mixtral.py 24.9 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

Pierre Stock's avatar
Pierre Stock committed
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# 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 Mixtral model."""
26
27
import typing
from collections.abc import Callable, Iterable
28
from itertools import islice
29
from typing import Optional, Union
Pierre Stock's avatar
Pierre Stock committed
30
31
32

import torch
from torch import nn
33
from transformers import MixtralConfig
Pierre Stock's avatar
Pierre Stock committed
34

35
from vllm.attention import Attention
36
from vllm.compilation.decorators import support_torch_compile
37
38
39
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (get_ep_group, get_pp_group,
                              get_tensor_model_parallel_world_size)
40
from vllm.model_executor.layers.fused_moe import FusedMoE
Pierre Stock's avatar
Pierre Stock committed
41
from vllm.model_executor.layers.layernorm import RMSNorm
42
from vllm.model_executor.layers.linear import (QKVParallelLinear,
Philipp Moritz's avatar
Philipp Moritz committed
43
                                               ReplicatedLinear,
Pierre Stock's avatar
Pierre Stock committed
44
                                               RowParallelLinear)
45
from vllm.model_executor.layers.logits_processor import LogitsProcessor
46
from vllm.model_executor.layers.quantization import QuantizationConfig
47
from vllm.model_executor.layers.rotary_embedding import get_rope
Pierre Stock's avatar
Pierre Stock committed
48
from vllm.model_executor.layers.vocab_parallel_embedding import (
49
    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
50
51
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader, maybe_remap_kv_scale_name)
Pierre Stock's avatar
Pierre Stock committed
52
from vllm.model_executor.sampling_metadata import SamplingMetadata
53
from vllm.sequence import IntermediateTensors
Pierre Stock's avatar
Pierre Stock committed
54

55
56
from .interfaces import MixtureOfExperts, SupportsLoRA, SupportsPP
from .utils import (AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter,
57
58
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
59

Pierre Stock's avatar
Pierre Stock committed
60

Philipp Moritz's avatar
Philipp Moritz committed
61
62
63
64
65
66
67
68
class MixtralMoE(nn.Module):
    """A tensor-parallel MoE implementation for Mixtral that shards each expert
    across all ranks.

    Each expert's weights are sharded across all ranks and a fused MoE
    kernel is used for the forward pass, and finally we reduce the outputs
    across ranks.
    """
69

70
71
72
73
74
75
76
    def __init__(self,
                 num_experts: int,
                 top_k: int,
                 hidden_size: int,
                 intermediate_size: int,
                 params_dtype: Optional[torch.dtype] = None,
                 quant_config: Optional[QuantizationConfig] = None,
77
                 tp_size: Optional[int] = None,
78
                 dp_size: Optional[int] = None,
79
80
                 prefix: str = "",
                 enable_eplb: bool = False):
81
        super().__init__()
Philipp Moritz's avatar
Philipp Moritz committed
82
        self.hidden_size = hidden_size
83

84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
        self.ep_group = get_ep_group().device_group
        self.ep_rank = self.ep_group.rank()
        self.ep_size = self.ep_group.size()

        # Expert Parallelism Load balancing settings.
        vllm_config = get_current_vllm_config()
        parallel_config = vllm_config.parallel_config
        self.enable_eplb = enable_eplb

        self.n_routed_experts = num_experts
        self.n_logical_experts = num_experts
        self.n_redundant_experts = (
            parallel_config.eplb_config.num_redundant_experts)
        self.n_physical_experts = (self.n_logical_experts +
                                   self.n_redundant_experts)
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size
        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)

105
        # Gate always runs at half / full precision for now.
106

107
108
        self.gate = ReplicatedLinear(hidden_size,
                                     num_experts,
109
                                     bias=False,
110
                                     params_dtype=params_dtype,
111
112
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")
113

114
115
116
117
118
119
120
121
        self.experts = FusedMoE(num_experts=num_experts,
                                top_k=top_k,
                                hidden_size=hidden_size,
                                intermediate_size=intermediate_size,
                                params_dtype=params_dtype,
                                reduce_results=True,
                                renormalize=True,
                                quant_config=quant_config,
122
                                tp_size=tp_size,
123
                                dp_size=dp_size,
124
125
126
                                prefix=f"{prefix}.experts",
                                enable_eplb=self.enable_eplb,
                                num_redundant_experts=self.n_redundant_experts)
127

128
    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
129
130
        # NOTE: hidden_states can have either 1D or 2D shape.
        orig_shape = hidden_states.shape
Philipp Moritz's avatar
Philipp Moritz committed
131
        hidden_states = hidden_states.view(-1, self.hidden_size)
132
        # router_logits: (num_tokens, n_experts)
133
        router_logits, _ = self.gate(hidden_states)
134
        final_hidden_states = self.experts(hidden_states, router_logits)
135
        return final_hidden_states.view(orig_shape)
Pierre Stock's avatar
Pierre Stock committed
136
137
138
139


class MixtralAttention(nn.Module):

140
141
    def __init__(
        self,
142
        config: MixtralConfig,
143
144
145
146
147
148
149
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position: int = 4096 * 32,
        rope_theta: float = 10000,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
150
        prefix: str = "",
151
    ) -> None:
Pierre Stock's avatar
Pierre Stock committed
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
        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)
168
        # MixtralConfig has an optional head_dim argument
169
170
171
        self.head_dim = getattr(config, "head_dim", None)
        if self.head_dim is None:
            self.head_dim = self.hidden_size // self.total_num_heads
Pierre Stock's avatar
Pierre Stock committed
172
173
174
175
176
        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

177
        self.qkv_proj = QKVParallelLinear(
Pierre Stock's avatar
Pierre Stock committed
178
179
180
181
182
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
183
            quant_config=quant_config,
184
            prefix=f"{prefix}.qkv_proj",
Pierre Stock's avatar
Pierre Stock committed
185
        )
186
        self.o_proj = RowParallelLinear(
Pierre Stock's avatar
Pierre Stock committed
187
188
189
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
190
            quant_config=quant_config,
191
            prefix=f"{prefix}.o_proj",
Pierre Stock's avatar
Pierre Stock committed
192
193
194
195
196
197
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position,
            base=int(self.rope_theta),
198
            is_neox_style=True,
Pierre Stock's avatar
Pierre Stock committed
199
        )
200
201
202
203
204
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
                              num_kv_heads=self.num_kv_heads,
                              cache_config=cache_config,
205
206
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")
Pierre Stock's avatar
Pierre Stock committed
207
208
209
210
211
212

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
213
        qkv, _ = self.qkv_proj(hidden_states)
Pierre Stock's avatar
Pierre Stock committed
214
215
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)
216
        attn_output = self.attn(q, k, v)
217
        output, _ = self.o_proj(attn_output)
Pierre Stock's avatar
Pierre Stock committed
218
219
220
221
222
223
224
        return output


class MixtralDecoderLayer(nn.Module):

    def __init__(
        self,
225
        config: MixtralConfig,
226
        cache_config: Optional[CacheConfig] = None,
227
        quant_config: Optional[QuantizationConfig] = None,
228
        prefix: str = "",
229
        enable_eplb: bool = False,
Pierre Stock's avatar
Pierre Stock committed
230
231
232
233
234
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        # Requires transformers > 4.32.0
        rope_theta = getattr(config, "rope_theta", 10000)
235
        self.self_attn = MixtralAttention(
236
            config=config,
Pierre Stock's avatar
Pierre Stock committed
237
238
239
240
241
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            max_position=config.max_position_embeddings,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
242
            cache_config=cache_config,
243
244
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn")
Philipp Moritz's avatar
Philipp Moritz committed
245
246
247
248
        self.block_sparse_moe = MixtralMoE(
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
249
            intermediate_size=config.intermediate_size,
250
            quant_config=quant_config,
251
252
            prefix=f"{prefix}.block_sparse_moe",
            enable_eplb=enable_eplb)
253
254
255
256
        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)
Pierre Stock's avatar
Pierre Stock committed
257
258
259
260

    def forward(
        self,
        positions: torch.Tensor,
261
262
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
Pierre Stock's avatar
Pierre Stock committed
263
    ) -> torch.Tensor:
264
265
266
267
268
269
270
271
        # 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(
Pierre Stock's avatar
Pierre Stock committed
272
            positions=positions,
273
            hidden_states=hidden_states,
Pierre Stock's avatar
Pierre Stock committed
274
275
        )

276
277
278
279
280
        # Fully Connected
        hidden_states, residual = self.post_attention_layernorm(
            hidden_states, residual)
        hidden_states = self.block_sparse_moe(hidden_states)
        return hidden_states, residual
Pierre Stock's avatar
Pierre Stock committed
281

282

283
@support_torch_compile
284
class MixtralModel(nn.Module):
Pierre Stock's avatar
Pierre Stock committed
285

286
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
Pierre Stock's avatar
Pierre Stock committed
287
        super().__init__()
288
289
290
291
292

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
293
        parallel_config = vllm_config.parallel_config
294

295
296
        self.config = config
        self.quant_config = quant_config
297
298
299
300
        lora_vocab = (lora_config.lora_extra_vocab_size *
                      (lora_config.max_loras or 1)) if lora_config else 0
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size
301
302

        self.embed_tokens = VocabParallelEmbedding(
303
            self.vocab_size,
Pierre Stock's avatar
Pierre Stock committed
304
            config.hidden_size,
305
            org_num_embeddings=config.vocab_size,
Pierre Stock's avatar
Pierre Stock committed
306
        )
307

308
309
310
311
        self.enable_eplb = parallel_config.enable_eplb
        self.num_redundant_experts = (
            parallel_config.eplb_config.num_redundant_experts)

312
        self.start_layer, self.end_layer, self.layers = make_layers(
313
314
            config.num_hidden_layers,
            lambda prefix: MixtralDecoderLayer(
315
316
317
318
319
                config,
                cache_config,
                quant_config=quant_config,
                prefix=prefix,
                enable_eplb=self.enable_eplb,
320
321
            ),
            prefix=f"{prefix}.layers")
322

323
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
324
325
326
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))
Pierre Stock's avatar
Pierre Stock committed
327

328
329
330
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

Pierre Stock's avatar
Pierre Stock committed
331
332
333
334
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
335
        intermediate_tensors: Optional[IntermediateTensors],
336
        inputs_embeds: Optional[torch.Tensor] = None,
337
    ) -> Union[torch.Tensor, IntermediateTensors]:
338
        if get_pp_group().is_first_rank:
339
340
341
342
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
343
344
345
346
347
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
348
        for layer in islice(self.layers, self.start_layer, self.end_layer):
349
            hidden_states, residual = layer(positions, hidden_states, residual)
350
351
352
353
354
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
355
356
357
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states

358
359
360
361
362
363
364
    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="w1",
            ckpt_down_proj_name="w2",
            ckpt_up_proj_name="w3",
365
366
            num_experts=self.config.num_local_experts,
            num_redundant_experts=self.num_redundant_experts)
367

368
369
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
Pierre Stock's avatar
Pierre Stock committed
370
371
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
372
373
374
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
Pierre Stock's avatar
Pierre Stock committed
375
        ]
376

Pierre Stock's avatar
Pierre Stock committed
377
        params_dict = dict(self.named_parameters())
378
        loaded_params: set[str] = set()
379
        expert_params_mapping = self.get_expert_mapping()
380
        for name, loaded_weight in weights:
381
382
            if (self.quant_config is not None and
                (scale_name := self.quant_config.get_cache_scale(name))):
383
                # Loading kv cache quantization scales
384
385
386
387
388
389
390
391
392
                param = params_dict[scale_name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                                 loaded_weight[0])
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue

Pierre Stock's avatar
Pierre Stock committed
393
394
395
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue
CHU Tianxiang's avatar
CHU Tianxiang committed
396
397
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
398
399
                if ((name.endswith(".bias") or name.endswith("_bias"))
                        and name not in params_dict):
CHU Tianxiang's avatar
CHU Tianxiang committed
400
                    continue
401
402
403
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    continue
404
405
406
407
408
                if name.endswith("scale"):
                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue
CHU Tianxiang's avatar
CHU Tianxiang committed
409
                param = params_dict[name]
Pierre Stock's avatar
Pierre Stock committed
410
411
412
413
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
414
                is_expert_weight = False
415
416
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
417

Philipp Moritz's avatar
Philipp Moritz committed
418
419
                    if weight_name not in name:
                        continue
420
421
422
423

                    is_expert_weight = True
                    name_mapped = name.replace(weight_name, param_name)

424
                    # Skip layers on other devices.
425
                    if is_pp_missing_parameter(name_mapped, self):
426
                        continue
427
428
429
430

                    if ((name_mapped.endswith(".bias")
                         or name_mapped.endswith("_bias"))
                            and name_mapped not in params_dict):
431
                        continue
432
433
434
435
436
437
438
439
440
441
442
443
444

                    param = params_dict[name_mapped]
                    weight_loader = typing.cast(Callable[..., bool],
                                                param.weight_loader)
                    success = weight_loader(param,
                                            loaded_weight,
                                            name_mapped,
                                            shard_id=shard_id,
                                            expert_id=expert_id,
                                            return_success=True)
                    if success:
                        name = name_mapped
                        break
Philipp Moritz's avatar
Philipp Moritz committed
445
                else:
446
447
                    if is_expert_weight:
                        continue
Philipp Moritz's avatar
Philipp Moritz committed
448
                    # Skip loading extra bias for GPTQ models.
449
450
                    if ((name.endswith(".bias") or name.endswith("_bias"))
                            and name not in params_dict):
Philipp Moritz's avatar
Philipp Moritz committed
451
                        continue
452
453
454
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
455
                    # Remapping the name of FP8 kv-scale.
456
457
458
459
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

Philipp Moritz's avatar
Philipp Moritz committed
460
461
462
463
                    param = params_dict[name]
                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)
464
465
            loaded_params.add(name)
        return loaded_params
466
467


468
469
class MixtralForCausalLM(nn.Module, SupportsLoRA, SupportsPP,
                         MixtureOfExperts):
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
    fall_back_to_pt_during_load = False

    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }

    # LoRA specific attributes
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }
    embedding_padding_modules = ["lm_head"]

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
        self.config = config
        self.lora_config = lora_config
        self.quant_config = quant_config

        self.model = MixtralModel(vllm_config=vllm_config,
                                  prefix=maybe_prefix(prefix, "model"))
        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
            if not lora_config else lora_config.lora_vocab_padding_size,
            quant_config=quant_config,
510
            prefix=maybe_prefix(prefix, "lm_head"),
511
512
513
514
515
516
517
518
        )
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
        self.expert_weights = []
        self.moe_layers: list[FusedMoE] = []
        example_moe = None

        for layer in self.model.layers:
            if isinstance(layer, PPMissingLayer):
                continue
            assert isinstance(layer, MixtralDecoderLayer)
            if hasattr(layer, "block_sparse_moe") and isinstance(
                    layer.block_sparse_moe, MixtralMoE):
                example_moe = layer.block_sparse_moe
                self.moe_layers.append(layer.block_sparse_moe.experts)

        self.num_moe_layers = len(self.moe_layers)

        if example_moe is None:
            raise RuntimeError("No MixtralMoE layer found  in model.layers.")

        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
        self.num_expert_groups = 1
        self.num_shared_experts = 0

    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
        self.num_redundant_experts = (num_physical_experts -
                                      self.num_logical_experts)
        for layer in self.model.layers:
            if hasattr(layer, "block_sparse_moe") and isinstance(
                    layer.block_sparse_moe, MixtralMoE):
                moe = layer.block_sparse_moe
                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()

580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
    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,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states,
                                       sampling_metadata)
        return logits

603
604
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
605
        loader = AutoWeightsLoader(self)
606
        return loader.load_weights(weights)
607
608
609

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