vllm¶
Helm chart for deploying vLLM Inference service.
Refer to Deploy with Helm Charts for global guides.
Installing the Chart¶
To install the chart, run the following:
Note that you cannot use vllm as the service release name due to environment variables conflict.
cd GenAIInfra/helm-charts/common
export MODELDIR=/mnt/opea-models
export MODELNAME="Intel/neural-chat-7b-v3-3"
export HFTOKEN="insert-your-huggingface-token-here"
helm install myvllm vllm --set global.modelUseHostPath=${MODELDIR} --set LLM_MODEL_ID=${MODELNAME} --set global.HUGGINGFACEHUB_API_TOKEN=${HFTOKEN}
# To deploy on Gaudi enabled kubernetes cluster
# helm install myvllm vllm --set global.modelUseHostPath=${MODELDIR} --set LLM_MODEL_ID=${MODELNAME} --set global.HUGGINGFACEHUB_API_TOKEN=${HFTOKEN} --values gaudi-values.yaml
By default, the vllm service will downloading the “Intel/neural-chat-7b-v3-3”.
If you already cached the model locally, you can pass it to container like this example:
MODELDIR=/mnt/opea-models
MODELNAME=”facebook/opt-125m”
Verify¶
To verify the installation, run the command kubectl get pod
to make sure all pods are runinng.
Then run the command kubectl port-forward svc/myvllm 2080:80
to expose the vllm service for access.
Open another terminal and run the following command to verify the service if working:
curl http://localhost:2080/v1/completions \
-H "Content-Type: application/json" \
-d '{"model": "Intel/neural-chat-7b-v3-3", "prompt": "What is Deep Learning?", "max_tokens": 32, "temperature": 0}'
Values¶
Key |
Type |
Default |
Description |
---|---|---|---|
LLM_MODEL_ID |
string |
|
Models id from https://huggingface.co/, or predownloaded model directory |
global.HUGGINGFACEHUB_API_TOKEN |
string |
|
Hugging Face API token |
global.modelUseHostPath |
string |
|
Cached models directory, vllm will not download if the model is cached here. The host path “modelUseHostPath” will be mounted to container as /data directory. Set this to null/empty will force it to download model. |
image.repository |
string |
|
|
image.tag |
string |
|