "model"))
684
685
686
687
688
689
        if get_pp_group().is_last_rank:
            self.lm_head = ParallelLMHead(config.vocab_size,
                                          config.hidden_size,
                                          quant_config=quant_config)
        else:
            self.lm_head = PPMissingLayer()
wangding zeng's avatar
wangding zeng committed
690
        self.logits_processor = LogitsProcessor(config.vocab_size)
691
692
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
693
694
695
696
697
        
        self.quant_method = None
        if quant_config is not None:
            self.quant_method=quant_config.get_name()
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
wangding zeng's avatar
wangding zeng committed
698

699
700
701
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

wangding zeng's avatar
wangding zeng committed
702
703
704
705
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
706
        intermediate_tensors: Optional[IntermediateTensors] = None,
707
        inputs_embeds: Optional[torch.Tensor] = None,
708
    ) -> Union[torch.Tensor, IntermediateTensors]:
709
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
710
                                   inputs_embeds)
wangding zeng's avatar
wangding zeng committed
711
712
        return hidden_states

713
714
715
716
717
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
718
        logits = self.logits_processor(self.lm_head, hidden_states,
wangding zeng's avatar
wangding zeng committed
719
720
721
                                       sampling_metadata)
        return logits

722
723
724
725
726
727
728
729
730
731
732
733
734
735
    def make_empty_intermediate_tensors(
            self, batch_size: int, dtype: torch.dtype,
            device: torch.device) -> IntermediateTensors:
        return IntermediateTensors({
            "hidden_states":
            torch.zeros((batch_size, self.config.hidden_size),
                        dtype=dtype,
                        device=device),
            "residual":
            torch.zeros((batch_size, self.config.hidden_size),
                        dtype=dtype,
                        device=device),
        })

736
737
    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
wangding zeng's avatar
wangding zeng committed
738
739
740
741
742
743
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]

744
745
746
747
748
749
750
751
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        expert_params_mapping = FusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
            num_experts=self.config.n_routed_experts)

wangding zeng's avatar
wangding zeng committed
752
        params_dict = dict(self.named_parameters())
753
        loaded_params: Set[str] = set()
wangding zeng's avatar
wangding zeng committed
754
755
756
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
757

758
759
760
            spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
            if spec_layer is not None:
                continue  # skip spec decode layers for main model
761

wangding zeng's avatar
wangding zeng committed
762
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
763
                # Skip non-stacked layers and experts (experts handled below).
wangding zeng's avatar
wangding zeng committed
764
765
                if weight_name not in name:
                    continue
766
767
768
769
770
771
772
773
                # We have mlp.experts[0].gate_proj in the checkpoint.
                # Since we handle the experts below in expert_params_mapping,
                # we need to skip here BEFORE we update the name, otherwise
                # name will be updated to mlp.experts[0].gate_up_proj, which
                # will then be updated below in expert_params_mapping
                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
                if (("mlp.experts." in name) and name not in params_dict):
                    continue
wangding zeng's avatar
wangding zeng committed
774
775
776
777
                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
778
779
780
781

                if is_pp_missing_parameter(name, self):
                    continue

wangding zeng's avatar
wangding zeng committed
782
783
784
785
786
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
787
788
789
790
791
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
                    name = name.replace(weight_name, param_name)
792
793
794
795

                    if is_pp_missing_parameter(name, self):
                        continue

796
797
798
799
                    param = params_dict[name]
                    weight_loader = param.weight_loader
                    weight_loader(param,
                                  loaded_weight,
800
                                  name,
801
802
803
804
805
806
807
808
                                  shard_id=shard_id,
                                  expert_id=expert_id)
                    break
                else:
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue

809
810
811
812
813
                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

814
815
816
                    if is_pp_missing_parameter(name, self):
                        continue

817
818
819
820
                    param = params_dict[name]
                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)
821
            loaded_params.add(name)
822
823
824
825
826
827
828
            
        if self.use_llama_nn and self.quant_method is None:
                lay_key_words = [
                    "self_attn.q_a_proj.weight",
                    "self_attn.kv_a_proj_with_mqa.weight",
                    "mlp.gate.weight",
                    "mlp.gate_up_proj.weight",
zhuwenwen's avatar
zhuwenwen committed
829
830
831
                    "mlp.down_proj.weight",
                    "shared_experts.gate_up_proj.weight",
                    "shared_experts.down_proj.weight",
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
                    "self_attn.q_proj.weight",
                    "self_attn.q_b_proj.weight",
                    "self_attn.kv_b_proj.weight",
                    "self_attn.o_proj.weight",
                    "lm_head.weight"
                ]

                combined_words = "|".join(lay_key_words)
                
                for layername in loaded_params:
                    weight = params_dict[layername]
                    matches = re.findall(combined_words, layername)
                    if matches:
                        _weight = torch.zeros_like(weight.data)
                        ori_shape =_weight.shape
                        
                        ops.trans_w16_gemm(_weight, weight.data, _weight.shape[0], _weight.shape[1])
                        weight.data.copy_(_weight)
                        
                        weight.data=weight.data.reshape(ori_shape[1],-1)
                        
853
        return loaded_params
854
855
856
857


class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass
858
859
860
861
862
863
864
865
866
867
868
869


def get_spec_layer_idx_from_weight_name(config: PretrainedConfig,
                                        weight_name: str) -> Optional[int]:
    if hasattr(config,
               "num_nextn_predict_layers") and (config.num_nextn_predict_layers
                                                > 0):
        layer_idx = config.num_hidden_layers
        for i in range(config.num_nextn_predict_layers):
            if weight_name.startswith(f"model.layers.{layer_idx+i}."):
                return layer_idx + i
    return None