gpt_bigcode.py 9.92 KB
Newer Older
1
# coding=utf-8
2
3
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gpt2/modeling_gpt2.py
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
# Copyright 2023 The vLLM team.
# Copyright 2023 CTranslate2, and Michael Feil
# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# 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.
Woosuk Kwon's avatar
Woosuk Kwon committed
20
"""Inference-only GPTBigCode model compatible with HuggingFace weights."""
21
from typing import List, Optional, Tuple
22
23
24
25
26
27
28
29

import torch
from torch import nn
from transformers import GPTBigCodeConfig

from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention import PagedAttention
30
31
32
33
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               LinearMethodBase,
                                               QKVParallelLinear,
                                               RowParallelLinear)
34
from vllm.model_executor.layers.sampler import Sampler
35
36
from vllm.model_executor.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding)
37
from vllm.model_executor.parallel_utils.parallel_state import (
38
39
40
    get_tensor_model_parallel_world_size)
from vllm.model_executor.weight_utils import (default_weight_loader,
                                              hf_model_weights_iterator)
41
from vllm.sequence import SamplerOutput
42
43
44
45
46
47

KVCache = Tuple[torch.Tensor, torch.Tensor]


class GPTBigCodeAttention(nn.Module):

48
49
50
51
52
    def __init__(
        self,
        config: GPTBigCodeConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
53
54
55
        super().__init__()
        self.hidden_size = config.hidden_size
        total_num_heads = config.num_attention_heads
56
        self.tensor_model_parallel_world_size = (
57
            get_tensor_model_parallel_world_size())
58
59
60
        assert total_num_heads % self.tensor_model_parallel_world_size == 0
        self.num_heads = (total_num_heads //
                          self.tensor_model_parallel_world_size)
61
        self.head_dim = self.hidden_size // total_num_heads
62
        self.scale = self.head_dim**-0.5
63

64
65
        self.multi_query = config.multi_query
        if self.multi_query:
66
            total_num_kv_heads = 1
67
68
            self.num_kv_heads = 1
        else:
69
            total_num_kv_heads = total_num_heads
70
            self.num_kv_heads = self.num_heads
71
72
73
74
75
76
77
78
79
        self.kv_dim = self.head_dim * self.num_kv_heads
        self.c_attn = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            total_num_heads,
            total_num_kv_heads,
            bias=True,
            linear_method=linear_method,
        )
80
81
82
83
84

        self.c_proj = RowParallelLinear(
            self.hidden_size,
            self.hidden_size,
            bias=True,
85
            linear_method=linear_method,
86
        )
87
88
        self.attn = PagedAttention(self.num_heads,
                                   self.head_dim,
Zhuohan Li's avatar
Zhuohan Li committed
89
90
                                   scale=self.scale,
                                   num_kv_heads=self.num_kv_heads)
91
92
93
94
95
96
97
98

    def forward(
        self,
        hidden_states: torch.Tensor,
        kv_cache: KVCache,
        input_metadata: InputMetadata,
        cache_event: Optional[torch.cuda.Event],
    ) -> torch.Tensor:
99
100
101
        qkv, _ = self.c_attn(hidden_states)
        q, k, v = qkv.split(
            [
102
103
104
                self.hidden_size // self.tensor_model_parallel_world_size,
                self.kv_dim, self.kv_dim
            ],
105
106
            dim=-1,
        )
107
        key_cache, value_cache = kv_cache
108
109
        attn_output = self.attn(q, k, v, key_cache, value_cache,
                                input_metadata, cache_event)
110
111
112
113
114
115
116
117
118
119
        attn_output, _ = self.c_proj(attn_output)
        return attn_output


class GPTBigMLP(nn.Module):

    def __init__(
        self,
        intermediate_size: int,
        config: GPTBigCodeConfig,
120
        linear_method: Optional[LinearMethodBase] = None,
121
122
123
    ):
        super().__init__()
        hidden_size = config.hidden_size
124
125
126
127
        self.c_fc = ColumnParallelLinear(
            hidden_size,
            intermediate_size,
            bias=True,
128
            linear_method=linear_method,
129
130
131
132
133
        )
        self.c_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=True,
134
            linear_method=linear_method,
135
        )
136
137
138
        quant_config = getattr(linear_method, "quant_config", None)
        self.act = get_act_fn(config.activation_function, quant_config,
                              intermediate_size)
139
140
141
142
143
144
145
146
147
148

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.c_fc(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.c_proj(hidden_states)
        return hidden_states


class GPTBigCodeBlock(nn.Module):

149
150
151
152
153
    def __init__(
        self,
        config: GPTBigCodeConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
154
155
        super().__init__()
        hidden_size = config.hidden_size
156
157
        inner_dim = (config.n_inner if config.n_inner is not None else 4 *
                     hidden_size)
158
159

        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
160
        self.attn = GPTBigCodeAttention(config, linear_method)
161
        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
162
        self.mlp = GPTBigMLP(inner_dim, config, linear_method)
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191

    def forward(
        self,
        hidden_states: torch.Tensor,
        kv_cache: KVCache,
        input_metadata: InputMetadata,
        cache_event: Optional[torch.cuda.Event],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)
        attn_output = self.attn(
            hidden_states=hidden_states,
            kv_cache=kv_cache,
            input_metadata=input_metadata,
            cache_event=cache_event,
        )
        # residual connection
        hidden_states = attn_output + residual

        residual = hidden_states
        hidden_states = self.ln_2(hidden_states)
        feed_forward_hidden_states = self.mlp(hidden_states)
        # residual connection
        hidden_states = residual + feed_forward_hidden_states
        return hidden_states


class GPTBigCodeModel(nn.Module):

192
193
194
195
196
    def __init__(
        self,
        config: GPTBigCodeConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
197
198
        super().__init__()
        self.config = config
199
        assert not config.add_cross_attention
200
201
202

        self.embed_dim = config.hidden_size

203
        self.wte = VocabParallelEmbedding(config.vocab_size, self.embed_dim)
204
        self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
205
206
207
208
        self.h = nn.ModuleList([
            GPTBigCodeBlock(config, linear_method)
            for _ in range(config.num_hidden_layers)
        ])
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
        kv_caches: List[KVCache],
        input_metadata: InputMetadata,
        cache_events: Optional[List[torch.cuda.Event]],
    ) -> torch.Tensor:
        inputs_embeds = self.wte(input_ids)
        position_embeds = self.wpe(position_ids)
        hidden_states = inputs_embeds + position_embeds

        for i in range(len(self.h)):
224
            cache_event = None if cache_events is None else cache_events[i]
225
            layer = self.h[i]
226
227
            hidden_states = layer(hidden_states, kv_caches[i], input_metadata,
                                  cache_event)
228
229
230
231
232
233
234

        hidden_states = self.ln_f(hidden_states)
        return hidden_states


class GPTBigCodeForCausalLM(nn.Module):

235
236
237
238
239
    def __init__(
        self,
        config: GPTBigCodeConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
240
241
        super().__init__()
        self.config = config
242
243
        self.linear_method = linear_method
        self.transformer = GPTBigCodeModel(config, linear_method)
244
245
246
247
248
249
250
251
252
253
        self.lm_head_weight = self.transformer.wte.weight
        self.sampler = Sampler(config.vocab_size)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[KVCache],
        input_metadata: InputMetadata,
        cache_events: Optional[List[torch.cuda.Event]],
254
    ) -> SamplerOutput:
255
256
257
258
        hidden_states = self.transformer(input_ids, positions, kv_caches,
                                         input_metadata, cache_events)
        next_tokens = self.sampler(self.lm_head_weight, hidden_states,
                                   input_metadata)
259
260
        return next_tokens

261
262
    def load_weights(self,
                     model_name_or_path: str,
263
                     cache_dir: Optional[str] = None,
Jasmond L's avatar
Jasmond L committed
264
265
                     load_format: str = "auto",
                     revision: Optional[str] = None):
266
        params_dict = dict(self.named_parameters(remove_duplicate=False))
267
        for name, loaded_weight in hf_model_weights_iterator(
Jasmond L's avatar
Jasmond L committed
268
                model_name_or_path, cache_dir, load_format, revision):
269
270
271
272
273
274
            if "lm_head.weight" in name:
                continue
            if ".attn.bias" in name:
                # Skip attention mask.
                # NOTE: "c_attn.bias" should not be skipped.
                continue
275
276
277
278
            param = params_dict[name]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, loaded_weight)