Model Card for Model ID

This model is an AWS Neuron compiled version of the Korean fine-tuned model MLP-KTLim/llama-3-Korean-Bllossom-8B, available at https://huggingface.co/MLP-KTLim/llama-3-Korean-Bllossom-8B. It is intended for deployment on Amazon EC2 Inferentia2 and Amazon SageMaker. For detailed information about the model and its license, please refer to the original MLP-KTLim/llama-3-Korean-Bllossom-8B model page

Model Details

This model is compiled with neuronx-cc version, 2.13.66.0. It can be deployed with neuronnx-tgi docker, ghcr.io/huggingface/neuronx-tgi:0.0.23

How to Get Started with the Model

You can pull the docker image ghcr.io/huggingface/neuronx-tgi:0.0.23, downlaod this model and run the command like this example:

docker run \
-p 8080:80 \
-v $(pwd)/data:/data \
--privileged \
ghcr.io/huggingface/neuronx-tgi:latest \
--model-id /data/AWS-Neuron-llama-3-Korean-Bllossom-8B

After deployment, you can inference like this

curl 127.0.0.1:8080/generate \
-X POST \
-d '{"inputs":"딥러닝이 뭐야?","parameters":{"max_new_tokens":512}}' \
-H 'Content-Type: application/json'

curl localhost:8080/v1/chat/completions \
    -X POST \
    -d '{
"model": "tgi",
"messages": [
    {
    "role": "system",
    "content": "당신은 인공지능 전문가 입니다."
    },
    {
    "role": "user",
    "content": "딥러닝이 무엇입니까?"
    }
],
"stream": false,
"max_tokens": 512
}' \
    -H 'Content-Type: application/json'

This model can be deployed to Amazon SageMaker Endtpoint with this guide, Deploy this model to SageMaker Endpoint

In order to do neuron-compilation and depoly in detail , you can refer to Amazon EC2 Inferentia2 기반 위에 한국어 파인 튜닝 모델을 서빙하기

Hardware

At a minimum hardware, you can use Amazon EC2 inf2.xlarge and more powerful family such as inf2.8xlarge, inf2.24xlarge and inf2.48xlarge. The detailed information is Amazon EC2 Inf2 Instances

Model Card Contact

Gonsoo Moon, gonsoomoon@gmail.com