deepseek_v2.py 76.5 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
from vllm import envs
zhuwenwen's avatar
zhuwenwen committed
29

30
31
import typing
from collections.abc import Callable, Iterable
32
from itertools import islice
wangding zeng's avatar
wangding zeng committed
33
34
35

import torch
from torch import nn
36
from transformers import DeepseekV2Config, DeepseekV3Config
wangding zeng's avatar
wangding zeng committed
37

38
from vllm._aiter_ops import rocm_aiter_ops
39
from vllm.attention.layer import Attention
40
from vllm.compilation.decorators import support_torch_compile
41
42
43
44
45
46
47
48
from vllm.config import CacheConfig, ParallelConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (
    get_ep_group,
    get_pp_group,
    get_tensor_model_parallel_rank,
    get_tensor_model_parallel_world_size,
    tensor_model_parallel_all_gather,
)
49
from vllm.forward_context import get_forward_context
王敏's avatar
王敏 committed
50
from vllm.distributed.communication_op import tensor_model_parallel_all_reduce, tensor_model_parallel_reduce_scatter
51
from vllm.logger import init_logger
wangding zeng's avatar
wangding zeng committed
52
from vllm.model_executor.layers.activation import SiluAndMul
53
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
54
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
55
from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm
56
57
58
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    MergedColumnParallelLinear,
59
    QKVParallelLinear,
60
61
62
    ReplicatedLinear,
    RowParallelLinear,
)
wangding zeng's avatar
wangding zeng committed
63
from vllm.model_executor.layers.logits_processor import LogitsProcessor
64
from vllm.model_executor.layers.mla import MLAModules, MultiHeadLatentAttentionWrapper
65
from vllm.model_executor.layers.quantization import QuantizationConfig
66
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
67
68
    per_token_group_quant_fp8,
)
wangding zeng's avatar
wangding zeng committed
69
from vllm.model_executor.layers.rotary_embedding import get_rope
70
from vllm.model_executor.layers.sparse_attn_indexer import SparseAttnIndexer
wangding zeng's avatar
wangding zeng committed
71
from vllm.model_executor.layers.vocab_parallel_embedding import (
72
73
74
    ParallelLMHead,
    VocabParallelEmbedding,
)
75
from vllm.model_executor.model_loader.weight_utils import (
76
77
78
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
79
from vllm.model_executor.models.utils import sequence_parallel_chunk
80
from vllm.platforms import current_platform
81
from vllm.sequence import IntermediateTensors
82
from vllm.v1.attention.backend import AttentionBackend
83
84
85
from vllm.v1.attention.backends.mla.indexer import (
    DeepseekV32IndexerBackend,
)
86
from vllm.v1.kv_cache_interface import KVCacheSpec, MLAAttentionSpec
wangding zeng's avatar
wangding zeng committed
87

88
from .interfaces import MixtureOfExperts, SupportsEagle, SupportsLoRA, SupportsPP
89
90
91
92
93
94
95
from .utils import (
    PPMissingLayer,
    is_pp_missing_parameter,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
96
from vllm import _custom_ops as ops
zhuwenwen's avatar
zhuwenwen committed
97
from vllm.utils import W8a8GetCacheJSON
98

99
100
from vllm.model_executor.layers.layernorm import FusedRMSNormQuant

101
102
logger = init_logger(__name__)

wangding zeng's avatar
wangding zeng committed
103

104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
class DeepseekAttention(nn.Module):
    """Normal MHA implementation used by Deepseek v1."""

    def __init__(
        self,
        vllm_config: VllmConfig,
        config: DeepseekV2Config | DeepseekV3Config,
        hidden_size: int,
        num_heads: int,
        max_position_embeddings: int = 8192,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
        **kwargs,
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = config.num_key_value_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = hidden_size // self.total_num_heads
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.max_position_embeddings = max_position_embeddings

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
        )

        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
        )

        self.rotary_emb = get_rope(
            self.head_dim,
            max_position=max_position_embeddings,
160
            rope_parameters=config.rope_parameters,
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
        )
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
176
        *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
177
178
179
180
181
182
183
184
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

wangding zeng's avatar
wangding zeng committed
185

186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
def eff_2d_iqis_all_gather(
    iqis: tuple[torch.Tensor, torch.Tensor],
    tp_size: int | None = None,
    tp_rank: int | None = None
) -> tuple[torch.Tensor, torch.Tensor]:
    assert iqis is not None
    iq_tensor, is_tensor = iqis
    assert isinstance(iq_tensor, torch.Tensor)
    assert isinstance(is_tensor, torch.Tensor)
    assert iq_tensor.dtype == torch.int8, f"iq_tensor dtype is {iq_tensor.dtype}"
    assert is_tensor.dtype == torch.float32, f"is_tensor dtype is {is_tensor.dtype}"
    assert iq_tensor.dim() == 2
    assert is_tensor.dim() == 2
    
    m_local, n = iq_tensor.shape
    assert is_tensor.shape[0] == m_local, f"{is_tensor.shape[0]} != {iq_tensor.shape[0]}"
    assert is_tensor.shape[1] == 1, f"is_tensor dim 1 ={is_tensor.shape[1]}"
    
    iq_int8_2d = iq_tensor.view(torch.int8)
    is_int8_2d = is_tensor.view(torch.int8)
    combined_2d = torch.cat([iq_int8_2d, is_int8_2d], dim=1) # [m_local, n + 4]
    
    if not combined_2d.is_contiguous():
        combined_2d = combined_2d.contiguous()
    
    combined_gathered = tensor_model_parallel_all_gather(combined_2d, dim=0)
    split_idx = n
    iq_gathered_int8 = combined_gathered[:, :split_idx].contiguous()
    is_gathered_int8 = combined_gathered[:, split_idx:].contiguous()
    
    iq_gathered = iq_gathered_int8.view(torch.int8)
    assert iq_gathered.shape[0] == m_local * tp_size, f"iq_gathered dim0= {iq_gathered.shape[0]}, expected {m_local * tp_size}"
    # is_gathered_int8 should be [m_local*tp_size, 4]
    assert is_gathered_int8.shape[0] == m_local * tp_size, f"is_gathered_int8 dim0= {is_gathered_int8.shape[0]}, expected {m_local * tp_size}"
    assert is_gathered_int8.shape[1] == 4, f"is_gathered_int8 dim1= {is_gathered_int8.shape[1]}"
    is_gathered = is_gathered_int8.view(torch.float32)
    return (iq_gathered, is_gathered)

wangding zeng's avatar
wangding zeng committed
224
225
226
227
228
229
class DeepseekV2MLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
230
        quant_config: QuantizationConfig | None = None,
wangding zeng's avatar
wangding zeng committed
231
        reduce_results: bool = True,
232
        is_sequence_parallel=False,
233
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
234
235
    ) -> None:
        super().__init__()
236
237
238
239
240

        # If is_sequence_parallel, the input and output tensors are sharded
        # across the ranks within the tp_group. In this case the weights are
        # replicated and no collective ops are needed.
        # Otherwise we use standard TP with an allreduce at the end.
wangding zeng's avatar
wangding zeng committed
241
        self.gate_up_proj = MergedColumnParallelLinear(
242
243
            hidden_size,
            [intermediate_size] * 2,
wangding zeng's avatar
wangding zeng committed
244
            bias=False,
245
            quant_config=quant_config,
246
            disable_tp=is_sequence_parallel,
247
248
249
250
251
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
wangding zeng's avatar
wangding zeng committed
252
            bias=False,
253
            quant_config=quant_config,
王敏's avatar
王敏 committed
254
255
            #reduce_results=reduce_results,
            reduce_results=False,
256
            disable_tp=is_sequence_parallel,
257
258
            prefix=f"{prefix}.down_proj",
        )
王敏's avatar
王敏 committed
259
260
261
262
263
264
265
266
267
268
269
        self.tp_size = get_tensor_model_parallel_world_size()
        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,
                *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
                ):
270
        enable_mla_cp = get_forward_context().enable_mla_cp #envs.VLLM_MLA_CP# and not get_forward_context().draft_model
271
272
        if enable_mla_cp: 
            if iqis is not None and iqis[0] is not None and iqis[1] is not None:
273
274
275
276
277
278
                if False:
                    i_q_gahter = tensor_model_parallel_all_gather(iqis[0].contiguous(), 0)
                    i_s_gather = tensor_model_parallel_all_gather(iqis[1].contiguous(), 0)
                    iqis = (i_q_gahter, i_s_gather)
                else:
                    iqis = eff_2d_iqis_all_gather(iqis, tp_size=self.tp_size, tp_rank=get_tensor_model_parallel_rank())
279
            else:
280
                x = tensor_model_parallel_all_gather(x.contiguous(), 0)
281

王敏's avatar
王敏 committed
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
        if envs.USE_FUSED_RMS_QUANT:
            gate_up, _ = self.gate_up_proj(x, iqis=iqis)
            if envs.USE_FUSED_SILU_MUL_QUANT:
                from lmslim.quantize.quant_ops import lm_fuse_silu_mul_quant
                xq, xs = lm_fuse_silu_mul_quant(gate_up)
                x, _ = self.down_proj(gate_up, iqis=(xq, xs))
            else:
                x = self.act_fn(gate_up)
                x, _ = self.down_proj(x)
        else:
            gate_up, _ = self.gate_up_proj(x)
            x = self.act_fn(gate_up)
            x, _ = self.down_proj(x)

        if enable_mla_cp:
            x = tensor_model_parallel_reduce_scatter(x.contiguous(), dim=0)
298
            return x
王敏's avatar
王敏 committed
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
        elif self.tp_size > 1:
            x = tensor_model_parallel_all_reduce(x)
        return x
    

class DeepseekV2SharedMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: QuantizationConfig | None = None,
        reduce_results: bool = True,
        is_sequence_parallel=False,
        prefix: str = "",
    ) -> None:
        super().__init__()

        # If is_sequence_parallel, the input and output tensors are sharded
        # across the ranks within the tp_group. In this case the weights are
        # replicated and no collective ops are needed.
        # Otherwise we use standard TP with an allreduce at the end.
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            disable_tp=is_sequence_parallel,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=reduce_results,
            disable_tp=is_sequence_parallel,
            prefix=f"{prefix}.down_proj"
        )
wangding zeng's avatar
wangding zeng committed
338
        if hidden_act != "silu":
339
340
341
            raise ValueError(
                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
            )
wangding zeng's avatar
wangding zeng committed
342
343
        self.act_fn = SiluAndMul()

344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
    def forward(self, 
                x,
                *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
                ):
        if envs.USE_FUSED_RMS_QUANT:
            gate_up, _ = self.gate_up_proj(x, iqis=iqis)
            if envs.USE_FUSED_SILU_MUL_QUANT:
                from lmslim.quantize.quant_ops import lm_fuse_silu_mul_quant
                xq, xs = lm_fuse_silu_mul_quant(gate_up)
                x, _ = self.down_proj(gate_up, iqis=(xq, xs))
            else:
                x = self.act_fn(gate_up)
                x, _ = self.down_proj(x)
        else:
            gate_up, _ = self.gate_up_proj(x)
            x = self.act_fn(gate_up)
            x, _ = self.down_proj(x)
361
        return x
wangding zeng's avatar
wangding zeng committed
362
363
364
365
366


class DeepseekV2MoE(nn.Module):
    def __init__(
        self,
367
        config: DeepseekV2Config | DeepseekV3Config,
368
        parallel_config: ParallelConfig,
369
        quant_config: QuantizationConfig | None = None,
370
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
371
372
373
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
374
375
        self.tp_rank = get_tensor_model_parallel_rank()

376
        self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
377
378

        self.ep_group = get_ep_group().device_group
379
        self.ep_rank = get_ep_group().rank_in_group
380
381
382
        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
383

384
        self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe
385

386
        if config.hidden_act != "silu":
387
388
389
390
391
392
393
394
395
396
397
398
            raise ValueError(
                f"Unsupported activation: {config.hidden_act}. "
                "Only silu is supported for now."
            )

        self.gate = ReplicatedLinear(
            config.hidden_size,
            config.n_routed_experts,
            bias=False,
            quant_config=None,
            prefix=f"{prefix}.gate",
        )
399
        if getattr(config, "topk_method", None) == "noaux_tc":
zhuwenwen's avatar
zhuwenwen committed
400
401
402
403
404
405
406
407
            if envs.VLLM_ENABLE_MOE_FUSED_GATE:
                # avoid moe_fused_gate precision error
                self.gate.e_score_correction_bias = nn.Parameter(
                torch.empty(config.n_routed_experts))
            else:
                self.gate.e_score_correction_bias = nn.Parameter(
                    torch.empty(config.n_routed_experts, dtype=torch.float32)
                )
408
409
410
        else:
            self.gate.e_score_correction_bias = None

411
        # Load balancing settings.
412
413
        eplb_config = parallel_config.eplb_config
        self.enable_eplb = parallel_config.enable_eplb
414

415
        self.n_redundant_experts = eplb_config.num_redundant_experts
416
        self.n_logical_experts = self.n_routed_experts
417
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
418
419
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

420
421
422
423
        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
        )
424

425
        self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
426
427
428
429
        self.is_fusion_moe_shared_experts_enabled = (
            rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
        )
        if config.n_shared_experts is None or self.is_fusion_moe_shared_experts_enabled:
430
431
            self.shared_experts = None
        else:
432
            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
433

王敏's avatar
王敏 committed
434
            self.shared_experts = DeepseekV2SharedMLP(
wangding zeng's avatar
wangding zeng committed
435
436
437
438
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
439
                is_sequence_parallel=self.is_sequence_parallel,
440
                reduce_results=False,
441
                prefix=f"{prefix}.shared_experts",
wangding zeng's avatar
wangding zeng committed
442
443
            )

444
445
        self.experts = SharedFusedMoE(
            shared_experts=self.shared_experts,
446
            gate=self.gate,
447
448
449
450
451
452
            # num_experts=config.n_routed_experts,
            # top_k=config.num_experts_per_tok,
            num_experts=config.n_routed_experts 
                    + (config.n_shared_experts if self.is_fusion_moe_shared_experts_enabled else 0),
            top_k = config.num_experts_per_tok
                    + (config.n_shared_experts if self.is_fusion_moe_shared_experts_enabled else 0),
453
454
455
456
457
458
            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,
459
460
            num_expert_group=getattr(config, "n_group", 1),
            topk_group=getattr(config, "topk_group", 1),
461
            prefix=f"{prefix}.experts",
462
            scoring_func=getattr(config, "scoring_func", "softmax"),
463
            # we do scaling outside, set factor to 1.0 to avoid double mul
464
465
            # aiter applies routed_scaling_factor internally
            routed_scaling_factor=1.0
466
            if not self.is_rocm_aiter_moe_enabled
467
            else self.routed_scaling_factor,
468
469
470
471
            e_score_correction_bias=self.gate.e_score_correction_bias,
            enable_eplb=self.enable_eplb,
            num_redundant_experts=self.n_redundant_experts,
            is_sequence_parallel=self.is_sequence_parallel,
472
            n_shared_experts=config.n_shared_experts
473
            if self.is_fusion_moe_shared_experts_enabled
474
            else None,
475
        )
476

477
478
479
    def forward(self, hidden_states: torch.Tensor,
                *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
                ) -> torch.Tensor:
480
        enable_mla_cp = get_forward_context().enable_mla_cp#envs.VLLM_MLA_CP #and not get_forward_context().draft_model
王敏's avatar
王敏 committed
481
482
483
484
        if enable_mla_cp:
            hidden_states = tensor_model_parallel_all_gather(
                hidden_states.contiguous(), 0
            )
wangding zeng's avatar
wangding zeng committed
485
486
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
487
488
489
490
491
492

        # Chunk the hidden states so they aren't replicated across TP ranks.
        # This avoids duplicate computation in self.experts.
        # TODO: We can replace the all_reduce at the end of attn with a
        # reduce_scatter instead of chunking here.
        if self.is_sequence_parallel:
493
            hidden_states = sequence_parallel_chunk(hidden_states)
494

495
496
497
498
499
500
501
502
503
504
505
        needs_post_moe_combine = (
            getattr(self.experts, "dp_size", 1) > 1
            or getattr(self.experts, "pcp_size", 1) > 1
        )

        if (
            envs.VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD
            and self.shared_experts is not None
            and not needs_post_moe_combine
        ):
            shared_output = self.shared_experts(hidden_states, iqis=iqis)
506
507
508
509
510
            if self.experts.is_internal_router:
                # In this case, the gate/router runs inside the FusedMoE class.
                router_logits = hidden_states
            else:
                router_logits, _ = self.gate(hidden_states)
511
512
513
514
            routed_scaling_factor = (
                1.0 if self.is_rocm_aiter_moe_enabled
                else self.routed_scaling_factor
            )
515
516
            # Keep shared-expert path intact and only fuse routed scale + add
            # in the downstream MoE kernel.
517
518
519
520
521
522
            _, final_hidden_states = self.experts(
                hidden_states=hidden_states,
                router_logits=router_logits,
                iqis=iqis,
                shared_output=shared_output,
                routed_scaling_factor=routed_scaling_factor,
523
            )
524
525
526
527
528
529
530
531
532
533
534
535
536
537
        else:
            if self.experts.is_internal_router:
                # In this case, the gate/router runs inside the FusedMoE class
                fused_moe_out = self.experts(
                    hidden_states=hidden_states,
                    router_logits=hidden_states,
                    iqis=iqis,
                )
            else:
                # router_logits: (num_tokens, n_experts)
                router_logits, _ = self.gate(hidden_states)
                fused_moe_out = self.experts(
                    hidden_states=hidden_states, router_logits=router_logits
                )
538

539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
            shared_output, final_hidden_states = fused_moe_out
            if self.shared_experts is None:
                assert shared_output is None

            # Fix FP16 overflow
            # See DeepseekV2DecoderLayer for more details.
            if hidden_states.dtype != torch.float16:
                if not self.is_rocm_aiter_moe_enabled:
                    final_hidden_states *= self.routed_scaling_factor
            elif self.shared_experts is not None:
                assert shared_output is not None
                shared_output *= 1.0 / self.routed_scaling_factor

            if self.shared_experts is not None:
                assert shared_output is not None
                final_hidden_states += shared_output
555

王敏's avatar
王敏 committed
556
557
558
559
560
561
        if enable_mla_cp:
            final_hidden_states = tensor_model_parallel_reduce_scatter(
                final_hidden_states.contiguous(), 0
            )
            return final_hidden_states
        elif self.is_sequence_parallel:
562
            final_hidden_states = tensor_model_parallel_all_gather(
563
564
                final_hidden_states, 0
            )
565
566
            final_hidden_states = final_hidden_states[:num_tokens]
        elif self.tp_size > 1:
567
568
569
            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
                final_hidden_states
            )
wangding zeng's avatar
wangding zeng committed
570
571
572
573
574
575

        return final_hidden_states.view(num_tokens, hidden_dim)


def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
    import math
576

wangding zeng's avatar
wangding zeng committed
577
578
579
580
581
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


582
583
584
585
586
587
588
589
590
def _get_llama_4_scaling(
    original_max_position_embeddings: int, scaling_beta: float, positions: torch.Tensor
) -> torch.Tensor:
    scaling = 1 + scaling_beta * torch.log(
        1 + torch.floor(positions / original_max_position_embeddings)
    )
    # Broadcast over num_heads and head_dim
    return scaling[..., None, None]

wangding zeng's avatar
wangding zeng committed
591
592
593
594

class DeepseekV2Attention(nn.Module):
    def __init__(
        self,
595
        vllm_config: VllmConfig,
596
        config: DeepseekV2Config | DeepseekV3Config,
wangding zeng's avatar
wangding zeng committed
597
598
599
600
601
602
603
604
        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,
        max_position_embeddings: int = 8192,
605
606
607
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        topk_indices_buffer: torch.Tensor | None = None,
608
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
    ) -> 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.max_position_embeddings = max_position_embeddings
624
625
        assert topk_indices_buffer is None, (
            "topk_indices_buffer is not \
626
        supported for DeepseekV2Attention"
627
        )
wangding zeng's avatar
wangding zeng committed
628
629

        if self.q_lora_rank is not None:
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
            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",
            )
wangding zeng's avatar
wangding zeng committed
645
        else:
646
647
648
649
650
651
652
            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",
            )
wangding zeng's avatar
wangding zeng committed
653

654
655
656
657
658
        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,
659
660
661
            prefix=f"{prefix}.kv_a_proj_with_mqa",
        )
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
wangding zeng's avatar
wangding zeng committed
662
663
664
665
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
666
            quant_config=quant_config,
667
668
            prefix=f"{prefix}.kv_b_proj",
        )
wangding zeng's avatar
wangding zeng committed
669
        # O projection.
670
671
672
673
674
675
676
        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",
        )
677
        if config.rope_parameters["rope_type"] != "default":
678
679
680
681
682
            config.rope_parameters["rope_type"] = (
                "deepseek_yarn"
                if config.rope_parameters.get("apply_yarn_scaling", True)
                else "deepseek_llama_scaling"
            )
683

684
685
686
        self.rotary_emb = get_rope(
            qk_rope_head_dim,
            max_position=max_position_embeddings,
687
            rope_parameters=config.rope_parameters,
688
689
            is_neox_style=False,
        )
wangding zeng's avatar
wangding zeng committed
690

691
692
693
694
        if (
            config.rope_parameters["rope_type"] != "default"
            and config.rope_parameters["rope_type"] == "deepseek_yarn"
        ):
695
696
            mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False)
            scaling_factor = config.rope_parameters["factor"]
wangding zeng's avatar
wangding zeng committed
697
698
699
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

700
701
702
703
704
705
706
707
708
        self.attn = Attention(
            self.num_local_heads,
            self.qk_head_dim,
            self.scaling,
            num_kv_heads=self.num_local_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )
wangding zeng's avatar
wangding zeng committed
709
710
711
712
713

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
714
        llama_4_scaling: torch.Tensor | None,
715
        *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
wangding zeng's avatar
wangding zeng committed
716
717
718
719
    ) -> torch.Tensor:
        if self.q_lora_rank is not None:
            q = self.q_a_proj(hidden_states)[0]
            q = self.q_a_layernorm(q)
720
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
wangding zeng's avatar
wangding zeng committed
721
        else:
722
723
724
725
            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)
wangding zeng's avatar
wangding zeng committed
726
        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
727
        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
wangding zeng's avatar
wangding zeng committed
728
        latent_cache = latent_cache.unsqueeze(1)
729
        kv_a = self.kv_a_layernorm(kv_a)
wangding zeng's avatar
wangding zeng committed
730
        kv = self.kv_b_proj(kv_a)[0]
731
        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
wangding zeng's avatar
wangding zeng committed
732
        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
733
        k_pe = latent_cache[:, :, self.kv_lora_rank :]
734

wangding zeng's avatar
wangding zeng committed
735
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
736

737
        q[..., self.qk_nope_head_dim :] = q_pe
wangding zeng's avatar
wangding zeng committed
738
        k = torch.empty_like(q)
739
740
        k[..., : self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim :] = k_pe
741
742
743
744
745

        # Apply llama 4 scaling if provided
        if llama_4_scaling is not None:
            q *= llama_4_scaling

746
747
        # padding value to qk_head_dim for alignment
        v = torch.nn.functional.pad(
748
749
            v, [0, self.qk_head_dim - self.v_head_dim], value=0
        ).view(-1, self.num_local_heads * self.qk_head_dim)
750
        attn_output = self.attn(q, k, v)
751
752
753
        attn_output = attn_output.view(-1, self.num_local_heads, self.qk_head_dim)[
            ..., : self.v_head_dim
        ].reshape(-1, self.num_local_heads * self.v_head_dim)
wangding zeng's avatar
wangding zeng committed
754
755
756
757
        output, _ = self.o_proj(attn_output)
        return output


758
class DeepseekV32IndexerCache(torch.nn.Module, AttentionLayerBase):
759
760
761
    def __init__(
        self, head_dim: int, dtype: torch.dtype, prefix: str, cache_config: CacheConfig
    ):
762
763
764
765
766
767
768
769
770
771
772
        super().__init__()
        self.kv_cache = [torch.tensor([])]
        self.head_dim = head_dim
        self.prefix = prefix
        self.cache_config = cache_config
        self.dtype = dtype
        compilation_config = get_current_vllm_config().compilation_config
        if prefix in compilation_config.static_forward_context:
            raise ValueError(f"Duplicate layer name: {prefix}")
        compilation_config.static_forward_context[prefix] = self

773
    def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
774
775
776
777
778
779
780
        return MLAAttentionSpec(  # Only has one vector instead of K + V
            block_size=self.cache_config.block_size,
            num_kv_heads=1,
            head_size=self.head_dim,
            dtype=self.dtype,
        )

781
    def forward(self): ...
782
783
784
785
786
787

    def get_attn_backend(self) -> AttentionBackend:
        return DeepseekV32IndexerBackend


class Indexer(nn.Module):
788
789
790
    def __init__(
        self,
        vllm_config: VllmConfig,
791
        config: DeepseekV2Config | DeepseekV3Config,
792
793
        hidden_size: int,
        q_lora_rank: int,
794
795
796
        quant_config: QuantizationConfig | None,
        cache_config: CacheConfig | None,
        topk_indices_buffer: torch.Tensor | None,
797
798
        prefix: str = "",
    ):
799
800
801
802
803
804
805
806
807
808
        super().__init__()
        self.vllm_config = vllm_config
        self.config = config
        # self.indexer_cfg = config.attn_module_list_cfg[0]["attn_index"]
        self.topk_tokens = config.index_topk
        self.n_head = config.index_n_heads  # 64
        self.head_dim = config.index_head_dim  # 128
        self.rope_dim = config.qk_rope_head_dim  # 64
        self.q_lora_rank = q_lora_rank  # 1536
        # no tensor parallel, just replicated
809
810
811
812
813
814
815
816
817
818
819
820
821
822
        self.wq_b = ReplicatedLinear(
            self.q_lora_rank,
            self.head_dim * self.n_head,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.wq_b",
        )
        self.wk = ReplicatedLinear(
            hidden_size,
            self.head_dim,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.wk",
        )
823
        self.k_norm = LayerNorm(self.head_dim, eps=1e-6)
824
        self.weights_proj = ReplicatedLinear(
825
826
827
828
829
            hidden_size,
            self.n_head,
            bias=False,
            quant_config=None,
            prefix=f"{prefix}.weights_proj",
830
        )
831
832
833
834
835
836
837
838
839
840
        self.softmax_scale = self.head_dim**-0.5

        self.scale_fmt = "ue8m0"
        self.quant_block_size = 128  # TODO: get from config
        self.topk_indices_buffer = topk_indices_buffer

        # NOTE: (zyongye) we use fp8 naive cache,
        #       where we store value in fp8 and scale in fp32
        #       per self.quant_block_size element
        self.k_cache = DeepseekV32IndexerCache(
841
842
            head_dim=self.head_dim + self.head_dim // self.quant_block_size * 4 if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" else self.head_dim,
            dtype=torch.uint8  if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" else torch.bfloat16,
843
            prefix=f"{prefix}.k_cache",
844
845
            cache_config=cache_config,
        )
846
847
        self.max_model_len = vllm_config.model_config.max_model_len
        self.prefix = prefix
848
849
        from vllm.v1.attention.backends.mla.indexer import get_max_prefill_buffer_size

850
        self.max_total_seq_len = get_max_prefill_buffer_size(vllm_config)
851
852
853
854
855
856
857
858
859
860
        self.indexer_op = SparseAttnIndexer(
            self.k_cache,
            self.quant_block_size,
            self.scale_fmt,
            self.topk_tokens,
            self.head_dim,
            self.max_model_len,
            self.max_total_seq_len,
            self.topk_indices_buffer,
        )
861

862
    def forward(
863
864
        self, hidden_states: torch.Tensor, qr: torch.Tensor, positions, rotary_emb,
        *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
865
    ) -> torch.Tensor:
866
867
868
        q, _ = self.wq_b(qr)
        q = q.view(-1, self.n_head, self.head_dim)
        q_pe, q_nope = torch.split(
869
870
            q, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1
        )
871
872
873
874
        if envs.USE_FUSED_RMS_QUANT and self.wk.weight.dtype == torch.int8 and iqis is not None:
            k, _ = self.wk(hidden_states, iqis=iqis)
        else:
            k, _ = self.wk(hidden_states)
875
876
        k = self.k_norm(k)
        k_pe, k_nope = torch.split(
877
878
            k, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1
        )
879
880

        q_pe, k_pe = rotary_emb(positions, q_pe, k_pe.unsqueeze(1))
881
882
883
884
885
        # Note: RoPE (NeoX) can introduce extra leading dimensions during compilation
        # so we need to reshape back to token-flattened shapes
        q_pe = q_pe.reshape(-1, self.n_head, self.rope_dim)
        k_pe = k_pe.reshape(-1, 1, self.rope_dim)

886
887
        # `rotary_emb` is shape-preserving; `q_pe` is already
        # [num_tokens, n_head, rope_dim].
888
889
        q = torch.cat([q_pe, q_nope], dim=-1)
        # `k_pe` is [num_tokens, 1, rope_dim] (MQA).
890
        k = torch.cat([k_pe.squeeze(-2), k_nope], dim=-1)
891

892
        enable_mla_cp = get_forward_context().enable_mla_cp#envs.VLLM_MLA_CP # and not get_forward_context().draft_model
王敏's avatar
王敏 committed
893
894
895
896
        if enable_mla_cp:
            k = tensor_model_parallel_all_gather(
                k.contiguous(), 0
            )
897
898
899
            gather_indexes_tensor = get_forward_context().gather_indexes_tensor
            if envs.VLLM_MLA_CPLB and gather_indexes_tensor is not None:
                k = torch.index_select(k, 0, gather_indexes_tensor)                
王敏's avatar
王敏 committed
900

901
        # we only quant q here since k quant is fused with cache insertion
902
903
904
905
906
907
908
909
910
911
        if not current_platform.is_rocm() or torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938":
            q = q.view(-1, self.head_dim)
            q_fp8, q_scale = per_token_group_quant_fp8(
                q,
                self.quant_block_size,
                column_major_scales=False,
                use_ue8m0=self.scale_fmt is not None,
            )
            q_fp8 = q_fp8.view(-1, self.n_head, self.head_dim)
            q_scale = q_scale.view(-1, self.n_head, 1)
zhuwenwen's avatar
zhuwenwen committed
912
913
        else:
            q_fp8 = q
914

915
916
917
918
        if envs.USE_FUSED_RMS_QUANT and self.weights_proj.weight.dtype == torch.int8 and iqis is not None:
            weights, _ = self.weights_proj(hidden_states, iqis=iqis)
        else:
            weights, _ = self.weights_proj(hidden_states)
919
920
921
922
923
        if not current_platform.is_rocm() or torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938":
            weights = (
                weights.unsqueeze(-1) * q_scale * self.softmax_scale * self.n_head**-0.5
            )
            weights = weights.squeeze(-1)
924

925
        return self.indexer_op(hidden_states, q_fp8, k, weights)
926
927


928
929
930
931
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).
932

933
934
        For more info see MLACommonImpl in:
        vllm/v1/attention/backends/mla/utils.py
935
936
937
938
    """

    def __init__(
        self,
939
        vllm_config: VllmConfig,
940
        config: DeepseekV2Config | DeepseekV3Config,
941
942
943
944
945
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
946
        q_lora_rank: int | None,
947
948
        kv_lora_rank: int,
        max_position_embeddings: int = 8192,
949
950
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
951
        prefix: str = "",
952
        topk_indices_buffer: torch.Tensor | None = None,
953
954
955
956
957
958
959
960
961
962
963
964
965
966
    ) -> 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
王敏's avatar
王敏 committed
967
968
        #self.num_local_heads = num_heads // tp_size
        self.num_local_heads = num_heads // tp_size if not envs.VLLM_MLA_CP else self.num_heads
969
970
971
972
973

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

        if self.q_lora_rank is not None:
974
            self.fused_qkv_a_proj = MergedColumnParallelLinear(
975
976
977
978
                self.hidden_size,
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim],
                bias=False,
                quant_config=quant_config,
979
                prefix=f"{prefix}.fused_qkv_a_proj",
980
981
                disable_tp=True,
            )
982
983
984
985
986
987
        else:
            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,
988
989
                prefix=f"{prefix}.kv_a_proj_with_mqa",
            )
990
991

        if self.q_lora_rank is not None:
wujl5's avatar
wujl5 committed
992
993
994
995
            if envs.USE_FUSED_RMS_QUANT:
                self.q_a_layernorm = FusedRMSNormQuant(self.q_lora_rank, eps=config.rms_norm_eps)
            else:
                self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
996
997
998
999
1000
1001
            self.q_b_proj = ColumnParallelLinear(
                self.q_lora_rank,
                self.num_heads * self.qk_head_dim,
                bias=False,
                quant_config=quant_config,
                prefix=f"{prefix}.q_b_proj",
王敏's avatar
王敏 committed
1002
                disable_tp=envs.VLLM_MLA_CP,
1003
            )
1004
        else:
1005
1006
1007
1008
1009
1010
            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",
王敏's avatar
王敏 committed
1011
                disable_tp=envs.VLLM_MLA_CP,
1012
1013
            )
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
1014
1015
1016
1017
1018
        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,
1019
            prefix=f"{prefix}.kv_b_proj",
王敏's avatar
王敏 committed
1020
            disable_tp=envs.VLLM_MLA_CP,
1021
1022
1023
1024
1025
1026
1027
        )
        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",
王敏's avatar
王敏 committed
1028
            disable_tp=envs.VLLM_MLA_CP,
1029
        )
1030

1031
        if config.rope_parameters["rope_type"] != "default":
1032
1033
1034
1035
1036
1037
            config.rope_parameters["rope_type"] = (
                "deepseek_yarn"
                if config.rope_parameters.get("apply_yarn_scaling", True)
                else "deepseek_llama_scaling"
            )

1038
1039
1040
        self.rotary_emb = get_rope(
            qk_rope_head_dim,
            max_position=max_position_embeddings,
1041
            rope_parameters=config.rope_parameters,
1042
1043
            is_neox_style=False,
        )
1044
1045
1046
1047
1048

        if (
            config.rope_parameters["rope_type"] != "default"
            and config.rope_parameters["rope_type"] == "deepseek_yarn"
        ):
1049
1050
            mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False)
            scaling_factor = config.rope_parameters["factor"]
1051
1052
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale
1053
        #添加判断,默认开启DSA
1054
        force_disable_dsa = envs.VLLM_DISABLE_DSA
1055
        self.is_v32 = hasattr(config, "index_topk") and not force_disable_dsa
1056
1057

        if self.is_v32:
1058
1059
1060
            self.indexer_rope_emb = get_rope(
                qk_rope_head_dim,
                max_position=max_position_embeddings,
1061
                rope_parameters=config.rope_parameters,
zhuwenwen's avatar
zhuwenwen committed
1062
                is_neox_style=not getattr(config, "indexer_rope_interleave", True),
1063
            )
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
            self.indexer = Indexer(
                vllm_config,
                config,
                hidden_size,
                q_lora_rank,
                quant_config,
                cache_config,
                topk_indices_buffer,
                f"{prefix}.indexer",
            )
1074
        else:
1075
            self.indexer_rope_emb = None
1076
1077
            self.indexer = None

1078
1079
        mla_modules = MLAModules(
            kv_a_layernorm=self.kv_a_layernorm,
1080
            kv_b_proj=self.kv_b_proj,
1081
1082
1083
            rotary_emb=self.rotary_emb,
            o_proj=self.o_proj,
            fused_qkv_a_proj=self.fused_qkv_a_proj
1084
1085
            if self.q_lora_rank is not None
            else None,
1086
            kv_a_proj_with_mqa=self.kv_a_proj_with_mqa
1087
1088
1089
            if self.q_lora_rank is None
            else None,
            q_a_layernorm=self.q_a_layernorm if self.q_lora_rank is not None else None,
1090
1091
            q_b_proj=self.q_b_proj if self.q_lora_rank is not None else None,
            q_proj=self.q_proj if self.q_lora_rank is None else None,
1092
            indexer=self.indexer,
1093
            indexer_rotary_emb=self.indexer_rope_emb,
1094
1095
            is_sparse=self.is_v32,
            topk_indices_buffer=topk_indices_buffer,
1096
        )
1097

1098
        self.mla_attn = MultiHeadLatentAttentionWrapper(
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
            self.hidden_size,
            self.num_local_heads,
            self.scaling,
            self.qk_nope_head_dim,
            self.qk_rope_head_dim,
            self.v_head_dim,
            self.q_lora_rank,
            self.kv_lora_rank,
            mla_modules,
            cache_config,
            quant_config,
            prefix,
1111
1112
1113
1114
1115
1116
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1117
        llama_4_scaling: torch.Tensor | None,
1118
        *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
1119
    ) -> torch.Tensor:
1120
        return self.mla_attn(positions, hidden_states, llama_4_scaling, iqis=iqis)
1121
1122


wangding zeng's avatar
wangding zeng committed
1123
class DeepseekV2DecoderLayer(nn.Module):
1124
1125
1126
1127
    def __init__(
        self,
        vllm_config: VllmConfig,
        prefix: str,
1128
1129
        config: DeepseekV2Config | None = None,
        topk_indices_buffer: torch.Tensor | None = None,
1130
    ) -> None:
wangding zeng's avatar
wangding zeng committed
1131
        super().__init__()
1132

1133
1134
        if config is None:
            config = vllm_config.model_config.hf_config
1135
1136
1137
1138
1139
        model_config = vllm_config.model_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        parallel_config = vllm_config.parallel_config

wangding zeng's avatar
wangding zeng committed
1140
        self.hidden_size = config.hidden_size
1141
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
1142
        moe_layer_freq = getattr(config, "moe_layer_freq", 1)
1143
1144
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
1145
        layer_idx = int(prefix.split(sep=".")[-1])
1146
        self.layer_idx = layer_idx
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156

        # verify MLA attention specific fields
        qk_nope_head_dim = getattr(config, "qk_nope_head_dim", 0)
        qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 0)
        v_head_dim = getattr(config, "v_head_dim", 0)
        kv_lora_rank = getattr(config, "kv_lora_rank", 0)
        use_mha = config.model_type == "deepseek" or all(
            dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim)
        )

1157
1158
        self.use_mha = use_mha

1159
1160
1161
        if use_mha:
            attn_cls = DeepseekAttention
        elif model_config.use_mla:
1162
1163
1164
1165
            attn_cls = DeepseekV2MLAAttention
        else:
            attn_cls = DeepseekV2Attention
        self.self_attn = attn_cls(
1166
            vllm_config=vllm_config,
wangding zeng's avatar
wangding zeng committed
1167
1168
1169
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
1170
1171
1172
            qk_nope_head_dim=qk_nope_head_dim,
            qk_rope_head_dim=qk_rope_head_dim,
            v_head_dim=v_head_dim,
1173
            q_lora_rank=config.q_lora_rank if hasattr(config, "q_lora_rank") else None,
1174
            kv_lora_rank=kv_lora_rank,
wangding zeng's avatar
wangding zeng committed
1175
1176
1177
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
1178
            prefix=f"{prefix}.self_attn",
1179
            topk_indices_buffer=topk_indices_buffer,
wangding zeng's avatar
wangding zeng committed
1180
        )
1181

1182
1183
1184
        if (
            config.n_routed_experts is not None
            and layer_idx >= config.first_k_dense_replace
1185
            and layer_idx % moe_layer_freq == 0
1186
        ):
1187
1188
            self.mlp = DeepseekV2MoE(
                config=config,
1189
                parallel_config=parallel_config,
1190
1191
1192
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )
wangding zeng's avatar
wangding zeng committed
1193
1194
1195
1196
1197
1198
        else:
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
1199
                prefix=f"{prefix}.mlp",
wangding zeng's avatar
wangding zeng committed
1200
            )
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
        
        if not envs.USE_FUSED_RMS_QUANT:
            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
            )
        else:
            self.input_layernorm = FusedRMSNormQuant(config.hidden_size, eps=config.rms_norm_eps)
            self.post_attention_layernorm = FusedRMSNormQuant(
                config.hidden_size, eps=config.rms_norm_eps
            )

1213
        self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
wangding zeng's avatar
wangding zeng committed
1214

1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
    def forward_RQ(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
        llama_4_scaling: torch.Tensor | None = None,
    ) -> torch.Tensor:
        # Self Attention
        # Fix residual FP16 overflow
        residual_fix_overflow = False
        assert self.input_layernorm.has_weight is True
1226
1227
        # DSA should set update_input True
        _dsa_flag = hasattr(self.self_attn, "indexer") and self.self_attn.indexer is not None
1228
1229
1230
1231
1232
        if residual is None:
            residual = hidden_states.clone()
            i_q, i_s, _ = self.input_layernorm(x=hidden_states, 
                                               residual=None, 
                                               quant_dtype=torch.int8,
1233
                                               update_input=_dsa_flag
1234
1235
1236
1237
1238
1239
                                               )
            residual_fix_overflow = True
        else:
            i_q, i_s, residual = self.input_layernorm(x=hidden_states, 
                                                      residual=residual, 
                                                      quant_dtype=torch.int8,
1240
                                                      update_input=_dsa_flag
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
                                                      )
        attn_kwargs = {
            "positions": positions,
            "hidden_states": hidden_states,
            "iqis": (i_q, i_s)
        }
        if not self.use_mha:
            attn_kwargs["llama_4_scaling"] = llama_4_scaling
        hidden_states = self.self_attn(**attn_kwargs)

        if (
            not isinstance(self.self_attn, DeepseekAttention)
            and 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 *= 1.0 / self.routed_scaling_factor
            if self.layer_idx == 0:
                # The residual is shared by all layers, we only scale it on
                # first layer.
                residual *= 1.0 / self.routed_scaling_factor

        # Fully Connected
1265
1266
        enable_mla_cp = get_forward_context().enable_mla_cp
        skip_moe_large_batch_size = enable_mla_cp
1267
1268
1269
1270
1271
1272
1273
1274
        update_hs = True if isinstance(self.mlp, DeepseekV2MoE) else False
        assert self.post_attention_layernorm.has_weight is True
        _i_q, _i_s, residual = self.post_attention_layernorm(x=hidden_states, 
                                                             residual=residual, 
                                                             quant_dtype=torch.int8,
                                                             update_input=update_hs
                                                             )
        new_resi = residual
1275
        if skip_moe_large_batch_size and isinstance(self.mlp, DeepseekV2MoE):
1276
1277
1278
            hidden_states = self.mlp(hidden_states)
        else:
            hidden_states = self.mlp(hidden_states, iqis=(_i_q, _i_s))
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290

        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 *= 1.0 / self.routed_scaling_factor

        return hidden_states, new_resi

    def forward_default(
wangding zeng's avatar
wangding zeng committed
1291
1292
1293
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1294
        residual: torch.Tensor | None,
1295
        llama_4_scaling: torch.Tensor | None = None,
wangding zeng's avatar
wangding zeng committed
1296
    ) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
1297
1298
1299
1300
1301
1302
1303
        # Self Attention
        # Fix residual FP16 overflow
        residual_fix_overflow = False
        if residual is None:
            residual = hidden_states.clone()
            hidden_states = self.input_layernorm(hidden_states)
            residual_fix_overflow = True
wangding zeng's avatar
wangding zeng committed
1304
        else:
zhuwenwen's avatar
zhuwenwen committed
1305
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
1306
            
1307

1308
1309
1310
1311
1312
1313
1314
        attn_kwargs = {
            "positions": positions,
            "hidden_states": hidden_states,
        }
        if not self.use_mha:
            attn_kwargs["llama_4_scaling"] = llama_4_scaling
        hidden_states = self.self_attn(**attn_kwargs)
wangding zeng's avatar
wangding zeng committed
1315

1316
1317
1318
1319
        if (
            not isinstance(self.self_attn, DeepseekAttention)
            and hidden_states.dtype == torch.float16
        ):
1320
1321
1322
            # Fix FP16 overflow
            # We scale both hidden_states and residual before
            # rmsnorm, and rmsnorm result would not affect by scale.
1323
            hidden_states *= 1.0 / self.routed_scaling_factor
1324
1325
1326
            if self.layer_idx == 0:
                # The residual is shared by all layers, we only scale it on
                # first layer.
1327
                residual *= 1.0 / self.routed_scaling_factor
1328
1329

        # Fully Connected
1330
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
wangding zeng's avatar
wangding zeng committed
1331
        hidden_states = self.mlp(hidden_states)
1332

1333
        if isinstance(self.mlp, DeepseekV2MLP) and hidden_states.dtype == torch.float16:
1334
1335
1336
1337
1338
            # 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
1339
            hidden_states *= 1.0 / self.routed_scaling_factor
1340

wangding zeng's avatar
wangding zeng committed
1341
1342
        return hidden_states, residual

1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
    def choose_forward(self):
        if envs.USE_FUSED_RMS_QUANT:
            return self.forward_RQ
        else:
            return self.forward_default

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
        llama_4_scaling: torch.Tensor | None = None,
    ) -> torch.Tensor:
        forward_func = self.choose_forward()
        return forward_func(positions=positions,
                            hidden_states=hidden_states,
                            residual=residual,
                            llama_4_scaling=llama_4_scaling)

wangding zeng's avatar
wangding zeng committed
1362

1363
@support_torch_compile
wangding zeng's avatar
wangding zeng committed
1364
1365
1366
class DeepseekV2Model(nn.Module):
    fall_back_to_pt_during_load = False

1367
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
1368
        super().__init__()
1369
1370
1371

        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
1372
        self.config = config
1373
        self.device = current_platform.device_type
1374

王敏's avatar
王敏 committed
1375
1376
1377
        self.tp_size = get_tensor_model_parallel_world_size()
        self.tp_rank = get_tensor_model_parallel_rank()

wangding zeng's avatar
wangding zeng committed
1378
        self.vocab_size = config.vocab_size
1379
        #添加判断,默认开启DSA
1380
        force_disable_dsa = envs.VLLM_DISABLE_DSA
1381
1382
        self.is_v32 = hasattr(config, "index_topk") and not force_disable_dsa        

1383
1384
1385
1386
1387
1388
        if self.is_v32:
            topk_tokens = config.index_topk
            topk_indices_buffer = torch.empty(
                vllm_config.scheduler_config.max_num_batched_tokens,
                topk_tokens,
                dtype=torch.int32,
1389
                device=self.device,
1390
            )
1391
1392
        else:
            topk_indices_buffer = None
wangding zeng's avatar
wangding zeng committed
1393

1394
1395
1396
1397
        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
1398
                quant_config=quant_config,
1399
1400
                prefix=f"{prefix}.embed_tokens",
            )
1401
1402
1403
1404
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
1405
            lambda prefix: DeepseekV2DecoderLayer(
1406
                vllm_config, prefix, topk_indices_buffer=topk_indices_buffer
1407
1408
1409
            ),
            prefix=f"{prefix}.layers",
        )
1410
1411
1412
1413
1414

        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
1415
1416
1417
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
wangding zeng's avatar
wangding zeng committed
1418

1419
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
1420
1421
        return self.embed_tokens(input_ids)

wangding zeng's avatar
wangding zeng committed
1422
1423
    def forward(
        self,
zhuwenwen's avatar
zhuwenwen committed
1424
        input_ids: torch.Tensor,
wangding zeng's avatar
wangding zeng committed
1425
        positions: torch.Tensor,
1426
1427
1428
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
1429
        if get_pp_group().is_first_rank:
1430
1431
1432
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
1433
                hidden_states = self.embed_input_ids(input_ids)
1434
1435
1436
1437
1438
1439
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

1440
        enable_mla_cp = get_forward_context().enable_mla_cp#envs.VLLM_MLA_CP # and not get_forward_context().draft_model
王敏's avatar
王敏 committed
1441
        if enable_mla_cp:
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
            scatter_indexes_tensor = get_forward_context().scatter_indexes_tensor
            if scatter_indexes_tensor is None:
                hidden_states_per_rank = torch.chunk(hidden_states, chunks=self.tp_size, dim=0)
                hidden_states = hidden_states_per_rank[self.tp_rank].contiguous()

                if residual is not None:
                    residual_per_rank = torch.chunk(residual, chunks=self.tp_size, dim=0)
                    residual = residual_per_rank[self.tp_rank].contiguous()

                if positions is not None:
                    positions_per_rank = torch.chunk(positions, chunks=self.tp_size, dim=0)
                    positions = positions_per_rank[self.tp_rank].contiguous()
            else:
                scatter_indexes_tensor = torch.where(scatter_indexes_tensor == -1, 0, scatter_indexes_tensor)
                hidden_states = torch.index_select(hidden_states, 0, scatter_indexes_tensor)
王敏's avatar
王敏 committed
1457

1458
1459
                if residual is not None:
                    residual = torch.index_select(residual, 0, scatter_indexes_tensor)
王敏's avatar
王敏 committed
1460

1461
1462
                if positions is not None:
                    positions = torch.index_select(positions, 0, scatter_indexes_tensor)
王敏's avatar
王敏 committed
1463

1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
        # Compute llama 4 scaling once per forward pass if enabled
        llama_4_scaling_config = getattr(self.config, "llama_4_scaling", None)
        llama_4_scaling: torch.Tensor | None
        if llama_4_scaling_config is not None:
            llama_4_scaling = _get_llama_4_scaling(
                original_max_position_embeddings=llama_4_scaling_config[
                    "original_max_position_embeddings"
                ],
                scaling_beta=llama_4_scaling_config["beta"],
                positions=positions,
            )
        else:
            llama_4_scaling = None

1478
        for layer in islice(self.layers, self.start_layer, self.end_layer):
1479
1480
1481
            hidden_states, residual = layer(
                positions, hidden_states, residual, llama_4_scaling
            )
1482
1483

        if not get_pp_group().is_last_rank:
王敏's avatar
王敏 committed
1484
1485
1486
            if enable_mla_cp:
                hidden_states = tensor_model_parallel_all_gather(hidden_states.contiguous(), dim=0)
                residual = tensor_model_parallel_all_gather(residual.contiguous(), dim=0)
1487
1488
1489
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
1490

wangding zeng's avatar
wangding zeng committed
1491
        hidden_states, _ = self.norm(hidden_states, residual)
王敏's avatar
王敏 committed
1492
1493
1494

        if enable_mla_cp:
            hidden_states = tensor_model_parallel_all_gather(hidden_states.contiguous(), dim=0)
1495
1496
1497
            gather_indexes_tensor = get_forward_context().gather_indexes_tensor
            if gather_indexes_tensor is not None:
                hidden_states = torch.index_select(hidden_states, 0, gather_indexes_tensor)
王敏's avatar
王敏 committed
1498

wangding zeng's avatar
wangding zeng committed
1499
1500
1501
        return hidden_states


1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
class DeepseekV2MixtureOfExperts(MixtureOfExperts):
    moe_mlp_layers: list[DeepseekV2MoE]
    """
    List of MoE MLP layers in the model.
    """

    def extract_moe_parameters(self, example_moe: DeepseekV2MoE | None):
        if example_moe is None:
            self.num_moe_layers = 0
            self.num_expert_groups = 0
            self.num_logical_experts = 0
            self.num_physical_experts = 0
            self.num_local_physical_experts = 0
            self.num_routed_experts = 0
            self.num_shared_experts = 0
            self.num_redundant_experts = 0
            logger.warning("DeepSeekV2: No DeepseekV2MoE layer found in model.layers.")
        else:
            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

    def update_physical_experts_metadata(
        self,
        num_physical_experts: int,
        num_local_physical_experts: int,
    ) -> None:
        assert self.num_local_physical_experts == num_local_physical_experts
        self.num_physical_experts = num_physical_experts
        self.num_local_physical_experts = num_local_physical_experts
        self.num_redundant_experts = num_physical_experts - self.num_logical_experts
        for moe in self.moe_mlp_layers:
            moe.n_local_physical_experts = num_local_physical_experts
            moe.n_physical_experts = num_physical_experts
            moe.n_redundant_experts = self.num_redundant_experts
            moe.experts.update_expert_map()


class DeepseekV2ForCausalLM(
1544
    nn.Module, SupportsPP, DeepseekV2MixtureOfExperts, SupportsLoRA, SupportsEagle
1545
):
1546
1547
1548
    packed_modules_mapping = {
        "gate_up_proj": ["gate_proj", "up_proj"],
    }
1549
    model_cls = DeepseekV2Model
wangding zeng's avatar
wangding zeng committed
1550

1551
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
1552
        super().__init__()
1553
1554
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
1555
1556
1557
1558
1559
1560
1561

        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'

wangding zeng's avatar
wangding zeng committed
1562
1563
        self.config = config
        self.quant_config = quant_config
王敏's avatar
王敏 committed
1564

1565
1566
1567
1568
1569
1570
1571
1572
1573
        qk_nope_head_dim = getattr(config, "qk_nope_head_dim", 0)
        qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 0)
        self.use_mha = config.model_type == "deepseek" or all(
            dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim)
        )

        if self.use_mha:
            self.packed_modules_mapping["qkv_proj"] = ["q_proj", "k_proj", "v_proj"]

1574
1575
1576
1577
        # `packed_modules_mapping` needs to be modified before
        # initializing DeepseekV2Model, as it is passed inplace to
        # quantization config init and may be used to select the
        # quant_method for relevant layers during initialization.
1578
1579
1580
        self.fuse_qkv_a_proj = (
            hasattr(config, "q_lora_rank") and config.q_lora_rank is not None
        )
1581
1582
1583
1584
1585
1586
        if self.fuse_qkv_a_proj:
            self.packed_modules_mapping["fused_qkv_a_proj"] = [
                "q_a_proj",
                "kv_a_proj_with_mqa",
            ]

1587
        self.model = self.model_cls(
1588
1589
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
1590
        if get_pp_group().is_last_rank:
1591
1592
1593
1594
1595
1596
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
1597
1598
        else:
            self.lm_head = PPMissingLayer()
wangding zeng's avatar
wangding zeng committed
1599
        self.logits_processor = LogitsProcessor(config.vocab_size)
1600
        self.make_empty_intermediate_tensors = (
1601
1602
            self.model.make_empty_intermediate_tensors
        )
1603
1604
1605
1606
1607
1608
1609
        # Set MoE hyperparameters
        self.num_moe_layers = (
            self.config.num_hidden_layers - self.config.first_k_dense_replace
        )
        self.set_moe_parameters()

    def set_moe_parameters(self):
1610
1611
        self.expert_weights = []

1612
        self.num_expert_groups = getattr(self.config, "n_group", 1)
1613

1614
1615
        self.moe_layers = []
        self.moe_mlp_layers = []
1616
        example_moe = None
1617
        for layer in self.model.layers:
1618
1619
1620
            if isinstance(layer, PPMissingLayer):
                continue

1621
1622
            assert isinstance(layer, DeepseekV2DecoderLayer)
            if isinstance(layer.mlp, DeepseekV2MoE):
1623
1624
                # Pick last one layer since the first ones may be dense layers.
                example_moe = layer.mlp
1625
                self.moe_mlp_layers.append(layer.mlp)
1626
1627
                self.moe_layers.append(layer.mlp.experts)

1628
        self.extract_moe_parameters(example_moe)
王敏's avatar
王敏 committed
1629
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
1630
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
1631
        self.tritonsingleton= W8a8GetCacheJSON() 
zhuwenwen's avatar
zhuwenwen committed
1632
        self.tritonsingleton.topk = self.config.num_experts_per_tok
1633
        self.tritonsingleton.quant_method=self.quant_method 
1634

1635
1636
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
1637

wangding zeng's avatar
wangding zeng committed
1638
1639
    def forward(
        self,
zhuwenwen's avatar
zhuwenwen committed
1640
        input_ids: torch.Tensor,
wangding zeng's avatar
wangding zeng committed
1641
        positions: torch.Tensor,
1642
1643
1644
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
1645
1646
1647
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
wangding zeng's avatar
wangding zeng committed
1648
1649
        return hidden_states

1650
1651
1652
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
1653
    ) -> torch.Tensor | None:
1654
        logits = self.logits_processor(self.lm_head, hidden_states)
wangding zeng's avatar
wangding zeng committed
1655
1656
        return logits

1657
1658
1659
1660
    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        return SharedFusedMoE.make_expert_params_mapping(
1661
            self,
1662
1663
1664
1665
1666
1667
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
            num_experts=self.config.n_routed_experts,
            num_redundant_experts=0,
        )
1668

1669
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1670
1671
1672
        rocm_aiter_moe_shared_expert_enabled = (
            rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
        )
wangding zeng's avatar
wangding zeng committed
1673
1674
1675
1676
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
1677
1678
        ]
        mla_params_mapping = [
1679
1680
            ("fused_qkv_a_proj", "q_a_proj", 0),
            ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1),
wangding zeng's avatar
wangding zeng committed
1681
        ]
1682
1683
1684
1685
1686
1687
1688
1689
1690
        mha_params_mapping = [
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]
        if self.use_mha:
            stacked_params_mapping.extend(mha_params_mapping)
        else:
            stacked_params_mapping.extend(mla_params_mapping)
wangding zeng's avatar
wangding zeng committed
1691

1692
1693
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
1694
        expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
1695
            self,
王敏's avatar
王敏 committed
1696
1697
1698
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
1699
1700
1701
            num_experts=self.config.n_routed_experts
            + (
                self.config.n_shared_experts
1702
                if rocm_aiter_moe_shared_expert_enabled
1703
1704
                else 0
            ),
1705
1706
            num_redundant_experts=self.num_redundant_experts,
        )
1707

wangding zeng's avatar
wangding zeng committed
1708
        params_dict = dict(self.named_parameters())
1709
        loaded_params: set[str] = set()
1710
1711
1712
        # 判断是否加载"indexer"权重
        model_has_indexer = any("indexer" in param_name for param_name in params_dict.keys())
        
wangding zeng's avatar
wangding zeng committed
1713
1714
1715
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
1716

1717
1718
1719
1720
1721
            #  跳过加载"indexer"权重
            if "indexer" in name and not model_has_indexer:
                logger.info(f"Skipping indexer weight (DSA disabled): {name}")
                continue

1722
1723
1724
            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
1725

1726
1727
            is_fusion_moe_shared_experts_layer = (
                rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
1728
1729
            )

1730
            for param_name, weight_name, shard_id in stacked_params_mapping:
1731
                # Skip non-stacked layers and experts (experts handled below).
wangding zeng's avatar
wangding zeng committed
1732
1733
                if weight_name not in name:
                    continue
1734
1735
1736
1737
1738
1739
                # 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.
1740
                if ("mlp.experts." in name) and name not in params_dict:
1741
                    continue
1742
                if is_fusion_moe_shared_experts_layer:
1743
                    continue
1744
                name_mapped = name.replace(weight_name, param_name)
1745
1746
1747

                # QKV fusion is optional, fall back to normal
                # weight loading if it's not enabled
1748
                # if go with fusion option, then update name
1749
1750
1751
                if (
                    param_name == "fused_qkv_a_proj"
                ) and name_mapped not in params_dict:
1752
                    continue
1753
1754
                else:
                    name = name_mapped
wangding zeng's avatar
wangding zeng committed
1755
1756
1757
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
1758
1759
1760
1761

                if is_pp_missing_parameter(name, self):
                    continue

wangding zeng's avatar
wangding zeng committed
1762
1763
1764
1765
1766
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
1767
1768
                is_expert_weight = False

1769
1770
1771
1772
1773
1774
1775
1776
1777
                # Special handling: when AITER fusion_shared_experts is enabled,
                # checkpoints may provide a single widened shared_experts tensor
                # without explicit expert indices
                # (e.g. ...mlp.shared_experts.gate_proj.weight).
                # For models with multiple shared experts, split that tensor
                # evenly into per-shared-expert slices and load them into
                # appended expert slots mlp.experts.{n_routed_experts + j}.*
                # accordingly.
                num_chunks = 1
1778
                if is_fusion_moe_shared_experts_layer:
1779
1780
1781
1782
                    num_chunks = getattr(self.config, "n_shared_experts", 1) or 1
                    # Determine split axis based on op type
                    # gate/up: ColumnParallel → split along dim 0
                    # down: RowParallel → split along dim 1
1783
1784
1785
1786
1787
                    split_dim = (
                        1
                        if ("down_proj.weight" in name and loaded_weight.ndim > 1)
                        else 0
                    )
1788
1789
1790
1791
                    total = loaded_weight.shape[split_dim]
                    assert total % num_chunks == 0, (
                        f"Shared expert weight dim {total} "
                        f"not divisible by num_chunks {num_chunks}"
1792
                    )
1793
1794
1795
1796
1797
1798
                    chunk_size = total // num_chunks

                for j in range(num_chunks):
                    chunk_name = name
                    weight_to_load = loaded_weight

1799
                    if is_fusion_moe_shared_experts_layer:
1800
1801
1802
1803
1804
                        chunk_slice = slice(j * chunk_size, (j + 1) * chunk_size)
                        if loaded_weight.ndim == 1:
                            weight_to_load = loaded_weight[chunk_slice]
                        elif split_dim == 0:
                            weight_to_load = loaded_weight[chunk_slice, :]
1805
                        else:
1806
                            weight_to_load = loaded_weight[:, chunk_slice]
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
                        # Synthesize an expert-style name so expert mapping
                        # can route it
                        chunk_name = name.replace(
                            "mlp.shared_experts",
                            f"mlp.experts.{self.config.n_routed_experts + j}",
                        )

                    # Use expert_params_mapping to locate the destination
                    # param and delegate to its expert-aware weight_loader
                    # with expert_id.
                    for mapping in expert_params_mapping:
                        param_name, weight_name, expert_id, shard_id = mapping
                        if weight_name not in chunk_name:
                            continue

                        # 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 = chunk_name.replace(weight_name, param_name)

                        if is_pp_missing_parameter(name_mapped, self):
                            continue

                        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,
                            weight_to_load,
                            name_mapped,
                            shard_id=shard_id,
                            expert_id=expert_id,
                            return_success=True,
                        )
                        if success:
1849
                            if not is_fusion_moe_shared_experts_layer:
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
                                name = name_mapped
                            else:
                                loaded_params.add(name_mapped)
                            break
                    else:
                        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

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

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

                        if is_pp_missing_parameter(name, self):
                            continue
1872

1873
                        param = params_dict[name]
1874
1875
1876
1877
                        weight_loader = getattr(
                            param, "weight_loader", default_weight_loader
                        )
                        weight_loader(param, loaded_weight)
1878
            if not is_fusion_moe_shared_experts_layer:
1879
                loaded_params.add(name)
1880
                
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
        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
1910
            
1911
        return loaded_params
1912
1913


1914
1915
1916
1917
class DeepseekForCausalLM(DeepseekV2ForCausalLM):
    pass


1918
1919
class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass
1920
1921


zhuwenwen's avatar
zhuwenwen committed
1922
1923
1924
1925
class GlmMoeDsaForCausalLM(DeepseekV2ForCausalLM):
    pass


1926
1927
# Compatibility with
# https://huggingface.co/deepseek-ai/DeepSeek-V3-Base/blob/main/configuration_deepseek.py
1928
def get_spec_layer_idx_from_weight_name(
1929
1930
    config: DeepseekV2Config | DeepseekV3Config, weight_name: str
) -> int | None:
1931
1932
1933
1934
    if (
        hasattr(config, "num_nextn_predict_layers")
        and config.num_nextn_predict_layers > 0
    ):
1935
1936
        layer_idx = config.num_hidden_layers
        for i in range(config.num_nextn_predict_layers):
1937
            if weight_name.startswith(f"model.layers.{layer_idx + i}."):
1938
1939
                return layer_idx + i
    return None