qwen3_next.py 50.8 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

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

13
from vllm.attention.layer import Attention
14
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
34
35
    chunk_gated_delta_rule,
    fused_recurrent_gated_delta_rule,
)
36
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
37
from vllm.model_executor.layers.layernorm import (
38
39
40
    GemmaRMSNorm as Qwen3NextRMSNorm,
)
from vllm.model_executor.layers.layernorm import RMSNormGated
41
42
43
44
45
46
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    QKVParallelLinear,
    ReplicatedLinear,
    RowParallelLinear,
)
47
48
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.abstract import MambaBase
49
from vllm.model_executor.layers.mamba.mamba_mixer2 import mamba_v2_sharded_weight_loader
50
from vllm.model_executor.layers.mamba.mamba_utils import (
51
52
    MambaStateCopyFunc,
    MambaStateCopyFuncCalculator,
53
54
55
    MambaStateDtypeCalculator,
    MambaStateShapeCalculator,
)
56
from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
57
58
59
    causal_conv1d_fn,
    causal_conv1d_update,
)
60
61
62
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 (
63
64
65
    ParallelLMHead,
    VocabParallelEmbedding,
)
66
from vllm.model_executor.model_loader.weight_utils import (
67
    default_weight_loader,
68
    maybe_remap_kv_scale_name,
69
70
    sharded_weight_loader,
)
71
from vllm.model_executor.models.qwen2_moe import Qwen2MoeMLP as Qwen3NextMLP
72
from vllm.model_executor.models.utils import sequence_parallel_chunk
73
74
75
76
77
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
78
from vllm.utils.torch_utils import direct_register_custom_op
79
from vllm.v1.attention.backend import AttentionMetadata
80
81
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata

82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
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,
)
zhuwenwen's avatar
zhuwenwen committed
98
import vllm.envs as envs
99
100
101
102
103
104
105

logger = init_logger(__name__)

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


class Qwen3NextSparseMoeBlock(nn.Module):
106
    def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
107
        super().__init__()
108

Rayyyyy's avatar
Rayyyyy committed
109
        config = vllm_config.model_config.hf_text_config
110
111
112
        parallel_config = vllm_config.parallel_config
        quant_config = vllm_config.quant_config

113
114
115
        self.tp_size = get_tensor_model_parallel_world_size()

        self.ep_group = get_ep_group().device_group
116
        self.ep_rank = get_ep_group().rank_in_group
117
118
119
        self.ep_size = self.ep_group.size()
        self.n_routed_experts = config.num_experts

120
121
        self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe

122
123
124
        if self.tp_size > config.num_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
125
126
                f"the number of experts {config.num_experts}."
            )
127
128
129
130

        # Load balancing settings.
        vllm_config = get_current_vllm_config()
        eplb_config = vllm_config.parallel_config.eplb_config
131
        self.enable_eplb = parallel_config.enable_eplb
132
133
134

        self.n_logical_experts = self.n_routed_experts
        self.n_redundant_experts = eplb_config.num_redundant_experts
135
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
136
137
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

138
139
140
141
142
143
144
145
146
147
148
149
        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",
        )
150

151
152
153
154
155
156
157
        self.shared_expert_gate = ReplicatedLinear(
            config.hidden_size,
            1,
            bias=False,
            quant_config=None,
            prefix=f"{prefix}.shared_expert_gate",
        )
158
159
160
161
162
163
164

        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,
165
166
                reduce_results=False,
                expert_gate=self.shared_expert_gate,
167
                prefix=f"{prefix}.shared_expert",
168
169
170
            )
        else:
            self.shared_expert = None
171
172
173

        self.experts = SharedFusedMoE(
            shared_experts=self.shared_expert,
174
            gate=self.gate,
175
176
177
178
179
            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,
Rayyyyy's avatar
Rayyyyy committed
180
            renormalize=getattr(config, "norm_topk_prob", True),
181
182
183
184
185
186
            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,
        )
187
188
189
190

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # NOTE: hidden_states can have either 1D or 2D shape.
        orig_shape = hidden_states.shape
191
        num_tokens, hidden_dim = hidden_states.shape
192
193
        hidden_states = hidden_states.view(-1, hidden_dim)

194
195
196
        if self.is_sequence_parallel:
            hidden_states = sequence_parallel_chunk(hidden_states)

197
198
199
200
201
202
203
204
205
206
207
        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
            )
208

209
210
        if self.shared_expert is not None:
            final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
211
212
213

        if self.is_sequence_parallel:
            final_hidden_states = tensor_model_parallel_all_gather(
214
215
                final_hidden_states, 0
            )
216
217
            final_hidden_states = final_hidden_states[:num_tokens]
        elif self.tp_size > 1:
218
            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
219
220
                final_hidden_states
            )
221
222
223
224
225
226
227

        return final_hidden_states.view(orig_shape)


class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
    @property
    def mamba_type(self) -> str:
228
        return "gdn_attention"
229
230
231

    def get_state_dtype(self) -> tuple[torch.dtype, torch.dtype]:
        return MambaStateDtypeCalculator.gated_delta_net_state_dtype(
232
233
            self.model_config.dtype, self.cache_config.mamba_cache_dtype
        )
234
235
236

    def get_state_shape(self) -> tuple[tuple[int, ...], tuple[int, ...]]:
        return MambaStateShapeCalculator.gated_delta_net_state_shape(
237
238
239
240
241
242
243
244
            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,
        )
245
246
247
248

    def __init__(
        self,
        config: Qwen3NextConfig,
249
250
251
252
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        speculative_config: SpeculativeConfig | None = None,
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
        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
278
279
280
281
282
        self.num_spec = (
            self.speculative_config.num_speculative_tokens
            if self.speculative_config
            else 0
        )
283
284
285
286
287
288
289
290
291
292
293
294
295
296

        # 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
297
        self.in_proj_qkvz = ColumnParallelLinear(
298
            input_size=self.hidden_size,
299
            output_size=self.projection_size_qkvz,
300
301
            bias=False,
            quant_config=quant_config,
302
303
304
305
306
307
308
309
310
            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",
311
312
313
314
315
316
317
        )

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

        delattr(self.conv1d.weight, "weight_loader")
        set_weight_attrs(
318
319
320
321
322
323
324
325
326
327
328
329
330
            self.conv1d.weight,
            {
                "weight_loader": mamba_v2_sharded_weight_loader(
                    [
                        query_key_settings,
                        query_key_settings,
                        value_settings,
                    ],
                    self.tp_size,
                    self.tp_rank,
                )
            },
        )
331
332
333
334
335
336

        # 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(
337
338
            torch.ones(self.num_v_heads // self.tp_size),
        )
339
340
341
        self.A_log = nn.Parameter(
            torch.empty(
                divide(self.num_v_heads, self.tp_size),
342
343
            )
        )
344

345
346
        set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(0)})
        set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)})
347
348
349
350
351
352

        self.norm = RMSNormGated(
            self.head_v_dim,
            eps=self.layer_norm_epsilon,
            group_size=None,
            norm_before_gate=True,
353
            device=current_platform.current_device(),
354
            dtype=config.dtype,
355
356
        )

357
358
359
360
361
362
363
364
        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",
        )
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380

        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,
381
382
383
384
385
386
387
            (
                self.head_k_dim
                + self.head_k_dim
                + (self.head_v_dim + self.head_v_dim)
                * self.num_v_heads
                // self.num_k_heads
            ),
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
        )
        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,
405
            self.num_v_heads // self.num_k_heads,
406
407
408
409
410
        ]

        # [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]
411
        (query, key, value, z) = torch.split(mixed_qkvz, split_arg_list_qkvz, dim=2)
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
        (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(
435
436
437
438
            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)
439
        return query.contiguous(), key.contiguous(), value.contiguous()
440
441
442
443
444
445

    def forward(
        self,
        hidden_states: torch.Tensor,
        output: torch.Tensor,
    ):
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
        """
        Forward pass with three parts:
        1. Input projection
        2. Core attention (custom op)
        3. Output projection
        """
        num_tokens = hidden_states.size(0)

        # ============================================================
        # Part 1: Input Projection
        # ============================================================
        projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states)
        projected_states_ba, _ = self.in_proj_ba(hidden_states)
        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)

        # ============================================================
        # Part 2: Core Attention (Custom Op)
        # ============================================================
470
471
        # Note: we should not use torch.empty here like other attention backends,
        # see discussions in https://github.com/vllm-project/vllm/pull/28182
472
473
474
475
476
477
478
479
480
481
482
        core_attn_out = torch.zeros(
            (num_tokens, self.num_v_heads // self.tp_size, self.head_v_dim),
            dtype=hidden_states.dtype,
            device=hidden_states.device,
        )

        torch.ops.vllm.gdn_attention_core(
            mixed_qkv,
            b,
            a,
            core_attn_out,
483
484
485
            self.prefix,
        )

486
487
488
489
490
491
492
493
494
495
496
497
498
        # ============================================================
        # Part 3: Output Projection
        # ============================================================
        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_tokens], _ = self.out_proj(core_attn_out)

    def _forward_core(
499
        self,
500
501
502
503
        mixed_qkv: torch.Tensor,
        b: torch.Tensor,
        a: torch.Tensor,
        core_attn_out: torch.Tensor,
504
    ):
505
506
507
        """
        Core attention computation (called by custom op).
        """
508
509
510
511
512
513
514
515
516
517
518
519
520
521
        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
522
523
        spec_token_indx = attn_metadata.spec_token_indx
        non_spec_token_indx = attn_metadata.non_spec_token_indx
524
525
526
527
528
        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]
529
        num_actual_tokens = attn_metadata.num_actual_tokens
530
531
        num_accepted_tokens = attn_metadata.num_accepted_tokens

532
533
534
        mixed_qkv = mixed_qkv[:num_actual_tokens]
        b = b[:num_actual_tokens]
        a = a[:num_actual_tokens]
535

536
        # 1. Convolution sequence transformation
zhuwenwen's avatar
zhuwenwen committed
537
538
539
540
541
542
543
        if envs.VLLM_USE_NN:
            conv_weights = self.conv1d.weight.squeeze(1).transpose(
                0, 1).contiguous()
        else:
            conv_weights = self.conv1d.weight.view(
                self.conv1d.weight.size(0), self.conv1d.weight.size(2)
            )
544
545

        if spec_sequence_masks is not None:
546
            if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0:
547
548
549
                mixed_qkv_spec = mixed_qkv
                mixed_qkv_non_spec = None
            else:
550
551
                mixed_qkv_spec = mixed_qkv.index_select(0, spec_token_indx)
                mixed_qkv_non_spec = mixed_qkv.index_select(0, non_spec_token_indx)
552
553
554
555
        else:
            mixed_qkv_spec = None
            mixed_qkv_non_spec = mixed_qkv

556
        # 1.1: Process the multi-query part
557
558
559
560
561
562
563
        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,
564
565
566
                conv_state_indices=spec_state_indices_tensor[:, 0][
                    : attn_metadata.num_spec_decodes
                ],
567
                num_accepted_tokens=num_accepted_tokens,
568
569
                query_start_loc=spec_query_start_loc,
                max_query_len=spec_state_indices_tensor.size(-1),
570
571
572
                validate_data=False,
            )

573
        # 1.2: Process the remaining part
574
        if attn_metadata.num_prefills > 0:
575
            mixed_qkv_non_spec_T = mixed_qkv_non_spec.transpose(0, 1)
576
            # - "cache_indices" updates the conv_state cache in positions
577
            #   pointed to by "state_indices_tensor"
578
            mixed_qkv_non_spec = causal_conv1d_fn(
579
                mixed_qkv_non_spec_T,
580
581
582
583
584
585
586
                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,
587
                metadata=attn_metadata,
588
589
590
591
592
593
594
595
            ).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,
596
                conv_state_indices=non_spec_state_indices_tensor[
597
                    : attn_metadata.num_actual_tokens
598
                ],
599
600
601
602
603
                validate_data=True,
            )
        else:
            mixed_qkv_non_spec = None

604
        query_spec, key_spec, value_spec = self.rearrange_mixed_qkv(mixed_qkv_spec)
605
        query_non_spec, key_non_spec, value_non_spec = self.rearrange_mixed_qkv(
606
607
            mixed_qkv_non_spec
        )
608

609
        g, beta = fused_gdn_gating(self.A_log, a, b, self.dt_bias)
610
611

        if spec_sequence_masks is not None:
612
            if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0:
613
614
615
616
617
                g_spec = g
                beta_spec = beta
                g_non_spec = None
                beta_non_spec = None
            else:
618
619
620
621
                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)
622
623
624
625
626
627
        else:
            g_spec = None
            beta_spec = None
            g_non_spec = g
            beta_non_spec = beta

628
        # 2. Recurrent attention
629

630
        # 2.1: Process the multi-query part
631
        if spec_sequence_masks is not None:
632
633
634
635
636
637
638
639
640
641
642
643
644
            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,
            )
645
646
647
        else:
            core_attn_out_spec, last_recurrent_state = None, None

648
        # 2.2: Process the remaining part
649
        if attn_metadata.num_prefills > 0:
650
            initial_state = ssm_state[non_spec_state_indices_tensor].contiguous()
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
            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(
669
670
                ssm_state.dtype
            )
671
672
673
674
675
676
677
678
679
680
        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,
681
682
683
                    cu_seqlens=non_spec_query_start_loc[
                        : attn_metadata.num_decodes + 1
                    ],
684
685
                    ssm_state_indices=non_spec_state_indices_tensor,
                    use_qk_l2norm_in_kernel=True,
686
687
                )
            )
688
689
690
        else:
            core_attn_out_non_spec, last_recurrent_state = None, None

691
        # 3. Merge core attention output
692
        if spec_sequence_masks is not None and core_attn_out_non_spec is not None:
693
            merged_out = torch.empty(
694
695
696
697
                (1, num_actual_tokens, *core_attn_out_spec.shape[2:]),
                dtype=core_attn_out_non_spec.dtype,
                device=core_attn_out_non_spec.device,
            )
698
699
700
            merged_out.index_copy_(1, spec_token_indx, core_attn_out_spec)
            merged_out.index_copy_(1, non_spec_token_indx, core_attn_out_non_spec)
            core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)
701
        elif spec_sequence_masks is not None:
702
            core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)
703
        else:
704
            core_attn_out[:num_actual_tokens] = core_attn_out_non_spec.squeeze(0)
705
706
707
708
709
710


class Qwen3NextAttention(nn.Module):
    def __init__(
        self,
        config: Qwen3NextConfig,
711
712
713
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
        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(
738
739
            config, "dual_chunk_attention_config", None
        )
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
        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,
            max_position=config.max_position_embeddings,
763
            rope_parameters=config.rope_parameters,
764
765
766
767
768
769
770
771
772
773
774
775
776
            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),
777
778
779
780
                "dual_chunk_attention_config": self.dual_chunk_attention_config,
            }
            if self.dual_chunk_attention_config
            else {},
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
        )

        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(
796
797
                [self.q_size * 2, self.kv_size, self.kv_size], dim=-1
            )
798
799
800
801
802
803
            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:
804
            q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
805
806

        q = self.q_norm(q.view(-1, self.num_heads, self.head_dim)).view(
807
808
            -1, self.num_heads * self.head_dim
        )
809
        k = self.k_norm(k.view(-1, self.num_kv_heads, self.head_dim)).view(
810
811
            -1, self.num_kv_heads * self.head_dim
        )
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826

        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,
827
        vllm_config: VllmConfig,
828
829
830
831
        layer_type: str,
        prefix: str = "",
    ) -> None:
        super().__init__()
832
833
834
835
836
837

        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
838
839
840
841
842
843
844
845
846
847
848

        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,
849
850
                prefix=f"{prefix}.linear_attn",
            )
851
852
853
854
855
856
        elif self.layer_type == "full_attention":
            self.self_attn = Qwen3NextAttention(
                config,
                model_config=model_config,
                cache_config=cache_config,
                quant_config=quant_config,
857
                prefix=f"{prefix}.self_attn",
858
859
860
861
            )
        else:
            raise ValueError(f"Invalid layer_type {self.layer_type}")

862
863
864
        mlp_only_layers = (
            [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
        )
865
        if (self.layer_idx not in mlp_only_layers) and (
866
867
868
            config.num_experts > 0
            and (self.layer_idx + 1) % config.decoder_sparse_step == 0
        ):
869
            self.mlp = Qwen3NextSparseMoeBlock(
870
                vllm_config=vllm_config,
871
872
873
874
875
876
877
878
                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,
879
                prefix=f"{prefix}.mlp",
880
881
            )

882
883
884
        self.input_layernorm = Qwen3NextRMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )
885
        self.post_attention_layernorm = Qwen3NextRMSNorm(
886
887
            config.hidden_size, eps=config.rms_norm_eps
        )
888
889
890
891
892
893
894

        self.layer_scale = getattr(config, "layer_scale", False)
        if self.layer_scale:
            self.attn_layer_scale = torch.nn.Parameter(
                torch.zeros(
                    1,
                    1,
895
                    config.hidden_size,
896
                    dtype=config.dtype,
897
898
                ),
            )
899
900
901
902
            self.ffn_layer_scale = torch.nn.Parameter(
                torch.zeros(
                    1,
                    1,
903
                    config.hidden_size,
904
                    dtype=config.dtype,
905
906
                ),
            )
907
908
909
910

    def forward(
        self,
        hidden_states: torch.Tensor,
911
        residual: torch.Tensor | None,
912
913
914
915
916
917
918
        positions: torch.Tensor = None,
        **kwargs: object,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
919
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939

        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 * (
940
941
                    self.attn_layer_scale.to(hidden_states.dtype)[0] + 1
                )
942
943
            else:
                hidden_states = hidden_states * (
944
945
                    self.attn_layer_scale.to(hidden_states.dtype) + 1
                )
946
947

        # Fully Connected
948
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
949
950
951
952
953
        hidden_states = self.mlp(hidden_states)

        if self.layer_scale:
            if len(hidden_states.shape) == 2:
                hidden_states = hidden_states * (
954
955
                    self.ffn_layer_scale.to(hidden_states.dtype)[0] + 1
                )
956
            else:
957
                assert len(hidden_states.shape) == len(self.ffn_layer_scale.shape), (
958
959
960
                    f"shape must be the same {len(hidden_states.shape)}, "
                    f"{len(self.ffn_layer_scale.shape)}"
                )
961
                hidden_states = hidden_states * (
962
963
                    self.ffn_layer_scale.to(hidden_states.dtype) + 1
                )
964
965
966
967
968
969
970
971
972

        return hidden_states, residual


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

Rayyyyy's avatar
Rayyyyy committed
973
        config: Qwen3NextConfig = vllm_config.model_config.hf_text_config
974
        parallel_config = vllm_config.parallel_config
975

976
977
978
979
        eplb_config = parallel_config.eplb_config
        self.num_redundant_experts = eplb_config.num_redundant_experts

        self.config = config
980
981

        self.vocab_size = config.vocab_size
982
983
984
985
986
987
988
989

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

        def get_layer(prefix: str):
            return Qwen3NextDecoderLayer(
990
                vllm_config,
991
992
993
994
995
                layer_type=config.layer_types[extract_layer_index(prefix)],
                prefix=prefix,
            )

        self.start_layer, self.end_layer, self.layers = make_layers(
996
997
998
999
1000
            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
        )
1001

1002
        if get_pp_group().is_last_rank:
1003
            self.norm = Qwen3NextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
1004
1005
        else:
            self.norm = PPMissingLayer()
1006

1007
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
1008
1009
1010
1011
        return self.embed_tokens(input_ids)

    def forward(
        self,
zhuwenwen's avatar
zhuwenwen committed
1012
        input_ids: torch.Tensor,
1013
        positions: torch.Tensor,
1014
1015
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
1016
1017
1018
1019
1020
    ) -> torch.Tensor:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
1021
                hidden_states = self.embed_input_ids(input_ids)
1022
1023
1024
1025
1026
1027
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

1028
        for layer in islice(self.layers, self.start_layer, self.end_layer):
1029
1030
1031
1032
1033
1034
1035
            hidden_states, residual = layer(
                positions=positions,
                hidden_states=hidden_states,
                residual=residual,
            )

        if not get_pp_group().is_last_rank:
1036
1037
1038
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
1039
1040
1041
1042
1043
1044
        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)
1045
        return SharedFusedMoE.make_expert_params_mapping(
1046
            self,
1047
1048
1049
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
Rayyyyy's avatar
Rayyyyy committed
1050
            num_experts=getattr(self.config, "num_experts", 0),
1051
1052
            num_redundant_experts=self.num_redundant_experts,
        )
1053

1054
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
        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

1074
1075
1076
1077
1078
1079
            # Remapping the name of FP8 kv-scale.
            if name.endswith("scale"):
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue

1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
            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.
1111
1112
1113
                    if (
                        name.endswith(".bias") or name.endswith("_bias")
                    ) and name not in params_dict:
1114
                        continue
1115
1116
                    if name not in params_dict:
                        continue
1117
1118
                    param = params_dict[name]
                    weight_loader = param.weight_loader
1119
1120
1121
1122
1123
1124
1125
                    weight_loader(
                        param,
                        loaded_weight,
                        name,
                        shard_id=shard_id,
                        expert_id=expert_id,
                    )
1126
1127
1128
1129
1130
1131
1132
                    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
1133
1134
1135
1136
1137
                    if name not in params_dict:
                        logger.warning_once(
                            f"Parameter {name} not found in params_dict, skip loading"
                        )
                        continue
1138
                    param = params_dict[name]
1139
1140
1141
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
1142
1143
1144
1145
1146
                    weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params


1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
class QwenNextMixtureOfExperts(MixtureOfExperts):
    def update_physical_experts_metadata(
        self,
        num_physical_experts: int,
        num_local_physical_experts: int,
    ) -> None:
        assert self.num_local_physical_experts == num_local_physical_experts
        self.num_physical_experts = num_physical_experts
        self.num_local_physical_experts = num_local_physical_experts
        self.num_redundant_experts = num_physical_experts - self.num_logical_experts
        for 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 set_moe_parameters(self):
        self.expert_weights = []

        self.moe_layers = []
        example_moe = None
        for layer in self.model.layers:
            if isinstance(layer, Qwen3NextDecoderLayer) and isinstance(
                layer.mlp, Qwen3NextSparseMoeBlock
            ):
                example_moe = layer.mlp
                self.moe_layers.append(layer.mlp.experts)

1177
1178
        if example_moe is None:
            raise RuntimeError("No Qwen3Next layer found in the model.layers.")
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190

        # Set MoE hyperparameters
        self.num_moe_layers = len(self.moe_layers)
        self.num_expert_groups = 1
        self.num_shared_experts = 0
        self.num_logical_experts = example_moe.n_logical_experts
        self.num_physical_experts = example_moe.n_physical_experts
        self.num_local_physical_experts = example_moe.n_local_physical_experts
        self.num_routed_experts = example_moe.n_routed_experts
        self.num_redundant_experts = example_moe.n_redundant_experts


1191
class Qwen3NextForCausalLM(
1192
1193
1194
1195
1196
1197
    nn.Module,
    HasInnerState,
    SupportsLoRA,
    SupportsPP,
    QwenNextMixtureOfExperts,
    IsHybrid,
1198
):
1199
1200
1201
1202
1203
1204
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
1205
        "gate_up_proj": ["gate_proj", "up_proj"],
1206
1207
1208
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
Rayyyyy's avatar
Rayyyyy committed
1209
        config = vllm_config.model_config.hf_text_config
1210
1211
1212
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        cache_config = vllm_config.cache_config
1213

1214
        scheduler_config = vllm_config.scheduler_config
1215
1216
1217
1218
1219
        if cache_config.mamba_cache_mode == "all":
            raise NotImplementedError(
                "Qwen3Next currently does not support 'all' prefix caching, "
                "please use '--mamba-cache-mode=align' instead"
            )
1220
1221
1222
1223
1224
        self.quant_config = vllm_config.quant_config

        super().__init__()
        self.config = config
        self.scheduler_config = scheduler_config
1225
1226
1227
        self.model = Qwen3NextModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
1228

1229
        self.lm_head = ParallelLMHead(
1230
            config.vocab_size,
1231
            config.hidden_size,
1232
1233
            prefix=maybe_prefix(prefix, "lm_head"),
        )
1234
        self.logits_processor = LogitsProcessor(config.vocab_size)
1235
        self.make_empty_intermediate_tensors = (
1236
1237
            self.model.make_empty_intermediate_tensors
        )
1238
1239

        # Set MoE hyperparameters
1240
        self.set_moe_parameters()
1241

1242
1243
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
1244
1245
1246

    def forward(
        self,
zhuwenwen's avatar
zhuwenwen committed
1247
        input_ids: torch.Tensor,
1248
        positions: torch.Tensor,
1249
1250
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
1251
1252
        **kwargs: object,
    ):
1253
1254
1255
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
1256
1257
1258
1259
1260
1261
1262
1263
1264

        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(
1265
1266
            vllm_config.model_config.dtype, vllm_config.cache_config.mamba_cache_dtype
        )
1267
1268
1269

    @classmethod
    def get_mamba_state_shape_from_config(
1270
        cls, vllm_config: "VllmConfig"
1271
1272
    ) -> tuple[tuple[int, int], tuple[int, int]]:
        parallel_config = vllm_config.parallel_config
Rayyyyy's avatar
Rayyyyy committed
1273
        hf_config = vllm_config.model_config.hf_text_config
1274
        tp_size = parallel_config.tensor_parallel_size
1275
1276
1277
1278
1279
        num_spec = (
            vllm_config.speculative_config.num_speculative_tokens
            if vllm_config.speculative_config
            else 0
        )
1280
        return MambaStateShapeCalculator.gated_delta_net_state_shape(
1281
1282
1283
1284
1285
1286
1287
1288
            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,
        )
1289

1290
1291
1292
1293
    @classmethod
    def get_mamba_state_copy_func(cls) -> tuple[MambaStateCopyFunc, MambaStateCopyFunc]:
        return MambaStateCopyFuncCalculator.gated_delta_net_state_copy_func()

1294
1295
1296
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
1297
    ) -> torch.Tensor | None:
1298
        return self.logits_processor(self.lm_head, hidden_states)
1299

1300
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
        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()


1311
1312
1313
1314
1315
def gdn_attention_core(
    mixed_qkv: torch.Tensor,
    b: torch.Tensor,
    a: torch.Tensor,
    core_attn_out: torch.Tensor,
1316
1317
    layer_name: str,
) -> None:
1318
1319
1320
1321
1322
    """
    Custom op for the core attention computation.
    Only handles the convolution + recurrent attention part.
    Input/output projections are handled outside this op.
    """
1323
1324
    forward_context: ForwardContext = get_forward_context()
    self = forward_context.no_compile_layers[layer_name]
1325
1326
1327
1328
1329
1330
    self._forward_core(
        mixed_qkv=mixed_qkv,
        b=b,
        a=a,
        core_attn_out=core_attn_out,
    )
1331
1332


1333
1334
1335
1336
1337
def gdn_attention_core_fake(
    mixed_qkv: torch.Tensor,
    b: torch.Tensor,
    a: torch.Tensor,
    core_attn_out: torch.Tensor,
1338
1339
    layer_name: str,
) -> None:
1340
    """Fake implementation for torch.compile."""
1341
1342
1343
1344
    return


direct_register_custom_op(
1345
1346
1347
1348
    op_name="gdn_attention_core",
    op_func=gdn_attention_core,
    mutates_args=["core_attn_out"],
    fake_impl=gdn_attention_core_fake,
1349
1350
1351
1352
1353
1354
)


@triton.jit
def fused_gdn_gating_kernel(
    g,
1355
    beta_output,
1356
1357
    A_log,
    a,
1358
    b,
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
    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)
1372
    blk_b = tl.load(b + off, mask=mask)
1373
1374
1375
    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)
1376
1377
1378
    softplus_x = tl.where(
        beta * x <= threshold, (1 / beta) * tl.log(1 + tl.exp(beta * x)), x
    )
1379
1380
    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)
1381
    # compute beta_output = sigmoid(b)
1382
1383
1384
1385
    blk_beta_output = tl.sigmoid(blk_b.to(tl.float32))
    tl.store(
        beta_output + off, blk_beta_output.to(beta_output.dtype.element_ty), mask=mask
    )
1386
1387
1388
1389
1390


def fused_gdn_gating(
    A_log: torch.Tensor,
    a: torch.Tensor,
1391
    b: torch.Tensor,
1392
1393
1394
    dt_bias: torch.Tensor,
    beta: float = 1.0,
    threshold: float = 20.0,
1395
1396
1397
1398
1399
1400
1401
) -> tuple[torch.Tensor, torch.Tensor]:
    """
    Fused computation of g and beta for Gated Delta Net.
    g = -self.A_log.float().exp() * F.softplus(a.float() + self.dt_bias)
    beta_output = b.sigmoid()
    TODO maybe use torch.compile to replace this triton kernel
    """
1402
1403
1404
    batch, num_heads = a.shape
    seq_len = 1
    grid = (batch, seq_len, triton.cdiv(num_heads, 8))
1405
    g = torch.empty(1, batch, num_heads, dtype=torch.float32, device=a.device)
1406
    beta_output = torch.empty(1, batch, num_heads, dtype=b.dtype, device=b.device)
1407
    fused_gdn_gating_kernel[grid](
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
        g,
        beta_output,
        A_log,
        a,
        b,
        dt_bias,
        seq_len,
        num_heads,
        beta,
        threshold,
        8,
        num_warps=1,
1420
    )
1421
    return g, beta_output