qwen2_moe.py 23.6 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py
# Copyright 2024 The Qwen team.
# Copyright 2023 The vLLM team.
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only Qwen2MoE model compatible with HuggingFace weights."""
27
from collections.abc import Iterable
28
from itertools import islice
29
from typing import Any, Optional, Union
30
31
32
33

import torch
import torch.nn.functional as F
from torch import nn
34
from transformers import Qwen2MoeConfig
35

36
from vllm.attention import Attention
37
from vllm.compilation.decorators import support_torch_compile
38
from vllm.config import CacheConfig, VllmConfig
39
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
40
from vllm.logger import init_logger
41
from vllm.model_executor.layers.activation import SiluAndMul
42
from vllm.model_executor.layers.fused_moe import FusedMoE
43
from vllm.model_executor.layers.layernorm import RMSNorm
44
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
45
46
47
48
                                               QKVParallelLinear,
                                               ReplicatedLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
49
from vllm.model_executor.layers.quantization import QuantizationConfig
50
51
52
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
    ParallelLMHead, VocabParallelEmbedding)
53
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
54
from vllm.sequence import IntermediateTensors
55

56
from .interfaces import SupportsLoRA, SupportsPP
57
58
from .utils import (AutoWeightsLoader, extract_layer_index,
                    is_pp_missing_parameter,
59
60
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
61

62
63
logger = init_logger(__name__)

64
65
66
67
68
69
70
71

class Qwen2MoeMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
72
        quant_config: Optional[QuantizationConfig] = None,
73
        reduce_results: bool = True,
74
        prefix: str = "",
75
76
77
78
79
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
80
81
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj")
82
83
84
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
85
                                           quant_config=quant_config,
86
87
                                           reduce_results=reduce_results,
                                           prefix=f"{prefix}.down_proj")
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
        if hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {hidden_act}. "
                             "Only silu is supported for now.")
        self.act_fn = SiluAndMul()

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


class Qwen2MoeSparseMoeBlock(nn.Module):

    def __init__(
        self,
104
        config: Qwen2MoeConfig,
105
        quant_config: Optional[QuantizationConfig] = None,
106
        prefix: str = "",
107
108
109
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
110
111

        if self.tp_size > config.num_experts:
112
113
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
114
115
116
117
118
119
120
121
                f"the number of experts {config.num_experts}.")

        self.experts = FusedMoE(num_experts=config.num_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,
122
123
                                quant_config=quant_config,
                                prefix=f"{prefix}.experts")
124
125

        self.gate = ReplicatedLinear(config.hidden_size,
126
                                     config.num_experts,
127
                                     bias=False,
128
129
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")
130
131
132
133
134
        if config.shared_expert_intermediate_size > 0:
            self.shared_expert = Qwen2MoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.shared_expert_intermediate_size,
                hidden_act=config.hidden_act,
135
                quant_config=quant_config,
136
137
                reduce_results=self.experts.must_reduce_shared_expert_outputs(
                ),
138
                prefix=f"{prefix}.shared_expert",
139
140
141
142
143
144
145
146
            )
        else:
            self.shared_expert = None
        self.shared_expert_gate = torch.nn.Linear(config.hidden_size,
                                                  1,
                                                  bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
147
148
149
        # NOTE: hidden_states can have either 1D or 2D shape.
        orig_shape = hidden_states.shape
        hidden_dim = hidden_states.shape[-1]
150
151
152
153
154
155
156
157
158
159
        hidden_states = hidden_states.view(-1, hidden_dim)
        shared_output = None
        if self.shared_expert is not None:
            shared_output = self.shared_expert(hidden_states)
            if self.shared_expert_gate is not None:
                shared_output = F.sigmoid(
                    self.shared_expert_gate(hidden_states)) * shared_output

        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
160
161
        final_hidden_states = self.experts(hidden_states=hidden_states,
                                           router_logits=router_logits)
162
163
        if shared_output is not None:
            final_hidden_states = final_hidden_states + shared_output
164
        if self.tp_size > 1:
165
            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
166
                final_hidden_states)
167

168
        return final_hidden_states.view(orig_shape)
169
170
171
172
173
174
175
176
177
178


class Qwen2MoeAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
179
        rope_scaling: Optional[dict[str, Any]] = None,
180
        max_position_embeddings: int = 8192,
181
        cache_config: Optional[CacheConfig] = None,
182
        quant_config: Optional[QuantizationConfig] = None,
183
        prefix: str = "",
184
        dual_chunk_attention_config: Optional[dict[str, Any]] = None,
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = hidden_size // self.total_num_heads
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings
208
        self.dual_chunk_attention_config = dual_chunk_attention_config
209

210
211
212
213
214
215
216
        self.qkv_proj = QKVParallelLinear(hidden_size,
                                          self.head_dim,
                                          self.total_num_heads,
                                          self.total_num_kv_heads,
                                          bias=True,
                                          quant_config=quant_config,
                                          prefix=f"{prefix}.qkv_proj")
217

218
219
220
221
222
        self.o_proj = RowParallelLinear(self.total_num_heads * self.head_dim,
                                        hidden_size,
                                        bias=False,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")
223
224
225
226
227
228
229

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
230
            dual_chunk_attention_config=dual_chunk_attention_config,
231
        )
232
233
234
235
236
237
238
239
240
241
242
243
        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),
                "dual_chunk_attention_config": dual_chunk_attention_config,
            } if dual_chunk_attention_config else {})
244
245
246
247
248
249
250
251
252

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)
253
        attn_output = self.attn(q, k, v)
254
255
256
257
258
259
260
261
        output, _ = self.o_proj(attn_output)
        return output


class Qwen2MoeDecoderLayer(nn.Module):

    def __init__(
        self,
262
        config: Qwen2MoeConfig,
263
        cache_config: Optional[CacheConfig] = None,
264
        quant_config: Optional[QuantizationConfig] = None,
265
        prefix: str = "",
266
267
268
269
270
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
271
272
273
        dual_chunk_attention_config = getattr(config,
                                              "dual_chunk_attention_config",
                                              None)
274
275
276
277
278
279
280
281
282
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        self.self_attn = Qwen2MoeAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
283
            cache_config=cache_config,
284
            quant_config=quant_config,
285
            prefix=f"{prefix}.self_attn",
286
            dual_chunk_attention_config=dual_chunk_attention_config,
287
        )
288
289
290

        # Note: Qwen/Qwen2-57B-A14B-Instruct does not have
        # `mlp_only_layers` in the config.
291
        layer_idx = extract_layer_index(prefix)
292
293
294
        mlp_only_layers = ([] if not hasattr(config, "mlp_only_layers") else
                           config.mlp_only_layers)
        if (layer_idx not in mlp_only_layers) and (
295
296
                config.num_experts > 0 and
            (layer_idx + 1) % config.decoder_sparse_step == 0):
297
            self.mlp = Qwen2MoeSparseMoeBlock(config=config,
298
299
                                              quant_config=quant_config,
                                              prefix=f"{prefix}.mlp")
300
        else:
301
302
303
304
305
            self.mlp = Qwen2MoeMLP(hidden_size=config.hidden_size,
                                   intermediate_size=config.intermediate_size,
                                   hidden_act=config.hidden_act,
                                   quant_config=quant_config,
                                   prefix=f"{prefix}.mlp")
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
        self.input_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(config.hidden_size,
                                                eps=config.rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> torch.Tensor:
        # Self Attention
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )

        # Fully Connected
        hidden_states, residual = self.post_attention_layernorm(
            hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual


336
@support_torch_compile
337
338
class Qwen2MoeModel(nn.Module):

339
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
340
        super().__init__()
341
342
343
344
345

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

346
        self.vocab_size = config.vocab_size
347
        self.config = config
348
349
350
351
352

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
353
354
355
356
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: Qwen2MoeDecoderLayer(config=config,
                                                cache_config=cache_config,
357
358
                                                quant_config=quant_config,
                                                prefix=prefix),
359
360
            prefix=f"{prefix}.layers",
        )
361
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
362
363
364
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))
365

366
367
368
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

369
370
371
372
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
373
        intermediate_tensors: Optional[IntermediateTensors] = None,
374
        inputs_embeds: Optional[torch.Tensor] = None,
375
    ) -> Union[torch.Tensor, IntermediateTensors]:
376
        if get_pp_group().is_first_rank:
377
378
379
380
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
381
382
383
384
385
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
386
        for layer in islice(self.layers, self.start_layer, self.end_layer):
387
            hidden_states, residual = layer(positions, hidden_states, residual)
388
389
390
391
392
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
393
394
395
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states

396
397
398
399
400
401
402
403
404
    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        return FusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
            num_experts=self.config.num_experts)

405
406
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
407
408
409
410
411
412
413
414
415
416
        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())
417
        loaded_params: set[str] = set()
418
        expert_params_mapping = self.get_expert_mapping()
419
        for name, loaded_weight in weights:
420
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
421
                # Skip non-stacked layers and experts (experts handled below).
422
423
                if weight_name not in name:
                    continue
424
425
426
427
428
429
430
431
                # We have mlp.experts[0].gate_proj in the checkpoint.
                # Since we handle the experts below in expert_params_mapping,
                # we need to skip here BEFORE we update the name, otherwise
                # name will be updated to mlp.experts[0].gate_up_proj, which
                # will then be updated below in expert_params_mapping
                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
                if "mlp.experts" in name:
                    continue
432
433
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
434
435
                if ((name.endswith(".bias") or name.endswith("_bias"))
                        and name not in params_dict):
436
                    continue
437
438
439
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    continue
440
441
442
                if name not in params_dict:
                    continue

443
444
445
446
447
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
448
                for mapping in expert_params_mapping:
449
450
451
452
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
                    name = name.replace(weight_name, param_name)
453

454
455
456
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
457
458
459
460
                    # Skip loading extra bias for GPTQ models.
                    if ((name.endswith(".bias") or name.endswith("_bias"))
                            and name not in params_dict):
                        continue
461
462
463
464
                    param = params_dict[name]
                    weight_loader = param.weight_loader
                    weight_loader(param,
                                  loaded_weight,
465
                                  name,
466
467
468
469
470
                                  shard_id=shard_id,
                                  expert_id=expert_id)
                    break
                else:
                    # Skip loading extra bias for GPTQ models.
471
472
                    if ((name.endswith(".bias") or name.endswith("_bias"))
                            and name not in params_dict):
473
                        continue
474
475
476
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
477
478
479
480
481
                    # Remapping the name of FP8 kv-scale.
                    if name.endswith("kv_scale"):
                        remapped_kv_scale_name = name.replace(
                            ".kv_scale", ".attn.kv_scale")
                        if remapped_kv_scale_name not in params_dict:
482
                            logger.warning_once(
483
484
485
486
                                "Found kv_scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv_scale is not loaded.",  #  noqa: E501
                                name,
                                remapped_kv_scale_name,
                            )
487
488
489
                            continue
                        else:
                            name = remapped_kv_scale_name
490
491
492
493
                    param = params_dict[name]
                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)
494
495
            loaded_params.add(name)
        return loaded_params
496
497


498
class Qwen2MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
499
500

    fall_back_to_pt_during_load = False
501
502
503
504
505
506
507
508
509
510
511
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }
512
513
514
515
516
517
518
519
520
521
522

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        self.config = config
        self.quant_config = quant_config
        self.model = Qwen2MoeModel(vllm_config=vllm_config,
                                   prefix=maybe_prefix(prefix, "model"))
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
523
524
                                      quant_config=quant_config,
                                      prefix=maybe_prefix(prefix, "lm_head"))
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

    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,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds)
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> Optional[torch.Tensor]:
549
        logits = self.logits_processor(self.lm_head, hidden_states)
550
551
        return logits

552
553
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
554
        loader = AutoWeightsLoader(self)
555
        return loader.load_weights(weights)
556
557
558

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