deepseek_v2.py 86.5 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
# Copyright 2023-2024 SGLang Team
# 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.
# ==============================================================================
14

Liangsheng Yin's avatar
Liangsheng Yin committed
15
16
17
# Adapted from:
# https://github.com/vllm-project/vllm/blob/fb6af8bc086328ca6659e72d11ffd4309ce4de22/vllm/model_executor/models/deepseek_v2.py
"""Inference-only DeepseekV2 model."""
18

19
import logging
20
import os
21
from enum import IntEnum, auto
Liangsheng Yin's avatar
Liangsheng Yin committed
22
23
24
from typing import Any, Dict, Iterable, Optional, Tuple

import torch
Ke Bao's avatar
Ke Bao committed
25
import torch.nn.functional as F
Liangsheng Yin's avatar
Liangsheng Yin committed
26
from torch import nn
27
from tqdm import tqdm
Liangsheng Yin's avatar
Liangsheng Yin committed
28
from transformers import PretrainedConfig
29
30

from sglang.srt.distributed import (
Liangsheng Yin's avatar
Liangsheng Yin committed
31
    get_tensor_model_parallel_world_size,
32
    parallel_state,
Liangsheng Yin's avatar
Liangsheng Yin committed
33
34
    tensor_model_parallel_all_reduce,
)
35
from sglang.srt.layers.activation import SiluAndMul
36
37
38
39
40
from sglang.srt.layers.communicator import (
    LayerCommunicator,
    LayerScatterModes,
    enable_moe_dense_fully_dp,
)
Lianmin Zheng's avatar
Lianmin Zheng committed
41
42
43
from sglang.srt.layers.dp_attention import (
    get_attention_tp_rank,
    get_attention_tp_size,
44
    get_local_attention_dp_size,
Lianmin Zheng's avatar
Lianmin Zheng committed
45
)
46
from sglang.srt.layers.layernorm import RMSNorm
47
48
49
50
51
52
from sglang.srt.layers.linear import (
    ColumnParallelLinear,
    MergedColumnParallelLinear,
    ReplicatedLinear,
    RowParallelLinear,
)
Liangsheng Yin's avatar
Liangsheng Yin committed
53
from sglang.srt.layers.logits_processor import LogitsProcessor
fzyzcjy's avatar
fzyzcjy committed
54
from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, get_moe_impl_class
Lianmin Zheng's avatar
Lianmin Zheng committed
55
from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
56
from sglang.srt.layers.moe.topk import select_experts
57
from sglang.srt.layers.quantization import deep_gemm_wrapper
58
from sglang.srt.layers.quantization.base_config import QuantizationConfig
59
from sglang.srt.layers.quantization.fp8_kernel import (
60
    is_fp8_fnuz,
61
    per_tensor_quant_mla_fp8,
62
    per_token_group_quant_mla_deep_gemm_masked_fp8,
63
)
HandH1998's avatar
HandH1998 committed
64
from sglang.srt.layers.quantization.fp8_utils import (
65
    block_quant_dequant,
HandH1998's avatar
HandH1998 committed
66
    block_quant_to_tensor_quant,
67
    channel_quant_to_tensor_quant,
68
    normalize_e4m3fn_to_e4m3fnuz,
69
    requant_weight_ue8m0_inplace,
HandH1998's avatar
HandH1998 committed
70
)
71
72
73
from sglang.srt.layers.quantization.int8_utils import (
    block_dequant as int8_block_dequant,
)
Liangsheng Yin's avatar
Liangsheng Yin committed
74
from sglang.srt.layers.radix_attention import RadixAttention
75
from sglang.srt.layers.rotary_embedding import get_rope, get_rope_wrapper
76
77
78
79
from sglang.srt.layers.vocab_parallel_embedding import (
    ParallelLMHead,
    VocabParallelEmbedding,
)
80
81
82
83
from sglang.srt.managers.expert_distribution import (
    get_global_expert_distribution_recorder,
)
from sglang.srt.managers.expert_location import ModelConfigForExpertLocation
84
from sglang.srt.managers.expert_location_dispatch import ExpertLocationDispatchInfo
85
from sglang.srt.managers.schedule_batch import global_server_args_dict
86
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
87
from sglang.srt.model_loader.weight_utils import default_weight_loader
88
89
90
91
from sglang.srt.two_batch_overlap import (
    MaybeTboDeepEPDispatcher,
    model_forward_maybe_tbo,
)
92
93
94
from sglang.srt.utils import (
    BumpAllocator,
    DeepEPMode,
95
    LazyValue,
96
    add_prefix,
97
    bind_or_assign,
98
    cpu_has_amx_support,
99
100
    get_bool_env_var,
    get_int_env_var,
101
    is_cpu,
102
103
    is_cuda,
    is_hip,
104
    is_non_idle_and_non_empty,
105
    log_info_on_rank0,
106
)
107

108
_is_hip = is_hip()
Yineng Zhang's avatar
Yineng Zhang committed
109
_is_cuda = is_cuda()
110
_is_fp8_fnuz = is_fp8_fnuz()
111
_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
112
113
_is_cpu_amx_available = cpu_has_amx_support()
_is_cpu = is_cpu()
114

Yineng Zhang's avatar
Yineng Zhang committed
115
if _is_cuda:
116
    from sgl_kernel import awq_dequantize, bmm_fp8, merge_state_v2
117
118
elif _is_cpu and _is_cpu_amx_available:
    pass
Yineng Zhang's avatar
Yineng Zhang committed
119
else:
Lianmin Zheng's avatar
Lianmin Zheng committed
120
    from vllm._custom_ops import awq_dequantize
Liangsheng Yin's avatar
Liangsheng Yin committed
121

122
123
124
125
126
if _is_hip:
    from sglang.srt.layers.attention.triton_ops.rocm_mla_decode_rope import (
        decode_attention_fwd_grouped_rope,
    )

127
128
129
if _use_aiter:
    from aiter.rotary_embedding import get_rope

130
131
logger = logging.getLogger(__name__)

Liangsheng Yin's avatar
Liangsheng Yin committed
132

133
134
135
136
137
138
139
140
141
142
143
class AttnForwardMethod(IntEnum):
    # Use multi-head attention
    MHA = auto()

    # Use absorbed multi-latent attention
    MLA = auto()

    # Use multi-head attention, but with KV cache chunked.
    # This method can avoid OOM when prefix lengths are long.
    MHA_CHUNKED_KV = auto()

144
145
146
    # Use MLA but with fused RoPE
    MLA_FUSED_ROPE = auto()

147

Liangsheng Yin's avatar
Liangsheng Yin committed
148
149
150
151
152
153
154
155
class DeepseekV2MLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: bool = True,
156
        prefix: str = "",
157
158
        tp_rank: Optional[int] = None,
        tp_size: Optional[int] = None,
Liangsheng Yin's avatar
Liangsheng Yin committed
159
160
    ) -> None:
        super().__init__()
161
162
        self.tp_size = tp_size

Liangsheng Yin's avatar
Liangsheng Yin committed
163
        self.gate_up_proj = MergedColumnParallelLinear(
164
165
166
167
168
            hidden_size,
            [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("gate_up_proj", prefix),
169
170
            tp_rank=tp_rank,
            tp_size=tp_size,
Liangsheng Yin's avatar
Liangsheng Yin committed
171
172
173
174
175
176
177
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=reduce_results,
178
            prefix=add_prefix("down_proj", prefix),
179
180
            tp_rank=tp_rank,
            tp_size=tp_size,
Liangsheng Yin's avatar
Liangsheng Yin committed
181
182
183
184
185
186
187
188
        )
        if hidden_act != "silu":
            raise ValueError(
                f"Unsupported activation: {hidden_act}. "
                "Only silu is supported for now."
            )
        self.act_fn = SiluAndMul()

189
190
191
192
    def forward(self, x, forward_batch=None):
        if (self.tp_size == 1) and x.shape[0] == 0:
            return x

Liangsheng Yin's avatar
Liangsheng Yin committed
193
194
195
196
197
198
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


Ke Bao's avatar
Ke Bao committed
199
class MoEGate(nn.Module):
200
201
202
203
204
    def __init__(
        self,
        config,
        prefix: str = "",
    ):
Ke Bao's avatar
Ke Bao committed
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
        super().__init__()
        self.weight = nn.Parameter(
            torch.empty((config.n_routed_experts, config.hidden_size))
        )
        if config.topk_method == "noaux_tc":
            self.e_score_correction_bias = nn.Parameter(
                torch.empty((config.n_routed_experts))
            )
        else:
            self.e_score_correction_bias = None

    def forward(self, hidden_states):
        logits = F.linear(hidden_states, self.weight, None)
        return logits


Liangsheng Yin's avatar
Liangsheng Yin committed
221
222
223
224
225
class DeepseekV2MoE(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
fzyzcjy's avatar
fzyzcjy committed
226
        layer_id: int,
Liangsheng Yin's avatar
Liangsheng Yin committed
227
        quant_config: Optional[QuantizationConfig] = None,
228
        prefix: str = "",
229
        alt_stream: Optional[torch.cuda.Stream] = None,
Liangsheng Yin's avatar
Liangsheng Yin committed
230
231
232
233
234
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.routed_scaling_factor = config.routed_scaling_factor
        self.n_shared_experts = config.n_shared_experts
235
236
237
238
239
        self.num_fused_shared_experts = (
            0
            if global_server_args_dict["disable_shared_experts_fusion"]
            else config.n_shared_experts
        )
240
        self.config = config
fzyzcjy's avatar
fzyzcjy committed
241
        self.layer_id = layer_id
242
        self.alt_stream = alt_stream
243

Liangsheng Yin's avatar
Liangsheng Yin committed
244
245
246
247
248
249
250
251
252
253
254
255
        if self.tp_size > config.n_routed_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {config.n_routed_experts}."
            )

        if config.hidden_act != "silu":
            raise ValueError(
                f"Unsupported activation: {config.hidden_act}. "
                "Only silu is supported for now."
            )

256
        self.gate = MoEGate(config=config, prefix=add_prefix("gate", prefix))
Ke Bao's avatar
Ke Bao committed
257

258
        self.experts = get_moe_impl_class()(
259
            num_experts=config.n_routed_experts
260
            + self.num_fused_shared_experts
261
            + global_server_args_dict["ep_num_redundant_experts"],
262
            top_k=config.num_experts_per_tok + self.num_fused_shared_experts,
263
264
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
fzyzcjy's avatar
fzyzcjy committed
265
            layer_id=self.layer_id,
266
267
268
269
            renormalize=config.norm_topk_prob,
            quant_config=quant_config,
            use_grouped_topk=True,
            num_expert_group=config.n_group,
270
            num_fused_shared_experts=self.num_fused_shared_experts,
271
272
            topk_group=config.topk_group,
            correction_bias=self.gate.e_score_correction_bias,
273
            routed_scaling_factor=self.routed_scaling_factor,
274
275
276
277
278
279
            prefix=add_prefix("experts", prefix),
            **(
                dict(deepep_mode=DeepEPMode[global_server_args_dict["deepep_mode"]])
                if global_server_args_dict["enable_deepep_moe"]
                else {}
            ),
280
281
282
283
284
285
286
287
288
            # Additional args for FusedMoE
            **(
                dict(
                    enable_flashinfer_moe=True,
                    enable_ep_moe=global_server_args_dict["enable_ep_moe"],
                )
                if global_server_args_dict["enable_flashinfer_moe"]
                else {}
            ),
289
        )
Liangsheng Yin's avatar
Liangsheng Yin committed
290

291
        if config.n_shared_experts is not None and self.num_fused_shared_experts == 0:
Liangsheng Yin's avatar
Liangsheng Yin committed
292
            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
293
            # disable tp for shared experts when enable deepep moe
294
295
296
297
298
299
300
301
302
303
304
305
306
            self.shared_experts = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                reduce_results=False,
                prefix=add_prefix("shared_experts", prefix),
                **(
                    dict(tp_rank=0, tp_size=1)
                    if global_server_args_dict["enable_deepep_moe"]
                    else {}
                ),
            )
307

308
309
        self.top_k = config.num_experts_per_tok

310
        if global_server_args_dict["enable_deepep_moe"]:
311
312
            # TODO: we will support tp < ep in the future
            self.ep_size = get_tensor_model_parallel_world_size()
313
314
315
316
            self.num_experts = (
                config.n_routed_experts
                + global_server_args_dict["ep_num_redundant_experts"]
            )
317
318
319
320
321
322
323
324
325
            self.renormalize = config.norm_topk_prob
            self.topk_group = config.topk_group
            self.num_expert_group = config.n_group
            self.correction_bias = (
                self.gate.e_score_correction_bias.data
                if self.gate.e_score_correction_bias is not None
                else None
            )

326
            self.deepep_dispatcher = MaybeTboDeepEPDispatcher(
327
328
329
                group=parallel_state.get_tp_group().device_group,
                router_topk=self.top_k,
                permute_fusion=True,
330
                num_experts=self.num_experts,
331
                num_local_experts=config.n_routed_experts // self.tp_size,
Liangsheng Yin's avatar
Liangsheng Yin committed
332
                hidden_size=config.hidden_size,
333
                params_dtype=config.torch_dtype,
334
                deepep_mode=DeepEPMode[global_server_args_dict["deepep_mode"]],
335
                async_finish=True,
336
                return_recv_hook=True,
Liangsheng Yin's avatar
Liangsheng Yin committed
337
338
            )

339
        self._enable_deepep_moe = global_server_args_dict["enable_deepep_moe"]
340

341
342
343
344
345
346
347
    def get_moe_weights(self):
        return [
            x.data
            for name, x in self.experts.named_parameters()
            if name not in ["correction_bias"]
        ]

348
349
350
351
    def forward(
        self, hidden_states: torch.Tensor, forward_batch: Optional[ForwardBatch] = None
    ) -> torch.Tensor:
        if not self._enable_deepep_moe:
352
353
354
355
356
357
358
359
360
            DUAL_STREAM_TOKEN_THRESHOLD = 1024
            if (
                self.alt_stream is not None
                and self.num_fused_shared_experts == 0
                and hidden_states.shape[0] <= DUAL_STREAM_TOKEN_THRESHOLD
            ):
                return self.forward_normal_dual_stream(hidden_states)
            else:
                return self.forward_normal(hidden_states)
361
362
363
        else:
            return self.forward_deepep(hidden_states, forward_batch)

364
    def forward_normal_dual_stream(self, hidden_states: torch.Tensor) -> torch.Tensor:
365
366
367
        # router_logits: (num_tokens, n_experts)
        router_logits = self.gate(hidden_states)

368
369
370
        current_stream = torch.cuda.current_stream()
        self.alt_stream.wait_stream(current_stream)
        shared_output = self._forward_shared_experts(hidden_states)
371

372
373
374
375
376
377
378
379
380
381
382
383
        with torch.cuda.stream(self.alt_stream):
            final_hidden_states = self.experts(
                hidden_states=hidden_states, router_logits=router_logits
            )
            if not _is_cuda:
                final_hidden_states *= self.routed_scaling_factor
        current_stream.wait_stream(self.alt_stream)
        final_hidden_states = final_hidden_states + shared_output
        if self.tp_size > 1:
            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
        return final_hidden_states

384
385
386
387
388
389
390
    def forward_normal(self, hidden_states: torch.Tensor) -> torch.Tensor:
        shared_output = self._forward_shared_experts(hidden_states)
        # router_logits: (num_tokens, n_experts)
        router_logits = self.gate(hidden_states)
        final_hidden_states = self.experts(
            hidden_states=hidden_states, router_logits=router_logits
        )
391
392
        if not _is_cuda:
            final_hidden_states *= self.routed_scaling_factor
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
        if shared_output is not None:
            final_hidden_states = final_hidden_states + shared_output
        if self.tp_size > 1:
            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
        return final_hidden_states

    def forward_deepep(
        self, hidden_states: torch.Tensor, forward_batch: ForwardBatch
    ) -> torch.Tensor:
        forward_mode = forward_batch.forward_mode
        shared_output = None
        if is_non_idle_and_non_empty(forward_mode, hidden_states):
            # router_logits: (num_tokens, n_experts)
            router_logits = self.gate(hidden_states)
            shared_output = self._forward_shared_experts(hidden_states)
            topk_weights, topk_idx = select_experts(
                hidden_states=hidden_states,
                router_logits=router_logits,
                top_k=self.top_k,
                use_grouped_topk=True,
                renormalize=self.renormalize,
                topk_group=self.topk_group,
                num_expert_group=self.num_expert_group,
416
                num_fused_shared_experts=self.num_fused_shared_experts,
417
418
419
                correction_bias=self.correction_bias,
                routed_scaling_factor=self.routed_scaling_factor,
                num_token_non_padded=forward_batch.num_token_non_padded,
420
421
422
                expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new(
                    layer_id=self.layer_id,
                ),
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
            )
        else:
            topk_idx = torch.full(
                (0, self.top_k), -1, dtype=torch.int, device=hidden_states.device
            )
            topk_weights = torch.empty(
                (0, self.top_k), dtype=torch.float32, device=hidden_states.device
            )
        if self.ep_size > 1:
            # TODO(ch-wan): allow users to set num_max_dispatch_tokens_per_rank value
            (
                hidden_states,
                topk_idx,
                topk_weights,
                reorder_topk_ids,
                num_recv_tokens_per_expert,
                seg_indptr,
                masked_m,
                expected_m,
            ) = self.deepep_dispatcher.dispatch(
                hidden_states=hidden_states,
                topk_idx=topk_idx,
                topk_weights=topk_weights,
                forward_mode=forward_mode,
            )
        final_hidden_states = self.experts(
            hidden_states=hidden_states,
            topk_idx=topk_idx,
            topk_weights=topk_weights,
            reorder_topk_ids=reorder_topk_ids,
            seg_indptr=seg_indptr,
            masked_m=masked_m,
            expected_m=expected_m,
            num_recv_tokens_per_expert=num_recv_tokens_per_expert,
            forward_mode=forward_mode,
        )
        if self.ep_size > 1:
            final_hidden_states = self.deepep_dispatcher.combine(
                hidden_states=final_hidden_states,
                topk_idx=topk_idx,
                topk_weights=topk_weights,
                forward_mode=forward_mode,
            )

        if shared_output is not None:
468
469
470
471
472
            x = shared_output
            x.add_(final_hidden_states, alpha=self.routed_scaling_factor)
            final_hidden_states = x
        else:
            final_hidden_states *= self.routed_scaling_factor
473
474
475
476

        return final_hidden_states

    def _forward_shared_experts(self, hidden_states):
477
        if self.num_fused_shared_experts == 0:
478
479
480
481
            return self.shared_experts(hidden_states)
        else:
            return None

482
    def op_gate(self, state):
483
        if is_non_idle_and_non_empty(
484
            state.forward_batch.forward_mode, state.hidden_states_mlp_input
485
        ):
486
            # router_logits: (num_tokens, n_experts)
487
            state.router_logits = self.gate(state.hidden_states_mlp_input)
488
        else:
489
            state.router_logits = None
490

491
    def op_shared_experts(self, state):
492
        hidden_states_mlp_input = state.pop("hidden_states_mlp_input")
493
        if (self.num_fused_shared_experts == 0) and is_non_idle_and_non_empty(
494
            state.forward_batch.forward_mode, hidden_states_mlp_input
495
        ):
496
            state.shared_output = self.shared_experts(hidden_states_mlp_input)
497
        else:
498
            state.shared_output = None
499

500
    def op_select_experts(self, state):
501
        router_logits = state.pop("router_logits")
502
503
        hidden_states = state.hidden_states_mlp_input

504
        if router_logits is not None:
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
            with get_global_expert_distribution_recorder().with_current_layer(
                self.layer_id
            ):
                state.topk_weights_local, state.topk_idx_local = select_experts(
                    hidden_states=hidden_states,
                    router_logits=router_logits,
                    top_k=self.top_k,
                    use_grouped_topk=True,
                    renormalize=self.renormalize,
                    topk_group=self.topk_group,
                    num_expert_group=self.num_expert_group,
                    num_fused_shared_experts=self.num_fused_shared_experts,
                    correction_bias=self.correction_bias,
                    routed_scaling_factor=self.routed_scaling_factor,
                    num_token_non_padded=state.forward_batch.num_token_non_padded,
                    expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new(
                        layer_id=self.layer_id,
                    ),
                )
524
525
526
527
528
529
530
        else:
            state.topk_idx_local = torch.full(
                (0, self.top_k), -1, dtype=torch.int, device=hidden_states.device
            )
            state.topk_weights_local = torch.empty(
                (0, self.top_k), dtype=torch.float32, device=hidden_states.device
            )
531

532
    def op_dispatch_a(self, state):
533
        if self.ep_size > 1:
534
            # TODO(ch-wan): allow users to set num_max_dispatch_tokens_per_rank value
535
            self.deepep_dispatcher.dispatch_a(
536
                hidden_states=state.hidden_states_mlp_input,
537
538
539
                topk_idx=state.pop("topk_idx_local"),
                topk_weights=state.pop("topk_weights_local"),
                forward_mode=state.forward_batch.forward_mode,
540
                tbo_subbatch_index=state.get("tbo_subbatch_index"),
541
            )
542

543
    def op_dispatch_b(self, state):
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
        if self.ep_size > 1:
            with get_global_expert_distribution_recorder().with_current_layer(
                self.layer_id
            ):
                (
                    state.hidden_states_experts_input,
                    state.topk_idx_dispatched,
                    state.topk_weights_dispatched,
                    state.reorder_topk_ids,
                    state.num_recv_tokens_per_expert,
                    state.seg_indptr,
                    state.masked_m,
                    state.expected_m,
                ) = self.deepep_dispatcher.dispatch_b(
                    tbo_subbatch_index=state.get("tbo_subbatch_index"),
                )
560
561

    def op_experts(self, state):
562
563
564
565
566
567
568
569
570
571
572
        state.hidden_states_experts_output = self.experts(
            hidden_states=state.pop("hidden_states_experts_input"),
            topk_idx=state.topk_idx_dispatched,
            topk_weights=state.topk_weights_dispatched,
            reorder_topk_ids=state.pop("reorder_topk_ids"),
            seg_indptr=state.pop("seg_indptr"),
            masked_m=state.pop("masked_m"),
            expected_m=state.pop("expected_m"),
            num_recv_tokens_per_expert=state.pop("num_recv_tokens_per_expert"),
            forward_mode=state.forward_batch.forward_mode,
        )
573

574
    def op_combine_a(self, state):
575
        if self.ep_size > 1:
576
            self.deepep_dispatcher.combine_a(
577
                hidden_states=state.pop("hidden_states_experts_output"),
578
579
580
                topk_idx=state.pop("topk_idx_dispatched"),
                topk_weights=state.pop("topk_weights_dispatched"),
                forward_mode=state.forward_batch.forward_mode,
581
                tbo_subbatch_index=state.get("tbo_subbatch_index"),
582
            )
583

584
    def op_combine_b(self, state):
585
586
587
588
        if self.ep_size > 1:
            state.hidden_states_after_combine = self.deepep_dispatcher.combine_b(
                tbo_subbatch_index=state.get("tbo_subbatch_index"),
            )
589
590

    def op_output(self, state):
591
        final_hidden_states = state.pop("hidden_states_after_combine")
592
593
594
595
596
597
598

        if (shared_output := state.pop("shared_output")) is not None:
            x = shared_output
            x.add_(final_hidden_states, alpha=self.routed_scaling_factor)
            final_hidden_states = x
        else:
            final_hidden_states *= self.routed_scaling_factor
Liangsheng Yin's avatar
Liangsheng Yin committed
599

600
        state.hidden_states_mlp_output = final_hidden_states
601

Liangsheng Yin's avatar
Liangsheng Yin committed
602
603
604
605
606
607
608
609
610

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

    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
class DeepseekV2AttentionMLA(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
        q_lora_rank: int,
        kv_lora_rank: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[Dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        quant_config: Optional[QuantizationConfig] = None,
Lianmin Zheng's avatar
Lianmin Zheng committed
627
628
        reduce_results: bool = True,
        layer_id: int = None,
629
        prefix: str = "",
630
        alt_stream: Optional[torch.cuda.Stream] = None,
631
632
633
634
635
636
637
638
639
640
    ) -> None:
        super().__init__()
        self.layer_id = layer_id
        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
Lianmin Zheng's avatar
Lianmin Zheng committed
641
642
643
        attn_tp_rank = get_attention_tp_rank()
        attn_tp_size = get_attention_tp_size()

644
        self.num_heads = num_heads
Lianmin Zheng's avatar
Lianmin Zheng committed
645
646
        assert num_heads % attn_tp_size == 0
        self.num_local_heads = num_heads // attn_tp_size
647
648
649
650
        self.scaling = self.qk_head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

Lianmin Zheng's avatar
Lianmin Zheng committed
651
652
        # For tensor parallel attention
        if self.q_lora_rank is not None:
653
            self.fused_qkv_a_proj_with_mqa = ReplicatedLinear(
Ke Bao's avatar
Ke Bao committed
654
                self.hidden_size,
655
                self.q_lora_rank + self.kv_lora_rank + self.qk_rope_head_dim,
656
657
                bias=False,
                quant_config=quant_config,
658
                prefix=add_prefix("fused_qkv_a_proj_with_mqa", prefix),
659
            )
Lianmin Zheng's avatar
Lianmin Zheng committed
660
661
662
663
            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,
Ke Bao's avatar
Ke Bao committed
664
665
                bias=False,
                quant_config=quant_config,
Lianmin Zheng's avatar
Lianmin Zheng committed
666
667
668
                prefix=add_prefix("q_b_proj", prefix),
                tp_rank=attn_tp_rank,
                tp_size=attn_tp_size,
Ke Bao's avatar
Ke Bao committed
669
            )
Lianmin Zheng's avatar
Lianmin Zheng committed
670
671
        else:
            self.q_proj = ColumnParallelLinear(
672
                self.hidden_size,
Lianmin Zheng's avatar
Lianmin Zheng committed
673
                self.num_heads * self.qk_head_dim,
674
675
                bias=False,
                quant_config=quant_config,
Lianmin Zheng's avatar
Lianmin Zheng committed
676
677
678
                prefix=add_prefix("q_proj", prefix),
                tp_rank=attn_tp_rank,
                tp_size=attn_tp_size,
679
            )
680
681
682
683
684
685
686
687
            self.kv_a_proj_with_mqa = ReplicatedLinear(
                self.hidden_size,
                self.kv_lora_rank + self.qk_rope_head_dim,
                bias=False,
                quant_config=quant_config,
                prefix=add_prefix("kv_a_proj_with_mqa", prefix),
            )

Lianmin Zheng's avatar
Lianmin Zheng committed
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("kv_b_proj", prefix),
            tp_rank=attn_tp_rank,
            tp_size=attn_tp_size,
        )
        # O projection.
        self.o_proj = RowParallelLinear(
            self.num_heads * self.v_head_dim,
            self.hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=add_prefix("o_proj", prefix),
            tp_rank=attn_tp_rank,
            tp_size=attn_tp_size,
        )
708
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
Ke Bao's avatar
Ke Bao committed
709
710
711
712

        if rope_scaling:
            rope_scaling["rope_type"] = "deepseek_yarn"

713
        self.rotary_emb = get_rope_wrapper(
714
715
716
717
718
719
            qk_rope_head_dim,
            rotary_dim=qk_rope_head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
            is_neox_style=False,
720
            device=global_server_args_dict["device"],
721
722
723
724
725
726
727
        )

        if rope_scaling:
            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
            scaling_factor = rope_scaling["factor"]
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale
Ke Bao's avatar
Ke Bao committed
728
729
        else:
            self.rotary_emb.forward = self.rotary_emb.forward_native
730

731
        self.attn_mqa = RadixAttention(
732
733
734
735
736
737
            self.num_local_heads,
            self.kv_lora_rank + self.qk_rope_head_dim,
            self.scaling,
            num_kv_heads=1,
            layer_id=layer_id,
            v_head_dim=self.kv_lora_rank,
738
            quant_config=quant_config,
739
            prefix=add_prefix("attn_mqa", prefix),
740
741
        )

742
743
744
745
746
747
748
        self.attn_mha = RadixAttention(
            self.num_local_heads,
            self.qk_nope_head_dim + self.qk_rope_head_dim,
            self.scaling,
            num_kv_heads=self.num_local_heads,
            layer_id=layer_id,
            v_head_dim=self.v_head_dim,
749
            quant_config=quant_config,
750
            prefix=add_prefix("attn_mha", prefix),
751
752
        )

753
        self.alt_stream = alt_stream
754
        self.attn_mha.kv_b_proj = None
755

Ke Bao's avatar
Ke Bao committed
756
757
        self.w_kc = None
        self.w_vc = None
758
        self.w_scale = 1.0
759

760
761
762
763
        self.w_scale_k = None
        self.w_scale_v = None
        self.use_deep_gemm_bmm = False

Lianmin Zheng's avatar
Lianmin Zheng committed
764
765
766
        self.flashinfer_mla_disable_ragged = global_server_args_dict[
            "flashinfer_mla_disable_ragged"
        ]
767
768
769
        self.disable_chunked_prefix_cache = global_server_args_dict[
            "disable_chunked_prefix_cache"
        ]
770
        self.attention_backend = global_server_args_dict["attention_backend"]
771
772
773
        self.rocm_fused_decode_mla = get_bool_env_var(
            "SGLANG_ROCM_FUSED_DECODE_MLA", "false"
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
774

775
        # TODO: Design a finer way to determine the threshold
776
777
778
        self.chunked_prefix_cache_threshold = get_int_env_var(
            "SGL_CHUNKED_PREFIX_CACHE_THRESHOLD", 8192
        )
779
780
781
782

    def dispatch_attn_forward_method(
        self, forward_batch: ForwardBatch
    ) -> AttnForwardMethod:
783
784
785
786
787
788
789
790
791
792
793
794
        def _dispatch_mla_subtype():
            if _is_hip:
                if (
                    self.rocm_fused_decode_mla
                    and forward_batch.forward_mode.is_decode()
                ):
                    return AttnForwardMethod.MLA_FUSED_ROPE
                else:
                    return AttnForwardMethod.MLA
            else:
                return AttnForwardMethod.MLA

795
        if self.attention_backend == "flashinfer":
Lianmin Zheng's avatar
Lianmin Zheng committed
796
            # Flashinfer MLA: Do not absorb when enabling ragged prefill
797
            if (
Lianmin Zheng's avatar
Lianmin Zheng committed
798
799
800
801
                not self.flashinfer_mla_disable_ragged
                and forward_batch.forward_mode.is_extend()
                and not forward_batch.forward_mode.is_target_verify()
                and not forward_batch.forward_mode.is_draft_extend()
802
                and sum(forward_batch.extend_prefix_lens_cpu) == 0
803
804
805
            ):
                return AttnForwardMethod.MHA
            else:
806
                return _dispatch_mla_subtype()
807
        elif self.attention_backend == "fa3":
808
            # Flash Attention: Use MHA with chunked KV cache when prefilling on long sequences.
809
810
            if forward_batch.extend_prefix_lens_cpu is not None:
                sum_extend_prefix_lens = sum(forward_batch.extend_prefix_lens_cpu)
811
812
813
814
815
            if (
                forward_batch.forward_mode.is_extend()
                and not self.disable_chunked_prefix_cache
                and not forward_batch.forward_mode.is_target_verify()
                and not forward_batch.forward_mode.is_draft_extend()
816
817
818
819
                and (
                    sum_extend_prefix_lens >= self.chunked_prefix_cache_threshold
                    or sum_extend_prefix_lens == 0
                )
820
821
822
            ):
                return AttnForwardMethod.MHA_CHUNKED_KV
            else:
823
                return _dispatch_mla_subtype()
824
825
826
827
828
829
830
831
832
        elif self.attention_backend == "aiter":
            if (
                forward_batch.forward_mode.is_extend()
                and not forward_batch.forward_mode.is_target_verify()
                and not forward_batch.forward_mode.is_draft_extend()
            ):
                return AttnForwardMethod.MHA
            else:
                return AttnForwardMethod.MLA
Lianmin Zheng's avatar
Lianmin Zheng committed
833
834
        else:
            # Triton: Use normal computation for prefill and use weight absorption for extend/decode
835
            if (
Lianmin Zheng's avatar
Lianmin Zheng committed
836
837
838
                forward_batch.forward_mode.is_extend()
                and not forward_batch.forward_mode.is_target_verify()
                and not forward_batch.forward_mode.is_draft_extend()
839
                and sum(forward_batch.extend_prefix_lens_cpu) == 0
840
841
842
            ):
                return AttnForwardMethod.MHA
            else:
843
                return _dispatch_mla_subtype()
Lianmin Zheng's avatar
Lianmin Zheng committed
844

845
846
847
848
849
850
851
852
853
854
855
856
857
    def op_prepare(self, state):
        state.attn_intermediate_state = self.forward_prepare(
            positions=state.positions,
            hidden_states=state.pop("hidden_states_after_comm_pre_attn"),
            forward_batch=state.forward_batch,
            zero_allocator=state.zero_allocator,
        )

    def op_core(self, state):
        state.hidden_states_after_attn = self.forward_core(
            state.pop("attn_intermediate_state")
        )

858
859
860
861
    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
862
        forward_batch: ForwardBatch,
863
        zero_allocator: BumpAllocator,
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
    ):
        s = self.forward_prepare(
            positions=positions,
            hidden_states=hidden_states,
            forward_batch=forward_batch,
            zero_allocator=zero_allocator,
        )
        return self.forward_core(s)

    def forward_prepare(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ):
880
881
882
        if self.attn_mha.kv_b_proj is None:
            self.attn_mha.kv_b_proj = self.kv_b_proj

Lianmin Zheng's avatar
Lianmin Zheng committed
883
884
885
886
        if hidden_states.shape[0] == 0:
            assert (
                not self.o_proj.reduce_results
            ), "short-circuiting allreduce will lead to hangs"
887
            return hidden_states, None, forward_batch, None
888

889
890
891
        attn_forward_method = self.dispatch_attn_forward_method(forward_batch)

        if attn_forward_method == AttnForwardMethod.MHA:
892
893
894
            inner_state = self.forward_normal_prepare(
                positions, hidden_states, forward_batch, zero_allocator
            )
895
        elif attn_forward_method == AttnForwardMethod.MHA_CHUNKED_KV:
896
897
            inner_state = self.forward_normal_chunked_kv_prepare(
                positions, hidden_states, forward_batch, zero_allocator
898
            )
899
        elif attn_forward_method == AttnForwardMethod.MLA:
900
            inner_state = self.forward_absorb_prepare(
901
902
903
                positions, hidden_states, forward_batch, zero_allocator
            )
        elif attn_forward_method == AttnForwardMethod.MLA_FUSED_ROPE:
904
905
            inner_state = self.forward_absorb_fused_mla_rope_prepare(
                positions, hidden_states, forward_batch, zero_allocator
906
            )
907
        else:
908
            raise NotImplementedError
909
        return None, attn_forward_method, forward_batch, inner_state
910

911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
    def forward_core(self, intermediate_state):
        hidden_states, attn_forward_method, forward_batch, inner_state = (
            intermediate_state
        )
        if inner_state is None:
            return hidden_states

        if attn_forward_method == AttnForwardMethod.MHA:
            return self.forward_normal_core(*inner_state)
        elif attn_forward_method == AttnForwardMethod.MHA_CHUNKED_KV:
            return self.forward_normal_chunked_kv_core(*inner_state)
        elif attn_forward_method == AttnForwardMethod.MLA:
            return self.forward_absorb_core(*inner_state)
        elif attn_forward_method == AttnForwardMethod.MLA_FUSED_ROPE:
            return self.forward_absorb_fused_mla_rope_core(*inner_state)
        else:
            raise NotImplementedError

    def forward_normal_prepare(
930
931
932
933
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
934
935
        zero_allocator: BumpAllocator,
    ):
936
        if self.q_lora_rank is not None:
937
938
939
            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
            )
940
941
942
943
944
945
            q = self.q_a_layernorm(q)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(
                -1, self.num_local_heads, self.qk_head_dim
            )
946
947
            latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]

948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        latent_cache = latent_cache.unsqueeze(1)
        kv_a = self.kv_a_layernorm(kv_a.contiguous())
        kv = self.kv_b_proj(kv_a)[0]
        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
        k_nope = kv[..., : self.qk_nope_head_dim]
        v = kv[..., self.qk_nope_head_dim :]
        k_pe = latent_cache[:, :, self.kv_lora_rank :]
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
        q[..., self.qk_nope_head_dim :] = q_pe
        k = torch.empty_like(q)
        k[..., : self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim :] = k_pe

        latent_cache[:, :, : self.kv_lora_rank] = kv_a.unsqueeze(1)
        latent_cache[:, :, self.kv_lora_rank :] = k_pe

        # Save latent cache
        forward_batch.token_to_kv_pool.set_kv_buffer(
            self.attn_mha, forward_batch.out_cache_loc, latent_cache, None
        )
970
971
972
973

        return q, k, v, forward_batch

    def forward_normal_core(self, q, k, v, forward_batch):
974
975
976
977
978
        attn_output = self.attn_mha(q, k, v, forward_batch, save_kv_cache=False)
        attn_output = attn_output.reshape(-1, self.num_local_heads * self.v_head_dim)
        output, _ = self.o_proj(attn_output)
        return output

979
    def forward_absorb_prepare(
980
981
982
983
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
984
        zero_allocator: BumpAllocator,
985
    ):
986
987
        from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode

988
        if self.q_lora_rank is not None:
989
990
991
            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
            )
992
993
994
            k_nope = latent_cache[..., : self.kv_lora_rank]

            # overlap qk norm
995
            if self.alt_stream is not None and get_is_capture_mode():
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
                current_stream = torch.cuda.current_stream()
                self.alt_stream.wait_stream(current_stream)
                q = self.q_a_layernorm(q)
                with torch.cuda.stream(self.alt_stream):
                    k_nope = self.kv_a_layernorm(k_nope)
                current_stream.wait_stream(self.alt_stream)
            else:
                q = self.q_a_layernorm(q)
                k_nope = self.kv_a_layernorm(k_nope)

            k_nope = k_nope.unsqueeze(1)
1007
1008
1009
1010
1011
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(
                -1, self.num_local_heads, self.qk_head_dim
            )
1012
            latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
1013
1014
1015
            k_nope = latent_cache[..., : self.kv_lora_rank]
            k_nope = self.kv_a_layernorm(k_nope).unsqueeze(1)

1016
        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
1017
        k_pe = latent_cache[..., self.kv_lora_rank :].unsqueeze(1)
1018

1019
1020
        if self.use_deep_gemm_bmm:
            q_nope_val, q_nope_scale, masked_m, expected_m, aligned_m = (
1021
                per_token_group_quant_mla_deep_gemm_masked_fp8(q_nope.transpose(0, 1))
1022
1023
1024
1025
            )
            q_nope_out = q_nope.new_empty(
                (self.num_local_heads, aligned_m, self.kv_lora_rank)
            )
1026
            deep_gemm_wrapper.grouped_gemm_nt_f8f8bf16_masked(
1027
1028
1029
1030
1031
1032
1033
                (q_nope_val, q_nope_scale),
                (self.w_kc, self.w_scale_k),
                q_nope_out,
                masked_m,
                expected_m,
            )
            q_nope_out = q_nope_out[:, :expected_m, :]
1034
1035
        elif _is_hip:
            # TODO(haishaw): add bmm_fp8 to ROCm
1036
1037
1038
1039
            q_nope_out = torch.bmm(
                q_nope.to(torch.bfloat16).transpose(0, 1),
                self.w_kc.to(torch.bfloat16) * self.w_scale,
            )
1040
        elif self.w_kc.dtype == torch.float8_e4m3fn:
1041
            q_nope_val, q_nope_scale = per_tensor_quant_mla_fp8(
Lianmin Zheng's avatar
Lianmin Zheng committed
1042
                q_nope.transpose(0, 1),
1043
                zero_allocator.allocate(1),
1044
1045
1046
1047
1048
1049
            )
            q_nope_out = bmm_fp8(
                q_nope_val, self.w_kc, q_nope_scale, self.w_scale, torch.bfloat16
            )
        else:
            q_nope_out = torch.bmm(q_nope.transpose(0, 1), self.w_kc)
1050
1051

        q_nope_out = q_nope_out.transpose(0, 1)
1052
1053
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)

1054
1055
1056
1057
1058
        return q_pe, k_pe, q_nope_out, k_nope, forward_batch, zero_allocator

    def forward_absorb_core(
        self, q_pe, k_pe, q_nope_out, k_nope, forward_batch, zero_allocator
    ):
1059
1060
1061
1062
1063
        if (
            self.attention_backend == "fa3"
            or self.attention_backend == "flashinfer"
            or self.attention_backend == "cutlass_mla"
        ):
1064
            attn_output = self.attn_mqa(
Ke Bao's avatar
Ke Bao committed
1065
                q_nope_out, k_nope, k_nope, forward_batch, q_rope=q_pe, k_rope=k_pe
1066
1067
1068
            )
        else:
            q = torch.cat([q_nope_out, q_pe], dim=-1)
Ke Bao's avatar
Ke Bao committed
1069
            k = torch.cat([k_nope, k_pe], dim=-1)
1070
            attn_output = self.attn_mqa(q, k, k_nope, forward_batch)
1071
1072
        attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)

1073
1074
        if self.use_deep_gemm_bmm:
            attn_output_val, attn_output_scale, masked_m, expected_m, aligned_m = (
1075
1076
                per_token_group_quant_mla_deep_gemm_masked_fp8(
                    attn_output.transpose(0, 1)
1077
1078
1079
1080
1081
                )
            )
            attn_bmm_output = attn_output.new_empty(
                (self.num_local_heads, aligned_m, self.v_head_dim)
            )
1082
            deep_gemm_wrapper.grouped_gemm_nt_f8f8bf16_masked(
1083
1084
1085
1086
1087
1088
                (attn_output_val, attn_output_scale),
                (self.w_vc, self.w_scale_v),
                attn_bmm_output,
                masked_m,
                expected_m,
            )
Ke Bao's avatar
Ke Bao committed
1089
1090
1091
            attn_bmm_output = (
                attn_bmm_output[:, :expected_m, :].transpose(0, 1).flatten(1, 2)
            )
1092
1093
        elif _is_hip:
            # TODO(haishaw): add bmm_fp8 to ROCm
1094
1095
1096
1097
            attn_bmm_output = torch.bmm(
                attn_output.to(torch.bfloat16).transpose(0, 1),
                self.w_vc.to(torch.bfloat16) * self.w_scale,
            )
Ke Bao's avatar
Ke Bao committed
1098
            attn_bmm_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)
1099
        elif self.w_vc.dtype == torch.float8_e4m3fn:
1100
            attn_output_val, attn_output_scale = per_tensor_quant_mla_fp8(
Lianmin Zheng's avatar
Lianmin Zheng committed
1101
                attn_output.transpose(0, 1),
1102
                zero_allocator.allocate(1),
1103
1104
1105
1106
1107
1108
1109
1110
            )
            attn_bmm_output = bmm_fp8(
                attn_output_val,
                self.w_vc,
                attn_output_scale,
                self.w_scale,
                torch.bfloat16,
            )
Ke Bao's avatar
Ke Bao committed
1111
            attn_bmm_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)
1112
        else:
Ke Bao's avatar
Ke Bao committed
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
            attn_bmm_output = torch.empty(
                (attn_output.shape[0], self.num_local_heads * self.v_head_dim),
                dtype=attn_output.dtype,
                device=attn_output.device,
            )
            torch.bmm(
                attn_output.transpose(0, 1),
                self.w_vc,
                out=attn_bmm_output.view(
                    -1, self.num_local_heads, self.v_head_dim
                ).transpose(0, 1),
            )
        output, _ = self.o_proj(attn_bmm_output)
1126
1127
1128

        return output

1129
    def forward_absorb_fused_mla_rope_prepare(
1130
1131
1132
1133
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
1134
        zero_allocator: BumpAllocator,
1135
    ):
1136
1137
1138
1139
1140
1141
1142
1143
        enable_rope_fusion = (
            os.getenv("SGLANG_FUSED_MLA_ENABLE_ROPE_FUSION", "1") == "1"
        )
        q_len = hidden_states.shape[0]
        q_input = hidden_states.new_empty(
            q_len, self.num_local_heads, self.kv_lora_rank + self.qk_rope_head_dim
        )
        if self.q_lora_rank is not None:
1144
1145
1146
            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
            )
1147
1148
1149
1150
1151
1152
            q = self.q_a_layernorm(q)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(
                -1, self.num_local_heads, self.qk_head_dim
            )
1153
            latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
1154
1155
        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)

1156
1157
        if _is_hip:
            # TODO(haishaw): add bmm_fp8 to ROCm
1158
1159
1160
1161
1162
            q_nope_out = torch.bmm(
                q_nope.to(torch.bfloat16).transpose(0, 1),
                self.w_kc.to(torch.bfloat16) * self.w_scale,
            )
        elif self.w_kc.dtype == torch.float8_e4m3fn:
1163
            q_nope_val, q_nope_scale = per_tensor_quant_mla_fp8(
1164
1165
1166
                q_nope.transpose(0, 1),
                zero_allocator.allocate(1),
                dtype=torch.float8_e4m3fn,
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
            )
            q_nope_out = bmm_fp8(
                q_nope_val, self.w_kc, q_nope_scale, self.w_scale, torch.bfloat16
            )
        else:
            q_nope_out = torch.bmm(q_nope.transpose(0, 1), self.w_kc)
        q_input[..., : self.kv_lora_rank] = q_nope_out.transpose(0, 1)
        v_input = latent_cache[..., : self.kv_lora_rank]
        v_input = self.kv_a_layernorm(v_input.contiguous()).unsqueeze(1)
        k_input = latent_cache.unsqueeze(1)
        k_input[..., : self.kv_lora_rank] = v_input

        if not enable_rope_fusion:
            k_pe = k_input[..., self.kv_lora_rank :]
            q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
            q_input[..., self.kv_lora_rank :] = q_pe
            k_input[..., self.kv_lora_rank :] = k_pe
            k_pe_output = None
        else:
            k_pe_output = torch.empty_like(k_input[..., self.kv_lora_rank :])

        q_input[..., self.kv_lora_rank :] = q_pe

        # attn_output = self.attn_mqa(q_input, k_input, v_input, forward_batch)
        # Use Fused ROPE with use_rope=OFF.
        attn_output = torch.empty(
            (q_len, self.num_local_heads, self.kv_lora_rank),
            dtype=q.dtype,
            device=q.device,
        )
        attn_logits, _, kv_indptr, kv_indices, _, _, _ = (
            forward_batch.attn_backend.forward_metadata
        )
        cos_sin_cache = self.rotary_emb.cos_sin_cache
        num_kv_split = forward_batch.attn_backend.num_kv_splits
        sm_scale = self.attn_mqa.scaling
        if attn_logits is None:
            attn_logits = torch.empty(
                (
                    forward_batch.batch_size,
                    self.num_local_heads,
                    num_kv_split,
                    self.kv_lora_rank + 1,
                ),
                dtype=torch.float32,
                device=q.device,
            )

        # save current latent cache.
        forward_batch.token_to_kv_pool.set_kv_buffer(
            self.attn_mqa, forward_batch.out_cache_loc, k_input, None
        )
        key_cache_buf = forward_batch.token_to_kv_pool.get_key_buffer(
            self.attn_mqa.layer_id
        )
        val_cache_buf = key_cache_buf[..., : self.kv_lora_rank]

1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
        return (
            q_input,
            key_cache_buf,
            val_cache_buf,
            attn_output,
            kv_indptr,
            kv_indices,
            k_pe_output,
            cos_sin_cache,
            positions,
            attn_logits,
            num_kv_split,
            sm_scale,
            enable_rope_fusion,
            k_input,
            forward_batch,
            zero_allocator,
        )

    def forward_absorb_fused_mla_rope_core(
        self,
        q_input,
        key_cache_buf,
        val_cache_buf,
        attn_output,
        kv_indptr,
        kv_indices,
        k_pe_output,
        cos_sin_cache,
        positions,
        attn_logits,
        num_kv_split,
        sm_scale,
        enable_rope_fusion,
        k_input,
        forward_batch,
        zero_allocator,
    ):
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
        decode_attention_fwd_grouped_rope(
            q_input,
            key_cache_buf,
            val_cache_buf,
            attn_output,
            kv_indptr,
            kv_indices,
            k_pe_output,
            self.kv_lora_rank,
            self.rotary_emb.rotary_dim,
            cos_sin_cache,
            positions,
            attn_logits,
            num_kv_split,
            sm_scale,
            logit_cap=self.attn_mqa.logit_cap,
            use_rope=enable_rope_fusion,
            is_neox_style=self.rotary_emb.is_neox_style,
        )

        if enable_rope_fusion:
            k_input[..., self.kv_lora_rank :] = k_pe_output
            forward_batch.token_to_kv_pool.set_kv_buffer(
                self.attn_mqa, forward_batch.out_cache_loc, k_input, None
            )

        attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)

1290
1291
        if _is_hip:
            # TODO(haishaw): add bmm_fp8 to ROCm
1292
1293
1294
1295
1296
            attn_bmm_output = torch.bmm(
                attn_output.to(torch.bfloat16).transpose(0, 1),
                self.w_vc.to(torch.bfloat16) * self.w_scale,
            )
        elif self.w_vc.dtype == torch.float8_e4m3fn:
1297
            attn_output_val, attn_output_scale = per_tensor_quant_mla_fp8(
1298
1299
1300
                attn_output.transpose(0, 1),
                zero_allocator.allocate(1),
                dtype=torch.float8_e4m3fn,
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
            )
            attn_bmm_output = bmm_fp8(
                attn_output_val,
                self.w_vc,
                attn_output_scale,
                self.w_scale,
                torch.bfloat16,
            )
        else:
            attn_bmm_output = torch.bmm(attn_output.transpose(0, 1), self.w_vc)
        attn_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)
1312
1313
1314
1315
        output, _ = self.o_proj(attn_output)

        return output

1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
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
    def _chunked_prefix_attn_mha(
        self,
        q: torch.Tensor,
        accum_output: torch.Tensor,
        accum_lse: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:

        assert forward_batch.num_prefix_chunks is not None
        for i in range(forward_batch.num_prefix_chunks):
            forward_batch.set_prefix_chunk_idx(i)

            # Fetch latent cache from memory pool with precomputed chunked kv indices
            latent_cache_buf = forward_batch.token_to_kv_pool.get_key_buffer(
                self.attn_mha.layer_id
            )
            latent_cache = latent_cache_buf[
                forward_batch.prefix_chunk_kv_indices[i]
            ].contiguous()

            kv_a_normed, k_pe = latent_cache.split(
                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
            )
            kv_a_normed = kv_a_normed.squeeze(1).contiguous()
            kv = self.kv_b_proj(kv_a_normed)[0]
            kv = kv.view(
                -1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim
            )
            v = kv[..., self.qk_nope_head_dim :]
            k_nope = kv[..., : self.qk_nope_head_dim]

            k = torch.empty(
                (
                    k_nope.shape[0],
                    self.num_local_heads,
                    self.qk_nope_head_dim + self.qk_rope_head_dim,
                ),
                dtype=v.dtype,
                device=v.device,
            )
            k[..., : self.qk_nope_head_dim] = k_nope
            k[..., self.qk_nope_head_dim :] = k_pe

            output, lse = self.attn_mha(q, k, v, forward_batch, save_kv_cache=False)
            lse = torch.transpose(lse, 0, 1).contiguous()
            tmp_output = torch.empty_like(accum_output)
            tmp_lse = torch.empty_like(accum_lse)
            merge_state_v2(output, lse, accum_output, accum_lse, tmp_output, tmp_lse)
            accum_output, accum_lse = tmp_output, tmp_lse

        return accum_output

1368
    def forward_normal_chunked_kv_prepare(
1369
1370
1371
1372
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
1373
1374
        zero_allocator: BumpAllocator,
    ):
1375
1376
1377
1378
1379
1380
1381
1382
        # In normal mha, the k and v tensors will become overly large when the prefix length is long.
        # To avoid this, we split the kv cache into chunks and process them one after another.
        # Since mha is compute friendly, the for loop induced here will not introduce significant overhead.
        # The top comments in https://github.com/vllm-project/vllm/blob/main/vllm/v1/attention/backends/mla/common.py
        # will be helpful for understanding the purpose of this function.

        # First do normal mha forward to get output for extended part
        if self.q_lora_rank is not None:
1383
1384
1385
            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
            )
1386
1387
1388
1389
1390
1391
            q = self.q_a_layernorm(q)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(
                -1, self.num_local_heads, self.qk_head_dim
            )
1392
            latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        latent_cache = latent_cache.unsqueeze(1)
        kv_a = self.kv_a_layernorm(kv_a.contiguous())
        kv = self.kv_b_proj(kv_a)[0]
        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
        k_nope = kv[..., : self.qk_nope_head_dim]
        v = kv[..., self.qk_nope_head_dim :]
        k_pe = latent_cache[:, :, self.kv_lora_rank :]

        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
        q[..., self.qk_nope_head_dim :] = q_pe
        k = torch.empty_like(q)
        k[..., : self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim :] = k_pe

        latent_cache[:, :, : self.kv_lora_rank] = kv_a.unsqueeze(1)
        latent_cache[:, :, self.kv_lora_rank :] = k_pe

        # Save latent cache
        forward_batch.token_to_kv_pool.set_kv_buffer(
            self.attn_mha, forward_batch.out_cache_loc, latent_cache, None
        )

1417
1418
1419
        return q, k, v, forward_batch

    def forward_normal_chunked_kv_core(self, q, k, v, forward_batch):
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
        # Do mha for extended part without prefix
        forward_batch.set_attn_attend_prefix_cache(False)
        attn_output, lse = self.attn_mha(q, k, v, forward_batch, save_kv_cache=False)
        lse = torch.transpose(lse, 0, 1).contiguous()

        # Do mha attention with chunked prefix cache if there are any sequence with prefix
        if any(forward_batch.extend_prefix_lens_cpu):
            # Only initialize the info once
            if forward_batch.num_prefix_chunks is None:
                forward_batch.prepare_chunked_prefix_cache_info(q.device)

            forward_batch.set_attn_attend_prefix_cache(True)
            attn_output = self._chunked_prefix_attn_mha(
                q=q,
                accum_output=attn_output,
                accum_lse=lse,
                forward_batch=forward_batch,
            )

        attn_output = attn_output.reshape(-1, self.num_local_heads * self.v_head_dim)
        output, _ = self.o_proj(attn_output)
        return output

1443

Liangsheng Yin's avatar
Liangsheng Yin committed
1444
1445
1446
1447
1448
1449
1450
class DeepseekV2DecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        layer_id: int,
        quant_config: Optional[QuantizationConfig] = None,
1451
        is_nextn: bool = False,
1452
        prefix: str = "",
1453
        alt_stream: Optional[torch.cuda.Stream] = None,
Liangsheng Yin's avatar
Liangsheng Yin committed
1454
1455
1456
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
1457
        self.config = config
Liangsheng Yin's avatar
Liangsheng Yin committed
1458
1459
1460
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
Lianmin Zheng's avatar
Lianmin Zheng committed
1461
        self.enable_dp_attention = global_server_args_dict["enable_dp_attention"]
1462
        self.speculative_algorithm = global_server_args_dict["speculative_algorithm"]
Lianmin Zheng's avatar
Lianmin Zheng committed
1463
        self.layer_id = layer_id
1464
        self.is_nextn = is_nextn
Baizhou Zhang's avatar
Baizhou Zhang committed
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
        self.self_attn = DeepseekV2AttentionMLA(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            qk_nope_head_dim=config.qk_nope_head_dim,
            qk_rope_head_dim=config.qk_rope_head_dim,
            v_head_dim=config.v_head_dim,
            q_lora_rank=(
                config.q_lora_rank if hasattr(config, "q_lora_rank") else None
            ),
            kv_lora_rank=config.kv_lora_rank,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            layer_id=layer_id,
            reduce_results=False,
            prefix=add_prefix("self_attn", prefix),
1483
            alt_stream=alt_stream,
Baizhou Zhang's avatar
Baizhou Zhang committed
1484
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1485

1486
1487
1488
1489
1490
        self.is_layer_sparse = self._is_layer_sparse(layer_id, is_nextn=is_nextn)
        is_previous_layer_sparse = self._is_layer_sparse(layer_id - 1, is_nextn=False)

        self.layer_scatter_modes = LayerScatterModes.init_new(
            layer_id=layer_id,
1491
            num_layers=1 if is_nextn else config.num_hidden_layers,
1492
1493
            is_layer_sparse=self.is_layer_sparse,
            is_previous_layer_sparse=is_previous_layer_sparse,
1494
1495
        )

1496
        if self.is_layer_sparse:
1497
1498
1499
1500
            self.mlp = DeepseekV2MoE(
                config=config,
                quant_config=quant_config,
                prefix=add_prefix("mlp", prefix),
fzyzcjy's avatar
fzyzcjy committed
1501
                layer_id=self.layer_id,
1502
                alt_stream=alt_stream,
1503
            )
Liangsheng Yin's avatar
Liangsheng Yin committed
1504
        else:
1505
            if enable_moe_dense_fully_dp():
1506
1507
1508
                mlp_tp_rank, mlp_tp_size = 0, 1
            else:
                mlp_tp_rank, mlp_tp_size = None, None
Liangsheng Yin's avatar
Liangsheng Yin committed
1509
1510
1511
1512
1513
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
1514
                prefix=add_prefix("mlp", prefix),
1515
1516
                tp_rank=mlp_tp_rank,
                tp_size=mlp_tp_size,
Liangsheng Yin's avatar
Liangsheng Yin committed
1517
            )
1518

Liangsheng Yin's avatar
Liangsheng Yin committed
1519
1520
1521
1522
1523
        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
        )

1524
1525
1526
1527
        self.layer_communicator = LayerCommunicator(
            layer_scatter_modes=self.layer_scatter_modes,
            input_layernorm=self.input_layernorm,
            post_attention_layernorm=self.post_attention_layernorm,
1528
        )
1529
1530
1531
1532
1533
1534

    def _is_layer_sparse(self, layer_id: int, is_nextn: bool) -> bool:
        return is_nextn or (
            self.config.n_routed_experts is not None
            and layer_id >= self.config.first_k_dense_replace
            and layer_id % self.config.moe_layer_freq == 0
1535
1536
        )

Liangsheng Yin's avatar
Liangsheng Yin committed
1537
1538
1539
1540
    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
1541
        forward_batch: ForwardBatch,
Liangsheng Yin's avatar
Liangsheng Yin committed
1542
        residual: Optional[torch.Tensor],
1543
        zero_allocator: BumpAllocator,
Liangsheng Yin's avatar
Liangsheng Yin committed
1544
    ) -> torch.Tensor:
1545

1546
1547
        hidden_states, residual = self.layer_communicator.prepare_attn(
            hidden_states, residual, forward_batch
1548
1549
        )

1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
            forward_batch=forward_batch,
            zero_allocator=zero_allocator,
        )

        hidden_states, residual = self.layer_communicator.prepare_mlp(
            hidden_states, residual, forward_batch
        )

        hidden_states = self.mlp(hidden_states, forward_batch)

        hidden_states, residual = self.layer_communicator.postprocess_layer(
            hidden_states, residual, forward_batch
        )

1567
1568
1569
1570
1571
        if self.enable_dp_attention and self.speculative_algorithm.is_eagle():
            # NOTE: this line resolves the degradation of MTP reception rate for non-zero DP ranks.
            # See discussion here (https://github.com/sgl-project/sglang/pull/6081#discussion_r2147452251).
            hidden_states = hidden_states.clone()

1572
1573
        return hidden_states, residual

1574
1575
1576
1577
1578
1579
1580
1581
    def op_comm_prepare_attn(
        self,
        state,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        residual: Optional[torch.Tensor],
        zero_allocator: BumpAllocator,
1582
        tbo_subbatch_index: Optional[int] = None,
1583
1584
    ):
        state.hidden_states_after_comm_pre_attn, state.residual_after_input_ln = (
fzyzcjy's avatar
fzyzcjy committed
1585
            self.layer_communicator.prepare_attn(hidden_states, residual, forward_batch)
1586
1587
1588
1589
1590
1591
        )
        state.update(
            dict(
                forward_batch=forward_batch,
                positions=positions,
                zero_allocator=zero_allocator,
1592
                tbo_subbatch_index=tbo_subbatch_index,
1593
            )
1594
        )
1595

1596
1597
1598
1599
1600
1601
1602
    def op_comm_prepare_mlp(self, state):
        state.hidden_states_mlp_input, state.residual_after_comm_pre_mlp = (
            self.layer_communicator.prepare_mlp(
                state.pop("hidden_states_after_attn"),
                state.pop("residual_after_input_ln"),
                state.forward_batch,
            )
1603
        )
1604

1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
    def op_mlp(self, state):
        hidden_states = state.pop("hidden_states_mlp_input")
        if not (
            enable_moe_dense_fully_dp()
            and (not self.is_layer_sparse)
            and hidden_states.shape[0] == 0
        ):
            state.hidden_states_mlp_output = self.mlp(
                hidden_states, state.forward_batch.forward_mode
            )
        else:
            state.hidden_states_mlp_output = hidden_states
1617

1618
    def op_comm_postprocess_layer(self, state):
1619
        hidden_states, residual = self.layer_communicator.postprocess_layer(
1620
1621
1622
            state.pop("hidden_states_mlp_output"),
            state.pop("residual_after_comm_pre_mlp"),
            state.forward_batch,
1623
        )
1624

1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
        output = dict(
            positions=state.positions,
            hidden_states=hidden_states,
            residual=residual,
            forward_batch=state.forward_batch,
            zero_allocator=state.zero_allocator,
            tbo_subbatch_index=state.tbo_subbatch_index,
        )

        state.clear(
            expect_keys={
                "positions",
                "forward_batch",
                "zero_allocator",
                "tbo_subbatch_index",
            }
        )
        return output
1643

Liangsheng Yin's avatar
Liangsheng Yin committed
1644
1645
1646
1647
1648
1649
1650
1651

class DeepseekV2Model(nn.Module):
    fall_back_to_pt_during_load = False

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
1652
        prefix: str = "",
Liangsheng Yin's avatar
Liangsheng Yin committed
1653
1654
1655
1656
    ) -> None:
        super().__init__()
        self.padding_id = config.pad_token_id
        self.vocab_size = config.vocab_size
1657
        self.first_k_dense_replace = config.first_k_dense_replace
Liangsheng Yin's avatar
Liangsheng Yin committed
1658
1659
1660
1661

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
Ke Bao's avatar
Ke Bao committed
1662
            enable_tp=not global_server_args_dict["enable_dp_attention"],
Liangsheng Yin's avatar
Liangsheng Yin committed
1663
        )
1664
        self.alt_stream = torch.cuda.Stream() if _is_cuda else None
Liangsheng Yin's avatar
Liangsheng Yin committed
1665
1666
1667
1668
1669
1670
        self.layers = nn.ModuleList(
            [
                DeepseekV2DecoderLayer(
                    config,
                    layer_id,
                    quant_config=quant_config,
1671
                    prefix=add_prefix(f"layers.{layer_id}", prefix),
1672
                    alt_stream=self.alt_stream,
Liangsheng Yin's avatar
Liangsheng Yin committed
1673
1674
1675
1676
1677
1678
                )
                for layer_id in range(config.num_hidden_layers)
            ]
        )
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

1679
1680
1681
    def get_input_embeddings(self) -> torch.Tensor:
        return self.embed_tokens

Liangsheng Yin's avatar
Liangsheng Yin committed
1682
1683
1684
1685
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1686
        forward_batch: ForwardBatch,
1687
        input_embeds: torch.Tensor = None,
Liangsheng Yin's avatar
Liangsheng Yin committed
1688
    ) -> torch.Tensor:
1689
1690
        total_num_layers = len(self.layers)
        device = input_embeds.device if input_embeds is not None else input_ids.device
1691
        zero_allocator = BumpAllocator(
1692
            buffer_size=total_num_layers * 2 * (2 if forward_batch.can_run_tbo else 1),
1693
            dtype=torch.float32,
1694
            device=device,
1695
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1696

1697
1698
1699
1700
1701
        if input_embeds is None:
            hidden_states = self.embed_tokens(input_ids)
        else:
            hidden_states = input_embeds

Liangsheng Yin's avatar
Liangsheng Yin committed
1702
        residual = None
1703
1704
1705
1706
1707
1708
1709

        normal_num_layers = (
            self.first_k_dense_replace
            if forward_batch.can_run_tbo
            else total_num_layers
        )
        for i in range(normal_num_layers):
1710
1711
1712
1713
1714
            with get_global_expert_distribution_recorder().with_current_layer(i):
                layer = self.layers[i]
                hidden_states, residual = layer(
                    positions, hidden_states, forward_batch, residual, zero_allocator
                )
1715
1716
1717
1718
1719
1720
1721
1722
1723

        if normal_num_layers != total_num_layers:
            hidden_states, residual = model_forward_maybe_tbo(
                layers=self.layers[normal_num_layers:],
                enable_tbo=True,
                positions=positions,
                forward_batch=forward_batch,
                hidden_states=hidden_states,
                residual=residual,
1724
1725
1726
                input_data_scatter_mode=self.layers[
                    normal_num_layers - 1
                ].layer_scatter_modes.layer_output_mode,
1727
1728
1729
                zero_allocator=zero_allocator,
            )

Ke Bao's avatar
Ke Bao committed
1730
        if not forward_batch.forward_mode.is_idle():
1731
1732
1733
1734
            if residual is None:
                hidden_states = self.norm(hidden_states)
            else:
                hidden_states, _ = self.norm(hidden_states, residual)
Liangsheng Yin's avatar
Liangsheng Yin committed
1735
1736
1737
1738
1739
1740
1741
1742
1743
        return hidden_states


class DeepseekV2ForCausalLM(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
1744
        prefix: str = "",
Liangsheng Yin's avatar
Liangsheng Yin committed
1745
1746
1747
    ) -> None:
        super().__init__()
        self.config = config
1748
        self.tp_size = get_tensor_model_parallel_world_size()
Liangsheng Yin's avatar
Liangsheng Yin committed
1749
        self.quant_config = quant_config
1750
        self.determine_num_fused_shared_experts()
1751
1752
1753
1754
1755
1756
1757
1758
        self.model = DeepseekV2Model(
            config, quant_config, prefix=add_prefix("model", prefix)
        )
        self.lm_head = ParallelLMHead(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=add_prefix("lm_head", prefix),
1759
            use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
1760
1761
1762
        )
        self.logits_processor = LogitsProcessor(config)

1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
        self._routed_experts_weights_of_layer = LazyValue(
            lambda: {
                layer_id: layer.mlp.get_moe_weights()
                for layer_id, layer in enumerate(self.model.layers)
                if isinstance(layer.mlp, DeepseekV2MoE)
            }
        )

    @property
    def routed_experts_weights_of_layer(self):
        return self._routed_experts_weights_of_layer.value

1775
    def determine_num_fused_shared_experts(
1776
1777
        self, architecture: str = "DeepseekV3ForCausalLM"
    ):
1778
1779
1780
1781
1782
1783
1784
1785
        self.num_fused_shared_experts = 0
        if global_server_args_dict["disable_shared_experts_fusion"]:
            return

        # Only Deepseek V3/R1 can use shared experts fusion optimization now.
        disable_reason = None
        if (
            not _is_cuda
1786
            or torch.cuda.get_device_capability("cuda") < (8, 0)
1787
1788
1789
1790
            or self.config.architectures[0] != architecture
            or self.config.n_routed_experts != 256
            or self.config.n_shared_experts != 1
        ):
1791
            disable_reason = "Only Deepseek V3/R1 on NV-platform with capability >= 80 can use shared experts fusion optimization."
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
        elif (
            global_server_args_dict["enable_deepep_moe"]
            or global_server_args_dict["enable_ep_moe"]
        ):
            disable_reason = "Deepseek V3/R1 can not use shared experts fusion optimization when in deepep_moe or ep_moe mode."

        if disable_reason is not None:
            global_server_args_dict["disable_shared_experts_fusion"] = True
            log_info_on_rank0(
                logger,
                f"{disable_reason} Shared experts fusion optimization is disabled.",
            )
            return

        self.num_fused_shared_experts = self.config.n_shared_experts
1807

Mick's avatar
Mick committed
1808
1809
1810
    def get_input_embeddings(self) -> nn.Embedding:
        return self.model.embed_tokens

1811
    @torch.no_grad()
Liangsheng Yin's avatar
Liangsheng Yin committed
1812
1813
1814
1815
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1816
        forward_batch: ForwardBatch,
1817
        input_embeds: torch.Tensor = None,
Liangsheng Yin's avatar
Liangsheng Yin committed
1818
    ) -> torch.Tensor:
1819
        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
Lianmin Zheng's avatar
Lianmin Zheng committed
1820

1821
1822
1823
        return self.logits_processor(
            input_ids, hidden_states, self.lm_head, forward_batch
        )
Liangsheng Yin's avatar
Liangsheng Yin committed
1824

1825
    def post_load_weights(self, is_nextn=False, weight_names=None):
inkcherry's avatar
inkcherry committed
1826
1827

        # Perform post-processing after loading weights
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
        if is_nextn:
            layer_ids = [self.config.num_hidden_layers]
        else:
            if weight_names is None:
                layer_ids = range(self.config.num_hidden_layers)
            else:
                layer_ids = set()
                for name in weight_names:
                    if "kv_b_proj" in name:
                        layer_id = int(name.split(".")[2])
1838
                        if layer_id < self.config.num_hidden_layers:
1839
1840
                            layer_ids.add(layer_id)

1841
1842
1843
1844
1845
1846
        for layer_id in layer_ids:
            self_attn = (
                self.model.layers[layer_id].self_attn
                if not is_nextn
                else self.model.decoder.self_attn
            )
Baizhou Zhang's avatar
Baizhou Zhang committed
1847
1848
1849
1850
1851
1852
1853
1854
            if hasattr(self_attn.kv_b_proj, "qweight"):
                # AWQ compatible
                if _is_cuda:
                    w = awq_dequantize(
                        self_attn.kv_b_proj.qweight,
                        self_attn.kv_b_proj.scales,
                        self_attn.kv_b_proj.qzeros,
                    ).T
inkcherry's avatar
inkcherry committed
1855
                else:
Baizhou Zhang's avatar
Baizhou Zhang committed
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
                    w = awq_dequantize(
                        self_attn.kv_b_proj.qweight,
                        self_attn.kv_b_proj.scales,
                        self_attn.kv_b_proj.qzeros,
                        0,
                        0,
                        0,
                    ).T
            else:
                w = self_attn.kv_b_proj.weight
            # NOTE(HandH1998): Since `bmm_fp8` only supports per-tensor scale, we have to requantize `self_attn.kv_b_proj`.
            # This may affect the accuracy of fp8 model.
1868
1869
1870
1871
            # Fix deepseek v3 blockwise bmm by using deep_gemm
            use_deep_gemm_bmm = False
            model_dtype = torch.get_default_dtype()

Baizhou Zhang's avatar
Baizhou Zhang committed
1872
1873
1874
1875
            if w.dtype in (
                torch.float8_e4m3fn,
                torch.float8_e4m3fnuz,
            ):
1876
1877
1878
1879
                if (
                    hasattr(self.quant_config, "weight_block_size")
                    and self.quant_config.weight_block_size is not None
                ):
Baizhou Zhang's avatar
Baizhou Zhang committed
1880
                    weight_block_size = self.quant_config.weight_block_size
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
                    assert hasattr(self_attn.kv_b_proj, "weight_scale_inv")
                    if _is_fp8_fnuz:
                        weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
                            weight=w,
                            weight_scale=self_attn.kv_b_proj.weight_scale_inv,
                            input_scale=None,
                        )
                    else:
                        weight = w
                        weight_scale = self_attn.kv_b_proj.weight_scale_inv

                    if (
                        _is_cuda
                        and weight_block_size[0] == 128
                        and weight_block_size[1] == 128
                        and model_dtype == torch.bfloat16
                    ):
1898
1899
1900
1901
                        if (
                            deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM
                            and not deep_gemm_wrapper.DEEPGEMM_BLACKWELL
                            and get_bool_env_var("SGL_USE_DEEPGEMM_BMM", "false")
1902
                        ):
1903
1904
                            block_scale = weight_scale
                            use_deep_gemm_bmm = True
1905
                        else:
1906
1907
1908
1909
1910
                            w = block_quant_dequant(
                                weight,
                                weight_scale,
                                weight_block_size,
                                model_dtype,
1911
                            )
1912
1913
1914
1915
1916
                    else:
                        w, scale = block_quant_to_tensor_quant(
                            weight, weight_scale, weight_block_size
                        )
                        self_attn.w_scale = scale
Baizhou Zhang's avatar
Baizhou Zhang committed
1917
                else:
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
                    if _is_fp8_fnuz:
                        weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
                            weight=w,
                            weight_scale=self_attn.kv_b_proj.weight_scale,
                            input_scale=None,
                        )
                    else:
                        weight = w
                        weight_scale = self_attn.kv_b_proj.weight_scale

Baizhou Zhang's avatar
Baizhou Zhang committed
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
                    w, scale = channel_quant_to_tensor_quant(weight, weight_scale)
                    self_attn.w_scale = scale

            if w.dtype == torch.int8:
                if hasattr(self.quant_config, "weight_block_size"):
                    # block-wise int8 need it
                    weight_block_size = self.quant_config.weight_block_size
                    if weight_block_size is not None:
                        assert hasattr(self_attn.kv_b_proj, "weight_scale_inv")
                        weight = w
                        weight_scale = self_attn.kv_b_proj.weight_scale_inv
                        w = int8_block_dequant(
                            weight, weight_scale, weight_block_size
                        ).to(torch.bfloat16)
                else:
                    # channel-wise int8 need it
                    w = w.to(torch.bfloat16) * self_attn.kv_b_proj.weight_scale.to(
                        torch.bfloat16
                    )
1947

Baizhou Zhang's avatar
Baizhou Zhang committed
1948
1949
1950
            w_kc, w_vc = w.unflatten(
                0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)
            ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)
1951
            if not use_deep_gemm_bmm:
1952
1953
1954
1955
1956
1957
                self_attn.w_kc = bind_or_assign(
                    self_attn.w_kc, w_kc.transpose(1, 2).contiguous().transpose(1, 2)
                )
                self_attn.w_vc = bind_or_assign(
                    self_attn.w_vc, w_vc.contiguous().transpose(1, 2)
                )
1958
1959
1960
1961
                if (
                    hasattr(self_attn.kv_b_proj, "weight_scale")
                    and self_attn.w_scale is None
                ):
1962
1963
1964
                    self_attn.w_scale = bind_or_assign(
                        self_attn.w_scale, self_attn.kv_b_proj.weight_scale
                    )
1965
1966
1967
1968
1969
1970
1971
1972
                    if _is_hip:
                        self_attn.w_scale *= 2.0
            else:
                num_tiles_k = self_attn.qk_nope_head_dim // weight_block_size[1]
                num_tiles_n = self_attn.v_head_dim // weight_block_size[0]
                ws_kc, ws_vc = block_scale.unflatten(
                    0, (-1, (num_tiles_k + num_tiles_n))
                ).split([num_tiles_k, num_tiles_n], dim=1)
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
                self_attn.w_scale_k = bind_or_assign(
                    self_attn.w_scale_k, ws_kc.transpose(1, 2).contiguous()
                )
                self_attn.w_scale_v = bind_or_assign(
                    self_attn.w_scale_v, ws_vc.contiguous()
                )
                self_attn.w_kc = bind_or_assign(
                    self_attn.w_kc, w_kc.transpose(1, 2).contiguous()
                )
                self_attn.w_vc = bind_or_assign(self_attn.w_vc, w_vc.contiguous())
1983
                self_attn.use_deep_gemm_bmm = True
inkcherry's avatar
inkcherry committed
1984

1985
1986
1987
        if (
            deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM
            and deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0
1988
1989
            and hasattr(self.quant_config, "weight_block_size")
            and self.quant_config.weight_block_size is not None
1990
        ):
1991
1992
1993
            self._weight_requant_ue8m0()

    def _weight_requant_ue8m0(self):
1994
1995
        if self.config.architectures[0] == "DeepseekV3ForCausalLMNextN":
            return
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
        weight_block_size = self.quant_config.weight_block_size

        moe_layers = list(
            range(
                self.config.first_k_dense_replace,
                self.config.num_hidden_layers,
                self.config.moe_layer_freq,
            )
        )

        for layer_id in range(self.config.num_hidden_layers):
            layer = self.model.layers[layer_id]

            for module in [
                layer.self_attn.fused_qkv_a_proj_with_mqa,
                layer.self_attn.q_b_proj,
                layer.self_attn.kv_b_proj,
                layer.self_attn.o_proj,
            ]:
                requant_weight_ue8m0_inplace(
                    module.weight, module.weight_scale_inv, weight_block_size
                )

            if layer_id in moe_layers:
2020
2021
2022
2023
2024
2025
2026
2027
2028
                shared_experts = getattr(layer.mlp, "shared_experts", None)
                if shared_experts is not None:
                    for module in [
                        shared_experts.gate_up_proj,
                        shared_experts.down_proj,
                    ]:
                        requant_weight_ue8m0_inplace(
                            module.weight, module.weight_scale_inv, weight_block_size
                        )
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047

                experts = layer.mlp.experts
                if isinstance(experts, DeepEPMoE):
                    for w in [
                        experts.w13_weight_fp8,
                        experts.w2_weight_fp8,
                    ]:
                        requant_weight_ue8m0_inplace(w[0], w[1], weight_block_size)
            else:
                mlp = layer.mlp
                assert isinstance(mlp, DeepseekV2MLP)
                for module in [
                    mlp.gate_up_proj,
                    mlp.down_proj,
                ]:
                    requant_weight_ue8m0_inplace(
                        module.weight, module.weight_scale_inv, weight_block_size
                    )

2048
    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]], is_nextn=False):
2049

2050
2051
2052
        if is_nextn:
            if hasattr(self.config, "num_nextn_predict_layers"):
                num_nextn_layers = self.config.num_nextn_predict_layers
2053
                assert num_nextn_layers == 1, "Only 1 nextn layer is supported"
2054
2055
2056
2057
2058
2059
2060
2061
2062
                # compatible with old design
                nextn_layer_id = (
                    0
                    if self.config.num_hidden_layers == 1
                    else self.config.num_hidden_layers
                )
            else:
                raise ValueError("num_nextn_predict_layers is not in the config")

Liangsheng Yin's avatar
Liangsheng Yin committed
2063
2064
2065
2066
2067
2068
2069
2070
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]

        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
2071
        expert_params_mapping = get_moe_impl_class().make_expert_params_mapping(
Liangsheng Yin's avatar
Liangsheng Yin committed
2072
2073
2074
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
2075
            num_experts=self.config.n_routed_experts + self.num_fused_shared_experts,
Liangsheng Yin's avatar
Liangsheng Yin committed
2076
2077
        )

2078
2079
2080
2081
2082
2083
        # Fuse q_a_proj and kv_a_proj_with_mqa along output dimension when q_lora_rank is not None
        fuse_qkv_a_proj = hasattr(self.config, "q_lora_rank") and (
            self.config.q_lora_rank is not None
        )
        cached_a_proj = {} if fuse_qkv_a_proj else None

2084
2085
2086
2087
2088
2089
2090
2091
2092
        if is_nextn:
            nextn_layer_prefix = f"model.layers.{nextn_layer_id}"
            nextn_spec_weight_names = [
                "shared_head.norm",
                "eh_proj",
                "enorm",
                "hnorm",
            ]

2093
2094
        if self.num_fused_shared_experts > 0:
            assert self.num_fused_shared_experts == 1
2095
            log_info_on_rank0(logger, "Shared experts fusion optimization enabled.")
2096

Liangsheng Yin's avatar
Liangsheng Yin committed
2097
        params_dict = dict(self.named_parameters())
2098
        weight_names = []
Liangsheng Yin's avatar
Liangsheng Yin committed
2099
        for name, loaded_weight in weights:
2100
2101
2102
2103
2104
2105
            if self.num_fused_shared_experts > 0 and "mlp.shared_experts" in name:
                name = name.replace(
                    "mlp.shared_experts",
                    f"mlp.experts.{self.config.n_routed_experts}",
                )

2106
2107
            weight_names.append(name)

2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
            if not is_nextn:
                if hasattr(self.config, "num_nextn_predict_layers"):
                    num_nextn_layers = self.config.num_nextn_predict_layers
                    if num_nextn_layers > 0 and name.startswith("model.layers"):
                        name_list = name.split(".")
                        if (
                            len(name_list) >= 3
                            and int(name_list[2]) >= self.config.num_hidden_layers
                        ):
                            continue
            else:
                if not name.startswith(nextn_layer_prefix):
                    continue

                # Use shared head and embed weights from target model
                if "shared_head.head" in name or "embed_tokens" in name:
                    continue

                is_decoder = True
                # For nextn specific weights
                for weight_name in nextn_spec_weight_names:
                    if weight_name in name:
                        name = name.replace(nextn_layer_prefix, "model")
                        is_decoder = False
                        break
                # For decoder layer weights
                if is_decoder:
                    name = name.replace(nextn_layer_prefix, "model.decoder")

Liangsheng Yin's avatar
Liangsheng Yin committed
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
            if "rotary_emb.inv_freq" in name:
                continue
            for param_name, weight_name, shard_id in stacked_params_mapping:
                # Skip non-stacked layers and experts (experts handled below).
                if weight_name not in name:
                    continue
                # We have mlp.experts[0].gate_proj in the checkpoint.
                # Since we handle the experts below in expert_params_mapping,
                # we need to skip here BEFORE we update the name, otherwise
                # name will be updated to mlp.experts[0].gate_up_proj, which
                # will then be updated below in expert_params_mapping
                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
                if ("mlp.experts." in name) and name not in params_dict:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
                    name = name.replace(weight_name, param_name)
                    param = params_dict[name]
                    weight_loader = param.weight_loader
                    weight_loader(
                        param,
                        loaded_weight,
2170
                        name,
Liangsheng Yin's avatar
Liangsheng Yin committed
2171
2172
2173
2174
2175
2176
2177
2178
                        shard_id=shard_id,
                        expert_id=expert_id,
                    )
                    break
                else:
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
                    if fuse_qkv_a_proj and (
                        "q_a_proj" in name or "kv_a_proj_with_mqa" in name
                    ):
                        cached_a_proj[name] = loaded_weight
                        q_a_proj_name = (
                            name
                            if "q_a_proj" in name
                            else name.replace("kv_a_proj_with_mqa", "q_a_proj")
                        )
                        kv_a_proj_name = (
                            name
                            if "kv_a_proj_with_mqa" in name
                            else name.replace("q_a_proj", "kv_a_proj_with_mqa")
                        )

                        # When both q_a_proj and kv_a_proj_with_mqa has been cached, load the fused weight to parameter
                        if (
                            q_a_proj_name in cached_a_proj
                            and kv_a_proj_name in cached_a_proj
                        ):
                            q_a_proj_weight = cached_a_proj[q_a_proj_name]
                            kv_a_proj_weight = cached_a_proj[kv_a_proj_name]
2201
2202
2203
2204
2205
2206
                            cat_dim = 0
                            if (
                                self.quant_config.get_name() == "awq"
                                or self.quant_config.get_name() == "moe_wna16"
                            ):
                                cat_dim = 1
2207
                            fused_weight = torch.cat(
2208
                                [q_a_proj_weight, kv_a_proj_weight], dim=cat_dim
2209
                            )
2210
2211
2212
2213
2214
2215
                            param_name = (
                                name.replace("q_a_proj", "fused_qkv_a_proj_with_mqa")
                                if "q_a_proj" in name
                                else name.replace(
                                    "kv_a_proj_with_mqa", "fused_qkv_a_proj_with_mqa"
                                )
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
                            )
                            param = params_dict[param_name]

                            weight_loader = getattr(
                                param, "weight_loader", default_weight_loader
                            )
                            weight_loader(param, fused_weight)
                            cached_a_proj.pop(q_a_proj_name)
                            cached_a_proj.pop(kv_a_proj_name)
                    else:
2226
2227
2228
2229
                        if (
                            "k_scale" in name or "v_scale" in name
                        ) and name not in params_dict:
                            # modelopt attn kv scale is named differently
2230
2231
2232
                            for scale in ["k_scale", "v_scale"]:
                                if scale in name:
                                    name = name.replace(f"{scale[0]}_proj", "attn_mqa")
2233
2234
2235
2236
2237
                        param = params_dict[name]
                        weight_loader = getattr(
                            param, "weight_loader", default_weight_loader
                        )
                        weight_loader(param, loaded_weight)
Liangsheng Yin's avatar
Liangsheng Yin committed
2238

2239
        self.post_load_weights(is_nextn=is_nextn, weight_names=weight_names)
Ke Bao's avatar
Ke Bao committed
2240

2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
    def get_embed_and_head(self):
        return self.model.embed_tokens.weight, self.lm_head.weight

    def set_embed_and_head(self, embed, head):
        del self.model.embed_tokens.weight
        del self.lm_head.weight
        self.model.embed_tokens.weight = embed
        self.lm_head.weight = head
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

2252
2253
2254
2255
2256
2257
2258
2259
    @classmethod
    def get_model_config_for_expert_location(cls, config):
        return ModelConfigForExpertLocation(
            num_layers=config.num_hidden_layers,
            num_logical_experts=config.n_routed_experts,
            num_groups=config.n_group,
        )

Liangsheng Yin's avatar
Liangsheng Yin committed
2260

HandH1998's avatar
HandH1998 committed
2261
2262
2263
2264
2265
class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass


EntryClass = [DeepseekV2ForCausalLM, DeepseekV3ForCausalLM]