qwen3_next.py 48.7 KB
Newer Older
1
2
3
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Inference-only Qwen3Next model."""
4

5
from collections.abc import Iterable
6
from itertools import islice
7
8
9
10
11
12
13
14

import torch
from einops import rearrange
from torch import nn
from transformers.activations import ACT2FN

from vllm.attention import Attention, AttentionBackend, AttentionMetadata
from vllm.compilation.decorators import support_torch_compile
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
from vllm.config import (
    CacheConfig,
    ModelConfig,
    SpeculativeConfig,
    VllmConfig,
    get_current_vllm_config,
)
from vllm.distributed import (
    divide,
    get_ep_group,
    get_pp_group,
    get_tensor_model_parallel_rank,
    get_tensor_model_parallel_world_size,
    tensor_model_parallel_all_gather,
)
30
31
32
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.layers.fla.ops import (
33
    RMSNormGated,
34
35
36
    chunk_gated_delta_rule,
    fused_recurrent_gated_delta_rule,
)
37
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
38
from vllm.model_executor.layers.layernorm import GemmaRMSNorm as Qwen3NextRMSNorm
39
40
41
42
43
44
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    QKVParallelLinear,
    ReplicatedLinear,
    RowParallelLinear,
)
45
46
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.abstract import MambaBase
47
from vllm.model_executor.layers.mamba.mamba_mixer2 import mamba_v2_sharded_weight_loader
48
from vllm.model_executor.layers.mamba.mamba_utils import (
49
50
51
    MambaStateDtypeCalculator,
    MambaStateShapeCalculator,
)
52
from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
53
54
55
    causal_conv1d_fn,
    causal_conv1d_update,
)
56
57
58
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
59
60
61
62
    DEFAULT_VOCAB_PADDING_SIZE,
    ParallelLMHead,
    VocabParallelEmbedding,
)
63
from vllm.model_executor.model_loader.weight_utils import (
64
65
66
    default_weight_loader,
    sharded_weight_loader,
)
67
from vllm.model_executor.models.qwen2_moe import Qwen2MoeMLP as Qwen3NextMLP
68
from vllm.model_executor.models.utils import sequence_parallel_chunk
69
70
71
72
73
from vllm.model_executor.utils import set_weight_attrs
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs import Qwen3NextConfig
from vllm.triton_utils import tl, triton
74
from vllm.utils.torch_utils import direct_register_custom_op
75
76
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata

77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
from .interfaces import (
    HasInnerState,
    IsHybrid,
    MixtureOfExperts,
    SupportsLoRA,
    SupportsPP,
)
from .utils import (
    AutoWeightsLoader,
    PPMissingLayer,
    extract_layer_index,
    is_pp_missing_parameter,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
93
94
95
96
97
98
99

logger = init_logger(__name__)

KVCache = tuple[torch.Tensor, torch.Tensor]


class Qwen3NextSparseMoeBlock(nn.Module):
100
    def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
101
        super().__init__()
102
103
104
105
106

        config = vllm_config.model_config.hf_config
        parallel_config = vllm_config.parallel_config
        quant_config = vllm_config.quant_config

107
108
109
110
111
112
113
        self.tp_size = get_tensor_model_parallel_world_size()

        self.ep_group = get_ep_group().device_group
        self.ep_rank = self.ep_group.rank()
        self.ep_size = self.ep_group.size()
        self.n_routed_experts = config.num_experts

114
115
        self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe

116
117
118
        if self.tp_size > config.num_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
119
120
                f"the number of experts {config.num_experts}."
            )
121
122
123
124

        # Load balancing settings.
        vllm_config = get_current_vllm_config()
        eplb_config = vllm_config.parallel_config.eplb_config
125
        self.enable_eplb = parallel_config.enable_eplb
126
127
128

        self.n_logical_experts = self.n_routed_experts
        self.n_redundant_experts = eplb_config.num_redundant_experts
129
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
130
131
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

132
133
134
135
136
137
138
139
140
141
142
143
        self.physical_expert_start = self.ep_rank * self.n_local_physical_experts
        self.physical_expert_end = (
            self.physical_expert_start + self.n_local_physical_experts
        )

        self.gate = ReplicatedLinear(
            config.hidden_size,
            config.num_experts,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate",
        )
144

145
146
        self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)

147
148
149
150
151
152
        if config.shared_expert_intermediate_size > 0:
            self.shared_expert = Qwen3NextMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.shared_expert_intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
153
154
                reduce_results=False,
                expert_gate=self.shared_expert_gate,
155
                prefix=f"{prefix}.shared_expert",
156
157
158
            )
        else:
            self.shared_expert = None
159
160
161

        self.experts = SharedFusedMoE(
            shared_experts=self.shared_expert,
162
            gate=self.gate,
163
164
165
166
167
168
169
170
171
172
173
174
            num_experts=self.n_routed_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            reduce_results=False,
            renormalize=config.norm_topk_prob,
            quant_config=quant_config,
            prefix=f"{prefix}.experts",
            enable_eplb=self.enable_eplb,
            num_redundant_experts=self.n_redundant_experts,
            is_sequence_parallel=self.is_sequence_parallel,
        )
175
176
177
178

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # NOTE: hidden_states can have either 1D or 2D shape.
        orig_shape = hidden_states.shape
179
        num_tokens, hidden_dim = hidden_states.shape
180
181
        hidden_states = hidden_states.view(-1, hidden_dim)

182
183
184
        if self.is_sequence_parallel:
            hidden_states = sequence_parallel_chunk(hidden_states)

185
186
187
188
189
190
191
192
193
194
195
        if self.experts.is_internal_router:
            # In this case, the gate/router runs inside the FusedMoE class
            final_hidden_states = self.experts(
                hidden_states=hidden_states, router_logits=hidden_states
            )
        else:
            # 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
            )
196

197
198
        if self.shared_expert is not None:
            final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
199
200
201

        if self.is_sequence_parallel:
            final_hidden_states = tensor_model_parallel_all_gather(
202
203
                final_hidden_states, 0
            )
204
205
            final_hidden_states = final_hidden_states[:num_tokens]
        elif self.tp_size > 1:
206
            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
207
208
                final_hidden_states
            )
209
210
211
212
213
214
215
216
217
218
219

        return final_hidden_states.view(orig_shape)


class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
    @property
    def mamba_type(self) -> str:
        return "linear_attention"

    def get_attn_backend(self) -> type["AttentionBackend"]:
        from vllm.v1.attention.backends.gdn_attn import GDNAttentionBackend
220

221
222
223
224
        return GDNAttentionBackend

    def get_state_dtype(self) -> tuple[torch.dtype, torch.dtype]:
        return MambaStateDtypeCalculator.gated_delta_net_state_dtype(
225
226
            self.model_config.dtype, self.cache_config.mamba_cache_dtype
        )
227
228
229

    def get_state_shape(self) -> tuple[tuple[int, ...], tuple[int, ...]]:
        return MambaStateShapeCalculator.gated_delta_net_state_shape(
230
231
232
233
234
235
236
237
            self.tp_size,
            self.num_k_heads,
            self.num_v_heads,
            self.head_k_dim,
            self.head_v_dim,
            self.conv_kernel_size,
            self.num_spec,
        )
238
239
240
241

    def __init__(
        self,
        config: Qwen3NextConfig,
242
243
244
245
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        speculative_config: SpeculativeConfig | None = None,
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.tp_rank = get_tensor_model_parallel_rank()
        self.hidden_size = config.hidden_size
        self.num_v_heads = config.linear_num_value_heads
        self.num_k_heads = config.linear_num_key_heads
        self.head_k_dim = config.linear_key_head_dim
        self.head_v_dim = config.linear_value_head_dim
        self.key_dim = self.head_k_dim * self.num_k_heads
        self.value_dim = self.head_v_dim * self.num_v_heads

        self.conv_kernel_size = config.linear_conv_kernel_dim
        self.layer_idx = extract_layer_index(prefix)
        self.activation = config.hidden_act
        self.act = ACT2FN[config.hidden_act]
        self.layer_norm_epsilon = config.rms_norm_eps
        self.prefix = prefix

        self.config = config
        self.model_config = model_config
        self.cache_config = cache_config
        self.quant_config = quant_config
        self.speculative_config = speculative_config
271
272
273
274
275
        self.num_spec = (
            self.speculative_config.num_speculative_tokens
            if self.speculative_config
            else 0
        )
276
277
278
279
280
281
282
283
284
285
286
287
288
289

        # QKV
        self.conv_dim = self.key_dim * 2 + self.value_dim
        self.conv1d = ColumnParallelLinear(
            input_size=self.conv_kernel_size,
            output_size=self.conv_dim,
            bias=False,
            prefix=f"{prefix}.conv1d",
        )
        self.conv1d.weight.data = self.conv1d.weight.data.unsqueeze(1)

        # projection of the input hidden states
        self.projection_size_qkvz = self.key_dim * 2 + self.value_dim * 2
        self.projection_size_ba = self.num_v_heads * 2
290
        self.in_proj_qkvz = ColumnParallelLinear(
291
            input_size=self.hidden_size,
292
            output_size=self.projection_size_qkvz,
293
294
            bias=False,
            quant_config=quant_config,
295
296
297
298
299
300
301
302
303
            prefix=f"{prefix}.in_proj_qkvz",
        )
        # ba_proj doesn't support blockwise fp8 quantization.
        self.in_proj_ba = ColumnParallelLinear(
            input_size=self.hidden_size,
            output_size=self.projection_size_ba,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.in_proj_ba",
304
305
306
307
308
309
310
        )

        query_key_settings = (self.key_dim, 0, False)
        value_settings = (self.value_dim, 0, False)

        delattr(self.conv1d.weight, "weight_loader")
        set_weight_attrs(
311
312
313
314
315
316
317
318
319
320
321
322
323
            self.conv1d.weight,
            {
                "weight_loader": mamba_v2_sharded_weight_loader(
                    [
                        query_key_settings,
                        query_key_settings,
                        value_settings,
                    ],
                    self.tp_size,
                    self.tp_rank,
                )
            },
        )
324
325
326
327
328
329

        # selective projection used to make dt, B and C input dependant

        # time step projection (discretization)
        # instantiate once and copy inv_dt in init_weights of PretrainedModel
        self.dt_bias = nn.Parameter(
330
331
            torch.ones(self.num_v_heads // self.tp_size),
        )
332
333
334
        self.A_log = nn.Parameter(
            torch.empty(
                divide(self.num_v_heads, self.tp_size),
335
336
            )
        )
337

338
339
        set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(0)})
        set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)})
340
341
342
343
344
345

        self.norm = RMSNormGated(
            self.head_v_dim,
            eps=self.layer_norm_epsilon,
            group_size=None,
            norm_before_gate=True,
346
            device=current_platform.current_device(),
347
            dtype=config.dtype,
348
349
        )

350
351
352
353
354
355
356
357
        self.out_proj = RowParallelLinear(
            self.value_dim,
            self.hidden_size,
            bias=False,
            input_is_parallel=True,
            quant_config=quant_config,
            prefix=f"{prefix}.out_proj",
        )
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373

        compilation_config = get_current_vllm_config().compilation_config
        if prefix in compilation_config.static_forward_context:
            raise ValueError(f"Duplicate layer name: {prefix}")
        compilation_config.static_forward_context[prefix] = self

    def fix_query_key_value_ordering(
        self,
        mixed_qkvz,
        mixed_ba,
    ):
        """
        Derives `query`, `key` and `value` tensors from `mixed_qkvzba`.
        """
        new_tensor_shape_qkvz = mixed_qkvz.size()[:-1] + (
            self.num_k_heads // self.tp_size,
374
375
376
377
378
379
380
            (
                self.head_k_dim
                + self.head_k_dim
                + (self.head_v_dim + self.head_v_dim)
                * self.num_v_heads
                // self.num_k_heads
            ),
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
        )
        new_tensor_shape_ba = mixed_qkvz.size()[:-1] + (
            self.num_k_heads // self.tp_size,
            2 * self.num_v_heads // self.num_k_heads,
        )

        mixed_qkvz = mixed_qkvz.view(*new_tensor_shape_qkvz)
        mixed_ba = mixed_ba.view(*new_tensor_shape_ba)

        split_arg_list_qkvz = [
            self.head_k_dim,
            self.head_k_dim,
            (self.num_v_heads // self.num_k_heads * self.head_v_dim),
            (self.num_v_heads // self.num_k_heads * self.head_v_dim),
        ]
        split_arg_list_ba = [
            self.num_v_heads // self.num_k_heads,
398
            self.num_v_heads // self.num_k_heads,
399
400
401
402
403
        ]

        # [b, sq, ng, (hn + hn + np/ng * hn + np/ng + np/ng)]
        # --> [b, sq, ng, hn], [b, sq, ng, hn], [b, sq, ng, np/ng * hn],
        #  [b, sq, ng, np/ng * hn], [b, sq, ng, np/ng], [b, sq, ng, np/ng]
404
        (query, key, value, z) = torch.split(mixed_qkvz, split_arg_list_qkvz, dim=2)
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
        (b, a) = torch.split(mixed_ba, split_arg_list_ba, dim=2)

        # [b, sq, ng, np/ng * hn] -> [b, sq, np, hn]
        value = value.reshape(value.size(0), -1, self.head_v_dim)
        z = z.reshape(z.size(0), -1, self.head_v_dim)
        b = b.reshape(b.size(0), self.num_v_heads // self.tp_size)
        a = a.reshape(a.size(0), self.num_v_heads // self.tp_size)

        return query, key, value, z, b, a

    def rearrange_mixed_qkv(self, mixed_qkv):
        if mixed_qkv is None:
            return None, None, None
        query, key, value = torch.split(
            mixed_qkv,
            [
                self.key_dim // self.tp_size,
                self.key_dim // self.tp_size,
                self.value_dim // self.tp_size,
            ],
            dim=-1,
        )
        query, key = map(
428
429
430
431
            lambda x: rearrange(x, "l (h d) -> 1 l h d", d=self.head_k_dim),
            (query, key),
        )
        value = rearrange(value, "l (h d) -> 1 l h d", d=self.head_v_dim)
432
        return query.contiguous(), key.contiguous(), value.contiguous()
433
434
435
436
437
438

    def forward(
        self,
        hidden_states: torch.Tensor,
        output: torch.Tensor,
    ):
439
440
441
        return torch.ops.vllm.gdn_attention(
            hidden_states,
            output,
442
443
444
            self.prefix,
        )

445
    def _forward(
446
        self,
447
448
        hidden_states: torch.Tensor,
        output: torch.Tensor,
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
    ):
        forward_context = get_forward_context()
        attn_metadata: AttentionMetadata = forward_context.attn_metadata

        if attn_metadata is None:
            # V1 profile run
            return

        assert isinstance(attn_metadata, dict)
        attn_metadata = attn_metadata[self.prefix]
        assert isinstance(attn_metadata, GDNAttentionMetadata)
        has_initial_state = attn_metadata.has_initial_state
        spec_query_start_loc = attn_metadata.spec_query_start_loc
        non_spec_query_start_loc = attn_metadata.non_spec_query_start_loc
        spec_sequence_masks = attn_metadata.spec_sequence_masks
464
465
        spec_token_indx = attn_metadata.spec_token_indx
        non_spec_token_indx = attn_metadata.non_spec_token_indx
466
467
468
469
470
        spec_state_indices_tensor = attn_metadata.spec_state_indices_tensor  # noqa: E501
        non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor  # noqa: E501
        self_kv_cache = self.kv_cache[forward_context.virtual_engine]
        conv_state = self_kv_cache[0].transpose(-1, -2)
        ssm_state = self_kv_cache[1]
471
        num_actual_tokens = attn_metadata.num_actual_tokens
472
        num_accepted_tokens = attn_metadata.num_accepted_tokens
473

474
475
476
477
478
479
480
481
482
483
        # 1. Set up dimensions for reshapes later
        projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states[:num_actual_tokens])
        projected_states_ba, _ = self.in_proj_ba(hidden_states[:num_actual_tokens])
        query, key, value, z, b, a = self.fix_query_key_value_ordering(
            projected_states_qkvz, projected_states_ba
        )
        query, key, value = map(
            lambda x: rearrange(x, "l p d -> l (p d)"), (query, key, value)
        )
        mixed_qkv = torch.cat((query, key, value), dim=-1)
484

485
        # 2. Convolution sequence transformation
486
487
488
        conv_weights = self.conv1d.weight.view(
            self.conv1d.weight.size(0), self.conv1d.weight.size(2)
        )
489
490

        if spec_sequence_masks is not None:
491
            if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0:
492
493
494
                mixed_qkv_spec = mixed_qkv
                mixed_qkv_non_spec = None
            else:
495
496
                mixed_qkv_spec = mixed_qkv.index_select(0, spec_token_indx)
                mixed_qkv_non_spec = mixed_qkv.index_select(0, non_spec_token_indx)
497
498
499
500
        else:
            mixed_qkv_spec = None
            mixed_qkv_non_spec = mixed_qkv

501
        # 2.1: process the mutli-query part
502
503
504
505
506
507
508
        if spec_sequence_masks is not None:
            mixed_qkv_spec = causal_conv1d_update(
                mixed_qkv_spec,
                conv_state,
                conv_weights,
                self.conv1d.bias,
                self.activation,
509
510
511
                conv_state_indices=spec_state_indices_tensor[:, 0][
                    : attn_metadata.num_spec_decodes
                ],
512
                num_accepted_tokens=num_accepted_tokens,
513
514
                query_start_loc=spec_query_start_loc,
                max_query_len=spec_state_indices_tensor.size(-1),
515
516
517
                validate_data=False,
            )

518
        # 2.2: process the remaining part
519
        if attn_metadata.num_prefills > 0:
520
            mixed_qkv_non_spec_T = mixed_qkv_non_spec.transpose(0, 1)
521
            # - "cache_indices" updates the conv_state cache in positions
522
            #   pointed to by "state_indices_tensor"
523
            mixed_qkv_non_spec = causal_conv1d_fn(
524
                mixed_qkv_non_spec_T,
525
526
527
528
529
530
531
                conv_weights,
                self.conv1d.bias,
                activation=self.activation,
                conv_states=conv_state,
                has_initial_state=has_initial_state,
                cache_indices=non_spec_state_indices_tensor,
                query_start_loc=non_spec_query_start_loc,
532
                metadata=attn_metadata,
533
534
535
536
537
538
539
540
            ).transpose(0, 1)
        elif attn_metadata.num_decodes > 0:
            mixed_qkv_non_spec = causal_conv1d_update(
                mixed_qkv_non_spec,
                conv_state,
                conv_weights,
                self.conv1d.bias,
                self.activation,
541
542
543
                conv_state_indices=non_spec_state_indices_tensor[
                    : attn_metadata.num_decodes
                ],
544
545
546
547
548
                validate_data=True,
            )
        else:
            mixed_qkv_non_spec = None

549
        query_spec, key_spec, value_spec = self.rearrange_mixed_qkv(mixed_qkv_spec)
550
        query_non_spec, key_non_spec, value_non_spec = self.rearrange_mixed_qkv(
551
552
            mixed_qkv_non_spec
        )
553
554
555
556

        beta = b.sigmoid()
        # g = -self.A_log.float().exp() * F.softplus(a.float() + self.dt_bias)
        g = fused_gdn_gating(self.A_log, a, self.dt_bias)
557
        g, beta = map(lambda x: rearrange(x, "l d -> 1 l d"), (g, beta))
558
559

        if spec_sequence_masks is not None:
560
            if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0:
561
562
563
564
565
                g_spec = g
                beta_spec = beta
                g_non_spec = None
                beta_non_spec = None
            else:
566
567
568
569
                g_spec = g.index_select(1, spec_token_indx)
                beta_spec = beta.index_select(1, spec_token_indx)
                g_non_spec = g.index_select(1, non_spec_token_indx)
                beta_non_spec = beta.index_select(1, non_spec_token_indx)
570
571
572
573
574
575
        else:
            g_spec = None
            beta_spec = None
            g_non_spec = g
            beta_non_spec = beta

576
        # 3. Recurrent attention
577

578
        # 3.1: process the mutlti-query part
579
        if spec_sequence_masks is not None:
580
581
582
583
584
585
586
587
588
589
590
591
592
            core_attn_out_spec, last_recurrent_state = fused_recurrent_gated_delta_rule(
                q=query_spec,
                k=key_spec,
                v=value_spec,
                g=g_spec,
                beta=beta_spec,
                initial_state=ssm_state,
                inplace_final_state=True,
                cu_seqlens=spec_query_start_loc[: attn_metadata.num_spec_decodes + 1],
                ssm_state_indices=spec_state_indices_tensor,
                num_accepted_tokens=num_accepted_tokens,
                use_qk_l2norm_in_kernel=True,
            )
593
594
595
        else:
            core_attn_out_spec, last_recurrent_state = None, None

596
        # 3.2: process the remaining part
597
        if attn_metadata.num_prefills > 0:
598
            initial_state = ssm_state[non_spec_state_indices_tensor].contiguous()
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
            initial_state[~has_initial_state, ...] = 0
            (
                core_attn_out_non_spec,
                last_recurrent_state,
            ) = chunk_gated_delta_rule(
                q=query_non_spec,
                k=key_non_spec,
                v=value_non_spec,
                g=g_non_spec,
                beta=beta_non_spec,
                initial_state=initial_state,
                output_final_state=True,
                cu_seqlens=non_spec_query_start_loc,
                head_first=False,
                use_qk_l2norm_in_kernel=True,
            )
            # Init cache
            ssm_state[non_spec_state_indices_tensor] = last_recurrent_state.to(
617
618
                ssm_state.dtype
            )
619
620
621
622
623
624
625
626
627
628
        elif attn_metadata.num_decodes > 0:
            core_attn_out_non_spec, last_recurrent_state = (
                fused_recurrent_gated_delta_rule(
                    q=query_non_spec,
                    k=key_non_spec,
                    v=value_non_spec,
                    g=g_non_spec,
                    beta=beta_non_spec,
                    initial_state=ssm_state,
                    inplace_final_state=True,
629
630
631
                    cu_seqlens=non_spec_query_start_loc[
                        : attn_metadata.num_decodes + 1
                    ],
632
633
                    ssm_state_indices=non_spec_state_indices_tensor,
                    use_qk_l2norm_in_kernel=True,
634
635
                )
            )
636
637
638
        else:
            core_attn_out_non_spec, last_recurrent_state = None, None

639
        # Merge core attention output
640
        if spec_sequence_masks is not None and core_attn_out_non_spec is not None:
641
            core_attn_out = torch.empty(
642
643
644
645
                (1, num_actual_tokens, *core_attn_out_spec.shape[2:]),
                dtype=core_attn_out_non_spec.dtype,
                device=core_attn_out_non_spec.device,
            )
646
647
648
            core_attn_out.index_copy_(1, spec_token_indx, core_attn_out_spec)
            core_attn_out.index_copy_(1, non_spec_token_indx, core_attn_out_non_spec)

649
        elif spec_sequence_masks is not None:
650
            core_attn_out = core_attn_out_spec
651
        else:
652
653
654
655
656
657
658
659
660
661
662
            core_attn_out = core_attn_out_non_spec

        z_shape_og = z.shape
        # reshape input data into 2D tensor
        core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1])
        z = z.reshape(-1, z.shape[-1])
        core_attn_out = self.norm(core_attn_out, z)
        core_attn_out = core_attn_out.reshape(z_shape_og)
        core_attn_out = rearrange(core_attn_out, "... h d -> ... (h d)")

        output[:num_actual_tokens], _ = self.out_proj(core_attn_out)
663
664
665
666
667
668


class Qwen3NextAttention(nn.Module):
    def __init__(
        self,
        config: Qwen3NextConfig,
669
670
671
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = config.num_attention_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = config.num_key_value_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = config.head_dim or (self.hidden_size // self.num_heads)
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.dual_chunk_attention_config = getattr(
696
697
            config, "dual_chunk_attention_config", None
        )
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
        self.attn_output_gate = getattr(config, "attn_output_gate", True)

        self.qkv_proj = QKVParallelLinear(
            config.hidden_size,
            self.head_dim,
            self.total_num_heads * (1 + self.attn_output_gate),
            self.total_num_kv_heads,
            bias=getattr(config, "qkv_bias", False),
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )

        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            config.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )

        self.rotary_emb = get_rope(
            head_size=self.head_dim,
            rotary_dim=self.head_dim,
            max_position=config.max_position_embeddings,
            base=config.rope_theta,
            rope_scaling=config.rope_scaling,
            partial_rotary_factor=config.partial_rotary_factor,
            dual_chunk_attention_config=self.dual_chunk_attention_config,
        )

        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
            **{
                "layer_idx": extract_layer_index(prefix),
738
739
740
741
                "dual_chunk_attention_config": self.dual_chunk_attention_config,
            }
            if self.dual_chunk_attention_config
            else {},
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
        )

        self.q_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.k_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        output: torch.Tensor,
        hidden_states: torch.Tensor,
    ):
        qkv, _ = self.qkv_proj(hidden_states)

        if self.attn_output_gate:
            q_gate, k, v = qkv.split(
757
758
                [self.q_size * 2, self.kv_size, self.kv_size], dim=-1
            )
759
760
761
762
763
764
            orig_shape = q_gate.shape[:-1]
            q_gate = q_gate.view(*orig_shape, self.num_heads, -1)
            q, gate = torch.chunk(q_gate, 2, dim=-1)
            q = q.reshape(*orig_shape, -1)
            gate = gate.reshape(*orig_shape, -1)
        else:
765
            q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
766
767

        q = self.q_norm(q.view(-1, self.num_heads, self.head_dim)).view(
768
769
            -1, self.num_heads * self.head_dim
        )
770
        k = self.k_norm(k.view(-1, self.num_kv_heads, self.head_dim)).view(
771
772
            -1, self.num_kv_heads * self.head_dim
        )
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787

        q, k = self.rotary_emb(positions, q, k)

        attn_output = self.attn(q, k, v)

        if self.attn_output_gate:
            gate = torch.sigmoid(gate)
            attn_output = attn_output * gate

        output[:], _ = self.o_proj(attn_output)


class Qwen3NextDecoderLayer(nn.Module):
    def __init__(
        self,
788
        vllm_config: VllmConfig,
789
790
791
792
        layer_type: str,
        prefix: str = "",
    ) -> None:
        super().__init__()
793
794
795
796
797
798

        config = vllm_config.model_config.hf_config
        model_config = vllm_config.model_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        speculative_config = vllm_config.speculative_config
799
800
801
802
803
804
805
806
807
808
809

        self.layer_type = layer_type
        self.layer_idx = extract_layer_index(prefix)

        if self.layer_type == "linear_attention":
            self.linear_attn = Qwen3NextGatedDeltaNet(
                config,
                model_config=model_config,
                cache_config=cache_config,
                quant_config=quant_config,
                speculative_config=speculative_config,
810
811
                prefix=f"{prefix}.linear_attn",
            )
812
813
814
815
816
817
        elif self.layer_type == "full_attention":
            self.self_attn = Qwen3NextAttention(
                config,
                model_config=model_config,
                cache_config=cache_config,
                quant_config=quant_config,
818
                prefix=f"{prefix}.self_attn",
819
820
821
822
            )
        else:
            raise ValueError(f"Invalid layer_type {self.layer_type}")

823
824
825
        mlp_only_layers = (
            [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
        )
826
        if (self.layer_idx not in mlp_only_layers) and (
827
828
829
            config.num_experts > 0
            and (self.layer_idx + 1) % config.decoder_sparse_step == 0
        ):
830
            self.mlp = Qwen3NextSparseMoeBlock(
831
                vllm_config=vllm_config,
832
833
834
835
836
837
838
839
840
841
                prefix=f"{prefix}.mlp",
            )
        else:
            self.mlp = Qwen3NextMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
            )

842
843
844
        self.input_layernorm = Qwen3NextRMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )
845
        self.post_attention_layernorm = Qwen3NextRMSNorm(
846
847
            config.hidden_size, eps=config.rms_norm_eps
        )
848
849
850
851
852
853
854

        self.layer_scale = getattr(config, "layer_scale", False)
        if self.layer_scale:
            self.attn_layer_scale = torch.nn.Parameter(
                torch.zeros(
                    1,
                    1,
855
                    config.hidden_size,
856
                    dtype=config.dtype,
857
858
                ),
            )
859
860
861
862
            self.ffn_layer_scale = torch.nn.Parameter(
                torch.zeros(
                    1,
                    1,
863
                    config.hidden_size,
864
                    dtype=config.dtype,
865
866
                ),
            )
867
868
869
870

    def forward(
        self,
        hidden_states: torch.Tensor,
871
        residual: torch.Tensor | None,
872
873
874
875
876
877
878
        positions: torch.Tensor = None,
        **kwargs: object,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
879
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899

        self_attention_output = torch.empty_like(hidden_states)
        if self.layer_type == "linear_attention":
            self.linear_attn(
                hidden_states=hidden_states,
                output=self_attention_output,
            )
        elif self.layer_type == "full_attention":
            self.self_attn(
                hidden_states=hidden_states,
                output=self_attention_output,
                positions=positions,
            )
        else:
            raise ValueError("Invalid layer_type")
        hidden_states = self_attention_output

        if self.layer_scale:
            if len(hidden_states.shape) == 2:
                hidden_states = hidden_states * (
900
901
                    self.attn_layer_scale.to(hidden_states.dtype)[0] + 1
                )
902
903
            else:
                hidden_states = hidden_states * (
904
905
                    self.attn_layer_scale.to(hidden_states.dtype) + 1
                )
906
907

        # Fully Connected
908
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
909
910
911
912
913
        hidden_states = self.mlp(hidden_states)

        if self.layer_scale:
            if len(hidden_states.shape) == 2:
                hidden_states = hidden_states * (
914
915
                    self.ffn_layer_scale.to(hidden_states.dtype)[0] + 1
                )
916
            else:
917
                assert len(hidden_states.shape) == len(self.ffn_layer_scale.shape), (
918
919
920
                    f"shape must be the same {len(hidden_states.shape)}, "
                    f"{len(self.ffn_layer_scale.shape)}"
                )
921
                hidden_states = hidden_states * (
922
923
                    self.ffn_layer_scale.to(hidden_states.dtype) + 1
                )
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939

        return hidden_states, residual


@support_torch_compile
class Qwen3NextModel(nn.Module):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config: Qwen3NextConfig = vllm_config.model_config.hf_config
        parallel_config = vllm_config.parallel_config
        lora_config = vllm_config.lora_config
        eplb_config = parallel_config.eplb_config
        self.num_redundant_experts = eplb_config.num_redundant_experts

        self.config = config
940
941
942
943
944
        lora_vocab = (
            (lora_config.lora_extra_vocab_size * (lora_config.max_loras or 1))
            if lora_config
            else 0
        )
945
946
947
948
949
950
951
952
953
954
        self.vocab_size = config.vocab_size + lora_vocab

        self.embed_tokens = VocabParallelEmbedding(
            self.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
        )

        def get_layer(prefix: str):
            return Qwen3NextDecoderLayer(
955
                vllm_config,
956
957
958
959
960
                layer_type=config.layer_types[extract_layer_index(prefix)],
                prefix=prefix,
            )

        self.start_layer, self.end_layer, self.layers = make_layers(
961
962
963
964
965
            config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers"
        )
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
966

967
        if get_pp_group().is_last_rank:
968
            self.norm = Qwen3NextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
969
970
        else:
            self.norm = PPMissingLayer()
971
972
973
974
975
976
977
978

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
979
980
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
981
982
983
984
985
986
987
988
989
990
991
992
    ) -> torch.Tensor:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

993
        for layer in islice(self.layers, self.start_layer, self.end_layer):
994
995
996
997
998
999
1000
            hidden_states, residual = layer(
                positions=positions,
                hidden_states=hidden_states,
                residual=residual,
            )

        if not get_pp_group().is_last_rank:
1001
1002
1003
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
1004
1005
1006
1007
1008
1009
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states

    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
1010
        return SharedFusedMoE.make_expert_params_mapping(
1011
1012
1013
1014
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
            num_experts=self.config.num_experts,
1015
1016
            num_redundant_experts=self.num_redundant_experts,
        )
1017

1018
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        expert_params_mapping = self.get_expert_mapping()
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue

            if name.startswith("mtp."):
                continue

            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue

                if "mlp.experts" in name:
                    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
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    continue
                # name = apply_attn_prefix(name, params_dict)
                if 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)
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
                    # Skip loading extra bias for GPTQ models.
1069
1070
1071
                    if (
                        name.endswith(".bias") or name.endswith("_bias")
                    ) and name not in params_dict:
1072
1073
1074
                        continue
                    param = params_dict[name]
                    weight_loader = param.weight_loader
1075
1076
1077
1078
1079
1080
1081
                    weight_loader(
                        param,
                        loaded_weight,
                        name,
                        shard_id=shard_id,
                        expert_id=expert_id,
                    )
1082
1083
1084
1085
1086
1087
1088
1089
                    break
                else:
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue
                    if is_pp_missing_parameter(name, self):
                        continue
                    param = params_dict[name]
1090
1091
1092
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
1093
1094
1095
1096
1097
                    weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params


1098
1099
1100
class Qwen3NextForCausalLM(
    nn.Module, HasInnerState, SupportsLoRA, SupportsPP, MixtureOfExperts, IsHybrid
):
1101
1102
1103
1104
1105
1106
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
1107
        "gate_up_proj": ["gate_proj", "up_proj"],
1108
1109
1110
1111
1112
1113
1114
1115
1116
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        config = vllm_config.model_config.hf_config
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        cache_config = vllm_config.cache_config
        lora_config = vllm_config.lora_config
        scheduler_config = vllm_config.scheduler_config
1117
        assert not cache_config.enable_prefix_caching, (
1118
            "Qwen3Next currently does not support prefix caching"
1119
        )
1120
1121
1122
1123
1124
        self.quant_config = vllm_config.quant_config

        super().__init__()
        self.config = config
        self.scheduler_config = scheduler_config
1125
1126
1127
        self.model = Qwen3NextModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
        self.unpadded_vocab_size = config.vocab_size
        if lora_config:
            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
        self.lm_head = ParallelLMHead(
            self.unpadded_vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            padding_size=DEFAULT_VOCAB_PADDING_SIZE
            # We need bigger padding if using lora for kernel
            # compatibility
1138
1139
1140
1141
1142
1143
1144
            if not lora_config
            else lora_config.lora_vocab_padding_size,
            prefix=maybe_prefix(prefix, "lm_head"),
        )
        self.logits_processor = LogitsProcessor(
            self.unpadded_vocab_size, config.vocab_size
        )
1145
        self.make_empty_intermediate_tensors = (
1146
1147
            self.model.make_empty_intermediate_tensors
        )
1148
1149
1150
1151

        # Set MoE hyperparameters
        self.expert_weights = []

1152
        self.moe_layers: list[SharedFusedMoE] = []
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
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
        example_layer = None
        for layer in self.model.layers:
            if isinstance(layer, PPMissingLayer):
                continue

            assert isinstance(layer, Qwen3NextDecoderLayer)
            if isinstance(layer.mlp, Qwen3NextSparseMoeBlock):
                example_layer = layer.mlp
                self.moe_layers.append(layer.mlp.experts)

        if example_layer is None:
            raise RuntimeError("No Qwen3Next layer found in the model.layers.")

        self.num_moe_layers = len(self.moe_layers)
        self.num_expert_groups = 1
        self.num_shared_experts = 0
        self.num_logical_experts = example_layer.n_logical_experts
        self.num_physical_experts = example_layer.n_physical_experts
        self.num_local_physical_experts = example_layer.n_local_physical_experts
        self.num_routed_experts = example_layer.n_routed_experts
        self.num_redundant_experts = example_layer.n_redundant_experts

    def set_eplb_state(
        self,
        expert_load_view: torch.Tensor,
        logical_to_physical_map: torch.Tensor,
        logical_replica_count: torch.Tensor,
    ) -> None:
        for layer_idx, layer in enumerate(self.moe_layers):
            # Register the expert weights.
            self.expert_weights.append(layer.get_expert_weights())
            layer.set_eplb_state(
                moe_layer_idx=layer_idx,
                expert_load_view=expert_load_view,
                logical_to_physical_map=logical_to_physical_map,
                logical_replica_count=logical_replica_count,
            )

    def update_physical_experts_metadata(
        self,
        num_physical_experts: int,
        num_local_physical_experts: int,
    ) -> None:
        assert self.num_local_physical_experts == num_local_physical_experts
        self.num_physical_experts = num_physical_experts
        self.num_local_physical_experts = num_local_physical_experts
1199
        self.num_redundant_experts = num_physical_experts - self.num_logical_experts
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
        for layer in self.model.layers:
            if isinstance(layer.mlp, Qwen3NextSparseMoeBlock):
                moe = layer.mlp
                moe.n_local_physical_experts = num_local_physical_experts
                moe.n_physical_experts = num_physical_experts
                moe.n_redundant_experts = self.num_redundant_experts
                moe.experts.update_expert_map()

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1215
1216
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
1217
1218
        **kwargs: object,
    ):
1219
1220
1221
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
1222
1223
1224
1225
1226
1227
1228
1229
1230

        return hidden_states

    @classmethod
    def get_mamba_state_dtype_from_config(
        cls,
        vllm_config: "VllmConfig",
    ) -> tuple[torch.dtype, torch.dtype]:
        return MambaStateDtypeCalculator.gated_delta_net_state_dtype(
1231
1232
            vllm_config.model_config.dtype, vllm_config.cache_config.mamba_cache_dtype
        )
1233
1234
1235

    @classmethod
    def get_mamba_state_shape_from_config(
1236
        cls, vllm_config: "VllmConfig"
1237
1238
1239
1240
    ) -> tuple[tuple[int, int], tuple[int, int]]:
        parallel_config = vllm_config.parallel_config
        hf_config = vllm_config.model_config.hf_config
        tp_size = parallel_config.tensor_parallel_size
1241
1242
1243
1244
1245
        num_spec = (
            vllm_config.speculative_config.num_speculative_tokens
            if vllm_config.speculative_config
            else 0
        )
1246
        return MambaStateShapeCalculator.gated_delta_net_state_shape(
1247
1248
1249
1250
1251
1252
1253
1254
            tp_size,
            hf_config.linear_num_key_heads,
            hf_config.linear_num_value_heads,
            hf_config.linear_key_head_dim,
            hf_config.linear_value_head_dim,
            hf_config.linear_conv_kernel_dim,
            num_spec,
        )
1255
1256
1257
1258

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
1259
    ) -> torch.Tensor | None:
1260
        return self.logits_processor(self.lm_head, hidden_states)
1261

1262
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=["mtp."],
        )
        return loader.load_weights(weights)

    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        return self.model.get_expert_mapping()


1273
1274
1275
def gdn_attention(
    hidden_states: torch.Tensor,
    output: torch.Tensor,
1276
1277
1278
1279
    layer_name: str,
) -> None:
    forward_context: ForwardContext = get_forward_context()
    self = forward_context.no_compile_layers[layer_name]
1280
1281
1282
1283
1284
1285
    self._forward(hidden_states=hidden_states, output=output)


def gdn_attention_fake(
    hidden_states: torch.Tensor,
    output: torch.Tensor,
1286
1287
1288
1289
1290
1291
    layer_name: str,
) -> None:
    return


direct_register_custom_op(
1292
1293
1294
1295
    op_name="gdn_attention",
    op_func=gdn_attention,
    mutates_args=["output"],
    fake_impl=gdn_attention_fake,
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
)


# g = -self.A_log.float().exp() * F.softplus(a.float() + self.dt_bias)
@triton.jit
def fused_gdn_gating_kernel(
    g,
    A_log,
    a,
    dt_bias,
    seq_len,
    NUM_HEADS: tl.constexpr,
    beta: tl.constexpr,
    threshold: tl.constexpr,
    BLK_HEADS: tl.constexpr,
):
    i_b, i_s, i_d = tl.program_id(0), tl.program_id(1), tl.program_id(2)
    head_off = i_d * BLK_HEADS + tl.arange(0, BLK_HEADS)
    off = i_b * seq_len * NUM_HEADS + i_s * NUM_HEADS + head_off
    mask = head_off < NUM_HEADS
    blk_A_log = tl.load(A_log + head_off, mask=mask)
    blk_a = tl.load(a + off, mask=mask)
    blk_bias = tl.load(dt_bias + head_off, mask=mask)
    # If the model is loaded in fp16, without the .float() here, A might be -inf
    x = blk_a.to(tl.float32) + blk_bias.to(tl.float32)
1321
1322
1323
    softplus_x = tl.where(
        beta * x <= threshold, (1 / beta) * tl.log(1 + tl.exp(beta * x)), x
    )
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
    blk_g = -tl.exp(blk_A_log.to(tl.float32)) * softplus_x
    tl.store(g + off, blk_g.to(g.dtype.element_ty), mask=mask)


def fused_gdn_gating(
    A_log: torch.Tensor,
    a: torch.Tensor,
    dt_bias: torch.Tensor,
    beta: float = 1.0,
    threshold: float = 20.0,
) -> torch.Tensor:
    batch, num_heads = a.shape
    seq_len = 1
    grid = (batch, seq_len, triton.cdiv(num_heads, 8))
    g = torch.empty_like(a, dtype=torch.float32)
1339
1340
1341
    fused_gdn_gating_kernel[grid](
        g, A_log, a, dt_bias, seq_len, num_heads, beta, threshold, 8, num_warps=1
    )
1342
    return g