File size: 10,955 Bytes
a0db2f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
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
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
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
<!--Copyright 2022 The HuggingFace Team. 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.

โš ๏ธ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.

-->

# TorchScript๋กœ ๋‚ด๋ณด๋‚ด๊ธฐ[[export-to-torchscript]]

<Tip>

TorchScript๋ฅผ ํ™œ์šฉํ•œ ์‹คํ—˜์€ ์•„์ง ์ดˆ๊ธฐ ๋‹จ๊ณ„๋กœ, ๊ฐ€๋ณ€์ ์ธ ์ž…๋ ฅ ํฌ๊ธฐ ๋ชจ๋ธ๋“ค์„ ํ†ตํ•ด ๊ทธ ๊ธฐ๋Šฅ์„ฑ์„ ๊ณ„์† ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. 
์ด ๊ธฐ๋Šฅ์€ ์ €ํฌ๊ฐ€ ๊ด€์‹ฌ์„ ๋‘๊ณ  ์žˆ๋Š” ๋ถ„์•ผ ์ค‘ ํ•˜๋‚˜์ด๋ฉฐ, 
์•ž์œผ๋กœ ์ถœ์‹œ๋  ๋ฒ„์ „์—์„œ ๋” ๋งŽ์€ ์ฝ”๋“œ ์˜ˆ์ œ, ๋” ์œ ์—ฐํ•œ ๊ตฌํ˜„, ๊ทธ๋ฆฌ๊ณ  Python ๊ธฐ๋ฐ˜ ์ฝ”๋“œ์™€ ์ปดํŒŒ์ผ๋œ TorchScript๋ฅผ ๋น„๊ตํ•˜๋Š” ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋“ฑ์„ ํ†ตํ•ด ๋ถ„์„์„ ์‹ฌํ™”ํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.

</Tip>

[TorchScript ๋ฌธ์„œ](https://pytorch.org/docs/stable/jit.html)์—์„œ๋Š” ์ด๋ ‡๊ฒŒ ๋งํ•ฉ๋‹ˆ๋‹ค.

> TorchScript๋Š” PyTorch ์ฝ”๋“œ์—์„œ ์ง๋ ฌํ™” ๋ฐ ์ตœ์ ํ™” ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.

[JIT๊ณผ TRACE](https://pytorch.org/docs/stable/jit.html)๋Š” ๊ฐœ๋ฐœ์ž๊ฐ€ ๋ชจ๋ธ์„ ๋‚ด๋ณด๋‚ด์„œ ํšจ์œจ ์ง€ํ–ฅ์ ์ธ C++ ํ”„๋กœ๊ทธ๋žจ๊ณผ ๊ฐ™์€ ๋‹ค๋ฅธ ํ”„๋กœ๊ทธ๋žจ์—์„œ ์žฌ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” PyTorch ๋ชจ๋“ˆ์ž…๋‹ˆ๋‹ค.

PyTorch ๊ธฐ๋ฐ˜ Python ํ”„๋กœ๊ทธ๋žจ๊ณผ ๋‹ค๋ฅธ ํ™˜๊ฒฝ์—์„œ ๋ชจ๋ธ์„ ์žฌ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก, ๐Ÿค— Transformers ๋ชจ๋ธ์„ TorchScript๋กœ ๋‚ด๋ณด๋‚ผ ์ˆ˜ ์žˆ๋Š” ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. 
์ด ๋ฌธ์„œ์—์„œ๋Š” TorchScript๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ๋‚ด๋ณด๋‚ด๊ณ  ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.

๋ชจ๋ธ์„ ๋‚ด๋ณด๋‚ด๋ ค๋ฉด ๋‘ ๊ฐ€์ง€๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค:

- `torchscript` ํ”Œ๋ž˜๊ทธ๋กœ ๋ชจ๋ธ ์ธ์Šคํ„ด์Šคํ™”
- ๋”๋ฏธ ์ž…๋ ฅ์„ ์‚ฌ์šฉํ•œ ์ˆœ์ „ํŒŒ(forward pass)

์ด ํ•„์ˆ˜ ์กฐ๊ฑด๋“ค์€ ์•„๋ž˜์— ์ž์„ธํžˆ ์„ค๋ช…๋œ ๊ฒƒ์ฒ˜๋Ÿผ ๊ฐœ๋ฐœ์ž๋“ค์ด ์ฃผ์˜ํ•ด์•ผ ํ•  ์—ฌ๋Ÿฌ ์‚ฌํ•ญ๋“ค์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.

## TorchScript ํ”Œ๋ž˜๊ทธ์™€ ๋ฌถ์ธ ๊ฐ€์ค‘์น˜(tied weights)[[torchscript-flag-and-tied-weights]]

`torchscript` ํ”Œ๋ž˜๊ทธ๊ฐ€ ํ•„์š”ํ•œ ์ด์œ ๋Š” ๋Œ€๋ถ€๋ถ„์˜ ๐Ÿค— Transformers ์–ธ์–ด ๋ชจ๋ธ์—์„œ `Embedding` ๋ ˆ์ด์–ด์™€ `Decoding` ๋ ˆ์ด์–ด ๊ฐ„์˜ ๋ฌถ์ธ ๊ฐ€์ค‘์น˜(tied weights)๊ฐ€ ์กด์žฌํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
TorchScript๋Š” ๋ฌถ์ธ ๊ฐ€์ค‘์น˜๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์„ ๋‚ด๋ณด๋‚ผ ์ˆ˜ ์—†์œผ๋ฏ€๋กœ, ๋ฏธ๋ฆฌ ๊ฐ€์ค‘์น˜๋ฅผ ํ’€๊ณ  ๋ณต์ œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

`torchscript` ํ”Œ๋ž˜๊ทธ๋กœ ์ธ์Šคํ„ด์Šคํ™”๋œ ๋ชจ๋ธ์€ `Embedding` ๋ ˆ์ด์–ด์™€ `Decoding` ๋ ˆ์ด์–ด๊ฐ€ ๋ถ„๋ฆฌ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ ์ดํ›„์— ํ›ˆ๋ จํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค.
ํ›ˆ๋ จ์„ ํ•˜๊ฒŒ ๋˜๋ฉด ๋‘ ๋ ˆ์ด์–ด ๊ฐ„ ๋™๊ธฐํ™”๊ฐ€ ํ•ด์ œ๋˜์–ด ์˜ˆ์ƒ์น˜ ๋ชปํ•œ ๊ฒฐ๊ณผ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์–ธ์–ด ๋ชจ๋ธ ํ—ค๋“œ๋ฅผ ๊ฐ–์ง€ ์•Š์€ ๋ชจ๋ธ์€ ๊ฐ€์ค‘์น˜๊ฐ€ ๋ฌถ์—ฌ ์žˆ์ง€ ์•Š์•„์„œ ์ด ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๋ชจ๋ธ๋“ค์€ `torchscript` ํ”Œ๋ž˜๊ทธ ์—†์ด ์•ˆ์ „ํ•˜๊ฒŒ ๋‚ด๋ณด๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

## ๋”๋ฏธ ์ž…๋ ฅ๊ณผ ํ‘œ์ค€ ๊ธธ์ด[[dummy-inputs-and-standard-lengths]]

๋”๋ฏธ ์ž…๋ ฅ(dummy inputs)์€ ๋ชจ๋ธ์˜ ์ˆœ์ „ํŒŒ(forward pass)์— ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. 
์ž…๋ ฅ ๊ฐ’์ด ๋ ˆ์ด์–ด๋ฅผ ํ†ตํ•ด ์ „ํŒŒ๋˜๋Š” ๋™์•ˆ, PyTorch๋Š” ๊ฐ ํ…์„œ์—์„œ ์‹คํ–‰๋œ ๋‹ค๋ฅธ ์—ฐ์‚ฐ์„ ์ถ”์ ํ•ฉ๋‹ˆ๋‹ค. 
์ด๋Ÿฌํ•œ ๊ธฐ๋ก๋œ ์—ฐ์‚ฐ์€ ๋ชจ๋ธ์˜ *์ถ”์ (trace)*์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.

์ถ”์ ์€ ์ž…๋ ฅ์˜ ์ฐจ์›์„ ๊ธฐ์ค€์œผ๋กœ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. 
๋”ฐ๋ผ์„œ ๋”๋ฏธ ์ž…๋ ฅ์˜ ์ฐจ์›์— ์ œํ•œ๋˜์–ด, ๋‹ค๋ฅธ ์‹œํ€€์Šค ๊ธธ์ด๋‚˜ ๋ฐฐ์น˜ ํฌ๊ธฐ์—์„œ๋Š” ์ž‘๋™ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. 
๋‹ค๋ฅธ ํฌ๊ธฐ๋กœ ์‹œ๋„ํ•  ๊ฒฝ์šฐ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค:

```
`The expanded size of the tensor (3) must match the existing size (7) at non-singleton dimension 2`
```
์ถ”๋ก  ์ค‘ ๋ชจ๋ธ์— ๊ณต๊ธ‰๋  ๊ฐ€์žฅ ํฐ ์ž…๋ ฅ๋งŒํผ ํฐ ๋”๋ฏธ ์ž…๋ ฅ ํฌ๊ธฐ๋กœ ๋ชจ๋ธ์„ ์ถ”์ ํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค. 
ํŒจ๋”ฉ์€ ๋ˆ„๋ฝ๋œ ๊ฐ’์„ ์ฑ„์šฐ๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 
๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋ธ์ด ๋” ํฐ ์ž…๋ ฅ ํฌ๊ธฐ๋กœ ์ถ”์ ๋˜๊ธฐ ๋•Œ๋ฌธ์—, ํ–‰๋ ฌ์˜ ์ฐจ์›์ด ์ปค์ง€๊ณ  ๊ณ„์‚ฐ๋Ÿ‰์ด ๋งŽ์•„์ง‘๋‹ˆ๋‹ค.

๋‹ค์–‘ํ•œ ์‹œํ€€์Šค ๊ธธ์ด ๋ชจ๋ธ์„ ๋‚ด๋ณด๋‚ผ ๋•Œ๋Š” ๊ฐ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์ˆ˜ํ–‰๋˜๋Š” ์ด ์—ฐ์‚ฐ ํšŸ์ˆ˜์— ์ฃผ์˜ํ•˜๊ณ  ์„ฑ๋Šฅ์„ ์ฃผ์˜ ๊นŠ๊ฒŒ ํ™•์ธํ•˜์„ธ์š”.

## Python์—์„œ TorchScript ์‚ฌ์šฉํ•˜๊ธฐ[[using-torchscript-in-python]]

์ด ์„น์…˜์—์„œ๋Š” ๋ชจ๋ธ์„ ์ €์žฅํ•˜๊ณ  ๊ฐ€์ ธ์˜ค๋Š” ๋ฐฉ๋ฒ•, ์ถ”์ ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

### ๋ชจ๋ธ ์ €์žฅํ•˜๊ธฐ[[saving-a-model]]

`BertModel`์„ TorchScript๋กœ ๋‚ด๋ณด๋‚ด๋ ค๋ฉด `BertConfig` ํด๋ž˜์Šค์—์„œ `BertModel`์„ ์ธ์Šคํ„ด์Šคํ™”ํ•œ ๋‹ค์Œ, `traced_bert.pt`๋ผ๋Š” ํŒŒ์ผ๋ช…์œผ๋กœ ๋””์Šคํฌ์— ์ €์žฅํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค.

```python
from transformers import BertModel, BertTokenizer, BertConfig
import torch

enc = BertTokenizer.from_pretrained("bert-base-uncased")

# ์ž…๋ ฅ ํ…์ŠคํŠธ ํ† ํฐํ™”ํ•˜๊ธฐ
text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
tokenized_text = enc.tokenize(text)

# ์ž…๋ ฅ ํ† ํฐ ์ค‘ ํ•˜๋‚˜๋ฅผ ๋งˆ์Šคํ‚นํ•˜๊ธฐ
masked_index = 8
tokenized_text[masked_index] = "[MASK]"
indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]

# ๋”๋ฏธ ์ž…๋ ฅ ๋งŒ๋“ค๊ธฐ
tokens_tensor = torch.tensor([indexed_tokens])
segments_tensors = torch.tensor([segments_ids])
dummy_input = [tokens_tensor, segments_tensors]

# torchscript ํ”Œ๋ž˜๊ทธ๋กœ ๋ชจ๋ธ ์ดˆ๊ธฐํ™”ํ•˜๊ธฐ
# ์ด ๋ชจ๋ธ์€ LM ํ—ค๋“œ๊ฐ€ ์—†์œผ๋ฏ€๋กœ ํ•„์š”ํ•˜์ง€ ์•Š์ง€๋งŒ, ํ”Œ๋ž˜๊ทธ๋ฅผ True๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
config = BertConfig(
    vocab_size_or_config_json_file=32000,
    hidden_size=768,
    num_hidden_layers=12,
    num_attention_heads=12,
    intermediate_size=3072,
    torchscript=True,
)

# ๋ชจ๋ธ์„ ์ธ์Šคํ„ดํŠธํ™”ํ•˜๊ธฐ
model = BertModel(config)

# ๋ชจ๋ธ์„ ํ‰๊ฐ€ ๋ชจ๋“œ๋กœ ๋‘์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
model.eval()

# ๋งŒ์•ฝ *from_pretrained*๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ์ธ์Šคํ„ด์Šคํ™”ํ•˜๋Š” ๊ฒฝ์šฐ, TorchScript ํ”Œ๋ž˜๊ทธ๋ฅผ ์‰ฝ๊ฒŒ ์„ค์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค
model = BertModel.from_pretrained("bert-base-uncased", torchscript=True)

# ์ถ”์  ์ƒ์„ฑํ•˜๊ธฐ
traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
torch.jit.save(traced_model, "traced_bert.pt")
```

### ๋ชจ๋ธ ๊ฐ€์ ธ์˜ค๊ธฐ[[loading-a-model]]

์ด์ œ ์ด์ „์— ์ €์žฅํ•œ `BertModel`, ์ฆ‰ `traced_bert.pt`๋ฅผ ๋””์Šคํฌ์—์„œ ๊ฐ€์ ธ์˜ค๊ณ , ์ด์ „์— ์ดˆ๊ธฐํ™”ํ•œ `dummy_input`์—์„œ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

```python
loaded_model = torch.jit.load("traced_bert.pt")
loaded_model.eval()

all_encoder_layers, pooled_output = loaded_model(*dummy_input)
```

### ์ถ”์ ๋œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก ํ•˜๊ธฐ[[using-a-traced-model-for-inference]]

`__call__` ์ด์ค‘ ์–ธ๋”์Šค์ฝ”์–ด(dunder) ๋ฉ”์†Œ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก ์— ์ถ”์ ๋œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์„ธ์š”:

```python
traced_model(tokens_tensor, segments_tensors)
```

## Neuron SDK๋กœ Hugging Face TorchScript ๋ชจ๋ธ์„ AWS์— ๋ฐฐํฌํ•˜๊ธฐ[[deploy-hugging-face-torchscript-models-to-aws-with-the-neuron-sdk]]

AWS๊ฐ€ ํด๋ผ์šฐ๋“œ์—์„œ ์ €๋น„์šฉ, ๊ณ ์„ฑ๋Šฅ ๋จธ์‹  ๋Ÿฌ๋‹ ์ถ”๋ก ์„ ์œ„ํ•œ [Amazon EC2 Inf1](https://aws.amazon.com/ec2/instance-types/inf1/) ์ธ์Šคํ„ด์Šค ์ œํ’ˆ๊ตฐ์„ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. 
Inf1 ์ธ์Šคํ„ด์Šค๋Š” ๋”ฅ๋Ÿฌ๋‹ ์ถ”๋ก  ์›Œํฌ๋กœ๋“œ์— ํŠนํ™”๋œ ๋งž์ถค ํ•˜๋“œ์›จ์–ด ๊ฐ€์†๊ธฐ์ธ AWS Inferentia ์นฉ์œผ๋กœ ๊ตฌ๋™๋ฉ๋‹ˆ๋‹ค. 
[AWS Neuron](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/#)์€ Inferentia๋ฅผ ์œ„ํ•œ SDK๋กœ, Inf1์— ๋ฐฐํฌํ•˜๊ธฐ ์œ„ํ•œ transformers ๋ชจ๋ธ ์ถ”์  ๋ฐ ์ตœ์ ํ™”๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. 
Neuron SDK๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค:

1. ์ฝ”๋“œ ํ•œ ์ค„๋งŒ ๋ณ€๊ฒฝํ•˜๋ฉด ํด๋ผ์šฐ๋“œ ์ถ”๋ก ๋ฅผ ์œ„ํ•ด TorchScript ๋ชจ๋ธ์„ ์ถ”์ ํ•˜๊ณ  ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์‰ฌ์šด API
2. ์ฆ‰์‹œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์„ฑ๋Šฅ ์ตœ์ ํ™”๋กœ [๋น„์šฉ ํšจ์œจ ํ–ฅ์ƒ](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/benchmark/>)
3. [PyTorch](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/bert_tutorial/tutorial_pretrained_bert.html) ๋˜๋Š” [TensorFlow](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/tensorflow/huggingface_bert/huggingface_bert.html)๋กœ ๊ตฌ์ถ•๋œ Hugging Face transformers ๋ชจ๋ธ ์ง€์›

### ์‹œ์‚ฌ์ [[implications]]

[BERT (Bidirectional Encoder Representations from Transformers)](https://huggingface.co/docs/transformers/main/model_doc/bert) ์•„ํ‚คํ…์ฒ˜ ๋˜๋Š” ๊ทธ ๋ณ€ํ˜•์ธ [distilBERT](https://huggingface.co/docs/transformers/main/model_doc/distilbert) ๋ฐ [roBERTa](https://huggingface.co/docs/transformers/main/model_doc/roberta)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ Transformers ๋ชจ๋ธ์€ ์ถ”์ถœ ๊ธฐ๋ฐ˜ ์งˆ์˜์‘๋‹ต, ์‹œํ€€์Šค ๋ถ„๋ฅ˜ ๋ฐ ํ† ํฐ ๋ถ„๋ฅ˜์™€ ๊ฐ™์€ ๋น„์ƒ์„ฑ ์ž‘์—… ์‹œ Inf1์—์„œ ์ตœ์ƒ์˜ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. 
๊ทธ๋Ÿฌ๋‚˜ ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…๋„ [AWS Neuron MarianMT ํŠœํ† ๋ฆฌ์–ผ](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/transformers-marianmt.html)์„ ๋”ฐ๋ผ Inf1์—์„œ ์‹คํ–‰๋˜๋„๋ก ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Inferentia์—์„œ ๋ฐ”๋กœ ๋ณ€ํ™˜ํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด๋Š” Neuron ๋ฌธ์„œ์˜ [Model Architecture Fit](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/models/models-inferentia.html#models-inferentia) ์„น์…˜์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

### ์ข…์†์„ฑ[[dependencies]]

AWS Neuron์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ๋ณ€ํ™˜ํ•˜๋ ค๋ฉด [Neuron SDK ํ™˜๊ฒฝ](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/neuron-frameworks/pytorch-neuron/index.html#installation-guide)์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
 ์ด๋Š” [AWS Deep Learning AMI](https://docs.aws.amazon.com/dlami/latest/devguide/tutorial-inferentia-launching.html)์— ๋ฏธ๋ฆฌ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.

### AWS Neuron์œผ๋กœ ๋ชจ๋ธ ๋ณ€ํ™˜ํ•˜๊ธฐ[[converting-a-model-for-aws-neuron]]

`BertModel`์„ ์ถ”์ ํ•˜๋ ค๋ฉด, [Python์—์„œ TorchScript ์‚ฌ์šฉํ•˜๊ธฐ](torchscript#using-torchscript-in-python)์—์„œ์™€ ๋™์ผํ•œ ์ฝ”๋“œ๋ฅผ ์‚ฌ์šฉํ•ด์„œ AWS NEURON์šฉ ๋ชจ๋ธ์„ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค. 
`torch.neuron` ํ”„๋ ˆ์ž„์›Œํฌ ์ต์Šคํ…์…˜์„ ๊ฐ€์ ธ์™€ Python API๋ฅผ ํ†ตํ•ด Neuron SDK์˜ ๊ตฌ์„ฑ ์š”์†Œ์— ์ ‘๊ทผํ•ฉ๋‹ˆ๋‹ค:

```python
from transformers import BertModel, BertTokenizer, BertConfig
import torch
import torch.neuron
```

๋‹ค์Œ ์ค„๋งŒ ์ˆ˜์ •ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค:

```diff
- torch.jit.trace(model, [tokens_tensor, segments_tensors])
+ torch.neuron.trace(model, [token_tensor, segments_tensors])
```

์ด๋กœ์จ Neuron SDK๊ฐ€ ๋ชจ๋ธ์„ ์ถ”์ ํ•˜๊ณ  Inf1 ์ธ์Šคํ„ด์Šค์— ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.

AWS Neuron SDK์˜ ๊ธฐ๋Šฅ, ๋„๊ตฌ, ์˜ˆ์ œ ํŠœํ† ๋ฆฌ์–ผ ๋ฐ ์ตœ์‹  ์—…๋ฐ์ดํŠธ์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์•Œ์•„๋ณด๋ ค๋ฉด [AWS NeuronSDK ๋ฌธ์„œ](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/index.html)๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.