mixtral.py 23.2 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."""
王敏's avatar
王敏 committed
26
27
import os
import re
zhuwenwen's avatar
zhuwenwen committed
28

29
30
import typing
from collections.abc import Callable, Iterable
31
from itertools import islice
Pierre Stock's avatar
Pierre Stock committed
32
33
34

import torch
from torch import nn
35
from transformers import MixtralConfig
Pierre Stock's avatar
Pierre Stock committed
36

37
from vllm.attention.layer import Attention
38
from vllm.compilation.decorators import support_torch_compile
39
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
40
41
42
43
44
from vllm.distributed import (
    get_ep_group,
    get_pp_group,
    get_tensor_model_parallel_world_size,
)
45
from vllm.model_executor.layers.fused_moe import FusedMoE
Pierre Stock's avatar
Pierre Stock committed
46
from vllm.model_executor.layers.layernorm import RMSNorm
47
48
49
50
51
from vllm.model_executor.layers.linear import (
    QKVParallelLinear,
    ReplicatedLinear,
    RowParallelLinear,
)
52
from vllm.model_executor.layers.logits_processor import LogitsProcessor
53
from vllm.model_executor.layers.quantization import QuantizationConfig
54
from vllm.model_executor.layers.rotary_embedding import get_rope
Pierre Stock's avatar
Pierre Stock committed
55
from vllm.model_executor.layers.vocab_parallel_embedding import (
56
57
58
    ParallelLMHead,
    VocabParallelEmbedding,
)
59
from vllm.model_executor.model_loader.weight_utils import (
60
61
62
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
63
from vllm.sequence import IntermediateTensors
王敏's avatar
王敏 committed
64
65
from vllm import _custom_ops as ops
from vllm.model_executor.utils import pad_weight, gemm_bank_conf
Pierre Stock's avatar
Pierre Stock committed
66

67
from .interfaces import MixtureOfExperts, SupportsLoRA, SupportsPP
68
69
70
71
72
73
74
75
from .utils import (
    AutoWeightsLoader,
    PPMissingLayer,
    is_pp_missing_parameter,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
76

Pierre Stock's avatar
Pierre Stock committed
77

Philipp Moritz's avatar
Philipp Moritz committed
78
79
80
81
82
83
84
85
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.
    """
86

87
88
89
90
91
92
    def __init__(
        self,
        num_experts: int,
        top_k: int,
        hidden_size: int,
        intermediate_size: int,
93
94
95
96
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
        tp_size: int | None = None,
        dp_size: int | None = None,
97
98
99
        prefix: str = "",
        enable_eplb: bool = False,
    ):
100
        super().__init__()
Philipp Moritz's avatar
Philipp Moritz committed
101
        self.hidden_size = hidden_size
102

103
        self.ep_group = get_ep_group().device_group
104
        self.ep_rank = get_ep_group().rank_in_group
105
106
107
108
109
110
111
112
113
        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
114
115
        self.n_redundant_experts = parallel_config.eplb_config.num_redundant_experts
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
116
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size
117
118
119
120
        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
        )
121

122
        # Gate always runs at half / full precision for now.
123

124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
        self.gate = ReplicatedLinear(
            hidden_size,
            num_experts,
            bias=False,
            params_dtype=params_dtype,
            quant_config=None,
            prefix=f"{prefix}.gate",
        )

        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,
            tp_size=tp_size,
            dp_size=dp_size,
            prefix=f"{prefix}.experts",
            enable_eplb=self.enable_eplb,
            num_redundant_experts=self.n_redundant_experts,
        )
148

149
    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
150
151
        # NOTE: hidden_states can have either 1D or 2D shape.
        orig_shape = hidden_states.shape
Philipp Moritz's avatar
Philipp Moritz committed
152
        hidden_states = hidden_states.view(-1, self.hidden_size)
153
        # router_logits: (num_tokens, n_experts)
154
        router_logits, _ = self.gate(hidden_states)
155
        final_hidden_states = self.experts(hidden_states, router_logits)
156
        return final_hidden_states.view(orig_shape)
Pierre Stock's avatar
Pierre Stock committed
157
158
159


class MixtralAttention(nn.Module):
160
161
    def __init__(
        self,
162
        config: MixtralConfig,
163
164
165
166
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position: int = 4096 * 32,
167
168
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
169
        prefix: str = "",
170
    ) -> None:
Pierre Stock's avatar
Pierre Stock committed
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
        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)
187
        # MixtralConfig has an optional head_dim argument
188
189
190
        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
191
192
193
194
        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

195
        self.qkv_proj = QKVParallelLinear(
Pierre Stock's avatar
Pierre Stock committed
196
197
198
199
200
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
201
            quant_config=quant_config,
202
            prefix=f"{prefix}.qkv_proj",
Pierre Stock's avatar
Pierre Stock committed
203
        )
204
        self.o_proj = RowParallelLinear(
Pierre Stock's avatar
Pierre Stock committed
205
206
207
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
208
            quant_config=quant_config,
209
            prefix=f"{prefix}.o_proj",
Pierre Stock's avatar
Pierre Stock committed
210
211
212
213
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            max_position=max_position,
214
            rope_parameters=config.rope_parameters,
215
            is_neox_style=True,
Pierre Stock's avatar
Pierre Stock committed
216
        )
217
218
219
220
221
222
223
224
225
        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",
        )
Pierre Stock's avatar
Pierre Stock committed
226
227
228
229
230
231

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
232
        qkv, _ = self.qkv_proj(hidden_states)
Pierre Stock's avatar
Pierre Stock committed
233
234
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)
235
        attn_output = self.attn(q, k, v)
236
        output, _ = self.o_proj(attn_output)
Pierre Stock's avatar
Pierre Stock committed
237
238
239
240
241
242
        return output


class MixtralDecoderLayer(nn.Module):
    def __init__(
        self,
243
        config: MixtralConfig,
244
245
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
246
        prefix: str = "",
247
        enable_eplb: bool = False,
Pierre Stock's avatar
Pierre Stock committed
248
249
250
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
251
        self.self_attn = MixtralAttention(
252
            config=config,
Pierre Stock's avatar
Pierre Stock committed
253
254
255
256
            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,
257
            cache_config=cache_config,
258
            quant_config=quant_config,
259
260
            prefix=f"{prefix}.self_attn",
        )
Philipp Moritz's avatar
Philipp Moritz committed
261
262
263
264
        self.block_sparse_moe = MixtralMoE(
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
265
            intermediate_size=config.intermediate_size,
266
            quant_config=quant_config,
267
            prefix=f"{prefix}.block_sparse_moe",
268
269
270
271
272
273
            enable_eplb=enable_eplb,
        )
        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
274
275
276
277

    def forward(
        self,
        positions: torch.Tensor,
278
        hidden_states: torch.Tensor,
279
        residual: torch.Tensor | None,
Pierre Stock's avatar
Pierre Stock committed
280
    ) -> torch.Tensor:
281
282
283
284
285
        # Self Attention
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
286
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
287
        hidden_states = self.self_attn(
Pierre Stock's avatar
Pierre Stock committed
288
            positions=positions,
289
            hidden_states=hidden_states,
Pierre Stock's avatar
Pierre Stock committed
290
291
        )

292
        # Fully Connected
293
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
294
295
        hidden_states = self.block_sparse_moe(hidden_states)
        return hidden_states, residual
Pierre Stock's avatar
Pierre Stock committed
296

297

298
@support_torch_compile
299
class MixtralModel(nn.Module):
300
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
Pierre Stock's avatar
Pierre Stock committed
301
        super().__init__()
302
303
304
305

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
306

307
        parallel_config = vllm_config.parallel_config
308

309
310
        self.config = config
        self.quant_config = quant_config
311
312

        self.vocab_size = config.vocab_size
313
        self.org_vocab_size = config.vocab_size
314
315

        self.embed_tokens = VocabParallelEmbedding(
316
            self.vocab_size,
Pierre Stock's avatar
Pierre Stock committed
317
318
            config.hidden_size,
        )
319

320
        self.enable_eplb = parallel_config.enable_eplb
321
        self.num_redundant_experts = parallel_config.eplb_config.num_redundant_experts
322

323
        self.start_layer, self.end_layer, self.layers = make_layers(
324
325
            config.num_hidden_layers,
            lambda prefix: MixtralDecoderLayer(
326
327
328
329
330
                config,
                cache_config,
                quant_config=quant_config,
                prefix=prefix,
                enable_eplb=self.enable_eplb,
331
            ),
332
333
            prefix=f"{prefix}.layers",
        )
334

335
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
336
337
338
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
zhuwenwen's avatar
zhuwenwen committed
339
340
341
342
343
        self.quant_method = None
        if quant_config is not None:
            self.quant_method=quant_config.get_name()

        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
Pierre Stock's avatar
Pierre Stock committed
344

345
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
346
        return self.embed_tokens(input_ids)
Pierre Stock's avatar
Pierre Stock committed
347
348
349

    def forward(
        self,
zhuwenwen's avatar
zhuwenwen committed
350
        input_ids: torch.Tensor,
Pierre Stock's avatar
Pierre Stock committed
351
        positions: torch.Tensor,
352
353
354
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
355
        if get_pp_group().is_first_rank:
356
357
358
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
359
                hidden_states = self.embed_input_ids(input_ids)
360
361
362
363
364
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
365
        for layer in islice(self.layers, self.start_layer, self.end_layer):
366
            hidden_states, residual = layer(positions, hidden_states, residual)
367
        if not get_pp_group().is_last_rank:
368
369
370
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
371
372
373
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states

374
375
376
377
    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(
378
            self,
379
380
381
            ckpt_gate_proj_name="w1",
            ckpt_down_proj_name="w2",
            ckpt_up_proj_name="w3",
382
            num_experts=self.config.num_local_experts,
383
384
            num_redundant_experts=self.num_redundant_experts,
        )
385

386
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
Pierre Stock's avatar
Pierre Stock committed
387
388
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
389
390
391
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
Pierre Stock's avatar
Pierre Stock committed
392
        ]
393

Pierre Stock's avatar
Pierre Stock committed
394
        params_dict = dict(self.named_parameters())
395
        loaded_params: set[str] = set()
396
        expert_params_mapping = self.get_expert_mapping()
397
        for name, loaded_weight in weights:
398
399
400
            if self.quant_config is not None and (
                scale_name := self.quant_config.get_cache_scale(name)
            ):
401
                # Loading kv cache quantization scales
402
                param = params_dict[scale_name]
403
404
405
406
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                loaded_weight = (
                    loaded_weight if loaded_weight.dim() == 0 else loaded_weight[0]
                )
407
408
409
410
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue

411
            for param_name, weight_name, shard_id in stacked_params_mapping:
Pierre Stock's avatar
Pierre Stock committed
412
413
                if weight_name not in name:
                    continue
CHU Tianxiang's avatar
CHU Tianxiang committed
414
415
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
416
417
418
                if (
                    name.endswith(".bias") or name.endswith("_bias")
                ) and name not in params_dict:
CHU Tianxiang's avatar
CHU Tianxiang committed
419
                    continue
420
421
422
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    continue
423
424
425
426
427
                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
428
                param = params_dict[name]
Pierre Stock's avatar
Pierre Stock committed
429
430
431
432
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
433
                is_expert_weight = False
434
435
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
436

Philipp Moritz's avatar
Philipp Moritz committed
437
438
                    if weight_name not in name:
                        continue
439
440
441
442

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

443
                    # Skip layers on other devices.
444
                    if is_pp_missing_parameter(name_mapped, self):
445
                        continue
446

447
448
449
                    if (
                        name_mapped.endswith(".bias") or name_mapped.endswith("_bias")
                    ) and name_mapped not in params_dict:
450
                        continue
451
452

                    param = params_dict[name_mapped]
453
454
455
456
457
458
459
460
461
462
463
                    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,
                    )
464
465
466
                    if success:
                        name = name_mapped
                        break
Philipp Moritz's avatar
Philipp Moritz committed
467
                else:
468
469
                    if is_expert_weight:
                        continue
Philipp Moritz's avatar
Philipp Moritz committed
470
                    # Skip loading extra bias for GPTQ models.
471
472
473
                    if (
                        name.endswith(".bias") or name.endswith("_bias")
                    ) and name not in params_dict:
Philipp Moritz's avatar
Philipp Moritz committed
474
                        continue
475
476
477
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
478
                    # Remapping the name of FP8 kv-scale.
479
480
481
482
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

Philipp Moritz's avatar
Philipp Moritz committed
483
                    param = params_dict[name]
484
485
486
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
Philipp Moritz's avatar
Philipp Moritz committed
487
                    weight_loader(param, loaded_weight)
488
            loaded_params.add(name)
zhuwenwen's avatar
zhuwenwen committed
489
        
王敏's avatar
王敏 committed
490
491
492
493
494
495
496
497
498
        if self.use_llama_nn and self.quant_method is None:
            lay_key_words = [
                "block_sparse_moe.gate.weight",
                "self_attn.qkv_proj.weight",
                "self_attn.o_proj.weight",
                "lm_head.weight",
            ]
            combined_words = "|".join(lay_key_words)
            
zhuwenwen's avatar
zhuwenwen committed
499
500
            for layername in loaded_params:
                weight = params_dict[layername]
zhuwenwen's avatar
zhuwenwen committed
501
                os.environ['LM_NN'] = '0' 
zhuwenwen's avatar
zhuwenwen committed
502

王敏's avatar
王敏 committed
503
504
505
506
507
508
509
510
511
512
                matches = re.findall(combined_words, layername)

                if matches:
                    _weight = torch.zeros_like(weight.data)
                    ori_shape =_weight.shape
                    
                    ops.trans_w16_gemm(_weight, weight.data, _weight.shape[0], _weight.shape[1])
                    weight.data.copy_(_weight)
                    
                    weight.data=weight.data.reshape(ori_shape[1],-1)
zhuwenwen's avatar
zhuwenwen committed
513
                    
514
        return loaded_params
515
516


517
class MixtralForCausalLM(nn.Module, SupportsLoRA, SupportsPP, MixtureOfExperts):
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
    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",
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
538

539
        self.config = config
540

541
542
        self.quant_config = quant_config

543
544
545
        self.model = MixtralModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
546

547
        self.lm_head = ParallelLMHead(
548
            config.vocab_size,
549
550
            config.hidden_size,
            quant_config=quant_config,
551
            prefix=maybe_prefix(prefix, "lm_head"),
552
553
554
        )
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
555
        self.logits_processor = LogitsProcessor(config.vocab_size)
556
        self.make_empty_intermediate_tensors = (
557
558
            self.model.make_empty_intermediate_tensors
        )
559

560
        self.expert_weights = []
561
        self.moe_layers = []
562
563
564
565
566
567
568
        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(
569
570
                layer.block_sparse_moe, MixtralMoE
            ):
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
                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 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
595
        self.num_redundant_experts = num_physical_experts - self.num_logical_experts
596
597
        for layer in self.model.layers:
            if hasattr(layer, "block_sparse_moe") and isinstance(
598
599
                layer.block_sparse_moe, MixtralMoE
            ):
600
601
602
603
604
605
                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()

606
607
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
608
609
610

    def forward(
        self,
zhuwenwen's avatar
zhuwenwen committed
611
        input_ids: torch.Tensor,
612
        positions: torch.Tensor,
613
614
615
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
616
617
618
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
619
620
621
622
623
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
624
    ) -> torch.Tensor | None:
625
        logits = self.logits_processor(self.lm_head, hidden_states)
626
627
        return logits

628
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
629
        loader = AutoWeightsLoader(self)
630
        return loader.load_weights(weights)
631
632
633

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