deepseek_v2.py 74 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
def iqis_all_gather(
187
    iqis: tuple[torch.Tensor, torch.Tensor],
188
    tp_size: int | None = None
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
) -> 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)

223

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
271
        enable_lightly_cp = get_forward_context().enable_lightly_cp
        if enable_lightly_cp:
272
            if iqis is not None and iqis[0] is not None and iqis[1] is not None:
273
                iqis = iqis_all_gather(iqis, tp_size=self.tp_size)
274
            else:
275
                x = tensor_model_parallel_all_gather(x.contiguous(), 0)
276

王敏's avatar
王敏 committed
277
278
279
280
281
282
283
284
285
286
287
288
289
290
        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)

291
        if enable_lightly_cp:
王敏's avatar
王敏 committed
292
            x = tensor_model_parallel_reduce_scatter(x.contiguous(), dim=0)
293
            return x
王敏's avatar
王敏 committed
294
295
296
        elif self.tp_size > 1:
            x = tensor_model_parallel_all_reduce(x)
        return x
wangding zeng's avatar
wangding zeng committed
297
298
299
300
301


class DeepseekV2MoE(nn.Module):
    def __init__(
        self,
302
        config: DeepseekV2Config | DeepseekV3Config,
303
        parallel_config: ParallelConfig,
304
        quant_config: QuantizationConfig | None = None,
305
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
306
307
308
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
309
310
        self.tp_rank = get_tensor_model_parallel_rank()

311
        self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
312
313

        self.ep_group = get_ep_group().device_group
314
        self.ep_rank = get_ep_group().rank_in_group
315
316
317
        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
318

319
        self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe
320

321
        if config.hidden_act != "silu":
322
323
324
325
326
327
328
329
330
331
332
333
            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",
        )
334
        if getattr(config, "topk_method", None) == "noaux_tc":
zhuwenwen's avatar
zhuwenwen committed
335
336
337
338
339
340
341
342
            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)
                )
343
344
345
        else:
            self.gate.e_score_correction_bias = None

346
        # Load balancing settings.
347
348
        eplb_config = parallel_config.eplb_config
        self.enable_eplb = parallel_config.enable_eplb
349

350
        self.n_redundant_experts = eplb_config.num_redundant_experts
351
        self.n_logical_experts = self.n_routed_experts
352
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
353
354
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

355
356
357
358
        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
        )
359

360
        self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
361
362
363
364
        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:
365
366
            self.shared_experts = None
        else:
367
            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
368

369
            self.shared_experts = DeepseekV2MLP(
wangding zeng's avatar
wangding zeng committed
370
371
372
373
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
374
                is_sequence_parallel=self.is_sequence_parallel,
375
                reduce_results=False,
376
                prefix=f"{prefix}.shared_experts",
wangding zeng's avatar
wangding zeng committed
377
378
            )

379
380
        self.experts = SharedFusedMoE(
            shared_experts=self.shared_experts,
381
            gate=self.gate,
382
383
384
385
386
387
            # 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),
388
389
390
391
392
393
            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,
394
395
            num_expert_group=getattr(config, "n_group", 1),
            topk_group=getattr(config, "topk_group", 1),
396
            prefix=f"{prefix}.experts",
397
            scoring_func=getattr(config, "scoring_func", "softmax"),
398
            # we do scaling outside, set factor to 1.0 to avoid double mul
399
400
            # aiter applies routed_scaling_factor internally
            routed_scaling_factor=1.0
401
            if not self.is_rocm_aiter_moe_enabled
402
            else self.routed_scaling_factor,
403
404
405
406
            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,
407
            n_shared_experts=config.n_shared_experts
408
            if self.is_fusion_moe_shared_experts_enabled
409
            else None,
410
        )
411

412
413
414
    def forward(self, hidden_states: torch.Tensor,
                *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
                ) -> torch.Tensor:
415
416
        enable_lightly_cp = get_forward_context().enable_lightly_cp
        if enable_lightly_cp:
王敏's avatar
王敏 committed
417
418
419
            hidden_states = tensor_model_parallel_all_gather(
                hidden_states.contiguous(), 0
            )
wangding zeng's avatar
wangding zeng committed
420
421
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
422
423
424
425
426
427

        # 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:
428
            hidden_states = sequence_parallel_chunk(hidden_states)
429

430
431
432
433
434
435
436
437
438
439
440
        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)
441
442
443
444
445
            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)
446
447
448
449
            routed_scaling_factor = (
                1.0 if self.is_rocm_aiter_moe_enabled
                else self.routed_scaling_factor
            )
450
451
            # Keep shared-expert path intact and only fuse routed scale + add
            # in the downstream MoE kernel.
452
453
454
455
456
457
            _, 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,
458
            )
459
460
461
462
463
464
465
466
467
468
469
470
471
472
        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
                )
473

474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
            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
490

491
        if enable_lightly_cp:
王敏's avatar
王敏 committed
492
493
494
495
496
            final_hidden_states = tensor_model_parallel_reduce_scatter(
                final_hidden_states.contiguous(), 0
            )
            return final_hidden_states
        elif self.is_sequence_parallel:
497
            final_hidden_states = tensor_model_parallel_all_gather(
498
499
                final_hidden_states, 0
            )
500
501
            final_hidden_states = final_hidden_states[:num_tokens]
        elif self.tp_size > 1:
502
503
504
            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
                final_hidden_states
            )
wangding zeng's avatar
wangding zeng committed
505
506
507
508
509
510

        return final_hidden_states.view(num_tokens, hidden_dim)


def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
    import math
511

wangding zeng's avatar
wangding zeng committed
512
513
514
515
516
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


517
518
519
520
521
522
523
524
525
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
526
527
528
529

class DeepseekV2Attention(nn.Module):
    def __init__(
        self,
530
        vllm_config: VllmConfig,
531
        config: DeepseekV2Config | DeepseekV3Config,
wangding zeng's avatar
wangding zeng committed
532
533
534
535
536
537
538
539
        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,
540
541
542
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        topk_indices_buffer: torch.Tensor | None = None,
543
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
    ) -> 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
559
560
        assert topk_indices_buffer is None, (
            "topk_indices_buffer is not \
561
        supported for DeepseekV2Attention"
562
        )
wangding zeng's avatar
wangding zeng committed
563
564

        if self.q_lora_rank is not None:
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
            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
580
        else:
581
582
583
584
585
586
587
            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
588

589
590
591
592
593
        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,
594
595
596
            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
597
598
599
600
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
601
            quant_config=quant_config,
602
603
            prefix=f"{prefix}.kv_b_proj",
        )
wangding zeng's avatar
wangding zeng committed
604
        # O projection.
605
606
607
608
609
610
611
        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",
        )
612
        if config.rope_parameters["rope_type"] != "default":
613
614
615
616
617
            config.rope_parameters["rope_type"] = (
                "deepseek_yarn"
                if config.rope_parameters.get("apply_yarn_scaling", True)
                else "deepseek_llama_scaling"
            )
618

619
620
621
        self.rotary_emb = get_rope(
            qk_rope_head_dim,
            max_position=max_position_embeddings,
622
            rope_parameters=config.rope_parameters,
623
624
            is_neox_style=False,
        )
wangding zeng's avatar
wangding zeng committed
625

626
627
628
629
        if (
            config.rope_parameters["rope_type"] != "default"
            and config.rope_parameters["rope_type"] == "deepseek_yarn"
        ):
630
631
            mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False)
            scaling_factor = config.rope_parameters["factor"]
wangding zeng's avatar
wangding zeng committed
632
633
634
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

635
636
637
638
639
640
641
642
643
        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
644
645
646
647
648

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
649
        llama_4_scaling: torch.Tensor | None,
650
        *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
wangding zeng's avatar
wangding zeng committed
651
652
653
654
    ) -> torch.Tensor:
        if self.q_lora_rank is not None:
            q = self.q_a_proj(hidden_states)[0]
            q = self.q_a_layernorm(q)
655
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
wangding zeng's avatar
wangding zeng committed
656
        else:
657
658
659
660
            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
661
        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
662
        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
wangding zeng's avatar
wangding zeng committed
663
        latent_cache = latent_cache.unsqueeze(1)
664
        kv_a = self.kv_a_layernorm(kv_a)
wangding zeng's avatar
wangding zeng committed
665
        kv = self.kv_b_proj(kv_a)[0]
666
        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
wangding zeng's avatar
wangding zeng committed
667
        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
668
        k_pe = latent_cache[:, :, self.kv_lora_rank :]
669

wangding zeng's avatar
wangding zeng committed
670
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
671

672
        q[..., self.qk_nope_head_dim :] = q_pe
wangding zeng's avatar
wangding zeng committed
673
        k = torch.empty_like(q)
674
675
        k[..., : self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim :] = k_pe
676
677
678
679
680

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

681
682
        # padding value to qk_head_dim for alignment
        v = torch.nn.functional.pad(
683
684
            v, [0, self.qk_head_dim - self.v_head_dim], value=0
        ).view(-1, self.num_local_heads * self.qk_head_dim)
685
        attn_output = self.attn(q, k, v)
686
687
688
        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
689
690
691
692
        output, _ = self.o_proj(attn_output)
        return output


693
class DeepseekV32IndexerCache(torch.nn.Module, AttentionLayerBase):
694
695
696
    def __init__(
        self, head_dim: int, dtype: torch.dtype, prefix: str, cache_config: CacheConfig
    ):
697
698
699
700
701
702
703
704
705
706
707
        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

708
    def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
709
710
711
712
713
714
715
        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,
        )

716
    def forward(self): ...
717
718
719
720
721
722

    def get_attn_backend(self) -> AttentionBackend:
        return DeepseekV32IndexerBackend


class Indexer(nn.Module):
723
724
725
    def __init__(
        self,
        vllm_config: VllmConfig,
726
        config: DeepseekV2Config | DeepseekV3Config,
727
728
        hidden_size: int,
        q_lora_rank: int,
729
730
731
        quant_config: QuantizationConfig | None,
        cache_config: CacheConfig | None,
        topk_indices_buffer: torch.Tensor | None,
732
733
        prefix: str = "",
    ):
734
735
736
737
738
739
740
741
742
743
        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
744
745
746
747
748
749
750
751
752
753
754
755
756
757
        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",
        )
758
        self.k_norm = LayerNorm(self.head_dim, eps=1e-6)
759
        self.weights_proj = ReplicatedLinear(
760
761
762
763
764
            hidden_size,
            self.n_head,
            bias=False,
            quant_config=None,
            prefix=f"{prefix}.weights_proj",
765
        )
766
767
768
769
770
771
772
773
774
775
        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(
776
777
            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,
778
            prefix=f"{prefix}.k_cache",
779
780
            cache_config=cache_config,
        )
781
782
        self.max_model_len = vllm_config.model_config.max_model_len
        self.prefix = prefix
783
784
        from vllm.v1.attention.backends.mla.indexer import get_max_prefill_buffer_size

785
        self.max_total_seq_len = get_max_prefill_buffer_size(vllm_config)
786
787
788
789
790
791
792
793
794
795
        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,
        )
796

797
    def forward(
798
799
        self, hidden_states: torch.Tensor, qr: torch.Tensor, positions, rotary_emb,
        *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
800
    ) -> torch.Tensor:
801
802
803
        q, _ = self.wq_b(qr)
        q = q.view(-1, self.n_head, self.head_dim)
        q_pe, q_nope = torch.split(
804
805
            q, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1
        )
806
807
808
809
        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)
810
811
        k = self.k_norm(k)
        k_pe, k_nope = torch.split(
812
813
            k, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1
        )
814
815

        q_pe, k_pe = rotary_emb(positions, q_pe, k_pe.unsqueeze(1))
816
817
818
819
820
        # 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)

821
822
        # `rotary_emb` is shape-preserving; `q_pe` is already
        # [num_tokens, n_head, rope_dim].
823
824
        q = torch.cat([q_pe, q_nope], dim=-1)
        # `k_pe` is [num_tokens, 1, rope_dim] (MQA).
825
        k = torch.cat([k_pe.squeeze(-2), k_nope], dim=-1)
826

827
828
        enable_lightly_cp = get_forward_context().enable_lightly_cp
        if enable_lightly_cp:
王敏's avatar
王敏 committed
829
830
831
            k = tensor_model_parallel_all_gather(
                k.contiguous(), 0
            )
832
            gather_indexes_tensor = get_forward_context().gather_indexes_tensor
833
834
            enable_lightly_cplb = get_forward_context().enable_lightly_cplb
            if enable_lightly_cplb and gather_indexes_tensor is not None:
835
                k = torch.index_select(k, 0, gather_indexes_tensor)                
王敏's avatar
王敏 committed
836

837
        # we only quant q here since k quant is fused with cache insertion
838
839
840
841
842
843
844
845
846
847
        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
848
849
        else:
            q_fp8 = q
850

851
852
853
854
        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)
855
856
857
858
859
        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)
860

861
        return self.indexer_op(hidden_states, q_fp8, k, weights)
862
863


864
865
866
867
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).
868

869
870
        For more info see MLACommonImpl in:
        vllm/v1/attention/backends/mla/utils.py
871
872
873
874
    """

    def __init__(
        self,
875
        vllm_config: VllmConfig,
876
        config: DeepseekV2Config | DeepseekV3Config,
877
878
879
880
881
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
882
        q_lora_rank: int | None,
883
884
        kv_lora_rank: int,
        max_position_embeddings: int = 8192,
885
886
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
887
        prefix: str = "",
888
        topk_indices_buffer: torch.Tensor | None = None,
889
890
891
892
893
894
895
896
897
898
899
900
901
902
    ) -> 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
903
904
        self.num_local_heads = num_heads // tp_size if not \
            vllm_config.parallel_config.enable_lightly_cp else self.num_heads
905
906
907
908
909

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

        if self.q_lora_rank is not None:
910
            self.fused_qkv_a_proj = MergedColumnParallelLinear(
911
912
913
914
                self.hidden_size,
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim],
                bias=False,
                quant_config=quant_config,
915
                prefix=f"{prefix}.fused_qkv_a_proj",
916
917
                disable_tp=True,
            )
918
919
920
921
922
923
        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,
924
925
                prefix=f"{prefix}.kv_a_proj_with_mqa",
            )
926
927

        if self.q_lora_rank is not None:
wujl5's avatar
wujl5 committed
928
929
930
931
            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)
932
933
934
935
936
937
            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",
938
                disable_tp=vllm_config.parallel_config.enable_lightly_cp
939
            )
940
        else:
941
942
943
944
945
946
            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",
947
                disable_tp=vllm_config.parallel_config.enable_lightly_cp,
948
949
            )
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
950
951
952
953
954
        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,
955
            prefix=f"{prefix}.kv_b_proj",
956
            disable_tp=vllm_config.parallel_config.enable_lightly_cp,
957
958
959
960
961
962
963
        )
        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",
964
            disable_tp=vllm_config.parallel_config.enable_lightly_cp,
965
        )
966

967
        if config.rope_parameters["rope_type"] != "default":
968
969
970
971
972
973
            config.rope_parameters["rope_type"] = (
                "deepseek_yarn"
                if config.rope_parameters.get("apply_yarn_scaling", True)
                else "deepseek_llama_scaling"
            )

974
975
976
        self.rotary_emb = get_rope(
            qk_rope_head_dim,
            max_position=max_position_embeddings,
977
            rope_parameters=config.rope_parameters,
978
979
            is_neox_style=False,
        )
980
981
982
983
984

        if (
            config.rope_parameters["rope_type"] != "default"
            and config.rope_parameters["rope_type"] == "deepseek_yarn"
        ):
985
986
            mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False)
            scaling_factor = config.rope_parameters["factor"]
987
988
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale
989
        #添加判断,默认开启DSA
990
        force_disable_dsa = envs.VLLM_DISABLE_DSA
991
        self.is_v32 = hasattr(config, "index_topk") and not force_disable_dsa
992
993

        if self.is_v32:
994
995
996
            self.indexer_rope_emb = get_rope(
                qk_rope_head_dim,
                max_position=max_position_embeddings,
997
                rope_parameters=config.rope_parameters,
zhuwenwen's avatar
zhuwenwen committed
998
                is_neox_style=not getattr(config, "indexer_rope_interleave", True),
999
            )
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
            self.indexer = Indexer(
                vllm_config,
                config,
                hidden_size,
                q_lora_rank,
                quant_config,
                cache_config,
                topk_indices_buffer,
                f"{prefix}.indexer",
            )
1010
        else:
1011
            self.indexer_rope_emb = None
1012
1013
            self.indexer = None

1014
1015
        mla_modules = MLAModules(
            kv_a_layernorm=self.kv_a_layernorm,
1016
            kv_b_proj=self.kv_b_proj,
1017
1018
1019
            rotary_emb=self.rotary_emb,
            o_proj=self.o_proj,
            fused_qkv_a_proj=self.fused_qkv_a_proj
1020
1021
            if self.q_lora_rank is not None
            else None,
1022
            kv_a_proj_with_mqa=self.kv_a_proj_with_mqa
1023
1024
1025
            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,
1026
1027
            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,
1028
            indexer=self.indexer,
1029
            indexer_rotary_emb=self.indexer_rope_emb,
1030
1031
            is_sparse=self.is_v32,
            topk_indices_buffer=topk_indices_buffer,
1032
        )
1033

1034
        self.mla_attn = MultiHeadLatentAttentionWrapper(
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
            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,
1047
1048
1049
1050
1051
1052
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1053
        llama_4_scaling: torch.Tensor | None,
1054
        *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None
1055
    ) -> torch.Tensor:
1056
        return self.mla_attn(positions, hidden_states, llama_4_scaling, iqis=iqis)
1057
1058


wangding zeng's avatar
wangding zeng committed
1059
class DeepseekV2DecoderLayer(nn.Module):
1060
1061
1062
1063
    def __init__(
        self,
        vllm_config: VllmConfig,
        prefix: str,
1064
1065
        config: DeepseekV2Config | None = None,
        topk_indices_buffer: torch.Tensor | None = None,
1066
    ) -> None:
wangding zeng's avatar
wangding zeng committed
1067
        super().__init__()
1068

1069
1070
        if config is None:
            config = vllm_config.model_config.hf_config
1071
1072
1073
1074
1075
        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
1076
        self.hidden_size = config.hidden_size
1077
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
1078
        moe_layer_freq = getattr(config, "moe_layer_freq", 1)
1079
1080
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
1081
        layer_idx = int(prefix.split(sep=".")[-1])
1082
        self.layer_idx = layer_idx
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092

        # 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)
        )

1093
1094
        self.use_mha = use_mha

1095
1096
1097
        if use_mha:
            attn_cls = DeepseekAttention
        elif model_config.use_mla:
1098
1099
1100
1101
            attn_cls = DeepseekV2MLAAttention
        else:
            attn_cls = DeepseekV2Attention
        self.self_attn = attn_cls(
1102
            vllm_config=vllm_config,
wangding zeng's avatar
wangding zeng committed
1103
1104
1105
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
1106
1107
1108
            qk_nope_head_dim=qk_nope_head_dim,
            qk_rope_head_dim=qk_rope_head_dim,
            v_head_dim=v_head_dim,
1109
            q_lora_rank=config.q_lora_rank if hasattr(config, "q_lora_rank") else None,
1110
            kv_lora_rank=kv_lora_rank,
wangding zeng's avatar
wangding zeng committed
1111
1112
1113
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
1114
            prefix=f"{prefix}.self_attn",
1115
            topk_indices_buffer=topk_indices_buffer,
wangding zeng's avatar
wangding zeng committed
1116
        )
1117

1118
1119
1120
        if (
            config.n_routed_experts is not None
            and layer_idx >= config.first_k_dense_replace
1121
            and layer_idx % moe_layer_freq == 0
1122
        ):
1123
1124
            self.mlp = DeepseekV2MoE(
                config=config,
1125
                parallel_config=parallel_config,
1126
1127
1128
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )
wangding zeng's avatar
wangding zeng committed
1129
1130
1131
1132
1133
1134
        else:
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
1135
                prefix=f"{prefix}.mlp",
wangding zeng's avatar
wangding zeng committed
1136
            )
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
        
        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
            )

1149
        self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
wangding zeng's avatar
wangding zeng committed
1150

1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
    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
1162
1163
        # DSA should set update_input True
        _dsa_flag = hasattr(self.self_attn, "indexer") and self.self_attn.indexer is not None
1164
1165
1166
1167
1168
        if residual is None:
            residual = hidden_states.clone()
            i_q, i_s, _ = self.input_layernorm(x=hidden_states, 
                                               residual=None, 
                                               quant_dtype=torch.int8,
1169
                                               update_input=_dsa_flag
1170
1171
1172
1173
1174
1175
                                               )
            residual_fix_overflow = True
        else:
            i_q, i_s, residual = self.input_layernorm(x=hidden_states, 
                                                      residual=residual, 
                                                      quant_dtype=torch.int8,
1176
                                                      update_input=_dsa_flag
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
                                                      )
        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
1201
        enable_lightly_cp = get_forward_context().enable_lightly_cp
1202
1203
1204
1205
1206
1207
1208
1209
        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
1210
        if enable_lightly_cp and isinstance(self.mlp, DeepseekV2MoE):
1211
1212
1213
            hidden_states = self.mlp(hidden_states)
        else:
            hidden_states = self.mlp(hidden_states, iqis=(_i_q, _i_s))
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225

        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
1226
1227
1228
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1229
        residual: torch.Tensor | None,
1230
        llama_4_scaling: torch.Tensor | None = None,
wangding zeng's avatar
wangding zeng committed
1231
    ) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
1232
1233
1234
1235
1236
1237
1238
        # 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
1239
        else:
zhuwenwen's avatar
zhuwenwen committed
1240
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
1241
            
1242

1243
1244
1245
1246
1247
1248
1249
        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
1250

1251
1252
1253
1254
        if (
            not isinstance(self.self_attn, DeepseekAttention)
            and hidden_states.dtype == torch.float16
        ):
1255
1256
1257
            # Fix FP16 overflow
            # We scale both hidden_states and residual before
            # rmsnorm, and rmsnorm result would not affect by scale.
1258
            hidden_states *= 1.0 / self.routed_scaling_factor
1259
1260
1261
            if self.layer_idx == 0:
                # The residual is shared by all layers, we only scale it on
                # first layer.
1262
                residual *= 1.0 / self.routed_scaling_factor
1263
1264

        # Fully Connected
1265
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
wangding zeng's avatar
wangding zeng committed
1266
        hidden_states = self.mlp(hidden_states)
1267

1268
        if isinstance(self.mlp, DeepseekV2MLP) and hidden_states.dtype == torch.float16:
1269
1270
1271
1272
1273
            # 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
1274
            hidden_states *= 1.0 / self.routed_scaling_factor
1275

wangding zeng's avatar
wangding zeng committed
1276
1277
        return hidden_states, residual

1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
    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
1297

1298
@support_torch_compile
wangding zeng's avatar
wangding zeng committed
1299
1300
1301
class DeepseekV2Model(nn.Module):
    fall_back_to_pt_during_load = False

1302
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
1303
        super().__init__()
1304
1305
1306

        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
1307
        self.config = config
1308
        self.device = current_platform.device_type
1309

王敏's avatar
王敏 committed
1310
1311
1312
        self.tp_size = get_tensor_model_parallel_world_size()
        self.tp_rank = get_tensor_model_parallel_rank()

wangding zeng's avatar
wangding zeng committed
1313
        self.vocab_size = config.vocab_size
1314
        #添加判断,默认开启DSA
1315
        force_disable_dsa = envs.VLLM_DISABLE_DSA
1316
1317
        self.is_v32 = hasattr(config, "index_topk") and not force_disable_dsa        

1318
1319
1320
1321
1322
1323
        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,
1324
                device=self.device,
1325
            )
1326
1327
        else:
            topk_indices_buffer = None
wangding zeng's avatar
wangding zeng committed
1328

1329
1330
1331
1332
        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
1333
                quant_config=quant_config,
1334
1335
                prefix=f"{prefix}.embed_tokens",
            )
1336
1337
1338
1339
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
1340
            lambda prefix: DeepseekV2DecoderLayer(
1341
                vllm_config, prefix, topk_indices_buffer=topk_indices_buffer
1342
1343
1344
            ),
            prefix=f"{prefix}.layers",
        )
1345
1346
1347
1348
1349

        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
1350
1351
1352
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
wangding zeng's avatar
wangding zeng committed
1353

1354
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
1355
1356
        return self.embed_tokens(input_ids)

wangding zeng's avatar
wangding zeng committed
1357
1358
    def forward(
        self,
zhuwenwen's avatar
zhuwenwen committed
1359
        input_ids: torch.Tensor,
wangding zeng's avatar
wangding zeng committed
1360
        positions: torch.Tensor,
1361
1362
1363
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
1364
        if get_pp_group().is_first_rank:
1365
1366
1367
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
1368
                hidden_states = self.embed_input_ids(input_ids)
1369
1370
1371
1372
1373
1374
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

1375
1376
        enable_lightly_cp = get_forward_context().enable_lightly_cp
        if enable_lightly_cp:
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
            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
1392

1393
1394
                if residual is not None:
                    residual = torch.index_select(residual, 0, scatter_indexes_tensor)
王敏's avatar
王敏 committed
1395

1396
1397
                if positions is not None:
                    positions = torch.index_select(positions, 0, scatter_indexes_tensor)
王敏's avatar
王敏 committed
1398

1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
        # 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

1413
        for layer in islice(self.layers, self.start_layer, self.end_layer):
1414
1415
1416
            hidden_states, residual = layer(
                positions, hidden_states, residual, llama_4_scaling
            )
1417
1418

        if not get_pp_group().is_last_rank:
1419
            if enable_lightly_cp:
王敏's avatar
王敏 committed
1420
1421
                hidden_states = tensor_model_parallel_all_gather(hidden_states.contiguous(), dim=0)
                residual = tensor_model_parallel_all_gather(residual.contiguous(), dim=0)
1422
1423
1424
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
1425

wangding zeng's avatar
wangding zeng committed
1426
        hidden_states, _ = self.norm(hidden_states, residual)
王敏's avatar
王敏 committed
1427

1428
        if enable_lightly_cp:
王敏's avatar
王敏 committed
1429
            hidden_states = tensor_model_parallel_all_gather(hidden_states.contiguous(), dim=0)
1430
1431
1432
            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
1433

wangding zeng's avatar
wangding zeng committed
1434
1435
1436
        return hidden_states


1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
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(
1479
    nn.Module, SupportsPP, DeepseekV2MixtureOfExperts, SupportsLoRA, SupportsEagle
1480
):
1481
1482
1483
    packed_modules_mapping = {
        "gate_up_proj": ["gate_proj", "up_proj"],
    }
1484
    model_cls = DeepseekV2Model
wangding zeng's avatar
wangding zeng committed
1485

1486
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
1487
        super().__init__()
1488
1489
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
1490
1491
1492
1493
1494
1495
1496

        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
1497
1498
        self.config = config
        self.quant_config = quant_config
王敏's avatar
王敏 committed
1499

1500
1501
1502
1503
1504
1505
1506
1507
1508
        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"]

1509
1510
1511
1512
        # `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.
1513
1514
1515
        self.fuse_qkv_a_proj = (
            hasattr(config, "q_lora_rank") and config.q_lora_rank is not None
        )
1516
1517
1518
1519
1520
1521
        if self.fuse_qkv_a_proj:
            self.packed_modules_mapping["fused_qkv_a_proj"] = [
                "q_a_proj",
                "kv_a_proj_with_mqa",
            ]

1522
        self.model = self.model_cls(
1523
1524
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
1525
        if get_pp_group().is_last_rank:
1526
1527
1528
1529
1530
1531
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
1532
1533
        else:
            self.lm_head = PPMissingLayer()
wangding zeng's avatar
wangding zeng committed
1534
        self.logits_processor = LogitsProcessor(config.vocab_size)
1535
        self.make_empty_intermediate_tensors = (
1536
1537
            self.model.make_empty_intermediate_tensors
        )
1538
1539
1540
1541
1542
1543
1544
        # 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):
1545
1546
        self.expert_weights = []

1547
        self.num_expert_groups = getattr(self.config, "n_group", 1)
1548

1549
1550
        self.moe_layers = []
        self.moe_mlp_layers = []
1551
        example_moe = None
1552
        for layer in self.model.layers:
1553
1554
1555
            if isinstance(layer, PPMissingLayer):
                continue

1556
1557
            assert isinstance(layer, DeepseekV2DecoderLayer)
            if isinstance(layer.mlp, DeepseekV2MoE):
1558
1559
                # Pick last one layer since the first ones may be dense layers.
                example_moe = layer.mlp
1560
                self.moe_mlp_layers.append(layer.mlp)
1561
1562
                self.moe_layers.append(layer.mlp.experts)

1563
        self.extract_moe_parameters(example_moe)
王敏's avatar
王敏 committed
1564
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
1565
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
1566
        self.tritonsingleton= W8a8GetCacheJSON() 
zhuwenwen's avatar
zhuwenwen committed
1567
        self.tritonsingleton.topk = self.config.num_experts_per_tok
1568
        self.tritonsingleton.quant_method=self.quant_method 
1569

1570
1571
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
1572

wangding zeng's avatar
wangding zeng committed
1573
1574
    def forward(
        self,
zhuwenwen's avatar
zhuwenwen committed
1575
        input_ids: torch.Tensor,
wangding zeng's avatar
wangding zeng committed
1576
        positions: torch.Tensor,
1577
1578
1579
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
1580
1581
1582
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
wangding zeng's avatar
wangding zeng committed
1583
1584
        return hidden_states

1585
1586
1587
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
1588
    ) -> torch.Tensor | None:
1589
        logits = self.logits_processor(self.lm_head, hidden_states)
wangding zeng's avatar
wangding zeng committed
1590
1591
        return logits

1592
1593
1594
1595
    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(
1596
            self,
1597
1598
1599
1600
1601
1602
            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,
        )
1603

1604
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1605
1606
1607
        rocm_aiter_moe_shared_expert_enabled = (
            rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
        )
wangding zeng's avatar
wangding zeng committed
1608
1609
1610
1611
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
1612
1613
        ]
        mla_params_mapping = [
1614
1615
            ("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
1616
        ]
1617
1618
1619
1620
1621
1622
1623
1624
1625
        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
1626

1627
1628
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
1629
        expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
1630
            self,
王敏's avatar
王敏 committed
1631
1632
1633
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
1634
1635
1636
            num_experts=self.config.n_routed_experts
            + (
                self.config.n_shared_experts
1637
                if rocm_aiter_moe_shared_expert_enabled
1638
1639
                else 0
            ),
1640
1641
            num_redundant_experts=self.num_redundant_experts,
        )
1642

wangding zeng's avatar
wangding zeng committed
1643
        params_dict = dict(self.named_parameters())
1644
        loaded_params: set[str] = set()
1645
1646
1647
        # 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
1648
1649
1650
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
1651

1652
1653
1654
1655
1656
            #  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

1657
1658
1659
            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
1660

1661
1662
            is_fusion_moe_shared_experts_layer = (
                rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
1663
1664
            )

1665
            for param_name, weight_name, shard_id in stacked_params_mapping:
1666
                # Skip non-stacked layers and experts (experts handled below).
wangding zeng's avatar
wangding zeng committed
1667
1668
                if weight_name not in name:
                    continue
1669
1670
1671
1672
1673
1674
                # 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.
1675
                if ("mlp.experts." in name) and name not in params_dict:
1676
                    continue
1677
                if is_fusion_moe_shared_experts_layer:
1678
                    continue
1679
                name_mapped = name.replace(weight_name, param_name)
1680
1681
1682

                # QKV fusion is optional, fall back to normal
                # weight loading if it's not enabled
1683
                # if go with fusion option, then update name
1684
1685
1686
                if (
                    param_name == "fused_qkv_a_proj"
                ) and name_mapped not in params_dict:
1687
                    continue
1688
1689
                else:
                    name = name_mapped
wangding zeng's avatar
wangding zeng committed
1690
1691
1692
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
1693
1694
1695
1696

                if is_pp_missing_parameter(name, self):
                    continue

wangding zeng's avatar
wangding zeng committed
1697
1698
1699
1700
1701
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
1702
1703
                is_expert_weight = False

1704
1705
1706
1707
1708
1709
1710
1711
1712
                # 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
1713
                if is_fusion_moe_shared_experts_layer:
1714
1715
1716
1717
                    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
1718
1719
1720
1721
1722
                    split_dim = (
                        1
                        if ("down_proj.weight" in name and loaded_weight.ndim > 1)
                        else 0
                    )
1723
1724
1725
1726
                    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}"
1727
                    )
1728
1729
1730
1731
1732
1733
                    chunk_size = total // num_chunks

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

1734
                    if is_fusion_moe_shared_experts_layer:
1735
1736
1737
1738
1739
                        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, :]
1740
                        else:
1741
                            weight_to_load = loaded_weight[:, chunk_slice]
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
                        # 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:
1784
                            if not is_fusion_moe_shared_experts_layer:
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
                                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
1807

1808
                        param = params_dict[name]
1809
1810
1811
1812
                        weight_loader = getattr(
                            param, "weight_loader", default_weight_loader
                        )
                        weight_loader(param, loaded_weight)
1813
            if not is_fusion_moe_shared_experts_layer:
1814
                loaded_params.add(name)
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
        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
1845
            
1846
        return loaded_params
1847
1848


1849
1850
1851
1852
class DeepseekForCausalLM(DeepseekV2ForCausalLM):
    pass


1853
1854
class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass
1855
1856


zhuwenwen's avatar
zhuwenwen committed
1857
1858
1859
1860
class GlmMoeDsaForCausalLM(DeepseekV2ForCausalLM):
    pass


1861
1862
# Compatibility with
# https://huggingface.co/deepseek-ai/DeepSeek-V3-Base/blob/main/configuration_deepseek.py
1863
def get_spec_layer_idx_from_weight_name(
1864
1865
    config: DeepseekV2Config | DeepseekV3Config, weight_name: str
) -> int | None:
1866
1867
1868
1869
    if (
        hasattr(config, "num_nextn_predict_layers")
        and config.num_nextn_predict_layers > 0
    ):
1870
1871
        layer_idx = config.num_hidden_layers
        for i in range(config.num_nextn_predict_layers):
1872
            if weight_name.startswith(f"model.layers.{layer_idx + i}."):
1873
1874
                return layer_idx + i
    return None