File size: 7,561 Bytes
a9bd396 | 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 | <!--Copyright 2025 The LG AI Research and 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.
โ ๏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.
-->
# EXAONE 4
## ๊ฐ์
**[EXAONE 4.0](https://github.com/LG-AI-EXAONE/EXAONE-4.0)** ๋ชจ๋ธ๊ตฐ์ [EXAONE 3.5](https://github.com/LG-AI-EXAONE/EXAONE-3.5) ๋ชจ๋ธ๊ตฐ์ ๋์ ์ค์ฉ์ฑ๊ณผ [EXAONE Deep](https://github.com/LG-AI-EXAONE/EXAONE-Deep) ๋ชจ๋ธ๊ตฐ์ ํฅ์๋ ์ฌ๊ณ ์ถ๋ก ๋ฅ๋ ฅ์ ๊ฐ๊ฐ Non-reasoning mode์ Reasoning mode๋ก ํตํฉํ ์์ฐ์ด ๋ชจ๋ธ(language model)์
๋๋ค. ์์ด์ ํฑ(agentic) AI ์๋์ ๋ฐ๋ง์ถฐ EXAONE 4.0์ ์์ด์ ํฑ ๋๊ตฌ ์ฌ์ฉ ๋ฅ๋ ฅ๊ณผ ๊ฐ์ ํต์ฌ ๊ธฐ๋ฅ์ ํตํฉํ๊ณ , ๊ธฐ์กด์ ๋ค๊ตญ์ด ๋ฅ๋ ฅ์ ์์ด, ํ๊ตญ์ด์ ๋๋ถ์ด ์คํ์ธ์ด๊น์ง ํ์ฅํ์ต๋๋ค.
EXAONE 4.0 ๋ชจ๋ธ๊ตฐ์ ๋ ๊ฐ์ ๋ชจ๋ธ: ๋์ ์ฑ๋ฅ์ ์ํด ์ต์ ํ๋ 32B ์คํ ๋ชจ๋ธ, ๊ทธ๋ฆฌ๊ณ ์จ-๋๋ฐ์ด์ค ํ์ฉ์ ์ํด ๋์์ธ๋ 1.2B ์ํ ๋ชจ๋ธ์ผ๋ก ๊ตฌ์ฑ๋์ด ์์ต๋๋ค.
EXAONE 4.0์ ๋ชจ๋ธ ๊ตฌ์กฐ๋ ์ด์ EXAONE ๋ชจ๋ธ๋ค๊ณผ ๋ค๋ฅธ ์ํคํ
์ฒ ๋์์ธ์ ์ฑํํ์ต๋๋ค.
1. **Hybrid Attention**: 32B ๋ชจ๋ธ์ *Local attention (sliding window attention)*๊ณผ *Global attention (full attention)*์ 3:1 ๋น์จ๋ก ์ฐ๊ฒฐํ hybrid attention ๊ตฌ์กฐ๋ฅผ ์ฑํํ์ต๋๋ค. ๋ํ ์ ์ฒด ๋ฌธ๋งฅ์ ๋ ์ ์ดํดํ ์ ์๋๋ก global attention์์ RoPE๋ฅผ ์ฌ์ฉํ์ง ์์์ต๋๋ค.
2. **QK-Reorder-Norm**: ๋ ๋์ downstream tasks ์ฑ๋ฅ์ ์ํด ์ฐ์ฐ๋์ ์ฆ๊ฐ๋ฅผ ๊ฐ์ํ๋ฉฐ ์ ํต์ ์ผ๋ก ์ฌ์ฉ๋๊ณ ์๋ Pre-LN ๋ฐฉ์์ ๋ณ๊ฒฝํ์ต๋๋ค. LayerNorm์ ์์น๋ฅผ attention๊ณผ MLP์ ์ถ๋ ฅ์ ์ ์ฉ๋๋๋ก ์ฌ๋ฐฐ์นํ๊ณ , Q์ K projection ์งํ์๋ RMS normalization์ ์ถ๊ฐํ์ต๋๋ค.
๋ ์์ธํ ์ ๋ณด๋ [๊ธฐ์ ๋ณด๊ณ ์](https://huggingface.co/papers/2507.11407), [HuggingFace ๋
ผ๋ฌธ](https://huggingface.co/papers/2507.11407), [๋ธ๋ก๊ทธ](https://www.lgresearch.ai/blog/view?seq=576), [๊ณต์ GitHub](https://github.com/LG-AI-EXAONE/EXAONE-4.0) ํ์ด์ง๋ฅผ ์ฐธ๊ณ ํด์ฃผ์๊ธธ ๋ฐ๋๋๋ค.
๊ณต๊ฐ๋ ๋ชจ๋ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํธ๋ [HuggingFace ์ฝ๋ ์
](https://huggingface.co/collections/LGAI-EXAONE/exaone-40-686b2e0069800c835ed48375)์์ ํ์ธํ ์ ์์ต๋๋ค.
## ๋ชจ๋ธ ์ธ๋ถ ์ ๋ณด
| Model Configuration | 32B | 1.2B |
|:-------------------|:-----:|:------:|
| d_model | 5,120 | 2,048 |
| Number of layers | 64 | 30 |
| Normalization | QK-Reorder-LN | QK-Reorder-LN |
| Non-linearity | SwiGLU | SwiGLU |
| Feedforward dimension | 27,392 | 4,096 |
| Attention type | Hybrid (3:1 Local-Global) | Global |
| Head type | GQA | GQA |
| Number of heads | 40 | 32 |
| Number of KV heads | 8 | 8 |
| Head size | 128 | 64 |
| Max sequence length | 131,072 | 65,536 |
| RoPE theta | 1,000,000 | 1,000,000 |
| Tokenizer | BBPE | BBPE |
| Vocab size | 102,400 | 102,400 |
| Tied word embedding | False | True |
| Knowledge cut-off | Nov. 2024 | Nov. 2024 |
## ์ฌ์ฉ ํ
### Non-reasoning mode
์ผ๋ฐ์ ์ธ ๋ํ์ ๊ฒฝ์ฐ ์๋ ์์ ์ ๊ฐ์ด EXAONE 4.0์ ์ฌ์ฉํ ์ ์์ต๋๋ค.
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "LGAI-EXAONE/EXAONE-4.0-32B"
model = AutoModelForCausalLM.from_pretrained(
model_name,
dtype="bfloat16",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
# ์ํ๋ ์
๋ ฅ์ ์ ํํ์ธ์
prompt = "Explain how wonderful you are"
prompt = "Explica lo increรญble que eres"
prompt = "๋๊ฐ ์ผ๋ง๋ ๋๋จํ์ง ์ค๋ช
ํด ๋ด"
messages = [
{"role": "user", "content": prompt}
]
input_ids = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt"
)
output = model.generate(
input_ids.to(model.device),
max_new_tokens=128,
do_sample=False,
)
print(tokenizer.decode(output[0]))
```
### Reasoning mode
The EXAONE 4.0 models have reasoning capabilities for handling complex problems. You can activate reasoning mode by using the `enable_thinking=True` argument with the tokenizer, which opens a reasoning block that starts with `<think>` tag without closing it.
EXAONE 4.0 ๋ชจ๋ธ๊ตฐ์ ๋ณต์กํ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ธฐ ์ํ ์ฌ๊ณ ์ถ๋ก ๋ฅ๋ ฅ์ ๊ฐ์ถ๊ณ ์์ต๋๋ค. ํ ํฌ๋์ด์ ์์ `enable_thinking=True` ์ธ์๋ฅผ ์ฌ์ฉํด์ reasoning mode๋ก ๋ชจ๋ธ์ ์ฌ์ฉํ ์ ์์ต๋๋ค. ์ด ๊ฒฝ์ฐ `<think>` ํ ํฐ์ผ๋ก ์ถ๋ก ๋ธ๋ก์ ์ฐ ๋ค, ๋ซ์ง ์๊ณ ์ถ๋ก ์ ์์ํฉ๋๋ค.
```python
messages = [
{"role": "user", "content": "Which one is bigger, 3.12 vs 3.9?"}
]
input_ids = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt",
enable_thinking=True,
)
output = model.generate(
input_ids.to(model.device),
max_new_tokens=128,
do_sample=True,
temperature=0.6,
top_p=0.95
)
print(tokenizer.decode(output[0]))
```
> [!IMPORTANT]
> ๋ชจ๋ธ์ reasoning mode๋ก ์ฌ์ฉํ ๊ฒฝ์ฐ, ์์ฑ๋๋ ๋ต๋ณ์ด sampling parameters์ ๊ต์ฅํ ๋ฏผ๊ฐํฉ๋๋ค. ๋ฐ๋ผ์ ๋ ๋์ ์์ฑ ํ์ง์ ์ํด ๊ณต์ [Usage Guideline](https://github.com/LG-AI-EXAONE/EXAONE-4.0#usage-guideline)๋ฅผ ์ฐธ์กฐํด ์ฃผ์๊ธธ ๋ฐ๋๋๋ค.
### Agentic tool use
EXAONE 4.0 ๋ชจ๋ธ์ ๋๊ตฌ ์ฌ์ฉ ๋ฅ๋ ฅ์ ๊ฐ์ถ ๋๋ถ์ Agent๋ก ์ฌ์ฉํ ์ ์์ต๋๋ค. ์ด๋ฅผ ์ํด์๋ ์๋ ์์ ์ ๊ฐ์ด ๋๊ตฌ ๋ช
์ธ๋ฅผ ๋ชจ๋ธ์๊ฒ ์ ๊ณตํด ์ฃผ์ด์ผ ํฉ๋๋ค.
```python
import random
def roll_dice(max_num: int):
return random.randint(1, max_num)
tools = [
{
"type": "function",
"function": {
"name": "roll_dice",
"description": "Roll a dice with the number 1 to N. User can select the number N.",
"parameters": {
"type": "object",
"required": ["max_num"],
"properties": {
"max_num": {
"type": "int",
"description": "Max number of the dice"
}
}
}
}
}
]
messages = [
{"role": "user", "content": "Roll D6 dice twice!"}
]
input_ids = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt",
tools=tools,
)
output = model.generate(
input_ids.to(model.device),
max_new_tokens=1024,
do_sample=True,
temperature=0.6,
top_p=0.95,
)
print(tokenizer.decode(output[0]))
```
## Exaone4Config
[[autodoc]] Exaone4Config
## Exaone4Model
[[autodoc]] Exaone4Model
- forward
## Exaone4ForCausalLM
[[autodoc]] Exaone4ForCausalLM
- forward
## Exaone4ForSequenceClassification
[[autodoc]] Exaone4ForSequenceClassification
- forward
## Exaone4ForTokenClassification
[[autodoc]] Exaone4ForTokenClassification
- forward
## Exaone4ForQuestionAnswering
[[autodoc]] Exaone4ForQuestionAnswering
- forward |