mixtral.py 19.9 KB
Newer Older
Pierre Stock's avatar
Pierre Stock committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
# coding=utf-8
# 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
24
25
import os
import re
26
from typing import Iterable, List, Optional, Tuple
Pierre Stock's avatar
Pierre Stock committed
27
28
29

import torch
from torch import nn
30
from transformers import MixtralConfig
Pierre Stock's avatar
Pierre Stock committed
31

32
from vllm.attention import Attention, AttentionMetadata
33
from vllm.config import CacheConfig, LoRAConfig
34
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
35
from vllm.model_executor.layers.fused_moe import FusedMoE
Pierre Stock's avatar
Pierre Stock committed
36
from vllm.model_executor.layers.layernorm import RMSNorm
37
from vllm.model_executor.layers.linear import (QKVParallelLinear,
Philipp Moritz's avatar
Philipp Moritz committed
38
                                               ReplicatedLinear,
Pierre Stock's avatar
Pierre Stock committed
39
                                               RowParallelLinear)
40
from vllm.model_executor.layers.logits_processor import LogitsProcessor
41
42
from vllm.model_executor.layers.quantization.base_config import (
    QuantizationConfig)
43
from vllm.model_executor.layers.rotary_embedding import get_rope
44
from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
Pierre Stock's avatar
Pierre Stock committed
45
from vllm.model_executor.layers.vocab_parallel_embedding import (
46
    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
47
48
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader, maybe_remap_kv_scale_name)
Pierre Stock's avatar
Pierre Stock committed
49
from vllm.model_executor.sampling_metadata import SamplingMetadata
50
from vllm.sequence import IntermediateTensors
王敏's avatar
王敏 committed
51
52
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
53

54
from .interfaces import SupportsLoRA
55
from .utils import is_pp_missing_parameter, make_layers
56

Pierre Stock's avatar
Pierre Stock committed
57

Philipp Moritz's avatar
Philipp Moritz committed
58
59
60
61
62
63
64
65
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.
    """
66

67
68
69
70
71
72
73
    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,
74
75
                 tp_size: Optional[int] = None,
                 prefix: str = ""):
76
        super().__init__()
Philipp Moritz's avatar
Philipp Moritz committed
77
        self.hidden_size = hidden_size
78

79
        # Gate always runs at half / full precision for now.
80

81
82
        self.gate = ReplicatedLinear(hidden_size,
                                     num_experts,
83
                                     bias=False,
84
                                     params_dtype=params_dtype,
85
86
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")
87

88
89
90
91
92
93
94
95
        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,
96
97
                                tp_size=tp_size,
                                prefix=f"{prefix}.experts")
98

99
    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
100
101
        # NOTE: hidden_states can have either 1D or 2D shape.
        orig_shape = hidden_states.shape
Philipp Moritz's avatar
Philipp Moritz committed
102
        hidden_states = hidden_states.view(-1, self.hidden_size)
103
        # router_logits: (num_tokens, n_experts)
104
        router_logits, _ = self.gate(hidden_states)
105
        final_hidden_states = self.experts(hidden_states, router_logits)
106
        return final_hidden_states.view(orig_shape)
Pierre Stock's avatar
Pierre Stock committed
107
108
109
110


class MixtralAttention(nn.Module):

111
112
113
114
115
116
117
118
119
    def __init__(
        self,
        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,
120
        prefix: str = "",
121
    ) -> None:
Pierre Stock's avatar
Pierre Stock committed
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
        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

144
        self.qkv_proj = QKVParallelLinear(
Pierre Stock's avatar
Pierre Stock committed
145
146
147
148
149
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
150
            quant_config=quant_config,
151
            prefix=f"{prefix}.qkv_proj",
Pierre Stock's avatar
Pierre Stock committed
152
        )
153
        self.o_proj = RowParallelLinear(
Pierre Stock's avatar
Pierre Stock committed
154
155
156
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
157
            quant_config=quant_config,
158
            prefix=f"{prefix}.o_proj",
Pierre Stock's avatar
Pierre Stock committed
159
160
161
162
163
164
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position,
            base=int(self.rope_theta),
165
            is_neox_style=True,
Pierre Stock's avatar
Pierre Stock committed
166
        )
167
168
169
170
171
172
        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)
Pierre Stock's avatar
Pierre Stock committed
173
174
175
176
177

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
178
179
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
Pierre Stock's avatar
Pierre Stock committed
180
    ) -> torch.Tensor:
181
        qkv, _ = self.qkv_proj(hidden_states)
Pierre Stock's avatar
Pierre Stock committed
182
183
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)
184
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
185
        output, _ = self.o_proj(attn_output)
Pierre Stock's avatar
Pierre Stock committed
186
187
188
189
190
191
192
        return output


class MixtralDecoderLayer(nn.Module):

    def __init__(
        self,
193
        config: MixtralConfig,
194
        cache_config: Optional[CacheConfig] = None,
195
        quant_config: Optional[QuantizationConfig] = None,
196
        prefix: str = "",
Pierre Stock's avatar
Pierre Stock committed
197
198
199
200
201
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        # Requires transformers > 4.32.0
        rope_theta = getattr(config, "rope_theta", 10000)
202
        self.self_attn = MixtralAttention(
Pierre Stock's avatar
Pierre Stock committed
203
204
205
206
207
            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,
208
            cache_config=cache_config,
209
210
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn")
Philipp Moritz's avatar
Philipp Moritz committed
211
212
213
214
        self.block_sparse_moe = MixtralMoE(
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
215
            intermediate_size=config.intermediate_size,
216
217
            quant_config=quant_config,
            prefix=f"{prefix}.block_sparse_moe")
218
219
220
221
        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
222
223
224
225

    def forward(
        self,
        positions: torch.Tensor,
226
        hidden_states: torch.Tensor,
227
228
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
229
        residual: Optional[torch.Tensor],
Pierre Stock's avatar
Pierre Stock committed
230
    ) -> torch.Tensor:
231
232
233
234
235
236
237
238
        # 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
239
            positions=positions,
240
            hidden_states=hidden_states,
Pierre Stock's avatar
Pierre Stock committed
241
            kv_cache=kv_cache,
242
            attn_metadata=attn_metadata,
Pierre Stock's avatar
Pierre Stock committed
243
244
        )

245
246
247
248
249
        # 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
250

251
252

class MixtralModel(nn.Module):
Pierre Stock's avatar
Pierre Stock committed
253
254
255

    def __init__(
        self,
256
        config: MixtralConfig,
257
        cache_config: Optional[CacheConfig] = None,
258
        quant_config: Optional[QuantizationConfig] = None,
259
        lora_config: Optional[LoRAConfig] = None,
260
        prefix: str = "",
Pierre Stock's avatar
Pierre Stock committed
261
262
263
    ) -> None:
        super().__init__()
        self.padding_idx = config.pad_token_id
264
265
266
267
        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
268
269

        self.embed_tokens = VocabParallelEmbedding(
270
            self.vocab_size,
Pierre Stock's avatar
Pierre Stock committed
271
            config.hidden_size,
272
            org_num_embeddings=config.vocab_size,
Pierre Stock's avatar
Pierre Stock committed
273
        )
274
275

        self.start_layer, self.end_layer, self.layers = make_layers(
276
277
278
279
280
            config.num_hidden_layers,
            lambda prefix: MixtralDecoderLayer(
                config, cache_config, quant_config=quant_config, prefix=prefix
            ),
            prefix=f"{prefix}.layers")
281

282
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
Pierre Stock's avatar
Pierre Stock committed
283
284
285
286
287

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
288
289
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
290
        intermediate_tensors: Optional[IntermediateTensors],
291
    ) -> torch.Tensor:
292
293
294
295
296
297
298
299
        if get_pp_group().is_first_rank:
            hidden_states = self.embed_tokens(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
        for i in range(self.start_layer, self.end_layer):
Pierre Stock's avatar
Pierre Stock committed
300
            layer = self.layers[i]
301
            hidden_states, residual = layer(positions, hidden_states,
302
303
304
305
306
307
308
                                            kv_caches[i - self.start_layer],
                                            attn_metadata, residual)
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
309
310
311
312
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


313
class MixtralForCausalLM(nn.Module, SupportsLoRA):
314
315
    fall_back_to_pt_during_load = False

Terry's avatar
Terry committed
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }

    # LoRA specific attributes
    supported_lora_modules = [
        "qkv_proj",
        "o_proj",
        "embed_tokens",
        "lm_head",
    ]
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }
    embedding_padding_modules = ["lm_head"]
336
337
338
339

    def __init__(
        self,
        config: MixtralConfig,
340
        cache_config: Optional[CacheConfig] = None,
341
        quant_config: Optional[QuantizationConfig] = None,
Terry's avatar
Terry committed
342
        lora_config: Optional[LoRAConfig] = None,
343
344
    ) -> None:
        super().__init__()
345

346
        self.config = config
347
348
        self.lora_config = lora_config

349
        self.model = MixtralModel(config,
350
                                  cache_config,
351
                                  quant_config,
352
353
                                  lora_config=lora_config,
                                  prefix="model")
Terry's avatar
Terry committed
354
355
356
357
358
359
360
361
362
363
364
        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,
365
            quant_config=quant_config,
Terry's avatar
Terry committed
366
        )
367
368
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
369
370
371
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size)
        self.sampler = Sampler()
372

王敏's avatar
王敏 committed
373
374
375
376
377
378
        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'

379
380
381
382
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
383
384
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
385
        intermediate_tensors: Optional[IntermediateTensors] = None,
386
387
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, kv_caches,
388
                                   attn_metadata, intermediate_tensors)
Pierre Stock's avatar
Pierre Stock committed
389
390
        return hidden_states

391
392
393
394
395
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
396
        logits = self.logits_processor(self.lm_head, hidden_states,
397
398
399
                                       sampling_metadata)
        return logits

400
401
402
403
404
405
406
407
408
409
410
411
412
413
    def make_empty_intermediate_tensors(
            self, batch_size: int, dtype: torch.dtype,
            device: torch.device) -> IntermediateTensors:
        return IntermediateTensors({
            "hidden_states":
            torch.zeros((batch_size, self.config.hidden_size),
                        dtype=dtype,
                        device=device),
            "residual":
            torch.zeros((batch_size, self.config.hidden_size),
                        dtype=dtype,
                        device=device),
        })

Pierre Stock's avatar
Pierre Stock committed
414
415
    def sample(
        self,
416
        logits: Optional[torch.Tensor],
Pierre Stock's avatar
Pierre Stock committed
417
        sampling_metadata: SamplingMetadata,
418
    ) -> Optional[SamplerOutput]:
419
        next_tokens = self.sampler(logits, sampling_metadata)
Pierre Stock's avatar
Pierre Stock committed
420
421
        return next_tokens

422
    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
Pierre Stock's avatar
Pierre Stock committed
423
424
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
425
426
427
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
Pierre Stock's avatar
Pierre Stock committed
428
        ]
429

430
431
432
433
434
435
436
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        expert_params_mapping = FusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="w1",
            ckpt_down_proj_name="w2",
            ckpt_up_proj_name="w3",
            num_experts=self.config.num_local_experts)
Philipp Moritz's avatar
Philipp Moritz committed
437

Pierre Stock's avatar
Pierre Stock committed
438
        params_dict = dict(self.named_parameters())
439
        for name, loaded_weight in weights:
Pierre Stock's avatar
Pierre Stock committed
440
441
            if "rotary_emb.inv_freq" in name:
                continue
Philipp Moritz's avatar
Philipp Moritz committed
442

Pierre Stock's avatar
Pierre Stock committed
443
444
445
            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
446
447
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
448
449
                if ((name.endswith(".bias") or name.endswith("_bias"))
                        and name not in params_dict):
CHU Tianxiang's avatar
CHU Tianxiang committed
450
                    continue
451
452
453
454
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    continue

CHU Tianxiang's avatar
CHU Tianxiang committed
455
                param = params_dict[name]
Pierre Stock's avatar
Pierre Stock committed
456
457
458
459
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
460
461
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
Philipp Moritz's avatar
Philipp Moritz committed
462
463
464
                    if weight_name not in name:
                        continue
                    name = name.replace(weight_name, param_name)
465
466
467
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
468
469
470
                    if ((name.endswith(".bias") or name.endswith("_bias"))
                            and name not in params_dict):
                        continue
Philipp Moritz's avatar
Philipp Moritz committed
471
472
473
474
                    param = params_dict[name]
                    weight_loader = param.weight_loader
                    weight_loader(param,
                                  loaded_weight,
475
                                  name,
476
                                  shard_id=shard_id,
Philipp Moritz's avatar
Philipp Moritz committed
477
478
479
480
                                  expert_id=expert_id)
                    break
                else:
                    # Skip loading extra bias for GPTQ models.
481
482
                    if ((name.endswith(".bias") or name.endswith("_bias"))
                            and name not in params_dict):
Philipp Moritz's avatar
Philipp Moritz committed
483
                        continue
484
485
486
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
487
                    # Remapping the name of FP8 kv-scale.
488
489
490
491
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

Philipp Moritz's avatar
Philipp Moritz committed
492
493
494
495
                    param = params_dict[name]
                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)
王敏's avatar
王敏 committed
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516

        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)
            
            for layername, weight in params_dict.items():
                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)