kishizaki-sci
commited on
Create README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,66 @@
|
|
1 |
-
---
|
2 |
-
license:
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
base_model:
|
4 |
+
- meta-llama/Llama-3.1-405B-Instruct
|
5 |
+
language:
|
6 |
+
- ja
|
7 |
+
- en
|
8 |
+
pipeline_tag: text-generation
|
9 |
+
library_name: transformers
|
10 |
+
tags:
|
11 |
+
- llama-3
|
12 |
+
- pytorch
|
13 |
+
- llama-3.1
|
14 |
+
- autoawq
|
15 |
+
- meta
|
16 |
+
---
|
17 |
+
# kishizaki-sci/Llama-3.1-405B-Instruct-AWQ-4bit-JP-EN
|
18 |
+
|
19 |
+
## model information
|
20 |
+
[Llama-3.1-405B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct)を[AutoAWQ](https://github.com/casper-hansen/AutoAWQ)で4bit 量子化したモデル。量子化の際のキャリブレーションデータに日本語と英語を含むデータを使用。
|
21 |
+
A model of Llama-3.1-405B-Instruct quantized to 4 bits using AutoAWQ. Calibration data containing Japanese and English was used during the quantization process.
|
22 |
+
|
23 |
+
## usage
|
24 |
+
|
25 |
+
### vLLM
|
26 |
+
```python
|
27 |
+
from vllm import LLM, SamplingParams
|
28 |
+
llm = LLM(
|
29 |
+
model="kishizaki-sci/Llama-3.1-405B-Instruct-AWQ-4bit-JP-EN",
|
30 |
+
tensor_parallel_size=4,
|
31 |
+
gpu_memory_utilization=0.97,
|
32 |
+
quantization="awq"
|
33 |
+
)
|
34 |
+
tokenizer = llm.get_tokenizer()
|
35 |
+
messages = [
|
36 |
+
{"role": "system", "content": "あなたは日本語で応答するAIチャットボットです。ユーザをサポートしてください。"},
|
37 |
+
{"role": "user", "content": "plotly.graph_objectsを使って散布図を作るサンプルコードを書いてください。"},
|
38 |
+
]
|
39 |
+
prompt = tokenizer.apply_chat_template(
|
40 |
+
messages,
|
41 |
+
tokenize=False,
|
42 |
+
add_generation_prompt=True
|
43 |
+
)
|
44 |
+
sampling_params = SamplingParams(
|
45 |
+
temperature=0.6,
|
46 |
+
top_p=0.9,
|
47 |
+
max_tokens=1024
|
48 |
+
)
|
49 |
+
outputs = llm.generate(prompt, sampling_params)
|
50 |
+
print(outputs[0].outputs[0].text)
|
51 |
+
```
|
52 |
+
H100 (94GB)を4基積んだインスタンスでの実行はこちらの[notebook](https://huggingface.co/kishizaki-sci/Llama-3.1-405B-Instruct-AWQ-4bit-JP-EN/blob/main/inference_vLLM.ipynb)をご覧ください。
|
53 |
+
Please refer to this notebook for execution on an instance equipped with a four H100 (94GB).
|
54 |
+
|
55 |
+
## calibration data
|
56 |
+
以下のデータセットから512個のデータ,プロンプトを抽出。1つのデータのトークン数は最大350制限。
|
57 |
+
Extract 512 data points and prompts from the following dataset. The maximum token limit per data point is 350.
|
58 |
+
- [TFMC/imatrix-dataset-for-japanese-llm](https://huggingface.co/datasets/TFMC/imatrix-dataset-for-japanese-llm)
|
59 |
+
- [meta-math/MetaMathQA](https://huggingface.co/datasets/meta-math/MetaMathQA)
|
60 |
+
- [m-a-p/CodeFeedback-Filtered-Instruction](https://huggingface.co/datasets/m-a-p/CodeFeedback-Filtered-Instruction)
|
61 |
+
- [kunishou/databricks-dolly-15k-ja](https://huggingface.co/datasets/kunishou/databricks-dolly-15k-ja)
|
62 |
+
- その他日本語版・英語版のwikipedia記事から作成したオリジナルデータ,有害プロンプト回避のためのオリジナルデータを使用。 Original data created from Japanese and English Wikipedia articles, as well as original data for avoiding harmful prompts, is used.
|
63 |
+
|
64 |
+
## License
|
65 |
+
[MIT License](https://opensource.org/license/mit)を適用する。ただし量子化のベースモデルに適用されている[Llama 3.1 Community License Agreement](https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/LICENSE)に従ってください。
|
66 |
+
The MIT License is applied. However, obey the Llama 3.1 Community License Agreement applied to the base model of quantization.
|