File size: 13,448 Bytes
ee6e328
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
<!--Copyright 2023 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.

⚠️ 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.

-->


# λŒ€κ·œλͺ¨ μ–Έμ–΄ λͺ¨λΈλ‘œ μƒμ„±ν•˜κΈ° [[generation-with-llms]]

[[open-in-colab]]

LLM λ˜λŠ” λŒ€κ·œλͺ¨ μ–Έμ–΄ λͺ¨λΈμ€ ν…μŠ€νŠΈ μƒμ„±μ˜ 핡심 ꡬ성 μš”μ†Œμž…λ‹ˆλ‹€. κ°„λ‹¨νžˆ λ§ν•˜λ©΄, 주어진 μž…λ ₯ ν…μŠ€νŠΈμ— λŒ€ν•œ λ‹€μŒ 단어(μ •ν™•ν•˜κ²ŒλŠ” 토큰)λ₯Ό μ˜ˆμΈ‘ν•˜κΈ° μœ„ν•΄ ν›ˆλ ¨λœ λŒ€κ·œλͺ¨ 사전 ν›ˆλ ¨ λ³€ν™˜κΈ° λͺ¨λΈλ‘œ κ΅¬μ„±λ©λ‹ˆλ‹€. 토큰을 ν•œ λ²ˆμ— ν•˜λ‚˜μ”© μ˜ˆμΈ‘ν•˜κΈ° λ•Œλ¬Έμ— μƒˆλ‘œμš΄ λ¬Έμž₯을 μƒμ„±ν•˜λ €λ©΄ λͺ¨λΈμ„ ν˜ΈμΆœν•˜λŠ” 것 외에 더 λ³΅μž‘ν•œ μž‘μ—…μ„ μˆ˜ν–‰ν•΄μ•Ό ν•©λ‹ˆλ‹€. 즉, μžκΈ°νšŒκ·€ 생성을 μˆ˜ν–‰ν•΄μ•Ό ν•©λ‹ˆλ‹€.

μžκΈ°νšŒκ·€ 생성은 λͺ‡ 개의 초기 μž…λ ₯값을 μ œκ³΅ν•œ ν›„, κ·Έ 좜λ ₯을 λ‹€μ‹œ λͺ¨λΈμ— μž…λ ₯으둜 μ‚¬μš©ν•˜μ—¬ 반볡적으둜 ν˜ΈμΆœν•˜λŠ” μΆ”λ‘  κ³Όμ •μž…λ‹ˆλ‹€. πŸ€— Transformersμ—μ„œλŠ” [`~generation.GenerationMixin.generate`] λ©”μ†Œλ“œκ°€ 이 역할을 ν•˜λ©°, μ΄λŠ” 생성 κΈ°λŠ₯을 가진 λͺ¨λ“  λͺ¨λΈμ—μ„œ μ‚¬μš© κ°€λŠ₯ν•©λ‹ˆλ‹€.

이 νŠœν† λ¦¬μ–Όμ—μ„œλŠ” λ‹€μŒ λ‚΄μš©μ„ λ‹€λ£¨κ²Œ λ©λ‹ˆλ‹€:

* LLM으둜 ν…μŠ€νŠΈ 생성
* 일반적으둜 λ°œμƒν•˜λŠ” 문제 ν•΄κ²°
* LLM을 μ΅œλŒ€ν•œ ν™œμš©ν•˜κΈ° μœ„ν•œ λ‹€μŒ 단계

μ‹œμž‘ν•˜κΈ° 전에 ν•„μš”ν•œ λͺ¨λ“  λΌμ΄λΈŒλŸ¬λ¦¬κ°€ μ„€μΉ˜λ˜μ–΄ μžˆλŠ”μ§€ ν™•μΈν•˜μ„Έμš”:

```bash
pip install transformers bitsandbytes>=0.39.0 -q
```


## ν…μŠ€νŠΈ 생성 [[generate-text]]

[인과적 μ–Έμ–΄ λͺ¨λΈλ§(causal language modeling)](tasks/language_modeling)을 λͺ©μ μœΌλ‘œ ν•™μŠ΅λœ μ–Έμ–΄ λͺ¨λΈμ€ 일련의 ν…μŠ€νŠΈ 토큰을 μž…λ ₯으둜 μ‚¬μš©ν•˜κ³ , κ·Έ 결과둜 λ‹€μŒ 토큰이 λ‚˜μ˜¬ ν™•λ₯  뢄포λ₯Ό μ œκ³΅ν•©λ‹ˆλ‹€.

<!-- [GIF 1 -- FWD PASS] -->
<figure class="image table text-center m-0 w-full">
    <video
        style="max-width: 90%; margin: auto;"
        autoplay loop muted playsinline
        src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/assisted-generation/gif_1_1080p.mov"
    ></video>
    <figcaption>"LLM의 μ „λ°© 패슀"</figcaption>
</figure>

LLMκ³Ό μžκΈ°νšŒκ·€ 생성을 ν•¨κ»˜ μ‚¬μš©ν•  λ•Œ 핡심적인 뢀뢄은 이 ν™•λ₯  λΆ„ν¬λ‘œλΆ€ν„° λ‹€μŒ 토큰을 μ–΄λ–»κ²Œ κ³ λ₯Ό κ²ƒμΈμ§€μž…λ‹ˆλ‹€. λ‹€μŒ 반볡 과정에 μ‚¬μš©λ  토큰을 κ²°μ •ν•˜λŠ” ν•œ, μ–΄λ– ν•œ 방법도 κ°€λŠ₯ν•©λ‹ˆλ‹€. ν™•λ₯  λΆ„ν¬μ—μ„œ κ°€μž₯ κ°€λŠ₯성이 높은 토큰을 μ„ νƒν•˜λŠ” κ²ƒμ²˜λŸΌ 간단할 μˆ˜λ„ 있고, κ²°κ³Ό λΆ„ν¬μ—μ„œ μƒ˜ν”Œλ§ν•˜κΈ° 전에 μˆ˜μ‹­ 가지 λ³€ν™˜μ„ μ μš©ν•˜λŠ” κ²ƒμ²˜λŸΌ λ³΅μž‘ν•  μˆ˜λ„ μžˆμŠ΅λ‹ˆλ‹€.

<!-- [GIF 2 -- TEXT GENERATION] -->
<figure class="image table text-center m-0 w-full">
    <video
        style="max-width: 90%; margin: auto;"
        autoplay loop muted playsinline
        src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/assisted-generation/gif_2_1080p.mov"
    ></video>
    <figcaption>"μžκΈ°νšŒκ·€ 생성은 ν™•λ₯  λΆ„ν¬μ—μ„œ λ‹€μŒ 토큰을 반볡적으둜 μ„ νƒν•˜μ—¬ ν…μŠ€νŠΈλ₯Ό μƒμ„±ν•©λ‹ˆλ‹€."</figcaption>
</figure>

μœ„μ—μ„œ μ„€λͺ…ν•œ 과정은 μ–΄λ–€ μ’…λ£Œ 쑰건이 좩쑱될 λ•ŒκΉŒμ§€ 반볡적으둜 μˆ˜ν–‰λ©λ‹ˆλ‹€. λͺ¨λΈμ΄ μ‹œν€€μŠ€μ˜ 끝(EOS 토큰)을 좜λ ₯ν•  λ•ŒκΉŒμ§€λ₯Ό μ’…λ£Œ 쑰건으둜 ν•˜λŠ” 것이 μ΄μƒμ μž…λ‹ˆλ‹€. 그렇지 μ•Šμ€ κ²½μš°μ—λŠ” 미리 μ •μ˜λœ μ΅œλŒ€ 길이에 λ„λ‹¬ν–ˆμ„ λ•Œ 생성이 μ€‘λ‹¨λ©λ‹ˆλ‹€.

λͺ¨λΈμ΄ μ˜ˆμƒλŒ€λ‘œ λ™μž‘ν•˜κΈ° μœ„ν•΄μ„  토큰 선택 단계와 정지 쑰건을 μ˜¬λ°”λ₯΄κ²Œ μ„€μ •ν•˜λŠ” 것이 μ€‘μš”ν•©λ‹ˆλ‹€. μ΄λŸ¬ν•œ 이유둜, 각 λͺ¨λΈμ—λŠ” κΈ°λ³Έ 생성 섀정이 잘 μ •μ˜λœ [`~generation.GenerationConfig`] 파일이 ν•¨κ»˜ μ œκ³΅λ©λ‹ˆλ‹€.

μ½”λ“œλ₯Ό ν™•μΈν•΄λ΄…μ‹œλ‹€!

<Tip>

κΈ°λ³Έ LLM μ‚¬μš©μ— 관심이 μžˆλ‹€λ©΄, 우리의 [`Pipeline`](pipeline_tutorial) μΈν„°νŽ˜μ΄μŠ€λ‘œ μ‹œμž‘ν•˜λŠ” 것을 μΆ”μ²œν•©λ‹ˆλ‹€. κ·ΈλŸ¬λ‚˜ LLM은 μ–‘μžν™”λ‚˜ 토큰 선택 λ‹¨κ³„μ—μ„œμ˜ λ―Έμ„Έν•œ μ œμ–΄μ™€ 같은 κ³ κΈ‰ κΈ°λŠ₯듀을 μ’…μ’… ν•„μš”λ‘œ ν•©λ‹ˆλ‹€. μ΄λŸ¬ν•œ μž‘μ—…μ€ [`~generation.GenerationMixin.generate`]λ₯Ό 톡해 κ°€μž₯ 잘 μˆ˜ν–‰λ  수 μžˆμŠ΅λ‹ˆλ‹€. LLM을 μ΄μš©ν•œ μžκΈ°νšŒκ·€ 생성은 μžμ›μ„ 많이 μ†Œλͺ¨ν•˜λ―€λ‘œ, μ μ ˆν•œ μ²˜λ¦¬λŸ‰μ„ μœ„ν•΄ GPUμ—μ„œ μ‹€ν–‰λ˜μ–΄μ•Ό ν•©λ‹ˆλ‹€.

</Tip>

<!-- TODO: update example to llama 2 (or a newer popular baseline) when it becomes ungated -->
λ¨Όμ €, λͺ¨λΈμ„ λΆˆλŸ¬μ˜€μ„Έμš”.

```py
>>> from transformers import AutoModelForCausalLM

>>> model = AutoModelForCausalLM.from_pretrained(
...     "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
... )
```

`from_pretrained` ν•¨μˆ˜λ₯Ό ν˜ΈμΆœν•  λ•Œ 2개의 ν”Œλž˜κ·Έλ₯Ό μ£Όλͺ©ν•˜μ„Έμš”:

- `device_map`은 λͺ¨λΈμ΄ GPU둜 μ΄λ™λ˜λ„λ‘ ν•©λ‹ˆλ‹€.
- `load_in_4bit`λŠ” λ¦¬μ†ŒμŠ€ μš”κ΅¬ 사항을 크게 쀄이기 μœ„ν•΄ [4λΉ„νŠΈ 동적 μ–‘μžν™”](main_classes/quantization)λ₯Ό μ μš©ν•©λ‹ˆλ‹€.

이 외에도 λͺ¨λΈμ„ μ΄ˆκΈ°ν™”ν•˜λŠ” λ‹€μ–‘ν•œ 방법이 μžˆμ§€λ§Œ, LLM을 처음 μ‹œμž‘ν•  λ•Œ 이 섀정을 μΆ”μ²œν•©λ‹ˆλ‹€.

μ΄μ–΄μ„œ ν…μŠ€νŠΈ μž…λ ₯을 [ν† ν¬λ‚˜μ΄μ €](tokenizer_summary)으둜 μ „μ²˜λ¦¬ν•˜μ„Έμš”.

```py
>>> from transformers import AutoTokenizer

>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b")
>>> model_inputs = tokenizer(["A list of colors: red, blue"], return_tensors="pt").to("cuda")
```

`model_inputs` λ³€μˆ˜μ—λŠ” ν† ν°ν™”λœ ν…μŠ€νŠΈ μž…λ ₯κ³Ό ν•¨κ»˜ μ–΄ν…μ…˜ λ§ˆμŠ€ν¬κ°€ λ“€μ–΄ μžˆμŠ΅λ‹ˆλ‹€. [`~generation.GenerationMixin.generate`]λŠ” μ–΄ν…μ…˜ λ§ˆμŠ€ν¬κ°€ μ œκ³΅λ˜μ§€ μ•Šμ•˜μ„ κ²½μš°μ—λ„ 이λ₯Ό μΆ”λ‘ ν•˜λ €κ³  λ…Έλ ₯ν•˜μ§€λ§Œ, μ΅œμƒμ˜ μ„±λŠ₯을 μœ„ν•΄μ„œλŠ” κ°€λŠ₯ν•˜λ©΄ μ–΄ν…μ…˜ 마슀크λ₯Ό μ „λ‹¬ν•˜λŠ” 것을 ꢌμž₯ν•©λ‹ˆλ‹€. 

λ§ˆμ§€λ§‰μœΌλ‘œ [`~generation.GenerationMixin.generate`] λ©”μ†Œλ“œλ₯Ό ν˜ΈμΆœν•΄ μƒμ„±λœ 토큰을 얻은 ν›„, 이λ₯Ό 좜λ ₯ν•˜κΈ° 전에 ν…μŠ€νŠΈ ν˜•νƒœλ‘œ λ³€ν™˜ν•˜μ„Έμš”.

```py
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'A list of colors: red, blue, green, yellow, black, white, and brown'
```

이게 μ „λΆ€μž…λ‹ˆλ‹€! λͺ‡ μ€„μ˜ μ½”λ“œλ§ŒμœΌλ‘œ LLM의 λŠ₯λ ₯을 ν™œμš©ν•  수 있게 λ˜μ—ˆμŠ΅λ‹ˆλ‹€.


## 일반적으둜 λ°œμƒν•˜λŠ” 문제 [[common-pitfalls]]

[생성 μ „λž΅](generation_strategies)이 많고, 기본값이 항상 μ‚¬μš© 사둀에 μ ν•©ν•˜μ§€ μ•Šμ„ 수 μžˆμŠ΅λ‹ˆλ‹€. 좜λ ₯이 μ˜ˆμƒκ³Ό λ‹€λ₯Ό λ•Œ ν”νžˆ λ°œμƒν•˜λŠ” λ¬Έμ œμ™€ 이λ₯Ό ν•΄κ²°ν•˜λŠ” 방법에 λŒ€ν•œ λͺ©λ‘μ„ λ§Œλ“€μ—ˆμŠ΅λ‹ˆλ‹€.

```py
>>> from transformers import AutoModelForCausalLM, AutoTokenizer

>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b")
>>> tokenizer.pad_token = tokenizer.eos_token  # Llama has no pad token by default
>>> model = AutoModelForCausalLM.from_pretrained(
...     "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
... )
```

### μƒμ„±λœ 좜λ ₯이 λ„ˆλ¬΄ μ§§κ±°λ‚˜ κΈΈλ‹€ [[generated-output-is-too-shortlong]]

[`~generation.GenerationConfig`] νŒŒμΌμ—μ„œ λ³„λ„λ‘œ μ§€μ •ν•˜μ§€ μ•ŠμœΌλ©΄, `generate`λŠ” 기본적으둜 μ΅œλŒ€ 20개의 토큰을 λ°˜ν™˜ν•©λ‹ˆλ‹€. `generate` ν˜ΈμΆœμ—μ„œ `max_new_tokens`을 μˆ˜λ™μœΌλ‘œ μ„€μ •ν•˜μ—¬ λ°˜ν™˜ν•  수 μžˆλŠ” μƒˆ ν† ν°μ˜ μ΅œλŒ€ 수λ₯Ό μ„€μ •ν•˜λŠ” 것이 μ’‹μŠ΅λ‹ˆλ‹€. LLM(μ •ν™•ν•˜κ²ŒλŠ” [디코더 μ „μš© λͺ¨λΈ](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt))은 μž…λ ₯ ν”„λ‘¬ν”„νŠΈλ„ 좜λ ₯의 μΌλΆ€λ‘œ λ°˜ν™˜ν•©λ‹ˆλ‹€.


```py
>>> model_inputs = tokenizer(["A sequence of numbers: 1, 2"], return_tensors="pt").to("cuda")

>>> # By default, the output will contain up to 20 tokens
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'A sequence of numbers: 1, 2, 3, 4, 5'

>>> # Setting `max_new_tokens` allows you to control the maximum length
>>> generated_ids = model.generate(**model_inputs, max_new_tokens=50)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'A sequence of numbers: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,'
```

### 잘λͺ»λœ 생성 λͺ¨λ“œ [[incorrect-generation-mode]]

기본적으둜 [`~generation.GenerationConfig`] νŒŒμΌμ—μ„œ λ³„λ„λ‘œ μ§€μ •ν•˜μ§€ μ•ŠμœΌλ©΄, `generate`λŠ” 각 λ°˜λ³΅μ—μ„œ κ°€μž₯ ν™•λ₯ μ΄ 높은 토큰을 μ„ νƒν•©λ‹ˆλ‹€(그리디 λ””μ½”λ”©). ν•˜λ €λŠ” μž‘μ—…μ— 따라 이 방법은 λ°”λžŒμ§ν•˜μ§€ μ•Šμ„ 수 μžˆμŠ΅λ‹ˆλ‹€. 예λ₯Ό λ“€μ–΄, μ±—λ΄‡μ΄λ‚˜ 에세이 μž‘μ„±κ³Ό 같은 창의적인 μž‘μ—…μ€ μƒ˜ν”Œλ§μ΄ 적합할 수 μžˆμŠ΅λ‹ˆλ‹€. 반면, μ˜€λ””μ˜€λ₯Ό ν…μŠ€νŠΈλ‘œ λ³€ν™˜ν•˜κ±°λ‚˜ λ²ˆμ—­κ³Ό 같은 μž…λ ₯ 기반 μž‘μ—…μ€ 그리디 디코딩이 더 적합할 수 μžˆμŠ΅λ‹ˆλ‹€. `do_sample=True`둜 μƒ˜ν”Œλ§μ„ ν™œμ„±ν™”ν•  수 있으며, 이 μ£Όμ œμ— λŒ€ν•œ μžμ„Έν•œ λ‚΄μš©μ€ 이 [λΈ”λ‘œκ·Έ 포슀트](https://huggingface.co/blog/how-to-generate)μ—μ„œ λ³Ό 수 μžˆμŠ΅λ‹ˆλ‹€.

```py
>>> # Set seed or reproducibility -- you don't need this unless you want full reproducibility
>>> from transformers import set_seed
>>> set_seed(0)

>>> model_inputs = tokenizer(["I am a cat."], return_tensors="pt").to("cuda")

>>> # LLM + greedy decoding = repetitive, boring output
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'I am a cat. I am a cat. I am a cat. I am a cat'

>>> # With sampling, the output becomes more creative!
>>> generated_ids = model.generate(**model_inputs, do_sample=True)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'I am a cat.\nI just need to be. I am always.\nEvery time'
```

### 잘λͺ»λœ νŒ¨λ”© [[wrong-padding-side]]

LLM은 [디코더 μ „μš©](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt) ꡬ쑰λ₯Ό 가지고 μžˆμ–΄, μž…λ ₯ ν”„λ‘¬ν”„νŠΈμ— λŒ€ν•΄ μ§€μ†μ μœΌλ‘œ 반볡 처리λ₯Ό ν•©λ‹ˆλ‹€. μž…λ ₯ λ°μ΄ν„°μ˜ 길이가 λ‹€λ₯΄λ©΄ νŒ¨λ”© μž‘μ—…μ΄ ν•„μš”ν•©λ‹ˆλ‹€. LLM은 νŒ¨λ”© ν† ν°μ—μ„œ μž‘λ™μ„ 이어가도둝 μ„€κ³„λ˜μ§€ μ•Šμ•˜κΈ° λ•Œλ¬Έμ—, μž…λ ₯ μ™Όμͺ½μ— νŒ¨λ”©μ΄ μΆ”κ°€ λ˜μ–΄μ•Ό ν•©λ‹ˆλ‹€. 그리고 μ–΄ν…μ…˜ λ§ˆμŠ€ν¬λ„ κΌ­ `generate` ν•¨μˆ˜μ— μ „λ‹¬λ˜μ–΄μ•Ό ν•©λ‹ˆλ‹€!

```py
>>> # The tokenizer initialized above has right-padding active by default: the 1st sequence,
>>> # which is shorter, has padding on the right side. Generation fails.
>>> model_inputs = tokenizer(
...     ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
... ).to("cuda")
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids[0], skip_special_tokens=True)[0]
''

>>> # With left-padding, it works as expected!
>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b", padding_side="left")
>>> tokenizer.pad_token = tokenizer.eos_token  # Llama has no pad token by default
>>> model_inputs = tokenizer(
...     ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
... ).to("cuda")
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'1, 2, 3, 4, 5, 6,'
```

<!-- TODO: when the prompting guide is ready, mention the importance of setting the right prompt in this section -->

## μΆ”κ°€ 자료 [[further-resources]]

μžκΈ°νšŒκ·€ 생성 ν”„λ‘œμ„ΈμŠ€λŠ” μƒλŒ€μ μœΌλ‘œ λ‹¨μˆœν•œ νŽΈμ΄μ§€λ§Œ, LLM을 μ΅œλŒ€ν•œ ν™œμš©ν•˜λ €λ©΄ μ—¬λŸ¬ 가지 μš”μ†Œλ₯Ό κ³ λ €ν•΄μ•Ό ν•˜λ―€λ‘œ 쉽지 μ•Šμ„ 수 μžˆμŠ΅λ‹ˆλ‹€. LLM에 λŒ€ν•œ 더 κΉŠμ€ 이해와 ν™œμš©μ„ μœ„ν•œ λ‹€μŒ λ‹¨κ³„λŠ” μ•„λž˜μ™€ κ°™μŠ΅λ‹ˆλ‹€:

<!-- TODO: complete with new guides -->
### κ³ κΈ‰ 생성 μ‚¬μš© [[advanced-generate-usage]]

1. [κ°€μ΄λ“œ](generation_strategies)λŠ” λ‹€μ–‘ν•œ 생성 방법을 μ œμ–΄ν•˜λŠ” 방법, 생성 μ„€μ • νŒŒμΌμ„ μ„€μ •ν•˜λŠ” 방법, 좜λ ₯을 μŠ€νŠΈλ¦¬λ°ν•˜λŠ” 방법에 λŒ€ν•΄ μ„€λͺ…ν•©λ‹ˆλ‹€.
2. [`~generation.GenerationConfig`]와 [`~generation.GenerationMixin.generate`], [generate-related classes](internal/generation_utils)λ₯Ό μ°Έμ‘°ν•΄λ³΄μ„Έμš”.

### LLM λ¦¬λ”λ³΄λ“œ [[llm-leaderboards]]

1. [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)λŠ” μ˜€ν”ˆ μ†ŒμŠ€ λͺ¨λΈμ˜ ν’ˆμ§ˆμ— 쀑점을 λ‘‘λ‹ˆλ‹€.
2. [Open LLM-Perf Leaderboard](https://huggingface.co/spaces/optimum/llm-perf-leaderboard)λŠ” LLM μ²˜λ¦¬λŸ‰μ— 쀑점을 λ‘‘λ‹ˆλ‹€.

### 지연 μ‹œκ°„ 및 μ²˜λ¦¬λŸ‰ [[latency-and-throughput]]

1. λ©”λͺ¨λ¦¬ μš”κ΅¬ 사항을 쀄이렀면, 동적 μ–‘μžν™”μ— λŒ€ν•œ [κ°€μ΄λ“œ](main_classes/quantization)λ₯Ό μ°Έμ‘°ν•˜μ„Έμš”.

### κ΄€λ ¨ 라이브러리 [[related-libraries]]

1. [`text-generation-inference`](https://github.com/huggingface/text-generation-inference)λŠ” LLM을 μœ„ν•œ μ‹€μ œ 운영 ν™˜κ²½μ— μ ν•©ν•œ μ„œλ²„μž…λ‹ˆλ‹€.
2. [`optimum`](https://github.com/huggingface/optimum)은 νŠΉμ • ν•˜λ“œμ›¨μ–΄ μž₯μΉ˜μ—μ„œ LLM을 μ΅œμ ν™”ν•˜κΈ° μœ„ν•΄ πŸ€— Transformersλ₯Ό ν™•μž₯ν•œ κ²ƒμž…λ‹ˆλ‹€.