"vllm/transformers_utils/tokenizer_group.py" did not exist on "c92acb9693c0504d7dabed2a0251b9f5d4ddaebb"
deepseek_v2.py 76.3 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, Tuple
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,
王敏's avatar
王敏 committed
43
44
45
46
                              get_tensor_model_parallel_world_size,
                              tensor_model_parallel_all_gather,
                              get_tensor_model_parallel_rank,
                              tensor_model_parallel_reduce_scatter)
wangding zeng's avatar
wangding zeng committed
47
from vllm.model_executor.layers.activation import SiluAndMul
王敏's avatar
王敏 committed
48
49
50
from vllm.model_executor.layers.fused_moe import FusedMoE, SharedFusedMoE

from vllm.model_executor.layers.fused_moe.utils import EPSharedExperts
wangding zeng's avatar
wangding zeng committed
51
52
53
54
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               MergedColumnParallelLinear,
                                               ReplicatedLinear,
55
                                               FusedQuantedReplicatedLinear,
wangding zeng's avatar
wangding zeng committed
56
57
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
58
from vllm.model_executor.layers.quantization import QuantizationConfig
zhuwenwen's avatar
zhuwenwen committed
59
from vllm.model_executor.layers.rotary_embedding import get_rope
wangding zeng's avatar
wangding zeng committed
60
61
from vllm.model_executor.layers.vocab_parallel_embedding import (
    ParallelLMHead, VocabParallelEmbedding)
王敏's avatar
王敏 committed
62
63
from vllm.model_executor.layers.dp_attention import (dp_gather, dp_reduce_scatter_tensor, 
    get_moe_tp_size, get_moe_tp_rank, get_attention_tp_size)
64
65
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader, maybe_remap_kv_scale_name)
wangding zeng's avatar
wangding zeng committed
66
from vllm.model_executor.sampling_metadata import SamplingMetadata
67
from vllm.sequence import IntermediateTensors
wangding zeng's avatar
wangding zeng committed
68

69
from .interfaces import MixtureOfExperts, SupportsPP
70
from .utils import (PPMissingLayer, is_pp_missing_parameter,
71
72
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
王敏's avatar
王敏 committed
73
from vllm import _custom_ops as ops
74
from vllm.utils import W8a8GetCacheJSON
75
from lmslim.quantize.quant_ops import lm_faster_rmsquant
76

wangding zeng's avatar
wangding zeng committed
77
78
79
80
81
82
83
84
85
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,
86
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
87
88
    ) -> None:
        super().__init__()
王敏's avatar
王敏 committed
89
90
91

        vllm_config = get_current_vllm_config()
        enable_dp_attention = vllm_config.parallel_config.enable_dp_attention
wangding zeng's avatar
wangding zeng committed
92
93
94
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
95
            quant_config=quant_config,
王敏's avatar
王敏 committed
96
97
            prefix=f"{prefix}.gate_up_proj",
            enable_dp_attn_moe=enable_dp_attention)
wangding zeng's avatar
wangding zeng committed
98
99
100
101
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
                                           quant_config=quant_config,
王敏's avatar
王敏 committed
102
103
104
                                           reduce_results=reduce_results if not enable_dp_attention else False,
                                           prefix=f"{prefix}.down_proj",
                                           enable_dp_attn_moe=enable_dp_attention)
wangding zeng's avatar
wangding zeng committed
105
106
107
108
109
        if hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {hidden_act}. "
                             "Only silu is supported for now.")
        self.act_fn = SiluAndMul()

110
    def forward(self, x,
111
112
113
                xqxs: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
                iqis: Optional[tuple[torch.Tensor, torch.Tensor]] = None
                ) -> torch.Tensor:
114
        if envs.USE_FUSED_RMS_QUANT:
115
116
            assert iqis is not None
            gate_up,  _   = self.gate_up_proj(x, iqis=iqis)
117
118
119
120
121
            if envs.USE_FUSED_SILU_MUL_QUANT:
                x, _ = self.down_proj(gate_up, use_fused_silu_mul_quant=True)
            else:
                x = self.act_fn(gate_up)
                x, _ = self.down_proj(x)
122
            return x
123
124
        elif envs.USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT and xqxs is not None:
            gate_up, _ = self.gate_up_proj(x, xqxs=xqxs)
125
126
127
128
129
            if envs.USE_FUSED_SILU_MUL_QUANT:
                x, _ = self.down_proj(gate_up, use_fused_silu_mul_quant=True)
            else:
                x = self.act_fn(gate_up)
                x, _ = self.down_proj(x)
130
            return x
131
132
133
134
135
        else:
            gate_up, _ = self.gate_up_proj(x)
            x = self.act_fn(gate_up)
            x, _ = self.down_proj(x)
            return x
wangding zeng's avatar
wangding zeng committed
136
137
138
139
140
141
142
143


class DeepseekV2MoE(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
144
        prefix: str = "",
145
        enable_eplb: bool = False,
wangding zeng's avatar
wangding zeng committed
146
147
148
149
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.routed_scaling_factor = config.routed_scaling_factor
150
151
152
153
154
155

        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
156
157
158
159
160

        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
161
        self.gate = ReplicatedLinear(config.hidden_size,
162
                                     config.n_routed_experts,
wangding zeng's avatar
wangding zeng committed
163
                                     bias=False,
164
165
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")
166
167
168
169
170
171
        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

172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
        # Load balancing settings.
        vllm_config = get_current_vllm_config()
        parallel_config = vllm_config.parallel_config
        self.enable_eplb = enable_eplb

        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)
187

王敏's avatar
王敏 committed
188
189
190
191
        dp_size = get_dp_group().world_size
        self.enable_expert_parallel = parallel_config.enable_expert_parallel
        self.use_deepep = dp_size > 1 and parallel_config.enable_expert_parallel and \
            (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput" or \
yangql's avatar
yangql committed
192
193
             envs.VLLM_ALL2ALL_BACKEND == "deepep_low_latency" or \
             envs.VLLM_ALL2ALL_BACKEND == "deepep_auto")
王敏's avatar
王敏 committed
194
        
195
        self.enable_shared_experts_overlap = False
196
197
        self.enable_shared_experts_fusion = (self.n_shared_experts != 0 and envs.VLLM_ENABLE_SHARED_EXPERTS_FUSION)

王敏's avatar
王敏 committed
198
        if not self.use_deepep:
199
            if config.n_shared_experts is not None and not envs.VLLM_ENABLE_SHARED_EXPERTS_FUSION:
王敏's avatar
王敏 committed
200
201
202
203
204
205
206
                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,
207
                    reduce_results = False,
王敏's avatar
王敏 committed
208
209
                    prefix=f"{prefix}.shared_experts",
                )
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
                self.enable_shared_experts_overlap = (not envs.VLLM_DISABLE_SHARED_EXPERTS_STREAM
                    and not envs.VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD
                    and config.n_shared_experts is not None)

            if self.enable_shared_experts_overlap:
                self.experts = SharedFusedMoE(
                    shared_experts=self.shared_experts,
                    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:
234
235
236
237
                num_fake_experts = (
                        config.n_shared_experts 
                        if self.enable_shared_experts_fusion
                        else 0)
238
                self.experts = FusedMoE(
239
240
                    num_experts=config.n_routed_experts + (num_fake_experts),
                    top_k=config.num_experts_per_tok + min(num_fake_experts, 1),
241
242
243
244
245
246
247
248
249
250
251
252
253
254
                    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)
王敏's avatar
王敏 committed
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
        else:
            if config.n_shared_experts is not None:
                intermediate_size = (config.moe_intermediate_size *
                                    config.n_shared_experts)
                self.shared_experts = EPSharedExperts(
                    hidden_size=config.hidden_size,
                    intermediate_size=intermediate_size,
                    hidden_act=config.hidden_act,
                    quant_config=quant_config,
                    reduce_results=False,
                    prefix=f"{prefix}.shared_experts",
                )
            self.experts = SharedFusedMoE(
                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,
                shared_experts=self.shared_experts)
王敏's avatar
王敏 committed
285

286
287
288
289
290
        self.run_shared_expert_singlely = (
                self.n_shared_experts is not None
                and not self.enable_shared_experts_overlap
                and not envs.VLLM_ENABLE_SHARED_EXPERTS_FUSION
                )
291

292
        from vllm.two_batch_overlap.two_batch_overlap import tbo_all_reduce
lizhigong's avatar
lizhigong committed
293
        self.tbo_all_reduce = tbo_all_reduce
wangding zeng's avatar
wangding zeng committed
294

295
    def forward(self, hidden_states: torch.Tensor,
296
                residual: Optional[torch.Tensor] = None,
297
                xqxs: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
298
                iqis: Optional[tuple[torch.Tensor, torch.Tensor]] = None  # iq = input quant, is = input scale
299
300
                ) -> Union[torch.Tensor,
                           Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]]:
王敏's avatar
王敏 committed
301
302
303
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)

304
        # For shared experts overlap optimization.
305
        def shared_exprts_overlap_pass(
306
307
308
309
310
311
312
                    hidden_states: torch.Tensor, router_logits: torch.Tensor, 
                    iqis: Optional[tuple[torch.Tensor, torch.Tensor]] = None
                ) -> tuple[torch.Tensor, torch.Tensor]:
            i_q, i_s = None, None
            if envs.USE_FUSED_RMS_QUANT:
                assert iqis is not None
                i_q, i_s = iqis[0], iqis[1]
313
314
            hidden_states_copy = hidden_states.clone()
            return self.experts(
315
316
317
318
319
                    hidden_states = hidden_states,
                    router_logits = router_logits,
                    hidden_states_copy = hidden_states_copy,
                    i_q = i_q,
                    i_s = i_s)
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
        
        # For shared experts fusion optimization.
        def shared_exprts_fusion_pass(
                    hidden_states: torch.Tensor, router_logits: torch.Tensor, 
                    iqis: Optional[tuple[torch.Tensor, torch.Tensor]] = None
                ) -> torch.Tensor:
            i_q, i_s = None, None
            if envs.USE_FUSED_RMS_QUANT:
                assert iqis is not None
                i_q, i_s = iqis[0], iqis[1]
            return self.experts(
                    hidden_states = hidden_states,
                    router_logits = router_logits,
                    i_q = i_q,
                    i_s = i_s)
335
336


337
        if envs.USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT and xqxs is not None:
338
            if self.n_shared_experts is not None and not self.enable_shared_experts_overlap:
339
                shared_output = self.shared_experts(hidden_states, xqxs=xqxs)
340

341
342
343
344
345
346
347
            router_logits, _ = self.gate(hidden_states)

            if envs.VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD:
                final_hidden_states = self.experts(
                    hidden_states=hidden_states,
                    router_logits=router_logits,
                    shared_output=shared_output)
王敏's avatar
王敏 committed
348
            else:
349
350
351
                if self.enable_shared_experts_fusion:
                    final_hidden_states = shared_exprts_fusion_pass(hidden_states, router_logits)
                elif self.enable_shared_experts_overlap:
352
353
                    assert self.shared_experts is not None
                    shared_output, final_hidden_states = shared_exprts_overlap_pass(hidden_states, router_logits)
354
355
356
                    # Fix FP16 overflow
                    # See DeepseekV2DecoderLayer for more details.
                    if hidden_states.dtype != torch.float16:
357
358
359
360
361
362
363
364
365
366
                        final_hidden_states *= self.routed_scaling_factor
                        final_hidden_states += shared_output
                    else:
                        assert shared_output is not None
                        final_hidden_states += (shared_output * (1.0 / self.routed_scaling_factor))
                else:
                    if hidden_states.dtype != torch.float16:
                        final_hidden_states = self.experts(
                            hidden_states=hidden_states,
                            router_logits=router_logits) * self.routed_scaling_factor
367
368
369
                    else:
                        # Fix FP16 overflow
                        # See DeepseekV2DecoderLayer for more details.
370
371
                        final_hidden_states = self.experts(hidden_states=hidden_states,
                                                        router_logits=router_logits)
372

373
374
375
376
377
378
379
380
                    if shared_output is not None:
                        if hidden_states.dtype != torch.float16:
                            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)
381
        else: # RQ
382
            if not self.enable_expert_parallel:
383
384
385
                i_q, i_s = None, None
                if iqis is not None:
                    i_q, i_s = iqis
386
                if self.run_shared_expert_singlely:
王敏's avatar
王敏 committed
387
                    if envs.USE_FUSED_RMS_QUANT:
388
                        shared_output = self.shared_experts(hidden_states, iqis=iqis)
王敏's avatar
王敏 committed
389
390
                    else:
                        shared_output = self.shared_experts(hidden_states)
391

王敏's avatar
王敏 committed
392
                router_logits, _ = self.gate(hidden_states)
393

394
395
396
                if self.enable_shared_experts_fusion:
                    final_hidden_states = shared_exprts_fusion_pass(hidden_states, router_logits, iqis = iqis)
                elif self.enable_shared_experts_overlap:
397
                    assert self.shared_experts is not None
398
                    shared_output, final_hidden_states = shared_exprts_overlap_pass(hidden_states, router_logits, iqis = iqis)
399
400
                    # Fix FP16 overflow
                    # See DeepseekV2DecoderLayer for more details.
401
                    if hidden_states.dtype != torch.float16:
402
403
404
405
406
407
408
                        final_hidden_states *= self.routed_scaling_factor
                        final_hidden_states += shared_output
                    else:
                        assert shared_output is not None
                        final_hidden_states += (shared_output * (1.0 / self.routed_scaling_factor))
                else:
                    if envs.VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD:
王敏's avatar
王敏 committed
409
410
                        final_hidden_states = self.experts(
                            hidden_states=hidden_states,
411
412
413
                            router_logits=router_logits,
                            shared_output=shared_output,
                            i_q=i_q, i_s=i_s)
414
                    else:
王敏's avatar
王敏 committed
415
                        if hidden_states.dtype != torch.float16:
416
417
418
419
                            final_hidden_states = self.experts(
                                hidden_states=hidden_states,
                                router_logits=router_logits,
                                i_q=i_q, i_s=i_s) * self.routed_scaling_factor
王敏's avatar
王敏 committed
420
421
422
                        else:
                            # Fix FP16 overflow
                            # See DeepseekV2DecoderLayer for more details.
423
424
                            # fp16 mode not fused quant
                            final_hidden_states = self.experts(hidden_states=hidden_states,
425
                                                            router_logits=router_logits,
426
                                                            i_q=i_q, i_s=i_s)
427
428
429
430
431
432
433
434
                        if shared_output is not None:
                            if hidden_states.dtype != torch.float16:
                                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)
435
            else:  # EP
王敏's avatar
王敏 committed
436
437
                router_logits, _ = self.gate(hidden_states)
                if self.use_deepep:
438
                    shared_output, final_hidden_states = self.experts(hidden_states=hidden_states,
王敏's avatar
王敏 committed
439
440
441
442
443
444
445
446
447
448
449
                                                router_logits=router_logits)

                    if shared_output is not None:
                        if hidden_states.dtype != torch.float16:
                            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)
                else:
450
                    if self.run_shared_expert_singlely:
王敏's avatar
王敏 committed
451
                        if envs.USE_FUSED_RMS_QUANT:
452
                            shared_output = self.shared_experts(hidden_states, iqis=iqis)
王敏's avatar
王敏 committed
453
454
455
                        else:
                            shared_output = self.shared_experts(hidden_states)

456
457
458
459
460
                    if self.enable_shared_experts_overlap:
                        assert self.shared_experts is not None
                        shared_output, final_hidden_states = shared_exprts_overlap_pass(hidden_states, router_logits)
                        # Fix FP16 overflow
                        # See DeepseekV2DecoderLayer for more details.
王敏's avatar
王敏 committed
461
                        if hidden_states.dtype != torch.float16:
462
                            final_hidden_states += shared_output
王敏's avatar
王敏 committed
463
                        else:
464
465
466
                            assert shared_output is not None
                            final_hidden_states += (shared_output * (1. / self.routed_scaling_factor))
                    else:
wujl5's avatar
wujl5 committed
467
468
                        if iqis is not None:
                            i_q, i_s = iqis
469
470
                        final_hidden_states = self.experts(hidden_states=hidden_states,
                                                           router_logits=router_logits,
471
                                                           i_q=i_q, i_s=i_s)
472
473
474
475
476
477
478
479
480
481
482
                        if shared_output is not None:
                            if hidden_states.dtype != torch.float16:
                                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)
lizhigong's avatar
lizhigong committed
483
            else:
484
485
486
487
                final_hidden_states = (
                    self.experts.maybe_all_reduce_tensor_model_parallel(
                        final_hidden_states))
            return final_hidden_states.view(num_tokens, hidden_dim)
wangding zeng's avatar
wangding zeng committed
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509


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,
510
        rope_scaling: Optional[dict[str, Any]] = None,
wangding zeng's avatar
wangding zeng committed
511
512
513
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
514
        prefix: str = "",
王敏's avatar
王敏 committed
515
        reduce_results: bool = True,
wangding zeng's avatar
wangding zeng committed
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
    ) -> 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,
537
538
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.q_a_proj")
wangding zeng's avatar
wangding zeng committed
539
540
541
542
543
544
            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,
545
546
                                                 quant_config=quant_config,
                                                 prefix=f"{prefix}.q_b_proj")
wangding zeng's avatar
wangding zeng committed
547
548
549
550
551
        else:
            self.q_proj = ColumnParallelLinear(self.hidden_size,
                                               self.num_heads *
                                               self.qk_head_dim,
                                               bias=False,
552
553
                                               quant_config=quant_config,
                                               prefix=f"{prefix}.q_proj")
wangding zeng's avatar
wangding zeng committed
554

555
556
557
558
559
560
        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
561
562
563
564
565
566
        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,
567
568
            quant_config=quant_config,
            prefix=f"{prefix}.kv_b_proj")
wangding zeng's avatar
wangding zeng committed
569
570
571
572
        # O projection.
        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
                                        self.hidden_size,
                                        bias=False,
573
                                        quant_config=quant_config,
王敏's avatar
王敏 committed
574
575
                                        prefix=f"{prefix}.o_proj",
                                        reduce_results=reduce_results)
576
577
        if rope_scaling:
            rope_scaling["rope_type"] = 'deepseek_yarn'
578

wangding zeng's avatar
wangding zeng committed
579
580
581
582
583
584
585
586
587
588
589
590
591
592
        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,
593
                              self.qk_head_dim,
wangding zeng's avatar
wangding zeng committed
594
595
596
                              self.scaling,
                              num_kv_heads=self.num_local_heads,
                              cache_config=cache_config,
597
598
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")
wangding zeng's avatar
wangding zeng committed
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624

    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:]
625

wangding zeng's avatar
wangding zeng committed
626
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
627

wangding zeng's avatar
wangding zeng committed
628
629
630
631
        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
632
633
634
635
        # 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)
636
        attn_output = self.attn(q, k, v)
wangding zeng's avatar
wangding zeng committed
637
        attn_output = attn_output.view(
638
639
            -1, self.num_local_heads,
            self.qk_head_dim)[..., :self.v_head_dim].reshape(
wangding zeng's avatar
wangding zeng committed
640
641
642
643
644
                -1, self.num_local_heads * self.v_head_dim)
        output, _ = self.o_proj(attn_output)
        return output


645
646
647
648
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).
649

650
651
652
653
654
655
656
657
658
659
660
661
662
663
    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,
664
        rope_scaling: Optional[dict[str, Any]] = None,
665
666
667
668
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
王敏's avatar
王敏 committed
669
        reduce_results: bool = True,
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
    ) -> 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:
691
            if envs.USE_FUSED_RMS_QUANT:
692
693
694
695
696
697
698
699
700
701
702
703
704
705
                if envs.VLLM_USE_FUSED_QA_KVA_GEMM:
                    self.qa_kva_proj = FusedQuantedReplicatedLinear(self.hidden_size,
                                                             self.q_lora_rank,
                                                             self.kv_lora_rank,
                                                             self.qk_rope_head_dim,
                                                             bias=False,
                                                             quant_config=quant_config,
                                                             prefix=f"{prefix}.qa_kva_proj")
                else:
                    self.q_a_proj = ReplicatedLinear(self.hidden_size,
                                                self.q_lora_rank,
                                                bias=False,
                                                quant_config=quant_config,
                                                prefix=f"{prefix}.q_a_proj")
706
                self.q_b_proj = ColumnParallelLinear(q_lora_rank,
707
708
709
710
                                                 self.num_heads *
                                                 self.qk_head_dim,
                                                 bias=False,
                                                 quant_config=quant_config,
711
                                                 eps=config.rms_norm_eps,
712
                                                 prefix=f"{prefix}.q_b_proj")
713
714
715
716
717
718
719
720
721
722
723
724
            else:
                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_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")
725

726
727
728
            self.q_a_layernorm = RMSNorm(self.q_lora_rank,
                                         eps=config.rms_norm_eps)

729
730
731
732
733
734
735
        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")
736
737
738
739
740
741
742
        if not envs.VLLM_USE_FUSED_QA_KVA_GEMM:
            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")
743
744
745
746
747
748
749
750
751
752
753
754
        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,
王敏's avatar
王敏 committed
755
756
                                        prefix=f"{prefix}.o_proj",
                                        reduce_results=reduce_results)
757

758
759
        if rope_scaling:
            rope_scaling["rope_type"] = 'deepseek_yarn'
760
761
762
763
764
765
766
767
768
769
770
771
        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

772
773
774
775
776
777
        # 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
778
779
        self.mla_attn = Attention(
            num_heads=self.num_local_heads,
780
            head_size=self.kv_lora_rank + self.qk_rope_head_dim,
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
            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])
799

800
801
802
803
    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
804
805
806
807
        pa_rms_weight: Optional[torch.Tensor] = None,
        pa_residual: Optional[torch.Tensor] = None,
        pa_rms_eps: Optional[float] = 1e-6,
        pa_quant_dtype: Optional[torch.dtype] = torch.int8,
808
809
810
        update_input: Optional[bool] = True,
        iqis: Optional[tuple[torch.Tensor, torch.Tensor]] = None
            
811
    ) -> Union[torch.Tensor,
812
               Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]]:
813
        if envs.USE_FUSED_RMS_QUANT and iqis is not None:
814
815
            if envs.VLLM_USE_FUSED_QA_KVA_GEMM:
                if self.q_lora_rank is not None:
816
                    qc_kvc_kpe, _bias = self.qa_kva_proj(hidden_states,  iqis)
817
818
819
820
821
822
823
                    q_c = qc_kvc_kpe[:, :self.q_lora_rank]
                    kvc_kpe = qc_kvc_kpe[:, self.q_lora_rank:]
                    q, _, _ = self.q_b_proj(q_c, rms_weight=self.q_a_layernorm.weight.data, residual=None, update_hd=False)
                    
                else:
                    q = self.q_proj(hidden_states)[0]
                kv_c, k_pe = kvc_kpe.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
824
            else:
825
                if self.q_lora_rank is not None:
826
                    q_c, _ = self.q_a_proj(hidden_states, iqis=iqis)
827
828
829
830
                    q, _, _ = self.q_b_proj(q_c, rms_weight=self.q_a_layernorm.weight.data, residual=None, update_hd=False)
                    
                else:
                    q = self.q_proj(hidden_states)[0]
831
                kvc_kpe = self.kv_a_proj_with_mqa(hidden_states, iqis=iqis)[0]
832
                kv_c, k_pe = kvc_kpe.split(
833
                                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
834

835
836
837
838
839
            if not envs.VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT:
                if envs.VLLM_USE_LIGHTOP:
                    kv_c_normed = self.kv_a_layernorm.forward_cuda_opt(kv_c)
                else:
                    kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())
840

841
842
843
                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)
844

845
846
                q[..., self.qk_nope_head_dim:], k_pe = self.rotary_emb(
                    positions, q[..., self.qk_nope_head_dim:], k_pe)
847

848
849
850
851
852
853
854
855
856
857
858
859
860
861
                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))
            else:
                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)
                weight = self.kv_a_layernorm.weight
                cos_sin_cache = self.rotary_emb.cos_sin_cache
                if cos_sin_cache.device != positions.device or cos_sin_cache.device != q.dtype:
                    cos_sin_cache = cos_sin_cache.to(positions.device, dtype=q.dtype)
862
                kv_c_normed = torch.empty(kv_c.shape, dtype=kv_c.dtype, device=kv_c.device)
863
864
865
866
867
868
                attn_out = self.mla_attn(
                    q[..., self.qk_nope_head_dim:],
                    kv_c,
                    k_pe,
                    output_shape=(hidden_states.shape[0],
                                self.num_local_heads * self.v_head_dim),
869
870
                    query_nope=q[..., :self.qk_nope_head_dim],
                    num_local_heads=self.num_local_heads,
871
872
873
874
875
                    q_ori=q,
                    key_normed=kv_c_normed,
                    positions=positions,
                    weight=weight,
                    cos_sin_cache=cos_sin_cache)
876
            return self.o_proj(attn_out)[0]
877
878
879
880
881
882
883
884
885
        elif envs.USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT and pa_rms_weight is not None and pa_residual is not None:
            if self.q_lora_rank is not None:
                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]
            else:
                q = self.q_proj(hidden_states)[0]
            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)
886
887
888
            if not envs.VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT:
                if envs.VLLM_USE_LIGHTOP:
                    kv_c_normed = self.kv_a_layernorm.forward_cuda_opt(kv_c)
889
                else:
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
                    kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())

                q = q.view(-1, self.num_local_heads, self.qk_head_dim)
                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)

                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))
            else:
                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)
                weight = self.kv_a_layernorm.weight
                cos_sin_cache = self.rotary_emb.cos_sin_cache
                if cos_sin_cache.device != positions.device or cos_sin_cache.device != q.dtype:
                    cos_sin_cache = cos_sin_cache.to(positions.device, dtype=q.dtype)
912
                kv_c_normed = torch.empty(kv_c.shape, dtype=kv_c.dtype, device=kv_c.device)
913
914
915
916
917
918
                attn_out = self.mla_attn(
                    q[..., self.qk_nope_head_dim:],
                    kv_c,
                    k_pe,
                    output_shape=(hidden_states.shape[0],
                                self.num_local_heads * self.v_head_dim),
919
920
                    query_nope=q[..., :self.qk_nope_head_dim],
                    num_local_heads=self.num_local_heads,
921
922
923
924
925
                    q_ori=q,
                    key_normed=kv_c_normed,
                    positions=positions,
                    weight=weight,
                    cos_sin_cache=cos_sin_cache)
926
            packages_ = self.o_proj(attn_out,
927
928
929
930
931
932
933
934
935
936
                                   pa_rms_weight=pa_rms_weight,
                                   pa_residual=pa_residual,
                                   pa_rms_eps=pa_rms_eps,
                                   pa_quant_dtype=pa_quant_dtype,
                                   update_input=update_input)[:4]
            assert len(packages_) == 4
            hs, resi, xq, xs = packages_
            assert xq is not None and xs is not None
            return hs, resi, xq, xs

937
        else:
938
939
940
941
942
943
944
945
            if self.q_lora_rank is not None:
                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]
            else:
                q = self.q_proj(hidden_states)[0]
            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)
946
947
948
949
950
            if not envs.VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT:
                if envs.VLLM_USE_LIGHTOP:
                    kv_c_normed = self.kv_a_layernorm.forward_cuda_opt(kv_c)
                else:
                    kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())
951

952
953
954
                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)
955

956
957
                q[..., self.qk_nope_head_dim:], k_pe = self.rotary_emb(
                    positions, q[..., self.qk_nope_head_dim:], k_pe)
958

959
960
961
962
963
964
965
966
967
968
                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))
            else:
                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)
zhuwenwen's avatar
zhuwenwen committed
969
970
971
972
                weight = self.kv_a_layernorm.weight
                cos_sin_cache = self.rotary_emb.cos_sin_cache
                if cos_sin_cache.device != positions.device or cos_sin_cache.device != q.dtype:
                    cos_sin_cache = cos_sin_cache.to(positions.device, dtype=q.dtype)
973
                kv_c_normed = torch.empty(kv_c.shape, dtype=kv_c.dtype, device=kv_c.device)
974
975
976
977
978
979
                attn_out = self.mla_attn(
                    q[..., self.qk_nope_head_dim:],
                    kv_c,
                    k_pe,
                    output_shape=(hidden_states.shape[0],
                                self.num_local_heads * self.v_head_dim),
980
981
                    query_nope=q[..., :self.qk_nope_head_dim],
                    num_local_heads=self.num_local_heads,
982
983
984
                    q_ori=q,
                    key_normed=kv_c_normed,
                    positions=positions,
zhuwenwen's avatar
zhuwenwen committed
985
986
                    weight=weight,
                    cos_sin_cache=cos_sin_cache)
987
            return self.o_proj(attn_out)[0]
988
989


wangding zeng's avatar
wangding zeng committed
990
991
992
993
994
class DeepseekV2DecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
995
        prefix: str,
996
        model_config: ModelConfig,
wangding zeng's avatar
wangding zeng committed
997
998
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
999
        enable_eplb: bool = False,
wangding zeng's avatar
wangding zeng committed
1000
1001
1002
1003
1004
1005
1006
    ) -> 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)
1007
1008
1009
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
        layer_idx = int(prefix.split(sep='.')[-1])
1010
        self.layer_idx = layer_idx
王敏's avatar
王敏 committed
1011
1012
1013
1014
1015
1016

        self.dp_size = get_dp_group().world_size
        vllm_config = get_current_vllm_config()
        parallel_config = vllm_config.parallel_config
        self.use_deepep = self.dp_size > 1 and parallel_config.enable_expert_parallel and \
            (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput" or \
yangql's avatar
yangql committed
1017
1018
             envs.VLLM_ALL2ALL_BACKEND == "deepep_low_latency" or \
             envs.VLLM_ALL2ALL_BACKEND == "deepep_auto")
王敏's avatar
王敏 committed
1019
1020
1021
        self.tp_size = get_tensor_model_parallel_world_size()
        self.config = config
        self.tp_rank = get_tensor_model_parallel_rank()
王敏's avatar
王敏 committed
1022
1023
        self.enable_dp_attention = vllm_config.parallel_config.enable_dp_attention
        self.dp_rank = vllm_config.parallel_config.data_parallel_rank
王敏's avatar
王敏 committed
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042

        if (config.n_routed_experts is not None
                and layer_idx >= config.first_k_dense_replace
                and layer_idx % config.moe_layer_freq == 0):
            self.mlp = DeepseekV2MoE(
                config=config,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
                enable_eplb=enable_eplb,
            )
        else:
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )

王敏's avatar
王敏 committed
1043
1044
        self.enable_ep_sp = isinstance(self.mlp,
                        DeepseekV2MoE) and self.use_deepep and self.tp_size > 1
王敏's avatar
王敏 committed
1045
1046
1047
1048
1049
1050
1051
1052
        self.is_mtp_layer = False
        if self.layer_idx == config.num_hidden_layers:
            self.is_mtp_layer = True
        reduce_results = True
        if isinstance(self.mlp,
                        DeepseekV2MoE) and self.use_deepep and \
                            self.tp_size > 1 and not self.is_mtp_layer:
            reduce_results = False
王敏's avatar
王敏 committed
1053
1054
1055
        else:
            if self.enable_dp_attention:
                reduce_results = False
王敏's avatar
王敏 committed
1056

1057
1058
1059
1060
1061
        if model_config.use_mla:
            attn_cls = DeepseekV2MLAAttention
        else:
            attn_cls = DeepseekV2Attention
        self.self_attn = attn_cls(
wangding zeng's avatar
wangding zeng committed
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
            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,
1076
            prefix=f"{prefix}.self_attn",
王敏's avatar
王敏 committed
1077
            reduce_results=reduce_results
wangding zeng's avatar
wangding zeng committed
1078
        )
1079

wangding zeng's avatar
wangding zeng committed
1080
1081
1082
1083
        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)
1084
        self.routed_scaling_factor = config.routed_scaling_factor
1085
1086
        self.use_fused_rms_quant = envs.USE_FUSED_RMS_QUANT
        self.use_fused_custom_all_reduce = envs.USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT
1087
        self._eps = config.rms_norm_eps
wangding zeng's avatar
wangding zeng committed
1088

1089

王敏's avatar
王敏 committed
1090

1091
    def forward_fused_RQ(
wangding zeng's avatar
wangding zeng committed
1092
1093
1094
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1095
1096
1097
1098
        residual: Optional[torch.Tensor]
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        # Fix residual FP16 overflow
        residual_fix_overflow = False
1099

1100
1101
1102
1103
        assert self.input_layernorm.has_weight is True
        if residual is None:
            residual = hidden_states
            residual_fix_overflow = True
1104
1105
1106
1107
1108
1109
            i_q, i_s = lm_faster_rmsquant(input=hidden_states,
                                         rms_weight = self.input_layernorm.weight.data,
                                         epsilon=self._eps,
                                         quant_dtype=torch.int8,
                                         residual=None,
                                         update_input=False)
1110
        else:
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
            i_q, i_s = lm_faster_rmsquant(input=hidden_states,
                                         rms_weight = self.input_layernorm.weight.data,
                                         epsilon=self._eps,
                                         quant_dtype=torch.int8,
                                         residual=residual,
                                         update_input=False)
            
        hidden_states = self.self_attn(positions=positions,
                                       hidden_states = hidden_states, # get attr
                                       iqis=(i_q, i_s))
            
1122
1123
1124
1125
1126
1127
1128
        if hidden_states.dtype == torch.float16:
            # rmsnorm, and rmsnorm result would not affect by scale.
            hidden_states *= 1. / self.routed_scaling_factor
            if self.layer_idx == 0 or residual_fix_overflow:
                # The residual is shared by all layers, we only scale it on
                # first layer.
                residual *= 1. / self.routed_scaling_factor
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
        
        update_hs = True if isinstance(self.mlp, DeepseekV2MoE) else False
        _i_q, _i_s = lm_faster_rmsquant(input=hidden_states,
                                        rms_weight=self.post_attention_layernorm.weight.data,
                                        epsilon=self._eps,
                                        quant_dtype=torch.int8,
                                        residual=residual,
                                        update_input=update_hs)
        new_resi = residual
        hidden_states = self.mlp(hidden_states, 
                                 iqis=(_i_q, _i_s))
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151

        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
            hidden_states *= 1. / self.routed_scaling_factor
        return hidden_states, new_resi

    def forward_fused_CRQ(
1152
1153
1154
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
        residual: Optional[torch.Tensor]
        ) -> Tuple[torch.Tensor, torch.Tensor]:
        residual_fix_overflow = False
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
            residual_fix_overflow = True
        else:
            hidden_states, resi_new = self.input_layernorm(
                hidden_states, residual)
1165
            residual = resi_new
1166
1167
1168
        new_hs, new_resi, xq, xs = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
1169
            pa_rms_weight=self.post_attention_layernorm.weight.data,
1170
1171
1172
1173
1174
            pa_residual=residual,
            pa_rms_eps=self.post_attention_layernorm.variance_epsilon,
            pa_quant_dtype = torch.int8,
            update_input=True
        )
1175
1176


1177
1178
1179
1180
1181
        assert xq is not None and xs is not None
        if new_hs.dtype == torch.float16: # overflow处理逻辑
            new_hs *= 1. / self.routed_scaling_factor
            if self.layer_idx == 0 or residual_fix_overflow:
                new_resi *= 1. / self.routed_scaling_factor
1182

1183
1184
1185
1186
1187
1188
        hidden_states = self.mlp(new_hs, xqxs=(xq, xs))

        if isinstance(self.mlp,
                    DeepseekV2MLP) and hidden_states.dtype == torch.float16:
            hidden_states *= 1. / self.routed_scaling_factor
        return hidden_states, new_resi
1189

1190
    def forward_default(
1191
        self,
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor]
        ) -> Tuple[torch.Tensor, torch.Tensor]:
        # Self Attention
        # Fix residual FP16 overflow
        residual_fix_overflow = False
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
            residual_fix_overflow = True
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)
1206

王敏's avatar
王敏 committed
1207
1208
1209
        if not self.is_mtp_layer and self.enable_ep_sp and \
            self.layer_idx > self.config.first_k_dense_replace:
            hidden_states = tensor_model_parallel_all_gather(hidden_states, dim=0)
王敏's avatar
王敏 committed
1210

1211
1212
1213
1214
1215
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )

王敏's avatar
王敏 committed
1216
1217
1218
        if not self.is_mtp_layer and self.enable_ep_sp:
            if self.layer_idx == self.config.first_k_dense_replace:
                residual = residual.tensor_split(self.tp_size)[self.tp_rank]
王敏's avatar
王敏 committed
1219

王敏's avatar
王敏 committed
1220
            hidden_states = tensor_model_parallel_reduce_scatter(hidden_states, dim=0)
王敏's avatar
王敏 committed
1221

王敏's avatar
王敏 committed
1222
1223
1224
1225
        if self.enable_dp_attention:
            if self.tp_rank == 0:
                hidden_states += residual
            hidden_states = dp_gather(hidden_states)
王敏's avatar
王敏 committed
1226

1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
        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.
            hidden_states *= 1. / self.routed_scaling_factor
            if self.layer_idx == 0 or residual_fix_overflow:
                # The residual is shared by all layers, we only scale it on
                # first layer.
                residual *= 1. / self.routed_scaling_factor

        # Fully Connected
王敏's avatar
王敏 committed
1238
1239
1240
1241
1242
1243
1244
1245
        if not self.enable_dp_attention:
            hidden_states, residual = self.post_attention_layernorm(
                hidden_states, residual)
        else:
            num_tokens = hidden_states.shape[0]
            new_bs = num_tokens // get_moe_tp_size() * get_attention_tp_size()
            residual = hidden_states[self.dp_rank*new_bs: (self.dp_rank+1)*new_bs, :]
            hidden_states = self.post_attention_layernorm(hidden_states)
1246

王敏's avatar
王敏 committed
1247
1248
1249
1250
1251
1252
1253
        if self.is_mtp_layer and self.enable_ep_sp:
            ori_bs = hidden_states.shape[0]
            pad_size = (ori_bs + self.tp_size - 1) // self.tp_size * self.tp_size - ori_bs
            if pad_size > 0:
                hidden_states = torch.nn.functional.pad(hidden_states, [0, 0, 0, pad_size], value=0)
            new_bs = (ori_bs+pad_size) // self.tp_size
            hidden_states = hidden_states[self.tp_rank*new_bs: (self.tp_rank+1)*new_bs, :]
1254

1255
1256
        hidden_states = self.mlp(hidden_states)

王敏's avatar
王敏 committed
1257
1258
1259
        if self.enable_dp_attention:
            hidden_states = dp_reduce_scatter_tensor(hidden_states)

王敏's avatar
王敏 committed
1260
1261
1262
        if self.is_mtp_layer and self.enable_ep_sp:
            hidden_states = tensor_model_parallel_all_gather(hidden_states, dim=0)
            hidden_states = hidden_states[:ori_bs, :]
王敏's avatar
王敏 committed
1263

1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
        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
            hidden_states *= 1. / self.routed_scaling_factor

        return hidden_states, residual
1274

1275
1276
    def choose_forward(self):
        if self.use_fused_rms_quant:
1277
            return self.forward_fused_RQ
1278

1279
1280
        elif self.use_fused_custom_all_reduce:
            return self.forward_fused_CRQ
wangding zeng's avatar
wangding zeng committed
1281
        else:
1282
            return self.forward_default
1283

1284
1285
1286
1287
1288
1289
1290
1291
1292
    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor]
    )  -> Tuple[torch.Tensor, torch.Tensor]:
        forward_func = self.choose_forward()
        return forward_func(positions=positions, hidden_states=hidden_states, residual=residual )

wangding zeng's avatar
wangding zeng committed
1293

1294
@support_torch_compile
wangding zeng's avatar
wangding zeng committed
1295
1296
1297
1298
class DeepseekV2Model(nn.Module):

    fall_back_to_pt_during_load = False

1299
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
1300
        super().__init__()
1301
1302

        config = vllm_config.model_config.hf_config
1303
        model_config = vllm_config.model_config
1304
1305
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
1306
        enable_eplb = vllm_config.parallel_config.enable_eplb
1307
        self.config = config
1308

wangding zeng's avatar
wangding zeng committed
1309
1310
        self.vocab_size = config.vocab_size

1311
1312
1313
1314
        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
1315
1316
                quant_config=quant_config,
                prefix=f"{prefix}.embed_tokens")
1317
1318
1319
1320
1321
1322
1323
1324
        else:
            self.embed_tokens = PPMissingLayer()

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: DeepseekV2DecoderLayer(
                config,
                prefix,
1325
                model_config=model_config,
1326
1327
                cache_config=cache_config,
                quant_config=quant_config,
1328
                enable_eplb=enable_eplb,
1329
1330
1331
1332
1333
1334
1335
            ),
            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()
1336
1337
1338
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))
1339

王敏's avatar
王敏 committed
1340
1341
1342
1343
1344
        self.dp_size = get_dp_group().world_size
        vllm_config = get_current_vllm_config()
        parallel_config = vllm_config.parallel_config
        self.use_deepep = self.dp_size > 1 and parallel_config.enable_expert_parallel and \
            (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput" or \
yangql's avatar
yangql committed
1345
1346
             envs.VLLM_ALL2ALL_BACKEND == "deepep_low_latency" or \
             envs.VLLM_ALL2ALL_BACKEND == "deepep_auto")
王敏's avatar
王敏 committed
1347
        self.tp_size = get_tensor_model_parallel_world_size()
wangding zeng's avatar
wangding zeng committed
1348

1349
1350
1351
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

wangding zeng's avatar
wangding zeng committed
1352
1353
1354
1355
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1356
        intermediate_tensors: Optional[IntermediateTensors],
1357
        inputs_embeds: Optional[torch.Tensor] = None,
1358
    ) -> Union[torch.Tensor, IntermediateTensors]:
1359
        if get_pp_group().is_first_rank:
1360
1361
1362
1363
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
1364
1365
1366
1367
1368
1369
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

1370
        for layer in self.layers[self.start_layer:self.end_layer]:
1371
            hidden_states, residual = layer(positions, hidden_states, residual)
1372
1373
1374
1375
1376
1377
1378

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

wangding zeng's avatar
wangding zeng committed
1379
        hidden_states, _ = self.norm(hidden_states, residual)
王敏's avatar
王敏 committed
1380
1381
1382
1383

        if self.use_deepep and self.tp_size > 1:
            hidden_states = tensor_model_parallel_all_gather(hidden_states, dim=0)

wangding zeng's avatar
wangding zeng committed
1384
1385
1386
        return hidden_states


1387
class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
wangding zeng's avatar
wangding zeng committed
1388

1389
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
1390
        super().__init__()
1391
1392
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
1393
1394
1395
1396
1397
1398
1399

        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'

1400
        self.use_w4a16_moe_sz = os.environ.get('AWQ_MOE_SZ') == '1'
wangding zeng's avatar
wangding zeng committed
1401
1402
        self.config = config
        self.quant_config = quant_config
王敏's avatar
王敏 committed
1403

1404
        self.model = DeepseekV2Model(vllm_config=vllm_config,
1405
                                     prefix=maybe_prefix(prefix, "model"))
1406
1407
1408
1409
1410
1411
        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
1412
        self.logits_processor = LogitsProcessor(config.vocab_size)
1413
1414
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
1415
1416
1417
1418
1419
1420
1421
1422
        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] = []
1423
        example_moe = None
1424
        for layer in self.model.layers:
1425
1426
1427
            if isinstance(layer, PPMissingLayer):
                continue

1428
1429
            assert isinstance(layer, DeepseekV2DecoderLayer)
            if isinstance(layer.mlp, DeepseekV2MoE):
1430
                example_moe = layer.mlp
1431
1432
1433
1434
1435
1436
1437
1438
1439
                self.moe_layers.append(layer.mlp.experts)

        # Pick last one layer since the first ones may be dense layers.
        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
1440

王敏's avatar
王敏 committed
1441
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
1442
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
1443
        self.tritonsingleton= W8a8GetCacheJSON()
1444
        self.tritonsingleton.topk = config.num_experts_per_tok
王敏's avatar
王敏 committed
1445
        self.tritonsingleton.quant_method=self.quant_method
王敏's avatar
王敏 committed
1446

1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
    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,
            )
1462

1463
1464
1465
    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
1466
1467
1468
1469
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1470
        intermediate_tensors: Optional[IntermediateTensors] = None,
1471
        inputs_embeds: Optional[torch.Tensor] = None,
1472
    ) -> Union[torch.Tensor, IntermediateTensors]:
1473
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
1474
                                   inputs_embeds)
wangding zeng's avatar
wangding zeng committed
1475
1476
        return hidden_states

1477
1478
1479
1480
1481
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
1482
        logits = self.logits_processor(self.lm_head, hidden_states,
wangding zeng's avatar
wangding zeng committed
1483
1484
1485
                                       sampling_metadata)
        return logits

1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
    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),
        })
1499

1500
1501
1502
1503
    def restore_qzeros_tensor(self, qzeros, qscales):

        low_bits = qzeros & 0x0F
        high_bits = qzeros >> 4
1504

1505
1506
1507
1508
1509
1510
        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)
1511

1512
1513
        uint32_tensor1 = uint16_tensor1.to(torch.int32) << 16
        uint32_tensor2 = uint16_tensor2.to(torch.int32)
1514

1515
1516
1517
1518
        result_tensor = uint32_tensor1 + uint32_tensor2
        result_tensor =result_tensor.view(torch.uint32)
        result_tensor = result_tensor.transpose(1, 2).contiguous()
        return result_tensor
1519

1520
1521
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
wangding zeng's avatar
wangding zeng committed
1522
1523
1524
1525
1526
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
1527
1528
1529
1530
1531
1532
        if envs.USE_FUSED_RMS_QUANT and envs.VLLM_USE_FUSED_QA_KVA_GEMM:
            fused_params_mapping = [
                ("qa_kva_proj", "q_a_proj", 0),
                ("qa_kva_proj", "kv_a_proj_with_mqa", 1)
            ]
            stacked_params_mapping += fused_params_mapping
wangding zeng's avatar
wangding zeng committed
1533

1534
1535
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
王敏's avatar
王敏 committed
1536
1537
1538
1539
        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",
1540
1541
1542
1543
1544
1545
            num_experts=self.config.n_routed_experts
            + ( 
                self.num_shared_experts
                if (envs.VLLM_ENABLE_SHARED_EXPERTS_FUSION and self.num_shared_experts > 0)
                else 0
            ),
1546
            num_redundant_experts=self.num_redundant_experts)
1547

wangding zeng's avatar
wangding zeng committed
1548
        params_dict = dict(self.named_parameters())
1549
        loaded_params: set[str] = set()
wangding zeng's avatar
wangding zeng committed
1550
1551
1552
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
1553

1554
1555
1556
            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
1557

1558
1559
1560
1561
1562
1563
1564
1565
1566
            # Assuame num of shared experts is only one.
            if envs.VLLM_ENABLE_SHARED_EXPERTS_FUSION \
                    and self.num_shared_experts > 0 \
                    and "mlp.shared_experts" in name:
                name = name.replace(
                    "mlp.shared_experts",
                    f"mlp.experts.{self.config.n_routed_experts}",
                )

wangding zeng's avatar
wangding zeng committed
1567
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
1568
                # Skip non-stacked layers and experts (experts handled below).
wangding zeng's avatar
wangding zeng committed
1569
1570
                if weight_name not in name:
                    continue
1571
1572
1573
1574
1575
1576
1577
1578
                # 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
1579
                old_weight_name = name
wangding zeng's avatar
wangding zeng committed
1580
                name = name.replace(weight_name, param_name)
王敏's avatar
王敏 committed
1581

wangding zeng's avatar
wangding zeng committed
1582
1583
1584
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
1585
1586
1587
1588

                if is_pp_missing_parameter(name, self):
                    continue

wangding zeng's avatar
wangding zeng committed
1589
1590
                param = params_dict[name]
                weight_loader = param.weight_loader
1591
1592
1593
1594
1595

                if envs.USE_FUSED_RMS_QUANT and envs.VLLM_USE_FUSED_QA_KVA_GEMM and (("q_a_proj"  in old_weight_name) or ("kv_a_proj_with_mqa" in old_weight_name)):
                    weight_loader(param, loaded_weight, old_weight_name)
                else:
                    weight_loader(param, loaded_weight, shard_id)
wangding zeng's avatar
wangding zeng committed
1596
1597
                break
            else:
1598
                is_expert_weight = False
1599
1600
1601
1602
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
1603

1604
1605
1606
1607
1608
1609
1610
1611
1612
                    # 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):
1613
1614
                        continue

1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
                    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:
1628
                        name = name_mapped
1629
                        break
1630
                else:
1631
1632
1633
1634
1635
                    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
1636

1637
1638
1639
1640
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue

1641
1642
1643
1644
1645
                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

1646
1647
1648
                    if is_pp_missing_parameter(name, self):
                        continue

zhuwenwen's avatar
zhuwenwen committed
1649
1650
1651
1652
                    try:
                        param = params_dict[name]
                    except Exception as e:
                        continue
1653
1654
1655
                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)
1656
            loaded_params.add(name)
1657

1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
        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)
1675

1676
1677
1678
1679
1680
1681
            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
1682

1683
1684
                    ops.trans_w16_gemm(_weight, weight.data, _weight.shape[0], _weight.shape[1])
                    weight.data.copy_(_weight)
1685

1686
                    weight.data=weight.data.reshape(ori_shape[1],-1)
1687

1688
        return loaded_params
1689
1690
1691
1692


class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704


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