mixtral_quant.py 17.8 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# 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."""
25
from typing import Iterable, Optional, Set, Tuple, Union
26
27
28
29
30
31
32

import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from transformers import MixtralConfig

33
from vllm.attention import Attention
34
from vllm.config import CacheConfig, VllmConfig
35
from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
36
37
                              get_tensor_model_parallel_world_size,
                              tensor_model_parallel_all_reduce)
38
from vllm.model_executor.layers.layernorm import RMSNorm
39
from vllm.model_executor.layers.linear import (QKVParallelLinear,
40
                                               ReplicatedLinear,
41
                                               RowParallelLinear)
42
from vllm.model_executor.layers.logits_processor import LogitsProcessor
43
from vllm.model_executor.layers.quantization import QuantizationConfig
44
from vllm.model_executor.layers.rotary_embedding import get_rope
Joe Runde's avatar
Joe Runde committed
45
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
46
from vllm.model_executor.layers.vocab_parallel_embedding import (
47
    ParallelLMHead, VocabParallelEmbedding)
48
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
49
from vllm.model_executor.sampling_metadata import SamplingMetadata
50
from vllm.sequence import IntermediateTensors
51

52
53
from .interfaces import SupportsPP
from .utils import (is_pp_missing_parameter,
54
55
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
56

57
58
59
60
61
62
63
64

class MixtralMLP(nn.Module):

    def __init__(
        self,
        num_experts: int,
        hidden_size: int,
        intermediate_size: int,
65
        quant_config: Optional[QuantizationConfig] = None,
66
67
68
69
70
71
72
73
74
    ) -> None:
        super().__init__()
        self.num_experts = num_experts
        self.ffn_dim = intermediate_size
        self.hidden_dim = hidden_size

        self.w1 = ReplicatedLinear(self.hidden_dim,
                                   self.ffn_dim,
                                   bias=False,
75
                                   quant_config=quant_config)
76
77
78
        self.w2 = ReplicatedLinear(self.ffn_dim,
                                   self.hidden_dim,
                                   bias=False,
79
                                   quant_config=quant_config)
80
81
82
        self.w3 = ReplicatedLinear(self.hidden_dim,
                                   self.ffn_dim,
                                   bias=False,
83
                                   quant_config=quant_config)
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101

        # TODO: Use vllm's SiluAndMul
        self.act_fn = nn.SiLU()

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        w1_out, _ = self.w1(hidden_states)
        w1_out = self.act_fn(w1_out)
        w3_out, _ = self.w3(hidden_states)
        current_hidden_states = w1_out * w3_out
        current_hidden_states, _ = self.w2(current_hidden_states)
        return current_hidden_states


class MixtralMoE(nn.Module):

    def __init__(
        self,
        config: MixtralConfig,
102
        quant_config: Optional[QuantizationConfig] = None,
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
    ):
        super().__init__()
        self.config = config
        self.rank = get_tensor_model_parallel_rank()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.num_total_experts = config.num_local_experts
        self.top_k = config.num_experts_per_tok
        if self.tp_size > self.num_total_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {self.num_total_experts}.")
        # Split experts equally between ranks
        self.expert_indicies = np.array_split(range(
            self.num_total_experts), self.tp_size)[self.rank].tolist()
        if not self.expert_indicies:
            raise ValueError(
                f"Rank {self.rank} has no experts assigned to it.")

        self.experts = nn.ModuleList([
            MixtralMLP(self.num_total_experts,
                       config.hidden_size,
                       config.intermediate_size,
125
                       quant_config=quant_config)
126
127
128
129
130
131
            if idx in self.expert_indicies else None
            for idx in range(self.num_total_experts)
        ])
        self.gate = ReplicatedLinear(config.hidden_size,
                                     self.num_total_experts,
                                     bias=False,
132
                                     quant_config=None)
133
134

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
135
        num_tokens, hidden_dim = hidden_states.shape
136
        hidden_states = hidden_states.view(-1, hidden_dim)
137
        # router_logits: (num_tokens, n_experts)
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
        router_logits, _ = self.gate(hidden_states)

        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
        routing_weights, selected_experts = torch.topk(routing_weights,
                                                       self.top_k,
                                                       dim=-1)
        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)

        final_hidden_states = None
        for expert_idx in self.expert_indicies:
            expert_layer = self.experts[expert_idx]
            expert_mask = (selected_experts == expert_idx)
            expert_weights = (routing_weights * expert_mask).sum(dim=-1,
                                                                 keepdim=True)

            current_hidden_states = expert_layer(hidden_states).mul_(
                expert_weights)
            if final_hidden_states is None:
                final_hidden_states = current_hidden_states
            else:
                final_hidden_states.add_(current_hidden_states)

        return tensor_model_parallel_all_reduce(final_hidden_states).view(
161
            num_tokens, hidden_dim)
162
163
164
165


class MixtralAttention(nn.Module):

166
167
168
169
170
171
172
173
174
    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position: int = 4096 * 32,
        rope_theta: float = 10000,
        quant_config: Optional[QuantizationConfig] = None,
        cache_config: Optional[CacheConfig] = None,
175
        prefix: str = "",
176
    ) -> None:
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
        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.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
205
            quant_config=quant_config,
206
207
208
209
210
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
211
            quant_config=quant_config,
212
213
214
215
216
217
218
219
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position,
            base=int(self.rope_theta),
            is_neox_style=True,
        )
220
221
222
223
224
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
                              num_kv_heads=self.num_kv_heads,
                              cache_config=cache_config,
225
226
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")
227
228
229
230
231
232
233
234
235

    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)
236
        attn_output = self.attn(q, k, v)
237
238
239
240
241
242
243
244
245
        output, _ = self.o_proj(attn_output)
        return output


class MixtralDecoderLayer(nn.Module):

    def __init__(
        self,
        config: MixtralConfig,
246
        cache_config: Optional[CacheConfig] = None,
247
        quant_config: Optional[QuantizationConfig] = None,
248
        prefix: str = "",
249
250
251
252
253
254
255
256
257
258
259
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        # Requires transformers > 4.32.0
        rope_theta = getattr(config, "rope_theta", 10000)
        self.self_attn = MixtralAttention(
            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,
260
            cache_config=cache_config,
261
262
263
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
        )
264
        self.block_sparse_moe = MixtralMoE(config=config,
265
                                           quant_config=quant_config)
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
        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.block_sparse_moe(hidden_states)
        return hidden_states, residual


class MixtralModel(nn.Module):

298
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
299
        super().__init__()
300
301
302
303
304

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

305
306
307
308
309
310
311
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
312
313
314
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: MixtralDecoderLayer(
315
316
                config, cache_config, quant_config=quant_config, prefix=prefix
            ),
317
            prefix=f"{prefix}.layers")
318
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
319
320
321
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))
322

323
324
325
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

326
327
328
329
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
330
        intermediate_tensors: Optional[IntermediateTensors],
331
        inputs_embeds: Optional[torch.Tensor] = None,
332
333
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
334
335
336
337
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
338
339
340
341
342
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
343
344
        for layer in self.layers[self.start_layer:self.end_layer]:
            hidden_states, residual = layer(positions, hidden_states, residual)
345
346
347
348
349
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
350
351
352
353
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


354
class MixtralForCausalLM(nn.Module, SupportsPP):
355
    fall_back_to_pt_during_load = False
356

357
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
358
        super().__init__()
359
360
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
361
        self.config = config
362
        self.quant_config = quant_config
363
364
        self.model = MixtralModel(vllm_config=vllm_config,
                                  prefix=maybe_prefix(prefix, "model"))
365
366
367
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
                                      quant_config=quant_config)
368
369
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
370
        self.logits_processor = LogitsProcessor(config.vocab_size)
Joe Runde's avatar
Joe Runde committed
371
        self.sampler = get_sampler()
372
373
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
374

375
376
377
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

378
379
380
381
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
382
        intermediate_tensors: Optional[IntermediateTensors] = None,
383
        inputs_embeds: Optional[torch.Tensor] = None,
384
    ) -> Union[torch.Tensor, IntermediateTensors]:
385
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
386
                                   inputs_embeds)
387
388
        return hidden_states

389
390
391
392
393
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
394
        logits = self.logits_processor(self.lm_head, hidden_states,
395
396
397
                                       sampling_metadata)
        return logits

398
399
    def sample(
        self,
400
        logits: Optional[torch.Tensor],
401
402
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
403
        next_tokens = self.sampler(logits, sampling_metadata)
404
405
        return next_tokens

406
407
    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
408
409
410
411
412
413
414
415
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]

        params_dict = dict(self.named_parameters())
416
        loaded_params: Set[str] = set()
417
        for name, loaded_weight in weights:
418
419
420
421
422
423
424
425
426
            if "rotary_emb.inv_freq" in name:
                continue
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
427
428
                if is_pp_missing_parameter(name, self):
                    continue
429
430
431
432
433
434
435
436
437
438
439
440
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Skip experts that are not assigned to this worker.
                if ("block_sparse_moe.experts." in name
                        and name not in params_dict):
                    continue
441
442
                if is_pp_missing_parameter(name, self):
                    continue
443
444
445
446
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
447
448
            loaded_params.add(name)
        return loaded_params