deepseek_v2.py 43.6 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

wangding zeng's avatar
wangding zeng committed
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# 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.
25
"""Inference-only DeepseekV2/DeepseekV3 model."""
王敏's avatar
王敏 committed
26
27
import os
import re
28
import vllm.envs as envs
zhuwenwen's avatar
zhuwenwen committed
29

30
31
import typing
from collections.abc import Callable, Iterable
32
from typing import Any, Optional, Union
wangding zeng's avatar
wangding zeng committed
33
34
35
36
37

import torch
from torch import nn
from transformers import PretrainedConfig

38
from vllm.attention import Attention
39
from vllm.compilation.decorators import support_torch_compile
40
41
from vllm.config import (CacheConfig, ModelConfig, VllmConfig,
                         get_current_vllm_config)
王敏's avatar
王敏 committed
42
from vllm.distributed import (get_ep_group, get_pp_group, get_dp_group,
43
                              get_tensor_model_parallel_world_size)
wangding zeng's avatar
wangding zeng committed
44
from vllm.model_executor.layers.activation import SiluAndMul
45
from vllm.model_executor.layers.fused_moe import FusedMoE
王敏's avatar
王敏 committed
46
47
from vllm.model_executor.layers.fused_moe.ep_moe.layer import EPMoE
from vllm.model_executor.layers.fused_moe.ep_moe.ep_moe_utlis import EPSharedExperts
wangding zeng's avatar
wangding zeng committed
48
49
50
51
52
53
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
54
from vllm.model_executor.layers.quantization import QuantizationConfig
wangding zeng's avatar
wangding zeng committed
55
56
57
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
    ParallelLMHead, VocabParallelEmbedding)
58
59
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader, maybe_remap_kv_scale_name)
wangding zeng's avatar
wangding zeng committed
60
from vllm.model_executor.sampling_metadata import SamplingMetadata
61
from vllm.sequence import IntermediateTensors
wangding zeng's avatar
wangding zeng committed
62

63
from .interfaces import MixtureOfExperts, SupportsPP
64
from .utils import (PPMissingLayer, is_pp_missing_parameter,
65
66
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
王敏's avatar
王敏 committed
67
from vllm import _custom_ops as ops
68
from vllm.utils import W8a8GetCacheJSON
69

70
os.environ['DPSK_FP16_QUICK'] = os.environ.get('DPSK_FP16_QUICK', '1')
wangding zeng's avatar
wangding zeng committed
71
72
73
74
75
76
77
78
79
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,
80
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
81
82
83
84
85
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
86
87
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj")
wangding zeng's avatar
wangding zeng committed
88
89
90
91
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
                                           quant_config=quant_config,
92
93
                                           reduce_results=reduce_results,
                                           prefix=f"{prefix}.down_proj")
wangding zeng's avatar
wangding zeng committed
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
        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,
112
        prefix: str = "",
113
        enable_eplb: bool = False,
wangding zeng's avatar
wangding zeng committed
114
115
116
117
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.routed_scaling_factor = config.routed_scaling_factor
118
119
120
121
122
123

        self.ep_group = get_ep_group().device_group
        self.ep_rank = self.ep_group.rank()
        self.ep_size = self.ep_group.size()
        self.n_routed_experts: int = config.n_routed_experts
        self.n_shared_experts: int = config.n_shared_experts
124
125
126
127
128

        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
129
        self.gate = ReplicatedLinear(config.hidden_size,
130
                                     config.n_routed_experts,
wangding zeng's avatar
wangding zeng committed
131
                                     bias=False,
132
133
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")
134
135
136
137
138
139
        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

140
141
142
143
        # Load balancing settings.
        vllm_config = get_current_vllm_config()
        parallel_config = vllm_config.parallel_config
        self.enable_eplb = enable_eplb
144
        self.dpsk_fp16_quick = os.environ.get('DPSK_FP16_QUICK') == '1'
145
146
147
148
149
150
151
152
153
154
155

        self.n_redundant_experts = parallel_config.num_redundant_experts
        self.n_logical_experts = self.n_routed_experts
        self.n_physical_experts = (self.n_logical_experts +
                                   self.n_redundant_experts)
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

        self.physical_expert_start = (self.ep_rank *
                                      self.n_local_physical_experts)
        self.physical_expert_end = (self.physical_expert_start +
                                    self.n_local_physical_experts)
王敏's avatar
王敏 committed
156
157
158
159
        
        dp_size = get_dp_group().world_size
        self.use_ep_opt = dp_size > 1 and parallel_config.enable_expert_parallel
        self.shared_experts = None
160

wangding zeng's avatar
wangding zeng committed
161
162
163
        if config.n_shared_experts is not None:
            intermediate_size = (config.moe_intermediate_size *
                                 config.n_shared_experts)
王敏's avatar
王敏 committed
164
165
            shared_expert_cls = DeepseekV2MLP if not self.use_ep_opt else EPSharedExperts
            self.shared_experts = shared_expert_cls(
wangding zeng's avatar
wangding zeng committed
166
167
168
169
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
王敏's avatar
王敏 committed
170
                reduce_results=False,
171
                prefix=f"{prefix}.shared_experts",
wangding zeng's avatar
wangding zeng committed
172
            )
王敏's avatar
王敏 committed
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211

        if not self.use_ep_opt:
            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,
                enable_eplb=self.enable_eplb,
                num_redundant_experts=self.n_redundant_experts,
                routed_scaling_factor=self.routed_scaling_factor)
        else:
            self.experts = EPMoE(
                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,
                routed_scaling_factor=self.routed_scaling_factor)

        
        if self.use_ep_opt:
            self.experts.set_shared_experts(self.shared_experts)
212
        from vllm.two_batch_overlap.two_batch_overlap import tbo_all_reduce
lizhigong's avatar
lizhigong committed
213
        self.tbo_all_reduce = tbo_all_reduce
wangding zeng's avatar
wangding zeng committed
214
215
216
217

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
王敏's avatar
王敏 committed
218
219
220
        if not self.use_ep_opt:
            if self.n_shared_experts is not None:
                shared_output = self.shared_experts(hidden_states)
wangding zeng's avatar
wangding zeng committed
221
222
        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
223

224
        if hidden_states.dtype != torch.float16 or self.dpsk_fp16_quick:
225
226
227
228
229
230
231
232
            final_hidden_states = self.experts(
                hidden_states=hidden_states,
                router_logits=router_logits) * self.routed_scaling_factor
        else:
            # Fix FP16 overflow
            # See DeepseekV2DecoderLayer for more details.
            final_hidden_states = self.experts(hidden_states=hidden_states,
                                               router_logits=router_logits)
zhuwenwen's avatar
zhuwenwen committed
233
        
王敏's avatar
王敏 committed
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
        if not self.use_ep_opt:
            if shared_output is not None:
                if hidden_states.dtype != torch.float16 or self.dpsk_fp16_quick:
                    final_hidden_states = final_hidden_states + shared_output
                else:
                    # Fix FP16 overflow
                    # See DeepseekV2DecoderLayer for more details.
                    final_hidden_states = final_hidden_states + shared_output \
                        * (1. / self.routed_scaling_factor)

            if self.tp_size > 1:
                if envs.VLLM_ENABLE_TBO:
                    final_hidden_states = self.tbo_all_reduce(final_hidden_states)
                else:
                    final_hidden_states = (
                        self.experts.maybe_all_reduce_tensor_model_parallel(
                            final_hidden_states))
wangding zeng's avatar
wangding zeng committed
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274

        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,
275
        rope_scaling: Optional[dict[str, Any]] = None,
wangding zeng's avatar
wangding zeng committed
276
277
278
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
279
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
    ) -> 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,
301
302
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.q_a_proj")
wangding zeng's avatar
wangding zeng committed
303
304
305
306
307
308
            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,
309
310
                                                 quant_config=quant_config,
                                                 prefix=f"{prefix}.q_b_proj")
wangding zeng's avatar
wangding zeng committed
311
312
313
314
315
        else:
            self.q_proj = ColumnParallelLinear(self.hidden_size,
                                               self.num_heads *
                                               self.qk_head_dim,
                                               bias=False,
316
317
                                               quant_config=quant_config,
                                               prefix=f"{prefix}.q_proj")
wangding zeng's avatar
wangding zeng committed
318

319
320
321
322
323
324
        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
325
326
327
328
329
330
        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,
331
332
            quant_config=quant_config,
            prefix=f"{prefix}.kv_b_proj")
wangding zeng's avatar
wangding zeng committed
333
334
335
336
        # O projection.
        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
                                        self.hidden_size,
                                        bias=False,
337
338
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")
339
340
        if rope_scaling:
            rope_scaling["rope_type"] = 'deepseek_yarn'
341

wangding zeng's avatar
wangding zeng committed
342
343
344
345
346
347
348
349
350
351
352
353
354
355
        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,
356
                              self.qk_head_dim,
wangding zeng's avatar
wangding zeng committed
357
358
359
                              self.scaling,
                              num_kv_heads=self.num_local_heads,
                              cache_config=cache_config,
360
361
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")
wangding zeng's avatar
wangding zeng committed
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

    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:]
388

wangding zeng's avatar
wangding zeng committed
389
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
390

wangding zeng's avatar
wangding zeng committed
391
392
393
394
        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
395
396
397
398
        # 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)
399
        attn_output = self.attn(q, k, v)
wangding zeng's avatar
wangding zeng committed
400
        attn_output = attn_output.view(
401
402
            -1, self.num_local_heads,
            self.qk_head_dim)[..., :self.v_head_dim].reshape(
wangding zeng's avatar
wangding zeng committed
403
404
405
406
407
                -1, self.num_local_heads * self.v_head_dim)
        output, _ = self.o_proj(attn_output)
        return output


408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
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,
427
        rope_scaling: Optional[dict[str, Any]] = None,
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
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
        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")

494
495
        if rope_scaling:
            rope_scaling["rope_type"] = 'deepseek_yarn'
496
497
498
499
500
501
502
503
504
505
506
507
        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

508
509
510
511
512
513
        # 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
514
515
        self.mla_attn = Attention(
            num_heads=self.num_local_heads,
516
            head_size=self.kv_lora_rank + self.qk_rope_head_dim,
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
            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,
            kv_b_proj=self.kv_b_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:
542
543
544
            q_c = self.q_a_proj(hidden_states)[0]
            q_c = self.q_a_layernorm(q_c)
            q = self.q_b_proj(q_c)[0]
545
        else:
546
            q = self.q_proj(hidden_states)[0]
547
548
549
        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())
550

551
552
553
554
555
556
557
        q = q.view(-1, self.num_local_heads, self.qk_head_dim)
        # Add head dim of 1 to k_pe
        k_pe = k_pe.unsqueeze(1)

        q[..., self.qk_nope_head_dim:], k_pe = self.rotary_emb(
            positions, q[..., self.qk_nope_head_dim:], k_pe)

558
559
560
561
562
563
564
        attn_out = self.mla_attn(
            q,
            kv_c_normed,
            k_pe,
            output_shape=(hidden_states.shape[0],
                          self.num_local_heads * self.v_head_dim))
        return self.o_proj(attn_out)[0]
565
566


wangding zeng's avatar
wangding zeng committed
567
568
569
570
571
class DeepseekV2DecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
572
        prefix: str,
573
        model_config: ModelConfig,
wangding zeng's avatar
wangding zeng committed
574
575
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
576
        enable_eplb: bool = False,
wangding zeng's avatar
wangding zeng committed
577
578
579
580
581
582
583
    ) -> 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)
584
585
586
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
        layer_idx = int(prefix.split(sep='.')[-1])
587
        self.layer_idx = layer_idx
588
589
590
591
592
        if model_config.use_mla:
            attn_cls = DeepseekV2MLAAttention
        else:
            attn_cls = DeepseekV2Attention
        self.self_attn = attn_cls(
wangding zeng's avatar
wangding zeng committed
593
594
595
596
597
598
599
600
601
602
603
604
605
606
            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,
607
            prefix=f"{prefix}.self_attn",
wangding zeng's avatar
wangding zeng committed
608
        )
609
        self.dpsk_fp16_quick = os.environ.get('DPSK_FP16_QUICK') == '1'
wangding zeng's avatar
wangding zeng committed
610
611
612
        if (config.n_routed_experts is not None
                and layer_idx >= config.first_k_dense_replace
                and layer_idx % config.moe_layer_freq == 0):
613
614
615
616
            self.mlp = DeepseekV2MoE(
                config=config,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
617
                enable_eplb=enable_eplb,
618
            )
wangding zeng's avatar
wangding zeng committed
619
620
621
622
623
624
        else:
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
625
                prefix=f"{prefix}.mlp",
wangding zeng's avatar
wangding zeng committed
626
627
628
629
630
            )
        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)
631
        self.routed_scaling_factor = config.routed_scaling_factor
wangding zeng's avatar
wangding zeng committed
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650

    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,
        )

王敏's avatar
王敏 committed
651
652
        #ops.print_tensor(hidden_states)

653
        if hidden_states.dtype == torch.float16 and not self.dpsk_fp16_quick:
654
655
656
657
658
659
660
661
            # Fix FP16 overflow
            # We scale both hidden_states and residual before
            # rmsnorm, and rmsnorm result would not affect by scale.
            hidden_states *= 1. / self.routed_scaling_factor
            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
662

wangding zeng's avatar
wangding zeng committed
663
664
665
666
        # Fully Connected
        hidden_states, residual = self.post_attention_layernorm(
            hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
667

668
        if isinstance(self.mlp,
669
                      DeepseekV2MLP) and hidden_states.dtype == torch.float16 and not self.dpsk_fp16_quick:
670
671
672
673
674
675
            # 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
            hidden_states *= 1. / self.routed_scaling_factor
676

wangding zeng's avatar
wangding zeng committed
677
678
679
        return hidden_states, residual


680
@support_torch_compile
wangding zeng's avatar
wangding zeng committed
681
682
683
684
class DeepseekV2Model(nn.Module):

    fall_back_to_pt_during_load = False

685
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
686
        super().__init__()
687
688

        config = vllm_config.model_config.hf_config
689
        model_config = vllm_config.model_config
690
691
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
692
        enable_eplb = vllm_config.parallel_config.enable_eplb
693
        self.config = config
694

wangding zeng's avatar
wangding zeng committed
695
696
        self.vocab_size = config.vocab_size

697
698
699
700
        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
701
702
                quant_config=quant_config,
                prefix=f"{prefix}.embed_tokens")
703
704
705
706
707
708
709
710
        else:
            self.embed_tokens = PPMissingLayer()

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: DeepseekV2DecoderLayer(
                config,
                prefix,
711
                model_config=model_config,
712
713
                cache_config=cache_config,
                quant_config=quant_config,
714
                enable_eplb=enable_eplb,
715
716
717
718
719
720
721
            ),
            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()
722
723
724
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))
wangding zeng's avatar
wangding zeng committed
725

726
727
728
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

wangding zeng's avatar
wangding zeng committed
729
730
731
732
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
733
        intermediate_tensors: Optional[IntermediateTensors],
734
        inputs_embeds: Optional[torch.Tensor] = None,
735
    ) -> Union[torch.Tensor, IntermediateTensors]:
736
        if get_pp_group().is_first_rank:
737
738
739
740
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
741
742
743
744
745
746
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

747
        for layer in self.layers[self.start_layer:self.end_layer]:
王敏's avatar
王敏 committed
748
749
750
            hidden_states, residual = layer(positions, hidden_states, residual)\
        
        #ops.print_tensor(hidden_states)
751
752
753
754
755
756
757

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

wangding zeng's avatar
wangding zeng committed
758
759
760
761
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


762
class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
wangding zeng's avatar
wangding zeng committed
763

764
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
765
        super().__init__()
766
767
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
768
769
770
771
772
773
774

        self.quant_method = None
        if quant_config is not None:
            self.quant_method = quant_config.get_name()
            os.environ['LLAMA_NN'] = '0'
            os.environ['LM_NN'] = '0'

775
        self.use_w4a16_moe_sz = os.environ.get('AWQ_MOE_SZ') == '1'
wangding zeng's avatar
wangding zeng committed
776
777
        self.config = config
        self.quant_config = quant_config
王敏's avatar
王敏 committed
778

779
        self.model = DeepseekV2Model(vllm_config=vllm_config,
780
                                     prefix=maybe_prefix(prefix, "model"))
781
782
783
784
785
786
        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
787
        self.logits_processor = LogitsProcessor(config.vocab_size)
788
789
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
        self.expert_weights = []

        # Set MoE hyperparameters
        self.num_moe_layers = (config.num_hidden_layers -
                               config.first_k_dense_replace)
        self.num_expert_groups = config.n_group

        self.moe_layers: list[FusedMoE] = []
        for layer in self.model.layers:
            assert isinstance(layer, DeepseekV2DecoderLayer)
            if isinstance(layer.mlp, DeepseekV2MoE):
                self.moe_layers.append(layer.mlp.experts)

        # Pick last one layer since the first ones may be dense layers.
        example_moe = typing.cast(
            DeepseekV2MoE, self.model.layers[config.num_hidden_layers - 1].mlp)
        self.num_logical_experts = example_moe.n_logical_experts
        self.num_physical_experts = example_moe.n_physical_experts
        self.num_local_physical_experts = example_moe.n_local_physical_experts
        self.num_routed_experts = example_moe.n_routed_experts
        self.num_shared_experts = example_moe.n_shared_experts
        self.num_redundant_experts = example_moe.n_redundant_experts
zhuwenwen's avatar
zhuwenwen committed
812
        
王敏's avatar
王敏 committed
813
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
814
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
815
816
817
        self.tritonsingleton= W8a8GetCacheJSON() 
        self.tritonsingleton.topk = config.num_experts_per_tok
        self.tritonsingleton.quant_method=self.quant_method 
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833

    def set_eplb_state(
        self,
        expert_load_view: torch.Tensor,
        logical_to_physical_map: torch.Tensor,
        logical_replica_count: torch.Tensor,
    ) -> None:
        for layer_idx, layer in enumerate(self.moe_layers):
            # Register the expert weights.
            self.expert_weights.append(layer.get_expert_weights())
            layer.set_eplb_state(
                moe_layer_idx=layer_idx,
                expert_load_view=expert_load_view,
                logical_to_physical_map=logical_to_physical_map,
                logical_replica_count=logical_replica_count,
            )
834

835
836
837
    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
838
839
840
841
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
842
        intermediate_tensors: Optional[IntermediateTensors] = None,
843
        inputs_embeds: Optional[torch.Tensor] = None,
844
    ) -> Union[torch.Tensor, IntermediateTensors]:
845
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
846
                                   inputs_embeds)
wangding zeng's avatar
wangding zeng committed
847
848
        return hidden_states

849
850
851
852
853
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
854
        logits = self.logits_processor(self.lm_head, hidden_states,
wangding zeng's avatar
wangding zeng committed
855
856
857
                                       sampling_metadata)
        return logits

858
859
860
861
862
863
864
865
866
867
868
869
870
    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),
        })
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
        
    def restore_qzeros_tensor(self, qzeros, qscales):

        low_bits = qzeros & 0x0F
        high_bits = qzeros >> 4
        
        zeors_tensor = torch.stack([low_bits, high_bits], dim=2).view(qzeros.shape[0], -1 , qzeros.shape[-1])
        zeors_int16 = zeors_tensor.to(torch.int16)
        assert zeors_int16.shape == qscales.shape

        uint16_tensor1 = zeors_int16.view(torch.uint16)
        uint16_tensor2 = qscales.view(torch.uint16)
        
        uint32_tensor1 = uint16_tensor1.to(torch.int32) << 16
        uint32_tensor2 = uint16_tensor2.to(torch.int32)
        
        result_tensor = uint32_tensor1 + uint32_tensor2
        result_tensor =result_tensor.view(torch.uint32)
        result_tensor = result_tensor.transpose(1, 2).contiguous()
        return result_tensor
891

892
893
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
wangding zeng's avatar
wangding zeng committed
894
895
896
897
898
899
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]

900
901
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
王敏's avatar
王敏 committed
902
903
904
905
        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",
906
907
            num_experts=self.config.n_routed_experts,
            num_redundant_experts=self.num_redundant_experts)
908

wangding zeng's avatar
wangding zeng committed
909
        params_dict = dict(self.named_parameters())
910
        loaded_params: set[str] = set()
wangding zeng's avatar
wangding zeng committed
911
912
913
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
914

915
916
917
            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
918

wangding zeng's avatar
wangding zeng committed
919
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
920
                # Skip non-stacked layers and experts (experts handled below).
wangding zeng's avatar
wangding zeng committed
921
922
                if weight_name not in name:
                    continue
923
924
925
926
927
928
929
930
                # 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
931
932
933
934
                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
935
936
937
938

                if is_pp_missing_parameter(name, self):
                    continue

wangding zeng's avatar
wangding zeng committed
939
940
941
942
943
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
944
                is_expert_weight = False
945
946
947
948
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
949

950
951
952
953
954
955
956
957
958
                    # Anyway, this is an expert weight and should not be
                    # attempted to load as other weights later
                    is_expert_weight = True

                    # Do not modify `name` since the loop may continue here
                    # Instead, create a new variable
                    name_mapped = name.replace(weight_name, param_name)

                    if is_pp_missing_parameter(name_mapped, self):
959
960
                        continue

961
962
963
964
965
966
967
968
969
970
971
972
973
                    param = params_dict[name_mapped]
                    # We should ask the weight loader to return success or not
                    # here since otherwise we may skip experts with other
                    # available replicas.
                    weight_loader = typing.cast(Callable[..., bool],
                                                param.weight_loader)
                    success = weight_loader(param,
                                            loaded_weight,
                                            name_mapped,
                                            shard_id=shard_id,
                                            expert_id=expert_id,
                                            return_success=True)
                    if success:
974
                        name = name_mapped
975
                        break
976
                else:
977
978
979
980
981
982
                    if is_expert_weight:
                        # We've checked that this is an expert weight
                        # However it's not mapped locally to this rank
                        # So we simply skip it
                        continue

983
984
985
986
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue

987
988
989
990
991
                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

992
993
994
                    if is_pp_missing_parameter(name, self):
                        continue

995
996
997
998
                    param = params_dict[name]
                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)
999
            loaded_params.add(name)
王敏's avatar
王敏 committed
1000
            
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
        if self.use_llama_nn and self.quant_method is None:
            lay_key_words = [
                "self_attn.q_proj.weight",
                "self_attn.q_a_proj.weight",
                "self_attn.q_b_proj.weight",
                "self_attn.kv_a_proj_with_mqa.weight",
                "self_attn.kv_b_proj.weight",
                "self_attn.o_proj.weight",
                "mlp.gate_up_proj.weight",
                "mlp.down_proj.weight",
                "mlp.gate.weight",
                "shared_experts.gate_up_proj.weight",
                "shared_experts.down_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)
zhuwenwen's avatar
zhuwenwen committed
1030
            
1031
        return loaded_params
1032
1033
1034
1035


class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047


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