deepseek_v2.py 69.7 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.backends.abstract import AttentionBackend
40
from vllm.attention.layer import Attention
41
from vllm.attention.ops.common import pack_seq_triton, unpack_seq_triton
42
from vllm.compilation.decorators import support_torch_compile
43
44
45
46
47
48
49
50
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,
)
51
52
from vllm.forward_context import get_forward_context
from vllm.logger import init_logger
wangding zeng's avatar
wangding zeng committed
53
from vllm.model_executor.layers.activation import SiluAndMul
54
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
55
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
56
from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm
57
58
59
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    MergedColumnParallelLinear,
60
    QKVParallelLinear,
61
62
63
    ReplicatedLinear,
    RowParallelLinear,
)
wangding zeng's avatar
wangding zeng committed
64
from vllm.model_executor.layers.logits_processor import LogitsProcessor
65
from vllm.model_executor.layers.mla import MLAModules, MultiHeadLatentAttentionWrapper
66
from vllm.model_executor.layers.quantization import QuantizationConfig
67
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
68
69
    per_token_group_quant_fp8,
)
wangding zeng's avatar
wangding zeng committed
70
71
from vllm.model_executor.layers.rotary_embedding import get_rope
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.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
88
from vllm.v1.worker.workspace import current_workspace_manager
wangding zeng's avatar
wangding zeng committed
89

90
from .interfaces import MixtureOfExperts, SupportsEagle, SupportsLoRA, SupportsPP
91
92
93
94
95
96
97
from .utils import (
    PPMissingLayer,
    is_pp_missing_parameter,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
98

99
100
101
102
103
if current_platform.is_rocm():
    from lightop import op, gemmopt
else:
    from vllm.utils.deep_gemm import fp8_mqa_logits, fp8_paged_mqa_logits

104
105
106
107
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
zhuwenwen's avatar
zhuwenwen committed
108
from vllm.utils import W8a8GetCacheJSON
109

110

111
112
logger = init_logger(__name__)

wangding zeng's avatar
wangding zeng committed
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
163
164
165
166
167
168
169
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,
170
            rope_parameters=config.rope_parameters,
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
        )
        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
194
195
196
197
198
199
200

class DeepseekV2MLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
201
        quant_config: QuantizationConfig | None = None,
wangding zeng's avatar
wangding zeng committed
202
        reduce_results: bool = True,
203
        is_sequence_parallel=False,
204
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
205
206
    ) -> None:
        super().__init__()
207
208
209
210
211

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

235
236
237
238
239
    def forward(self, x):
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x
wangding zeng's avatar
wangding zeng committed
240
241
242
243
244


class DeepseekV2MoE(nn.Module):
    def __init__(
        self,
245
        config: DeepseekV2Config | DeepseekV3Config,
246
        parallel_config: ParallelConfig,
247
        quant_config: QuantizationConfig | None = None,
248
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
249
250
251
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
252
253
        self.tp_rank = get_tensor_model_parallel_rank()

254
        self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
255
256

        self.ep_group = get_ep_group().device_group
257
        self.ep_rank = get_ep_group().rank_in_group
258
259
260
        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
261

262
        self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe
263

264
        if config.hidden_act != "silu":
265
266
267
268
269
270
271
272
273
274
275
276
            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",
        )
277
        if getattr(config, "topk_method", None) == "noaux_tc":
278
            self.gate.e_score_correction_bias = nn.Parameter(
279
280
                torch.empty(config.n_routed_experts, dtype=torch.float32)
            )
281
282
283
        else:
            self.gate.e_score_correction_bias = None

284
        # Load balancing settings.
285
286
        eplb_config = parallel_config.eplb_config
        self.enable_eplb = parallel_config.enable_eplb
287

288
        self.n_redundant_experts = eplb_config.num_redundant_experts
289
        self.n_logical_experts = self.n_routed_experts
290
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
291
292
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

293
294
295
296
        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
        )
297

298
        self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
299
300
301
302
        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:
303
304
            self.shared_experts = None
        else:
305
            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
306

wangding zeng's avatar
wangding zeng committed
307
308
309
310
311
            self.shared_experts = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
312
                is_sequence_parallel=self.is_sequence_parallel,
313
                reduce_results=False,
314
                prefix=f"{prefix}.shared_experts",
wangding zeng's avatar
wangding zeng committed
315
316
            )

317
318
        self.experts = SharedFusedMoE(
            shared_experts=self.shared_experts,
319
            gate=self.gate,
320
321
322
323
324
325
326
327
            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,
328
329
            num_expert_group=getattr(config, "n_group", 1),
            topk_group=getattr(config, "topk_group", 1),
330
            prefix=f"{prefix}.experts",
331
            scoring_func=getattr(config, "scoring_func", "softmax"),
332
            # we do scaling outside, set factor to 1.0 to avoid double mul
333
334
            # aiter applies routed_scaling_factor internally
            routed_scaling_factor=1.0
335
            if not self.is_rocm_aiter_moe_enabled
336
            else self.routed_scaling_factor,
337
338
339
340
            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,
341
            n_shared_experts=config.n_shared_experts
342
            if self.is_fusion_moe_shared_experts_enabled
343
            else None,
344
        )
345

346
    def forward(self, hidden_states: torch.Tensor,
zhuwenwen's avatar
zhuwenwen committed
347
348
                rms_weight: torch.Tensor | None = None,
                residual: torch.Tensor | None = None
349
                ) -> torch.Tensor:
wangding zeng's avatar
wangding zeng committed
350
351
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
352
353
354
355
356
357

        # 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:
358
            hidden_states = sequence_parallel_chunk(hidden_states)
359

360
361
362
363
364
        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
            )
365
        else:
366
367
368
369
370
            # 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
            )
371

372
373
374
        shared_output, final_hidden_states = fused_moe_out
        if self.shared_experts is None:
            assert shared_output is None
375
376
377

        # Fix FP16 overflow
        # See DeepseekV2DecoderLayer for more details.
zhuwenwen's avatar
zhuwenwen committed
378
        if hidden_states.dtype != torch.float16:
379
            if not self.is_rocm_aiter_moe_enabled:
380
                final_hidden_states *= self.routed_scaling_factor
381
382
        elif self.shared_experts is not None:
            assert shared_output is not None
383
            shared_output *= 1.0 / self.routed_scaling_factor
384
385
386
387

        if self.shared_experts is not None:
            assert shared_output is not None
            final_hidden_states += shared_output
388

389
390
        if self.is_sequence_parallel:
            final_hidden_states = tensor_model_parallel_all_gather(
391
392
                final_hidden_states, 0
            )
393
394
            final_hidden_states = final_hidden_states[:num_tokens]
        elif self.tp_size > 1:
395
396
397
            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
                final_hidden_states
            )
wangding zeng's avatar
wangding zeng committed
398
399
400
401
402
403

        return final_hidden_states.view(num_tokens, hidden_dim)


def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
    import math
404

wangding zeng's avatar
wangding zeng committed
405
406
407
408
409
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


410
411
412
413
414
415
416
417
418
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
419
420
421
422

class DeepseekV2Attention(nn.Module):
    def __init__(
        self,
423
        vllm_config: VllmConfig,
424
        config: DeepseekV2Config | DeepseekV3Config,
wangding zeng's avatar
wangding zeng committed
425
426
427
428
429
430
431
432
        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,
433
434
435
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        topk_indices_buffer: torch.Tensor | None = None,
436
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
    ) -> 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
452
453
        assert topk_indices_buffer is None, (
            "topk_indices_buffer is not \
454
        supported for DeepseekV2Attention"
455
        )
wangding zeng's avatar
wangding zeng committed
456
457

        if self.q_lora_rank is not None:
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
            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
473
        else:
474
475
476
477
478
479
480
            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
481

482
483
484
485
486
        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,
487
488
489
            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
490
491
492
493
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
494
            quant_config=quant_config,
495
496
            prefix=f"{prefix}.kv_b_proj",
        )
wangding zeng's avatar
wangding zeng committed
497
        # O projection.
498
499
500
501
502
503
504
        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",
        )
505
        if config.rope_parameters["rope_type"] != "default":
506
507
508
509
510
            config.rope_parameters["rope_type"] = (
                "deepseek_yarn"
                if config.rope_parameters.get("apply_yarn_scaling", True)
                else "deepseek_llama_scaling"
            )
511

512
513
514
        self.rotary_emb = get_rope(
            qk_rope_head_dim,
            max_position=max_position_embeddings,
515
            rope_parameters=config.rope_parameters,
516
517
            is_neox_style=False,
        )
wangding zeng's avatar
wangding zeng committed
518

519
520
521
522
        if (
            config.rope_parameters["rope_type"] != "default"
            and config.rope_parameters["rope_type"] == "deepseek_yarn"
        ):
523
524
            mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False)
            scaling_factor = config.rope_parameters["factor"]
wangding zeng's avatar
wangding zeng committed
525
526
527
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

528
529
530
531
532
533
534
535
536
        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
537
538
539
540
541

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
542
        llama_4_scaling: torch.Tensor | None,
wangding zeng's avatar
wangding zeng committed
543
544
545
546
    ) -> torch.Tensor:
        if self.q_lora_rank is not None:
            q = self.q_a_proj(hidden_states)[0]
            q = self.q_a_layernorm(q)
547
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
wangding zeng's avatar
wangding zeng committed
548
        else:
549
550
551
552
            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
553
        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
554
        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
wangding zeng's avatar
wangding zeng committed
555
        latent_cache = latent_cache.unsqueeze(1)
556
        kv_a = self.kv_a_layernorm(kv_a)
wangding zeng's avatar
wangding zeng committed
557
        kv = self.kv_b_proj(kv_a)[0]
558
        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
wangding zeng's avatar
wangding zeng committed
559
        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
560
        k_pe = latent_cache[:, :, self.kv_lora_rank :]
561

wangding zeng's avatar
wangding zeng committed
562
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
563

564
        q[..., self.qk_nope_head_dim :] = q_pe
wangding zeng's avatar
wangding zeng committed
565
        k = torch.empty_like(q)
566
567
        k[..., : self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim :] = k_pe
568
569
570
571
572

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

573
574
        # padding value to qk_head_dim for alignment
        v = torch.nn.functional.pad(
575
576
            v, [0, self.qk_head_dim - self.v_head_dim], value=0
        ).view(-1, self.num_local_heads * self.qk_head_dim)
577
        attn_output = self.attn(q, k, v)
578
579
580
        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
581
582
583
584
        output, _ = self.o_proj(attn_output)
        return output


585
class DeepseekV32IndexerCache(torch.nn.Module, AttentionLayerBase):
586
587
588
    def __init__(
        self, head_dim: int, dtype: torch.dtype, prefix: str, cache_config: CacheConfig
    ):
589
590
591
592
593
594
595
596
597
598
599
        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

600
    def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
601
602
603
604
605
606
607
        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,
        )

608
    def forward(self): ...
609
610
611
612
613
614
615
616
617
618
619
620
621
622

    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,
623
    scale_fmt: str | None,
624
625
626
627
    topk_tokens: int,
    head_dim: int,
    max_model_len: int,
    total_seq_lens: int,
628
    topk_indices_buffer: torch.Tensor | None,
629
630
631
) -> torch.Tensor:
    # careful! this will be None in dummy run
    attn_metadata = get_forward_context().attn_metadata
632
    fp8_dtype = current_platform.fp8_dtype()
633

634
635
    # assert isinstance(attn_metadata, dict)
    if not isinstance(attn_metadata, dict):
636
637
638
639
640
641
        # Reserve workspace for indexer during profiling run
        current_workspace_manager().get_simultaneous(
            ((total_seq_lens, head_dim), torch.float8_e4m3fn),
            ((total_seq_lens, 4), torch.uint8),
        )

642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
        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

664
665
666
667
668
669
670
671
    if not current_platform.is_rocm() or torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938":
        ops.indexer_k_quant_and_cache(
            k,
            kv_cache,
            slot_mapping,
            quant_block_size,
            scale_fmt,
        )
zhuwenwen's avatar
zhuwenwen committed
672
673
674
675
676
677
678
    # else:
    #     ops.indexer_k_cache(
    #         k,
    #         kv_cache,
    #         slot_mapping,
    #         scale_fmt,
    #     )
679

680
    topk_indices_buffer[: hidden_states.shape[0]] = -1
681
682
    if has_prefill:
        prefill_metadata = attn_metadata.prefill
683
684
685
686
687
688
689
690

        # Get the full shared workspace buffers once (will allocate on first use)
        workspace_manager = current_workspace_manager()
        k_fp8_full, k_scale_full = workspace_manager.get_simultaneous(
            ((total_seq_lens, head_dim), fp8_dtype),
            ((total_seq_lens, 4), torch.uint8),
        )

691
        for chunk in prefill_metadata.chunks:
692
693
694
695
696
697
698
699
700
701
            if not current_platform.is_rocm() or torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938":
                k_fp8 = k_fp8_full[: chunk.total_seq_lens]
                k_scale = k_scale_full[: chunk.total_seq_lens]
                ops.cp_gather_indexer_k_quant_cache(
                    kv_cache,
                    k_fp8,
                    k_scale,
                    chunk.block_table,
                    chunk.cu_seq_lens,
                )
702
703
            fp8_mqa_logits_func = fp8_mqa_logits
            if current_platform.is_rocm():
704
                # from vllm.attention.ops.rocm_aiter_mla_sparse import rocm_fp8_mqa_logits
705

706
707
708
                # fp8_mqa_logits_func = rocm_fp8_mqa_logits
                fp8_mqa_logits_func = op.mqa_logits
                
709
            logits = fp8_mqa_logits_func(
710
                q_fp8[chunk.token_start : chunk.token_end],
711
712
                (k_fp8, k_scale.view(torch.float32)) if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" else k_fp8,
                weights[chunk.token_start : chunk.token_end] if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" else weights[chunk.token_start : chunk.token_end].to(torch.float32),
713
714
                chunk.cu_seqlen_ks,
                chunk.cu_seqlen_ke,
715
716
717
718
719
                q_fp8[chunk.token_start : chunk.token_end].shape[0], 
                k.shape[0],
                64,
                128,
                True,
720
            )
721
            num_rows = logits.shape[0]
722
723
724
            topk_indices = topk_indices_buffer[
                chunk.token_start : chunk.token_end, :topk_tokens
            ]
725
            torch.ops._C.top_k_per_row_prefill(
726
727
728
729
730
731
732
                logits,
                chunk.cu_seqlen_ks,
                chunk.cu_seqlen_ke,
                topk_indices,
                num_rows,
                logits.stride(0),
                logits.stride(1),
733
                topk_tokens,
734
            )
735
736
737
738
739
740
741
742
743
744
745
746
747

    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(
748
749
                q_fp8[:num_decode_tokens], decode_lens
            )
750
751
        else:
            padded_q_fp8_decode_tokens = q_fp8[:num_decode_tokens].reshape(
752
753
                decode_lens.shape[0], -1, *q_fp8.shape[1:]
            )
754
755
756
757
758
        # 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
759
760
        fp8_paged_mqa_logits_func = fp8_paged_mqa_logits
        if current_platform.is_rocm():
761
762
763
            # from vllm.attention.ops.rocm_aiter_mla_sparse import (
            #     rocm_fp8_paged_mqa_logits,
            # )
764

765
766
            # fp8_paged_mqa_logits_func = rocm_fp8_paged_mqa_logits
            fp8_paged_mqa_logits_func = gemmopt.paged_mqa_logits
767
        logits = fp8_paged_mqa_logits_func(
768
769
            padded_q_fp8_decode_tokens,
            kv_cache,
770
            weights[:num_padded_tokens] if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" else weights[:num_padded_tokens].to(torch.float32),
771
772
773
            decode_metadata.seq_lens,
            decode_metadata.block_table,
            decode_metadata.schedule_metadata,
774
            max_model_len,
775
        )
776
        num_rows = logits.shape[0]
777
778
779
        topk_indices = topk_indices_buffer[:num_decode_tokens, :topk_tokens]

        torch.ops._C.top_k_per_row_decode(
780
            logits,
781
782
            next_n,
            decode_metadata.seq_lens,
783
784
785
786
            topk_indices,
            num_rows,
            logits.stride(0),
            logits.stride(1),
787
            topk_tokens,
788
        )
789
790
791
792
793
        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]),
794
795
                decode_lens,
            )
796
797
798
            topk_indices_buffer[:num_decode_tokens, : topk_indices.shape[-1]] = (
                topk_indices
            )
799
800
801
802
803
804
805
806
807
808
809
810

    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,
811
    scale_fmt: str | None,
812
813
814
815
    topk_tokens: int,
    head_dim: int,
    max_model_len: int,
    total_seq_lens: int,
816
    topk_indices_buffer: torch.Tensor | None,
817
818
819
820
821
822
823
824
825
826
827
828
829
830
) -> torch.Tensor:
    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):
831
832
833
    def __init__(
        self,
        vllm_config: VllmConfig,
834
        config: DeepseekV2Config | DeepseekV3Config,
835
836
        hidden_size: int,
        q_lora_rank: int,
837
838
839
        quant_config: QuantizationConfig | None,
        cache_config: CacheConfig | None,
        topk_indices_buffer: torch.Tensor | None,
840
841
        prefix: str = "",
    ):
842
843
844
845
846
847
848
849
850
851
        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
852
853
854
855
856
857
858
859
860
861
862
863
864
865
        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",
        )
866
        self.k_norm = LayerNorm(self.head_dim, eps=1e-6)
867
868
869
        self.weights_proj = ReplicatedLinear(
            hidden_size, self.n_head, quant_config=None, prefix=f"{prefix}.weights_proj"
        )
870
871
872
873
874
875
876
877
878
879
        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(
880
881
            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,
882
            prefix=f"{prefix}.k_cache",
883
884
            cache_config=cache_config,
        )
885
886
        self.max_model_len = vllm_config.model_config.max_model_len
        self.prefix = prefix
887
888
        from vllm.v1.attention.backends.mla.indexer import get_max_prefill_buffer_size

889
890
        self.max_total_seq_len = get_max_prefill_buffer_size(vllm_config)

891
892
893
    def forward(
        self, hidden_states: torch.Tensor, qr: torch.Tensor, positions, rotary_emb
    ) -> torch.Tensor:
894
895
896
        q, _ = self.wq_b(qr)
        q = q.view(-1, self.n_head, self.head_dim)
        q_pe, q_nope = torch.split(
897
898
            q, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1
        )
899
900
901
902

        k, _ = self.wk(hidden_states)
        k = self.k_norm(k)
        k_pe, k_nope = torch.split(
903
904
            k, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1
        )
905
906

        q_pe, k_pe = rotary_emb(positions, q_pe, k_pe.unsqueeze(1))
907
908
        q = torch.cat([q_pe.squeeze(0), q_nope], dim=-1)
        k = torch.cat([k_pe.squeeze((0, 2)), k_nope], dim=-1)
909
910

        # we only quant q here since k quant is fused with cache insertion
911
912
913
914
915
916
917
918
919
920
        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)
921
922

        weights, _ = self.weights_proj(hidden_states)
923
924
925
926
927
        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)
928
929
930
931
932

        return torch.ops.vllm.sparse_attn_indexer(
            hidden_states,
            self.k_cache.prefix,
            self.k_cache.kv_cache[0],
933
            q_fp8 if not current_platform.is_rocm() or torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" else q,
934
935
936
937
938
939
940
941
942
943
944
945
            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,
        )


946
947
948
949
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).
950

951
952
        For more info see MLACommonImpl in:
        vllm/v1/attention/backends/mla/utils.py
953
954
955
956
    """

    def __init__(
        self,
957
        vllm_config: VllmConfig,
958
        config: DeepseekV2Config | DeepseekV3Config,
959
960
961
962
963
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
964
        q_lora_rank: int | None,
965
966
        kv_lora_rank: int,
        max_position_embeddings: int = 8192,
967
968
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
969
        prefix: str = "",
970
        topk_indices_buffer: torch.Tensor | None = None,
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
    ) -> 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:
991
            self.fused_qkv_a_proj = MergedColumnParallelLinear(
992
993
994
995
                self.hidden_size,
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim],
                bias=False,
                quant_config=quant_config,
996
                prefix=f"{prefix}.fused_qkv_a_proj",
997
998
                disable_tp=True,
            )
999
1000
1001
1002
1003
1004
        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,
1005
1006
                prefix=f"{prefix}.kv_a_proj_with_mqa",
            )
1007
1008

        if self.q_lora_rank is not None:
1009
1010
1011
1012
1013
1014
1015
1016
            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",
            )
1017
        else:
1018
1019
1020
1021
1022
1023
1024
1025
            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)
1026
1027
1028
1029
1030
        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,
1031
1032
1033
1034
1035
1036
1037
1038
1039
            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",
        )
1040

1041
        if config.rope_parameters["rope_type"] != "default":
1042
1043
1044
1045
1046
1047
            config.rope_parameters["rope_type"] = (
                "deepseek_yarn"
                if config.rope_parameters.get("apply_yarn_scaling", True)
                else "deepseek_llama_scaling"
            )

1048
1049
1050
        self.rotary_emb = get_rope(
            qk_rope_head_dim,
            max_position=max_position_embeddings,
1051
            rope_parameters=config.rope_parameters,
1052
1053
            is_neox_style=False,
        )
1054
1055
1056
1057
1058

        if (
            config.rope_parameters["rope_type"] != "default"
            and config.rope_parameters["rope_type"] == "deepseek_yarn"
        ):
1059
1060
            mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False)
            scaling_factor = config.rope_parameters["factor"]
1061
1062
1063
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

zhuwenwen's avatar
zhuwenwen committed
1064
        self.is_v32 = hasattr(config, "index_topk")
1065
1066

        if self.is_v32:
1067
1068
1069
            self.indexer_rope_emb = get_rope(
                qk_rope_head_dim,
                max_position=max_position_embeddings,
1070
                rope_parameters=config.rope_parameters,
1071
1072
                is_neox_style=True,
            )
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
            self.indexer = Indexer(
                vllm_config,
                config,
                hidden_size,
                q_lora_rank,
                quant_config,
                cache_config,
                topk_indices_buffer,
                f"{prefix}.indexer",
            )
1083
        else:
1084
            self.indexer_rope_emb = None
1085
1086
            self.indexer = None

1087
1088
        mla_modules = MLAModules(
            kv_a_layernorm=self.kv_a_layernorm,
1089
            kv_b_proj=self.kv_b_proj,
1090
1091
1092
            rotary_emb=self.rotary_emb,
            o_proj=self.o_proj,
            fused_qkv_a_proj=self.fused_qkv_a_proj
1093
1094
            if self.q_lora_rank is not None
            else None,
1095
            kv_a_proj_with_mqa=self.kv_a_proj_with_mqa
1096
1097
1098
            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,
1099
1100
            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,
1101
            indexer=self.indexer,
1102
            indexer_rotary_emb=self.indexer_rope_emb,
1103
1104
            is_sparse=self.is_v32,
            topk_indices_buffer=topk_indices_buffer,
1105
        )
1106

1107
        self.mla_attn = MultiHeadLatentAttentionWrapper(
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
            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,
1120
1121
1122
1123
1124
1125
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1126
        llama_4_scaling: torch.Tensor | None,
1127
    ) -> torch.Tensor:
1128
        return self.mla_attn(positions, hidden_states, llama_4_scaling)
1129
1130


wangding zeng's avatar
wangding zeng committed
1131
class DeepseekV2DecoderLayer(nn.Module):
1132
1133
1134
1135
    def __init__(
        self,
        vllm_config: VllmConfig,
        prefix: str,
1136
1137
        config: DeepseekV2Config | None = None,
        topk_indices_buffer: torch.Tensor | None = None,
1138
    ) -> None:
wangding zeng's avatar
wangding zeng committed
1139
        super().__init__()
1140

1141
1142
        if config is None:
            config = vllm_config.model_config.hf_config
1143
1144
1145
1146
1147
        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
1148
        self.hidden_size = config.hidden_size
1149
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
1150
        moe_layer_freq = getattr(config, "moe_layer_freq", 1)
1151
1152
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
1153
        layer_idx = int(prefix.split(sep=".")[-1])
1154
        self.layer_idx = layer_idx
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164

        # 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)
        )

1165
1166
        self.use_mha = use_mha

1167
1168
1169
        if use_mha:
            attn_cls = DeepseekAttention
        elif model_config.use_mla:
1170
1171
1172
1173
            attn_cls = DeepseekV2MLAAttention
        else:
            attn_cls = DeepseekV2Attention
        self.self_attn = attn_cls(
1174
            vllm_config=vllm_config,
wangding zeng's avatar
wangding zeng committed
1175
1176
1177
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
1178
1179
1180
            qk_nope_head_dim=qk_nope_head_dim,
            qk_rope_head_dim=qk_rope_head_dim,
            v_head_dim=v_head_dim,
1181
            q_lora_rank=config.q_lora_rank if hasattr(config, "q_lora_rank") else None,
1182
            kv_lora_rank=kv_lora_rank,
wangding zeng's avatar
wangding zeng committed
1183
1184
1185
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
1186
            prefix=f"{prefix}.self_attn",
1187
            topk_indices_buffer=topk_indices_buffer,
wangding zeng's avatar
wangding zeng committed
1188
        )
1189

1190
1191
1192
        if (
            config.n_routed_experts is not None
            and layer_idx >= config.first_k_dense_replace
1193
            and layer_idx % moe_layer_freq == 0
1194
        ):
1195
1196
            self.mlp = DeepseekV2MoE(
                config=config,
1197
                parallel_config=parallel_config,
1198
1199
1200
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )
wangding zeng's avatar
wangding zeng committed
1201
1202
1203
1204
1205
1206
        else:
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
1207
                prefix=f"{prefix}.mlp",
wangding zeng's avatar
wangding zeng committed
1208
            )
1209
1210
1211
1212
        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
        )
1213
        self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
wangding zeng's avatar
wangding zeng committed
1214
1215
1216
1217
1218

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1219
        residual: torch.Tensor | None,
1220
        llama_4_scaling: torch.Tensor | None = None,
wangding zeng's avatar
wangding zeng committed
1221
    ) -> torch.Tensor:
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
        if envs.USE_FUSED_RMS_QUANT:
            # Fix residual FP16 overflow
            residual_fix_overflow = False
            
            assert self.input_layernorm.has_weight is True
            if residual is None:
                residual = hidden_states
                hidden_states, _ = self.self_attn(
                    positions = positions,
                    hidden_states = hidden_states,
                    rms_weight = self.input_layernorm.weight.data,
                    residual = None
                )
                residual_fix_overflow = True
            else:
                hidden_states, new_residual = self.self_attn(
                    positions = positions,
                    hidden_states = hidden_states,
                    rms_weight = self.input_layernorm.weight.data,
                    residual = residual
                )
                residual = new_residual
                
zhuwenwen's avatar
zhuwenwen committed
1245
            if hidden_states.dtype == torch.float16:
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
                # rmsnorm, and rmsnorm result would not affect by scale.
                hidden_states *= 1. / self.routed_scaling_factor
                if self.layer_idx == 0 or residual_fix_overflow:
                    # The residual is shared by all layers, we only scale it on
                    # first layer.
                    residual *= 1. / self.routed_scaling_factor

            hidden_states, new_resi = self.mlp(hidden_states, self.post_attention_layernorm.weight.data, residual)

            if isinstance(self.mlp,
zhuwenwen's avatar
zhuwenwen committed
1256
                        DeepseekV2MLP) and hidden_states.dtype == torch.float16:
1257
1258
1259
1260
1261
1262
1263
                # Fix FP16 overflow
                # Scaling the DeepseekV2MLP output, it is the input of
                # input_layernorm of next decoder layer.
                # The scaling of DeepseekV2MOE output would be done in the forward
                # of DeepseekV2MOE
                hidden_states *= 1. / self.routed_scaling_factor
            return hidden_states, new_resi
wangding zeng's avatar
wangding zeng committed
1264
        else:
1265
1266
1267
1268
            # Self Attention
            # Fix residual FP16 overflow
            residual_fix_overflow = False
            if residual is None:
1269
                residual = hidden_states.clone()
1270
1271
1272
                hidden_states = self.input_layernorm(hidden_states)
                residual_fix_overflow = True
            else:
1273
1274
                hidden_states, residual = self.input_layernorm(hidden_states, residual)
            
1275

1276
1277
1278
1279
1280
1281
1282
        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
1283

1284
1285
1286
1287
        if (
            not isinstance(self.self_attn, DeepseekAttention)
            and hidden_states.dtype == torch.float16
        ):
1288
1289
1290
            # Fix FP16 overflow
            # We scale both hidden_states and residual before
            # rmsnorm, and rmsnorm result would not affect by scale.
1291
            hidden_states *= 1.0 / self.routed_scaling_factor
1292
1293
1294
            if self.layer_idx == 0:
                # The residual is shared by all layers, we only scale it on
                # first layer.
1295
                residual *= 1.0 / self.routed_scaling_factor
1296
1297

        # Fully Connected
1298
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
wangding zeng's avatar
wangding zeng committed
1299
        hidden_states = self.mlp(hidden_states)
1300

1301
        if isinstance(self.mlp, DeepseekV2MLP) and hidden_states.dtype == torch.float16:
1302
1303
1304
1305
1306
            # 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
1307
            hidden_states *= 1.0 / self.routed_scaling_factor
1308

wangding zeng's avatar
wangding zeng committed
1309
1310
1311
        return hidden_states, residual


1312
@support_torch_compile
wangding zeng's avatar
wangding zeng committed
1313
1314
1315
class DeepseekV2Model(nn.Module):
    fall_back_to_pt_during_load = False

1316
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
1317
        super().__init__()
1318
1319
1320

        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
1321
        self.config = config
1322
        self.device = current_platform.device_type
1323

wangding zeng's avatar
wangding zeng committed
1324
        self.vocab_size = config.vocab_size
zhuwenwen's avatar
zhuwenwen committed
1325
        self.is_v32 = hasattr(config, "index_topk")
1326
1327
1328
1329
1330
1331
        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,
1332
                device=self.device,
1333
            )
1334
1335
        else:
            topk_indices_buffer = None
wangding zeng's avatar
wangding zeng committed
1336

1337
1338
1339
1340
        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
1341
                quant_config=quant_config,
1342
1343
                prefix=f"{prefix}.embed_tokens",
            )
1344
1345
1346
1347
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
1348
            lambda prefix: DeepseekV2DecoderLayer(
1349
                vllm_config, prefix, topk_indices_buffer=topk_indices_buffer
1350
1351
1352
            ),
            prefix=f"{prefix}.layers",
        )
1353
1354
1355
1356
1357

        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
1358
1359
1360
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
wangding zeng's avatar
wangding zeng committed
1361

1362
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
1363
1364
        return self.embed_tokens(input_ids)

wangding zeng's avatar
wangding zeng committed
1365
1366
1367
1368
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1369
1370
1371
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
1372
        if get_pp_group().is_first_rank:
1373
1374
1375
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
1376
                hidden_states = self.embed_input_ids(input_ids)
1377
1378
1379
1380
1381
1382
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
        # 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

1397
        for layer in islice(self.layers, self.start_layer, self.end_layer):
1398
1399
1400
            hidden_states, residual = layer(
                positions, hidden_states, residual, llama_4_scaling
            )
1401
1402

        if not get_pp_group().is_last_rank:
1403
1404
1405
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
1406

wangding zeng's avatar
wangding zeng committed
1407
1408
1409
1410
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
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(
1453
    nn.Module, SupportsPP, DeepseekV2MixtureOfExperts, SupportsLoRA, SupportsEagle
1454
):
1455
1456
1457
    packed_modules_mapping = {
        "gate_up_proj": ["gate_proj", "up_proj"],
    }
1458
    model_cls = DeepseekV2Model
wangding zeng's avatar
wangding zeng committed
1459

1460
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
1461
        super().__init__()
1462
1463
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
1464
1465
1466
1467
1468
1469
1470

        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'

1471
        self.use_w4a16_moe_sz = os.environ.get('AWQ_MOE_SZ') == '1'
wangding zeng's avatar
wangding zeng committed
1472
1473
        self.config = config
        self.quant_config = quant_config
王敏's avatar
王敏 committed
1474

1475
1476
1477
1478
1479
1480
1481
1482
1483
        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"]

1484
1485
1486
1487
        # `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.
1488
1489
1490
        self.fuse_qkv_a_proj = (
            hasattr(config, "q_lora_rank") and config.q_lora_rank is not None
        )
1491
1492
1493
1494
1495
1496
        if self.fuse_qkv_a_proj:
            self.packed_modules_mapping["fused_qkv_a_proj"] = [
                "q_a_proj",
                "kv_a_proj_with_mqa",
            ]

1497
        self.model = self.model_cls(
1498
1499
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
1500
        if get_pp_group().is_last_rank:
1501
1502
1503
1504
1505
1506
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
1507
1508
        else:
            self.lm_head = PPMissingLayer()
wangding zeng's avatar
wangding zeng committed
1509
        self.logits_processor = LogitsProcessor(config.vocab_size)
1510
        self.make_empty_intermediate_tensors = (
1511
1512
            self.model.make_empty_intermediate_tensors
        )
1513
1514
1515
1516
1517
1518
1519
        # 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):
1520
1521
        self.expert_weights = []

1522
        self.num_expert_groups = getattr(self.config, "n_group", 1)
1523

1524
1525
        self.moe_layers = []
        self.moe_mlp_layers = []
1526
        example_moe = None
1527
        for layer in self.model.layers:
1528
1529
1530
            if isinstance(layer, PPMissingLayer):
                continue

1531
1532
            assert isinstance(layer, DeepseekV2DecoderLayer)
            if isinstance(layer.mlp, DeepseekV2MoE):
1533
1534
                # Pick last one layer since the first ones may be dense layers.
                example_moe = layer.mlp
1535
                self.moe_mlp_layers.append(layer.mlp)
1536
1537
                self.moe_layers.append(layer.mlp.experts)

1538
        self.extract_moe_parameters(example_moe)
王敏's avatar
王敏 committed
1539
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
1540
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
1541
        self.tritonsingleton= W8a8GetCacheJSON() 
zhuwenwen's avatar
zhuwenwen committed
1542
        self.tritonsingleton.topk = self.config.num_experts_per_tok
1543
        self.tritonsingleton.quant_method=self.quant_method 
1544

1545
1546
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
1547

wangding zeng's avatar
wangding zeng committed
1548
1549
1550
1551
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1552
1553
1554
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
1555
1556
1557
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
wangding zeng's avatar
wangding zeng committed
1558
1559
        return hidden_states

1560
1561
1562
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
1563
    ) -> torch.Tensor | None:
1564
        logits = self.logits_processor(self.lm_head, hidden_states)
wangding zeng's avatar
wangding zeng committed
1565
1566
        return logits

1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
    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,
        )
1577

1578
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1579
1580
1581
        rocm_aiter_moe_shared_expert_enabled = (
            rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
        )
wangding zeng's avatar
wangding zeng committed
1582
1583
1584
1585
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
1586
1587
        ]
        mla_params_mapping = [
1588
1589
            ("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
1590
        ]
1591
1592
1593
1594
1595
1596
1597
1598
1599
        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
1600

1601
1602
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
1603
        expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
王敏's avatar
王敏 committed
1604
1605
1606
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
1607
1608
1609
            num_experts=self.config.n_routed_experts
            + (
                self.config.n_shared_experts
1610
                if rocm_aiter_moe_shared_expert_enabled
1611
1612
                else 0
            ),
1613
1614
            num_redundant_experts=self.num_redundant_experts,
        )
1615

wangding zeng's avatar
wangding zeng committed
1616
        params_dict = dict(self.named_parameters())
1617
        loaded_params: set[str] = set()
wangding zeng's avatar
wangding zeng committed
1618
1619
1620
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
1621

1622
1623
1624
            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
1625

1626
1627
            is_fusion_moe_shared_experts_layer = (
                rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
1628
1629
            )

1630
            for param_name, weight_name, shard_id in stacked_params_mapping:
1631
                # Skip non-stacked layers and experts (experts handled below).
wangding zeng's avatar
wangding zeng committed
1632
1633
                if weight_name not in name:
                    continue
1634
1635
1636
1637
1638
1639
                # 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.
1640
                if ("mlp.experts." in name) and name not in params_dict:
1641
                    continue
1642
                if is_fusion_moe_shared_experts_layer:
1643
                    continue
1644
                name_mapped = name.replace(weight_name, param_name)
1645
1646
1647

                # QKV fusion is optional, fall back to normal
                # weight loading if it's not enabled
1648
                # if go with fusion option, then update name
1649
1650
1651
                if (
                    param_name == "fused_qkv_a_proj"
                ) and name_mapped not in params_dict:
1652
                    continue
1653
1654
                else:
                    name = name_mapped
wangding zeng's avatar
wangding zeng committed
1655
1656
1657
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
1658
1659
1660
1661

                if is_pp_missing_parameter(name, self):
                    continue

wangding zeng's avatar
wangding zeng committed
1662
1663
1664
1665
1666
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
1667
1668
                is_expert_weight = False

1669
1670
1671
1672
1673
1674
1675
1676
1677
                # 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
1678
                if is_fusion_moe_shared_experts_layer:
1679
1680
1681
1682
1683
1684
1685
1686
1687
                    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}"
1688
                    )
1689
1690
1691
1692
1693
1694
                    chunk_size = total // num_chunks

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

1695
                    if is_fusion_moe_shared_experts_layer:
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
                        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:
1746
                            if not is_fusion_moe_shared_experts_layer:
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
                                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
1769

1770
                        param = params_dict[name]
1771
1772
1773
1774
                        weight_loader = getattr(
                            param, "weight_loader", default_weight_loader
                        )
                        weight_loader(param, loaded_weight)
1775
            if not is_fusion_moe_shared_experts_layer:
1776
                loaded_params.add(name)
1777
                
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
        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
1807
            
1808
        return loaded_params
1809
1810


1811
1812
1813
1814
class DeepseekForCausalLM(DeepseekV2ForCausalLM):
    pass


1815
1816
class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass
1817
1818


1819
1820
# Compatibility with
# https://huggingface.co/deepseek-ai/DeepSeek-V3-Base/blob/main/configuration_deepseek.py
1821
def get_spec_layer_idx_from_weight_name(
1822
1823
    config: DeepseekV2Config | DeepseekV3Config, weight_name: str
) -> int | None:
1824
1825
1826
1827
    if (
        hasattr(config, "num_nextn_predict_layers")
        and config.num_nextn_predict_layers > 0
    ):
1828
1829
        layer_idx = config.num_hidden_layers
        for i in range(config.num_nextn_predict_layers):
1830
            if weight_name.startswith(f"model.layers.{layer_idx + i}."):
1831
1832
                return layer_idx + i
    return None