deepseek_v2.py 65.5 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."""
26

27
28
import typing
from collections.abc import Callable, Iterable
29
from itertools import islice
wangding zeng's avatar
wangding zeng committed
30

31
32
import os
import re
wangding zeng's avatar
wangding zeng committed
33
34
import torch
from torch import nn
35
from transformers import DeepseekV2Config, DeepseekV3Config
wangding zeng's avatar
wangding zeng committed
36

37
from vllm._aiter_ops import rocm_aiter_ops
38
from vllm.attention.backends.abstract import AttentionBackend
39
from vllm.attention.layer import Attention
40
from vllm.attention.ops.common import pack_seq_triton, unpack_seq_triton
41
from vllm.compilation.decorators import support_torch_compile
42
43
44
45
46
47
48
49
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,
)
50
51
from vllm.forward_context import get_forward_context
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
70
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
71
72
73
    ParallelLMHead,
    VocabParallelEmbedding,
)
74
from vllm.model_executor.model_loader.weight_utils import (
75
76
77
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
78
from vllm.model_executor.models.utils import sequence_parallel_chunk
79
from vllm.platforms import current_platform
80
from vllm.sequence import IntermediateTensors
81
from vllm.utils.deep_gemm import fp8_mqa_logits, fp8_paged_mqa_logits
82
from vllm.utils.torch_utils import direct_register_custom_op
83
84
85
86
from vllm.v1.attention.backends.mla.indexer import (
    DeepseekV32IndexerBackend,
    DeepseekV32IndexerMetadata,
)
87
from vllm.v1.kv_cache_interface import KVCacheSpec, MLAAttentionSpec
wangding zeng's avatar
wangding zeng committed
88

89
from .interfaces import MixtureOfExperts, SupportsEagle, SupportsLoRA, SupportsPP
90
91
92
93
94
95
96
from .utils import (
    PPMissingLayer,
    is_pp_missing_parameter,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
97

98
99
100
101
102
103
104
if current_platform.is_cuda_alike():
    from vllm import _custom_ops as ops
elif current_platform.is_xpu():
    from vllm._ipex_ops import ipex_ops as ops

logger = init_logger(__name__)

wangding zeng's avatar
wangding zeng committed
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
160
161
162
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,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
163
            rope_parameters=config.rope_parameters,
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
        )
        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,
    ) -> 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
187
188
189
190
191
192
193

class DeepseekV2MLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
194
        quant_config: QuantizationConfig | None = None,
wangding zeng's avatar
wangding zeng committed
195
        reduce_results: bool = True,
196
        is_sequence_parallel=False,
197
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
198
199
    ) -> None:
        super().__init__()
200
201
202
203
204

        # 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
205
        self.gate_up_proj = MergedColumnParallelLinear(
206
207
            hidden_size,
            [intermediate_size] * 2,
wangding zeng's avatar
wangding zeng committed
208
            bias=False,
209
            quant_config=quant_config,
210
            disable_tp=is_sequence_parallel,
211
212
213
214
215
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
wangding zeng's avatar
wangding zeng committed
216
            bias=False,
217
            quant_config=quant_config,
218
            reduce_results=reduce_results,
219
            disable_tp=is_sequence_parallel,
220
221
            prefix=f"{prefix}.down_proj",
        )
wangding zeng's avatar
wangding zeng committed
222
        if hidden_act != "silu":
223
224
225
            raise ValueError(
                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
            )
wangding zeng's avatar
wangding zeng committed
226
227
228
229
230
231
232
233
234
235
236
237
        self.act_fn = SiluAndMul()

    def forward(self, x):
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class DeepseekV2MoE(nn.Module):
    def __init__(
        self,
238
        config: DeepseekV2Config | DeepseekV3Config,
239
        parallel_config: ParallelConfig,
240
        quant_config: QuantizationConfig | None = None,
241
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
242
243
244
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
245
246
        self.tp_rank = get_tensor_model_parallel_rank()

247
        self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
248
249

        self.ep_group = get_ep_group().device_group
250
        self.ep_rank = get_ep_group().rank_in_group
251
252
253
        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
254

255
        self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe
256

257
        if config.hidden_act != "silu":
258
259
260
261
262
263
264
265
266
267
268
269
            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",
        )
270
        if getattr(config, "topk_method", None) == "noaux_tc":
271
            self.gate.e_score_correction_bias = nn.Parameter(
272
273
                torch.empty(config.n_routed_experts, dtype=torch.float32)
            )
274
275
276
        else:
            self.gate.e_score_correction_bias = None

277
        # Load balancing settings.
278
279
        eplb_config = parallel_config.eplb_config
        self.enable_eplb = parallel_config.enable_eplb
280

281
        self.n_redundant_experts = eplb_config.num_redundant_experts
282
        self.n_logical_experts = self.n_routed_experts
283
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
284
285
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

286
287
288
289
        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
        )
290

291
        self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
292
293
294
295
        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:
296
297
            self.shared_experts = None
        else:
298
            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
299

wangding zeng's avatar
wangding zeng committed
300
301
302
303
304
            self.shared_experts = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
305
                is_sequence_parallel=self.is_sequence_parallel,
306
                reduce_results=False,
307
                prefix=f"{prefix}.shared_experts",
wangding zeng's avatar
wangding zeng committed
308
309
            )

310
311
        self.experts = SharedFusedMoE(
            shared_experts=self.shared_experts,
312
            gate=self.gate,
313
314
315
316
317
318
319
320
            num_experts=config.n_routed_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            reduce_results=False,
            renormalize=config.norm_topk_prob,
            quant_config=quant_config,
            use_grouped_topk=True,
321
322
            num_expert_group=getattr(config, "n_group", 1),
            topk_group=getattr(config, "topk_group", 1),
323
            prefix=f"{prefix}.experts",
324
            scoring_func=getattr(config, "scoring_func", "softmax"),
325
            # we do scaling outside, set factor to 1.0 to avoid double mul
326
327
            # aiter applies routed_scaling_factor internally
            routed_scaling_factor=1.0
328
            if not self.is_rocm_aiter_moe_enabled
329
            else self.routed_scaling_factor,
330
331
332
333
            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,
334
            n_shared_experts=config.n_shared_experts
335
            if self.is_fusion_moe_shared_experts_enabled
336
            else None,
337
        )
338

wangding zeng's avatar
wangding zeng committed
339
340
341
    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
342
343
344
345
346
347

        # 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:
348
            hidden_states = sequence_parallel_chunk(hidden_states)
349

350
351
352
353
354
        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
            )
355
        else:
356
357
358
359
360
            # 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
            )
361

362
363
364
        shared_output, final_hidden_states = fused_moe_out
        if self.shared_experts is None:
            assert shared_output is None
365
366
367
368

        # Fix FP16 overflow
        # See DeepseekV2DecoderLayer for more details.
        if hidden_states.dtype != torch.float16:
369
            if not self.is_rocm_aiter_moe_enabled:
370
                final_hidden_states *= self.routed_scaling_factor
371
372
        elif self.shared_experts is not None:
            assert shared_output is not None
373
            shared_output *= 1.0 / self.routed_scaling_factor
374
375
376
377

        if self.shared_experts is not None:
            assert shared_output is not None
            final_hidden_states += shared_output
378

379
380
        if self.is_sequence_parallel:
            final_hidden_states = tensor_model_parallel_all_gather(
381
382
                final_hidden_states, 0
            )
383
384
            final_hidden_states = final_hidden_states[:num_tokens]
        elif self.tp_size > 1:
385
386
387
            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
                final_hidden_states
            )
wangding zeng's avatar
wangding zeng committed
388
389
390
391
392
393

        return final_hidden_states.view(num_tokens, hidden_dim)


def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
    import math
394

wangding zeng's avatar
wangding zeng committed
395
396
397
398
399
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


400
401
402
403
404
405
406
407
408
409
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
410
411
412
class DeepseekV2Attention(nn.Module):
    def __init__(
        self,
413
        vllm_config: VllmConfig,
414
        config: DeepseekV2Config | DeepseekV3Config,
wangding zeng's avatar
wangding zeng committed
415
416
417
418
419
420
421
422
        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,
423
424
425
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        topk_indices_buffer: torch.Tensor | None = None,
426
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
    ) -> 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
442
443
        assert topk_indices_buffer is None, (
            "topk_indices_buffer is not \
444
        supported for DeepseekV2Attention"
445
        )
wangding zeng's avatar
wangding zeng committed
446
447

        if self.q_lora_rank is not None:
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
            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
463
        else:
464
465
466
467
468
469
470
            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
471

472
473
474
475
476
        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,
477
478
479
            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
480
481
482
483
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
484
            quant_config=quant_config,
485
486
            prefix=f"{prefix}.kv_b_proj",
        )
wangding zeng's avatar
wangding zeng committed
487
        # O projection.
488
489
490
491
492
493
494
        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",
        )
495
        if config.rope_parameters["rope_type"] != "default":
496
497
498
499
500
            config.rope_parameters["rope_type"] = (
                "deepseek_yarn"
                if config.rope_parameters.get("apply_yarn_scaling", True)
                else "deepseek_llama_scaling"
            )
501

502
503
504
505
        self.rotary_emb = get_rope(
            qk_rope_head_dim,
            rotary_dim=qk_rope_head_dim,
            max_position=max_position_embeddings,
506
            rope_parameters=config.rope_parameters,
507
508
            is_neox_style=False,
        )
wangding zeng's avatar
wangding zeng committed
509

510
511
512
513
        if (
            config.rope_parameters["rope_type"] != "default"
            and config.rope_parameters["rope_type"] == "deepseek_yarn"
        ):
514
515
            mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False)
            scaling_factor = config.rope_parameters["factor"]
wangding zeng's avatar
wangding zeng committed
516
517
518
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

519
520
521
522
523
524
525
526
527
        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
528
529
530
531
532

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
533
        llama_4_scaling: torch.Tensor | None,
wangding zeng's avatar
wangding zeng committed
534
535
536
537
    ) -> torch.Tensor:
        if self.q_lora_rank is not None:
            q = self.q_a_proj(hidden_states)[0]
            q = self.q_a_layernorm(q)
538
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
wangding zeng's avatar
wangding zeng committed
539
        else:
540
541
542
543
            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
544
        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
545
        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
wangding zeng's avatar
wangding zeng committed
546
        latent_cache = latent_cache.unsqueeze(1)
547
        kv_a = self.kv_a_layernorm(kv_a)
wangding zeng's avatar
wangding zeng committed
548
        kv = self.kv_b_proj(kv_a)[0]
549
        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
wangding zeng's avatar
wangding zeng committed
550
        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
551
        k_pe = latent_cache[:, :, self.kv_lora_rank :]
552

wangding zeng's avatar
wangding zeng committed
553
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
554

555
        q[..., self.qk_nope_head_dim :] = q_pe
wangding zeng's avatar
wangding zeng committed
556
        k = torch.empty_like(q)
557
558
        k[..., : self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim :] = k_pe
559
560
561
562
563

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

564
565
        # padding value to qk_head_dim for alignment
        v = torch.nn.functional.pad(
566
567
            v, [0, self.qk_head_dim - self.v_head_dim], value=0
        ).view(-1, self.num_local_heads * self.qk_head_dim)
568
        attn_output = self.attn(q, k, v)
569
570
571
        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
572
573
574
575
        output, _ = self.o_proj(attn_output)
        return output


576
class DeepseekV32IndexerCache(torch.nn.Module, AttentionLayerBase):
577
578
579
    def __init__(
        self, head_dim: int, dtype: torch.dtype, prefix: str, cache_config: CacheConfig
    ):
580
581
582
583
584
585
586
587
588
589
590
        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

591
    def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
592
593
594
595
596
597
598
        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,
        )

599
    def forward(self): ...
600
601
602
603
604
605
606
607
608
609
610
611
612

    def get_attn_backend(self) -> AttentionBackend:
        return DeepseekV32IndexerBackend


def sparse_attn_indexer(
    hidden_states: torch.Tensor,
    k_cache_prefix: str,
    kv_cache: torch.Tensor,
    q_fp8: torch.Tensor,
    k: torch.Tensor,
    weights: torch.Tensor,
    quant_block_size: int,
613
    scale_fmt: str | None,
614
615
616
617
    topk_tokens: int,
    head_dim: int,
    max_model_len: int,
    total_seq_lens: int,
618
    topk_indices_buffer: torch.Tensor | None,
619
620
621
) -> torch.Tensor:
    # careful! this will be None in dummy run
    attn_metadata = get_forward_context().attn_metadata
622
    fp8_dtype = current_platform.fp8_dtype()
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
    # assert isinstance(attn_metadata, dict)
    if not isinstance(attn_metadata, dict):
        return sparse_attn_indexer_fake(
            hidden_states,
            k_cache_prefix,
            kv_cache,
            q_fp8,
            k,
            weights,
            quant_block_size,
            scale_fmt,
            topk_tokens,
            head_dim,
            max_model_len,
            total_seq_lens,
            topk_indices_buffer,
        )
    attn_metadata = attn_metadata[k_cache_prefix]
    assert isinstance(attn_metadata, DeepseekV32IndexerMetadata)
    slot_mapping = attn_metadata.slot_mapping
    has_decode = attn_metadata.num_decodes > 0
    has_prefill = attn_metadata.num_prefills > 0
    num_decode_tokens = attn_metadata.num_decode_tokens

    ops.indexer_k_quant_and_cache(
        k,
        kv_cache,
        slot_mapping,
        quant_block_size,
        scale_fmt,
    )

655
    topk_indices_buffer[: hidden_states.shape[0]] = -1
656
657
    if has_prefill:
        prefill_metadata = attn_metadata.prefill
658
        for chunk in prefill_metadata.chunks:
659
660
661
            k_fp8 = torch.empty(
                [chunk.total_seq_lens, head_dim],
                device=k.device,
662
                dtype=fp8_dtype,
663
664
            )
            k_scale = torch.empty(
665
666
667
                [chunk.total_seq_lens, 4],
                device=k.device,
                dtype=torch.uint8,
668
            )
669
            ops.cp_gather_indexer_k_quant_cache(
670
671
672
673
674
675
                kv_cache,
                k_fp8,
                k_scale,
                chunk.block_table,
                chunk.cu_seq_lens,
            )
676
677
678
679
680
681
            fp8_mqa_logits_func = fp8_mqa_logits
            if current_platform.is_rocm():
                from vllm.attention.ops.rocm_aiter_mla_sparse import rocm_fp8_mqa_logits

                fp8_mqa_logits_func = rocm_fp8_mqa_logits
            logits = fp8_mqa_logits_func(
682
                q_fp8[chunk.token_start : chunk.token_end],
683
                (k_fp8, k_scale.view(torch.float32)),
684
                weights[chunk.token_start : chunk.token_end],
685
686
687
                chunk.cu_seqlen_ks,
                chunk.cu_seqlen_ke,
            )
688
            num_rows = logits.shape[0]
689
690
691
            topk_indices = topk_indices_buffer[
                chunk.token_start : chunk.token_end, :topk_tokens
            ]
692
            torch.ops._C.top_k_per_row_prefill(
693
694
695
696
697
698
699
                logits,
                chunk.cu_seqlen_ks,
                chunk.cu_seqlen_ke,
                topk_indices,
                num_rows,
                logits.stride(0),
                logits.stride(1),
700
                topk_tokens,
701
            )
702
703
704
705
706
707
708
709
710
711
712
713
714

    if has_decode:
        decode_metadata = attn_metadata.decode
        # kv_cache size requirement [num_block, block_size, n_head, head_dim],
        # we only have [num_block, block_size, head_dim],
        kv_cache = kv_cache.unsqueeze(-2)
        decode_lens = decode_metadata.decode_lens
        if decode_metadata.requires_padding:
            # pad in edge case where we have short chunked prefill length <
            # decode_threshold since we unstrictly split
            # prefill and decode by decode_threshold
            # (currently set to 1 + speculative tokens)
            padded_q_fp8_decode_tokens = pack_seq_triton(
715
716
                q_fp8[:num_decode_tokens], decode_lens
            )
717
718
        else:
            padded_q_fp8_decode_tokens = q_fp8[:num_decode_tokens].reshape(
719
720
                decode_lens.shape[0], -1, *q_fp8.shape[1:]
            )
721
722
723
724
725
        # TODO: move and optimize below logic with triton kernels
        batch_size = padded_q_fp8_decode_tokens.shape[0]
        next_n = padded_q_fp8_decode_tokens.shape[1]
        assert batch_size == decode_metadata.seq_lens.shape[0]
        num_padded_tokens = batch_size * next_n
726
727
728
729
730
731
732
733
        fp8_paged_mqa_logits_func = fp8_paged_mqa_logits
        if current_platform.is_rocm():
            from vllm.attention.ops.rocm_aiter_mla_sparse import (
                rocm_fp8_paged_mqa_logits,
            )

            fp8_paged_mqa_logits_func = rocm_fp8_paged_mqa_logits
        logits = fp8_paged_mqa_logits_func(
734
735
736
737
738
739
740
741
            padded_q_fp8_decode_tokens,
            kv_cache,
            weights[:num_padded_tokens],
            decode_metadata.seq_lens,
            decode_metadata.block_table,
            decode_metadata.schedule_metadata,
            max_model_len=max_model_len,
        )
742
        num_rows = logits.shape[0]
743
744
745
        topk_indices = topk_indices_buffer[:num_decode_tokens, :topk_tokens]

        torch.ops._C.top_k_per_row_decode(
746
            logits,
747
748
            next_n,
            decode_metadata.seq_lens,
749
750
751
752
            topk_indices,
            num_rows,
            logits.stride(0),
            logits.stride(1),
753
            topk_tokens,
754
        )
755
756
757
758
759
        if decode_metadata.requires_padding:
            # if padded, we need to unpack
            # the topk indices removing padded tokens
            topk_indices = unpack_seq_triton(
                topk_indices.reshape(batch_size, -1, topk_indices.shape[-1]),
760
761
                decode_lens,
            )
762
763
764
            topk_indices_buffer[:num_decode_tokens, : topk_indices.shape[-1]] = (
                topk_indices
            )
765
766
767
768
769
770
771
772
773
774
775
776

    return topk_indices_buffer


def sparse_attn_indexer_fake(
    hidden_states: torch.Tensor,
    k_cache_prefix: str,
    kv_cache: torch.Tensor,
    q_fp8: torch.Tensor,
    k: torch.Tensor,
    weights: torch.Tensor,
    quant_block_size: int,
777
    scale_fmt: str | None,
778
779
780
781
    topk_tokens: int,
    head_dim: int,
    max_model_len: int,
    total_seq_lens: int,
782
    topk_indices_buffer: torch.Tensor | None,
783
784
785
786
) -> torch.Tensor:
    # profile run
    # NOTE(Chen): create the max possible flattened_kv. So that
    # profile_run can get correct memory usage.
787
788
789
    _flattened_kv = torch.empty(
        [total_seq_lens, head_dim + 4], device=k.device, dtype=torch.uint8
    )
790
791
    fp8_dtype = current_platform.fp8_dtype()
    _k_fp8 = _flattened_kv[..., :head_dim].view(fp8_dtype).contiguous()
792
793
794
795
796
797
798
799
800
801
802
803
804
805
    _k_scale = _flattened_kv[..., head_dim:].view(torch.float32).contiguous()
    return topk_indices_buffer


direct_register_custom_op(
    op_name="sparse_attn_indexer",
    op_func=sparse_attn_indexer,
    mutates_args=["topk_indices_buffer"],
    fake_impl=sparse_attn_indexer_fake,
    dispatch_key=current_platform.dispatch_key,
)


class Indexer(nn.Module):
806
807
808
    def __init__(
        self,
        vllm_config: VllmConfig,
809
        config: DeepseekV2Config | DeepseekV3Config,
810
811
        hidden_size: int,
        q_lora_rank: int,
812
813
814
        quant_config: QuantizationConfig | None,
        cache_config: CacheConfig | None,
        topk_indices_buffer: torch.Tensor | None,
815
816
        prefix: str = "",
    ):
817
818
819
820
821
822
823
824
825
826
        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
827
828
829
830
831
832
833
834
835
836
837
838
839
840
        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",
        )
841
        self.k_norm = LayerNorm(self.head_dim, eps=1e-6)
842
843
844
        self.weights_proj = ReplicatedLinear(
            hidden_size, self.n_head, quant_config=None, prefix=f"{prefix}.weights_proj"
        )
845
846
847
848
849
850
851
852
853
854
        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(
855
            head_dim=self.head_dim + self.head_dim // self.quant_block_size * 4,
856
857
            dtype=torch.uint8,
            prefix=f"{prefix}.k_cache",
858
859
            cache_config=cache_config,
        )
860
861
        self.max_model_len = vllm_config.model_config.max_model_len
        self.prefix = prefix
862
863
        from vllm.v1.attention.backends.mla.indexer import get_max_prefill_buffer_size

864
865
        self.max_total_seq_len = get_max_prefill_buffer_size(vllm_config)

866
867
868
    def forward(
        self, hidden_states: torch.Tensor, qr: torch.Tensor, positions, rotary_emb
    ) -> torch.Tensor:
869
870
871
        q, _ = self.wq_b(qr)
        q = q.view(-1, self.n_head, self.head_dim)
        q_pe, q_nope = torch.split(
872
873
            q, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1
        )
874
875
876
877

        k, _ = self.wk(hidden_states)
        k = self.k_norm(k)
        k_pe, k_nope = torch.split(
878
879
            k, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1
        )
880
881

        q_pe, k_pe = rotary_emb(positions, q_pe, k_pe.unsqueeze(1))
882
883
        q = torch.cat([q_pe.squeeze(0), q_nope], dim=-1)
        k = torch.cat([k_pe.squeeze((0, 2)), k_nope], dim=-1)
884
885
886

        # we only quant q here since k quant is fused with cache insertion
        q = q.view(-1, self.head_dim)
887
888
889
890
891
892
        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,
        )
893
894
895
896
        q_fp8 = q_fp8.view(-1, self.n_head, self.head_dim)
        q_scale = q_scale.view(-1, self.n_head, 1)

        weights, _ = self.weights_proj(hidden_states)
897
898
899
        weights = (
            weights.unsqueeze(-1) * q_scale * self.softmax_scale * self.n_head**-0.5
        )
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
        weights = weights.squeeze(-1)

        return torch.ops.vllm.sparse_attn_indexer(
            hidden_states,
            self.k_cache.prefix,
            self.k_cache.kv_cache[0],
            q_fp8,
            k,
            weights,
            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,
        )


919
920
921
922
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).
923

924
925
        For more info see MLACommonImpl in:
        vllm/v1/attention/backends/mla/utils.py
926
927
928
929
    """

    def __init__(
        self,
930
        vllm_config: VllmConfig,
931
        config: DeepseekV2Config | DeepseekV3Config,
932
933
934
935
936
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
937
        q_lora_rank: int | None,
938
939
        kv_lora_rank: int,
        max_position_embeddings: int = 8192,
940
941
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
942
        prefix: str = "",
943
        topk_indices_buffer: torch.Tensor | None = None,
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
    ) -> 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

        if self.q_lora_rank is not None:
964
            self.fused_qkv_a_proj = MergedColumnParallelLinear(
965
966
967
968
                self.hidden_size,
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim],
                bias=False,
                quant_config=quant_config,
969
                prefix=f"{prefix}.fused_qkv_a_proj",
970
971
                disable_tp=True,
            )
972
973
974
975
976
977
        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,
978
979
                prefix=f"{prefix}.kv_a_proj_with_mqa",
            )
980
981

        if self.q_lora_rank is not None:
982
983
984
985
986
987
988
989
            self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
            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",
            )
990
        else:
991
992
993
994
995
996
997
998
            self.q_proj = ColumnParallelLinear(
                self.hidden_size,
                self.num_heads * self.qk_head_dim,
                bias=False,
                quant_config=quant_config,
                prefix=f"{prefix}.q_proj",
            )
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
999
1000
1001
1002
1003
        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,
1004
1005
1006
1007
1008
1009
1010
1011
1012
            prefix=f"{prefix}.kv_b_proj",
        )
        self.o_proj = RowParallelLinear(
            self.num_heads * self.v_head_dim,
            self.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )
1013

1014
        if config.rope_parameters["rope_type"] != "default":
1015
1016
1017
1018
1019
1020
            config.rope_parameters["rope_type"] = (
                "deepseek_yarn"
                if config.rope_parameters.get("apply_yarn_scaling", True)
                else "deepseek_llama_scaling"
            )

1021
1022
1023
1024
        self.rotary_emb = get_rope(
            qk_rope_head_dim,
            rotary_dim=qk_rope_head_dim,
            max_position=max_position_embeddings,
1025
            rope_parameters=config.rope_parameters,
1026
1027
            is_neox_style=False,
        )
1028
1029
1030
1031
1032

        if (
            config.rope_parameters["rope_type"] != "default"
            and config.rope_parameters["rope_type"] == "deepseek_yarn"
        ):
1033
1034
            mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False)
            scaling_factor = config.rope_parameters["factor"]
1035
1036
1037
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

1038
        self.is_v32 = hasattr(config, "index_topk")
1039
1040

        if self.is_v32:
1041
1042
1043
1044
            self.indexer_rope_emb = get_rope(
                qk_rope_head_dim,
                rotary_dim=qk_rope_head_dim,
                max_position=max_position_embeddings,
1045
                rope_parameters=config.rope_parameters,
1046
1047
                is_neox_style=True,
            )
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
            self.indexer = Indexer(
                vllm_config,
                config,
                hidden_size,
                q_lora_rank,
                quant_config,
                cache_config,
                topk_indices_buffer,
                f"{prefix}.indexer",
            )
1058
        else:
1059
            self.indexer_rope_emb = None
1060
1061
            self.indexer = None

1062
1063
        mla_modules = MLAModules(
            kv_a_layernorm=self.kv_a_layernorm,
1064
            kv_b_proj=self.kv_b_proj,
1065
1066
1067
            rotary_emb=self.rotary_emb,
            o_proj=self.o_proj,
            fused_qkv_a_proj=self.fused_qkv_a_proj
1068
1069
            if self.q_lora_rank is not None
            else None,
1070
            kv_a_proj_with_mqa=self.kv_a_proj_with_mqa
1071
1072
1073
            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,
1074
1075
            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,
1076
            indexer=self.indexer,
1077
            indexer_rotary_emb=self.indexer_rope_emb,
1078
1079
            is_sparse=self.is_v32,
            topk_indices_buffer=topk_indices_buffer,
1080
        )
1081

1082
        self.mla_attn = MultiHeadLatentAttentionWrapper(
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
            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,
1095
1096
1097
1098
1099
1100
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1101
        llama_4_scaling: torch.Tensor | None,
1102
    ) -> torch.Tensor:
1103
        return self.mla_attn(positions, hidden_states, llama_4_scaling)
1104
1105


wangding zeng's avatar
wangding zeng committed
1106
class DeepseekV2DecoderLayer(nn.Module):
1107
1108
1109
1110
    def __init__(
        self,
        vllm_config: VllmConfig,
        prefix: str,
1111
1112
        config: DeepseekV2Config | None = None,
        topk_indices_buffer: torch.Tensor | None = None,
1113
    ) -> None:
wangding zeng's avatar
wangding zeng committed
1114
        super().__init__()
1115

1116
1117
        if config is None:
            config = vllm_config.model_config.hf_config
1118
1119
1120
1121
1122
        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
1123
        self.hidden_size = config.hidden_size
1124
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
1125
        moe_layer_freq = getattr(config, "moe_layer_freq", 1)
1126
1127
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
1128
        layer_idx = int(prefix.split(sep=".")[-1])
1129
        self.layer_idx = layer_idx
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139

        # 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)
        )

1140
1141
        self.use_mha = use_mha

1142
1143
1144
        if use_mha:
            attn_cls = DeepseekAttention
        elif model_config.use_mla:
1145
1146
1147
1148
            attn_cls = DeepseekV2MLAAttention
        else:
            attn_cls = DeepseekV2Attention
        self.self_attn = attn_cls(
1149
            vllm_config=vllm_config,
wangding zeng's avatar
wangding zeng committed
1150
1151
1152
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
1153
1154
1155
            qk_nope_head_dim=qk_nope_head_dim,
            qk_rope_head_dim=qk_rope_head_dim,
            v_head_dim=v_head_dim,
1156
            q_lora_rank=config.q_lora_rank if hasattr(config, "q_lora_rank") else None,
1157
            kv_lora_rank=kv_lora_rank,
wangding zeng's avatar
wangding zeng committed
1158
1159
1160
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
1161
            prefix=f"{prefix}.self_attn",
1162
            topk_indices_buffer=topk_indices_buffer,
wangding zeng's avatar
wangding zeng committed
1163
        )
1164

1165
1166
1167
        if (
            config.n_routed_experts is not None
            and layer_idx >= config.first_k_dense_replace
1168
            and layer_idx % moe_layer_freq == 0
1169
        ):
1170
1171
            self.mlp = DeepseekV2MoE(
                config=config,
1172
                parallel_config=parallel_config,
1173
1174
1175
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )
wangding zeng's avatar
wangding zeng committed
1176
1177
1178
1179
1180
1181
        else:
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
1182
                prefix=f"{prefix}.mlp",
wangding zeng's avatar
wangding zeng committed
1183
            )
1184
1185
1186
1187
        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
        )
1188
        self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
wangding zeng's avatar
wangding zeng committed
1189
1190
1191
1192
1193

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1194
        residual: torch.Tensor | None,
1195
        llama_4_scaling: torch.Tensor | None = None,
wangding zeng's avatar
wangding zeng committed
1196
1197
1198
    ) -> torch.Tensor:
        # Self Attention
        if residual is None:
1199
            residual = hidden_states.clone()
wangding zeng's avatar
wangding zeng committed
1200
1201
            hidden_states = self.input_layernorm(hidden_states)
        else:
1202
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
1203
1204
1205
1206
1207
1208
1209
1210

        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
1211

1212
1213
1214
1215
        if (
            not isinstance(self.self_attn, DeepseekAttention)
            and hidden_states.dtype == torch.float16
        ):
1216
1217
1218
            # Fix FP16 overflow
            # We scale both hidden_states and residual before
            # rmsnorm, and rmsnorm result would not affect by scale.
1219
            hidden_states *= 1.0 / self.routed_scaling_factor
1220
1221
1222
            if self.layer_idx == 0:
                # The residual is shared by all layers, we only scale it on
                # first layer.
1223
                residual *= 1.0 / self.routed_scaling_factor
1224
1225

        # Fully Connected
1226
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
wangding zeng's avatar
wangding zeng committed
1227
        hidden_states = self.mlp(hidden_states)
1228

1229
        if isinstance(self.mlp, DeepseekV2MLP) and hidden_states.dtype == torch.float16:
1230
1231
1232
1233
1234
            # 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
1235
            hidden_states *= 1.0 / self.routed_scaling_factor
1236

wangding zeng's avatar
wangding zeng committed
1237
1238
1239
        return hidden_states, residual


1240
@support_torch_compile
wangding zeng's avatar
wangding zeng committed
1241
1242
1243
class DeepseekV2Model(nn.Module):
    fall_back_to_pt_during_load = False

1244
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
1245
        super().__init__()
1246
1247
1248

        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
1249
        self.config = config
1250
        self.device = current_platform.device_type
1251

wangding zeng's avatar
wangding zeng committed
1252
        self.vocab_size = config.vocab_size
1253
        self.is_v32 = hasattr(config, "index_topk")
1254
1255
1256
1257
1258
1259
        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,
1260
                device=self.device,
1261
            )
1262
1263
        else:
            topk_indices_buffer = None
wangding zeng's avatar
wangding zeng committed
1264

1265
1266
1267
1268
        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
1269
                quant_config=quant_config,
1270
1271
                prefix=f"{prefix}.embed_tokens",
            )
1272
1273
1274
1275
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
1276
            lambda prefix: DeepseekV2DecoderLayer(
1277
                vllm_config, prefix, topk_indices_buffer=topk_indices_buffer
1278
1279
1280
            ),
            prefix=f"{prefix}.layers",
        )
1281
1282
1283
1284
1285

        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
1286
1287
1288
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
wangding zeng's avatar
wangding zeng committed
1289

1290
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
1291
1292
        return self.embed_tokens(input_ids)

wangding zeng's avatar
wangding zeng committed
1293
1294
1295
1296
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1297
1298
1299
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
1300
        if get_pp_group().is_first_rank:
1301
1302
1303
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
1304
                hidden_states = self.embed_input_ids(input_ids)
1305
1306
1307
1308
1309
1310
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
        # 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

1325
        for layer in islice(self.layers, self.start_layer, self.end_layer):
1326
1327
1328
            hidden_states, residual = layer(
                positions, hidden_states, residual, llama_4_scaling
            )
1329
1330

        if not get_pp_group().is_last_rank:
1331
1332
1333
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
1334

wangding zeng's avatar
wangding zeng committed
1335
1336
1337
1338
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
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(
1381
    nn.Module, SupportsPP, DeepseekV2MixtureOfExperts, SupportsLoRA, SupportsEagle
1382
):
1383
1384
1385
    packed_modules_mapping = {
        "gate_up_proj": ["gate_proj", "up_proj"],
    }
1386
    model_cls = DeepseekV2Model
wangding zeng's avatar
wangding zeng committed
1387

1388
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
1389
        super().__init__()
1390
1391
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
wangding zeng's avatar
wangding zeng committed
1392
1393
        self.config = config
        self.quant_config = quant_config
1394

1395
1396
1397
1398
1399
1400
1401
1402
1403
        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"]

1404
1405
1406
1407
        # `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.
1408
1409
1410
        self.fuse_qkv_a_proj = (
            hasattr(config, "q_lora_rank") and config.q_lora_rank is not None
        )
1411
1412
1413
1414
1415
1416
        if self.fuse_qkv_a_proj:
            self.packed_modules_mapping["fused_qkv_a_proj"] = [
                "q_a_proj",
                "kv_a_proj_with_mqa",
            ]

1417
        self.model = self.model_cls(
1418
1419
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
1420
        if get_pp_group().is_last_rank:
1421
1422
1423
1424
1425
1426
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
1427
1428
        else:
            self.lm_head = PPMissingLayer()
wangding zeng's avatar
wangding zeng committed
1429
        self.logits_processor = LogitsProcessor(config.vocab_size)
1430
        self.make_empty_intermediate_tensors = (
1431
1432
            self.model.make_empty_intermediate_tensors
        )
1433
1434
1435
1436
1437
        # Set MoE hyperparameters
        self.num_moe_layers = (
            self.config.num_hidden_layers - self.config.first_k_dense_replace
        )
        self.set_moe_parameters()
1438
1439
1440
1441
        self.quant_method = None
        if quant_config is not None:
            self.quant_method=quant_config.get_name()
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
1442
1443

    def set_moe_parameters(self):
1444
1445
        self.expert_weights = []

1446
        self.num_expert_groups = getattr(self.config, "n_group", 1)
1447

1448
1449
        self.moe_layers = []
        self.moe_mlp_layers = []
1450
        example_moe = None
1451
        for layer in self.model.layers:
1452
1453
1454
            if isinstance(layer, PPMissingLayer):
                continue

1455
1456
            assert isinstance(layer, DeepseekV2DecoderLayer)
            if isinstance(layer.mlp, DeepseekV2MoE):
1457
1458
                # Pick last one layer since the first ones may be dense layers.
                example_moe = layer.mlp
1459
                self.moe_mlp_layers.append(layer.mlp)
1460
1461
                self.moe_layers.append(layer.mlp.experts)

1462
        self.extract_moe_parameters(example_moe)
wangding zeng's avatar
wangding zeng committed
1463

1464
1465
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
1466

wangding zeng's avatar
wangding zeng committed
1467
1468
1469
1470
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1471
1472
1473
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
1474
1475
1476
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
wangding zeng's avatar
wangding zeng committed
1477
1478
        return hidden_states

1479
1480
1481
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
1482
    ) -> torch.Tensor | None:
1483
        logits = self.logits_processor(self.lm_head, hidden_states)
wangding zeng's avatar
wangding zeng committed
1484
1485
        return logits

1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
    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(
            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,
        )

1497
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1498
1499
1500
        rocm_aiter_moe_shared_expert_enabled = (
            rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
        )
wangding zeng's avatar
wangding zeng committed
1501
1502
1503
1504
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
1505
1506
        ]
        mla_params_mapping = [
1507
1508
            ("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
1509
        ]
1510
1511
1512
1513
1514
1515
1516
1517
1518
        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
1519

1520
1521
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
1522
        expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
1523
1524
1525
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
1526
1527
1528
            num_experts=self.config.n_routed_experts
            + (
                self.config.n_shared_experts
1529
                if rocm_aiter_moe_shared_expert_enabled
1530
1531
                else 0
            ),
1532
1533
            num_redundant_experts=self.num_redundant_experts,
        )
1534

wangding zeng's avatar
wangding zeng committed
1535
        params_dict = dict(self.named_parameters())
1536
        loaded_params: set[str] = set()
wangding zeng's avatar
wangding zeng committed
1537
1538
1539
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
1540

1541
1542
1543
            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
1544

1545
1546
            is_fusion_moe_shared_experts_layer = (
                rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
1547
1548
            )

1549
            for param_name, weight_name, shard_id in stacked_params_mapping:
1550
                # Skip non-stacked layers and experts (experts handled below).
wangding zeng's avatar
wangding zeng committed
1551
1552
                if weight_name not in name:
                    continue
1553
1554
1555
1556
1557
1558
                # 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.
1559
                if ("mlp.experts." in name) and name not in params_dict:
1560
                    continue
1561
                if is_fusion_moe_shared_experts_layer:
1562
                    continue
1563
                name_mapped = name.replace(weight_name, param_name)
1564
1565
1566

                # QKV fusion is optional, fall back to normal
                # weight loading if it's not enabled
1567
                # if go with fusion option, then update name
1568
1569
1570
                if (
                    param_name == "fused_qkv_a_proj"
                ) and name_mapped not in params_dict:
1571
                    continue
1572
1573
                else:
                    name = name_mapped
wangding zeng's avatar
wangding zeng committed
1574
1575
1576
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
1577
1578
1579
1580

                if is_pp_missing_parameter(name, self):
                    continue

wangding zeng's avatar
wangding zeng committed
1581
1582
1583
1584
1585
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
1586
                is_expert_weight = False
1587
1588
1589
1590
1591
1592
1593
1594
1595
                # 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
1596
                if is_fusion_moe_shared_experts_layer:
1597
1598
1599
1600
1601
1602
1603
1604
1605
                    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
                    split_dim = 1 if "down_proj.weight" in name else 0
                    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}"
1606
                    )
1607
1608
1609
1610
1611
1612
                    chunk_size = total // num_chunks

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

1613
                    if is_fusion_moe_shared_experts_layer:
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
                        if split_dim == 0:
                            weight_to_load = loaded_weight[
                                j * chunk_size : (j + 1) * chunk_size, :
                            ]
                        else:
                            weight_to_load = loaded_weight[
                                :, j * chunk_size : (j + 1) * chunk_size
                            ]
                        # 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:
1664
                            if not is_fusion_moe_shared_experts_layer:
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
                                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
1687

zhuwenwen's avatar
zhuwenwen committed
1688
                        param = params_dict[name]
1689
1690
1691
1692
                        weight_loader = getattr(
                            param, "weight_loader", default_weight_loader
                        )
                        weight_loader(param, loaded_weight)
1693
            if not is_fusion_moe_shared_experts_layer:
1694
                loaded_params.add(name)
1695
                
1696
1697
1698
1699
1700
1701
        if self.use_llama_nn and self.quant_method is None:
                lay_key_words = [
                    "self_attn.q_a_proj.weight",
                    "self_attn.kv_a_proj_with_mqa.weight",
                    "mlp.gate.weight",
                    "mlp.gate_up_proj.weight",
zhuwenwen's avatar
zhuwenwen committed
1702
1703
1704
                    "mlp.down_proj.weight",
                    "shared_experts.gate_up_proj.weight",
                    "shared_experts.down_proj.weight",
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
                    "self_attn.q_proj.weight",
                    "self_attn.q_b_proj.weight",
                    "self_attn.kv_b_proj.weight",
                    "self_attn.o_proj.weight",
                    "lm_head.weight"
                ]

                combined_words = "|".join(lay_key_words)
                
                for layername in loaded_params:
                    weight = params_dict[layername]
                    matches = re.findall(combined_words, layername)
                    if matches:
                        _weight = torch.zeros_like(weight.data)
                        ori_shape =_weight.shape
                        
                        ops.trans_w16_gemm(_weight, weight.data, _weight.shape[0], _weight.shape[1])
                        weight.data.copy_(_weight)
                        
                        weight.data=weight.data.reshape(ori_shape[1],-1)
1725
        return loaded_params
1726
1727


1728
1729
1730
1731
class DeepseekForCausalLM(DeepseekV2ForCausalLM):
    pass


1732
1733
class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass
1734
1735


1736
1737
# Compatibility with
# https://huggingface.co/deepseek-ai/DeepSeek-V3-Base/blob/main/configuration_deepseek.py
1738
def get_spec_layer_idx_from_weight_name(
1739
1740
    config: DeepseekV2Config | DeepseekV3Config, weight_name: str
) -> int | None:
1741
1742
1743
1744
    if (
        hasattr(config, "num_nextn_predict_layers")
        and config.num_nextn_predict_layers > 0
    ):
1745
1746
        layer_idx = config.num_hidden_layers
        for i in range(config.num_nextn_predict_layers):
1747
            if weight_name.startswith(f"model.layers.{layer_idx + i}."):
1748
1749
                return layer_idx + i
    return None