[Artificial Intelligence] Deploying LLMs with WasmEdge in HomeLab

By See Hiong

January 13, 2024 - 2 minutes read - 402 words

In this post, we delve into the deployment of Lightweight Language Models (LLMs) using WasmEdge, a lightweight, high-performance, and extensible WebAssembly runtime. This setup is tailored to run LLMs in our previously configured HomeLab environment.

Preparation

To establish an OpenAI-compatible API server, begin by downloading the API server application:

curl -LO https://github.com/second-state/LlamaEdge/releases/latest/download/llama-api-server.wasm

For Rust-based Llama 2 inference, we require the Wasi-NN plugin. The Dockerfile below reflects this configuration:

FROM wasmedge/wasmedge:ubuntu-build-gcc

ENV template sample_template
ENV model sample_model
RUN mkdir /models

RUN apt update && apt install -y curl git ninja-build
RUN git clone https://github.com/WasmEdge/WasmEdge.git -b0.14.0-alpha.1

# Ubuntu/Debian with OpenBLAS
RUN apt update
RUN apt install -y software-properties-common lsb-release cmake unzip pkg-config
RUN apt install -y libopenblas-dev

RUN cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release \
  -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="GGML" \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_BLAS=ON \
  .

RUN cmake --build build
RUN cmake --install build

COPY ./llama-api-server.wasm /WasmEdge
RUN export WASMEDGE_PLUGIN_PATH=/WasmEdge/build/plugins/wasi_nn
WORKDIR /WasmEdge 

CMD ["sh", "-c", "wasmedge --dir .:. --nn-preload default:GGML:AUTO:/models/\"$model\" llama-api-server.wasm --prompt-template ${template} --log-stat"]

The template and model environment variables in the Dockerfile are parameters that can be customized later in the K3s YAML file.

Build and push the image to registry:

docker build . -t wasmedge-wasi-nn:0.14.0-alpha.1

docker tag wasmedge-wasi-nn:0.14.0-alpha.1 registry.local:5000/wasmedge-wasi-nn:0.14.0-alpha.1
docker push registry.local:5000/wasmedge-wasi-nn:0.14.0-alpha.1

Download the LLM:

wget https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf

Run the image:

docker run -p 8080:8080 -v $PWD:/models -e "model=llama-2-7b-chat.Q4_K_M.gguf" -e "template=llama-2-chat" wasmedge-wasi-nn:0.14.0-alpha.1

wasm-start-open-api-server

Test the API server:

curl -X POST http://localhost:8080/v1/models -H 'accept:application/json' | jq

wasm-wasi-nn-get-model

Deploying to HomeLab

After moving the downloaded LLM to the NAS server, create the deploy.yaml file:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: wasmedge-llama2-7b-chat
  namespace: llm
spec:
  replicas: 1
  selector:
    matchLabels:
      app: wasmedge-llama2-7b-chat
  template:
    metadata:
      labels:
        app: wasmedge-llama2-7b-chat
    spec:
      containers:
      - name: wasmedge-llama2-7b-chat
        image: registry.local:5000/wasmedge-wasi-nn:0.14.0-alpha.1
        imagePullPolicy: IfNotPresent
        ports:
        - containerPort: 8080
        env:
        - name: model
          value: llama-2-7b-chat.Q4_K_M.gguf
        - name: template
          value: llama-2-chat
        resources:
          requests:
            memory: "8Gi"
          limits:
            memory: "8Gi"
        volumeMounts:
        - name: models-store
          mountPath: /models
      imagePullSecrets:
      - name: regcred
      volumes:
      - name: models-store
        persistentVolumeClaim:
            claimName: nfs-models

Create the corresponding svc.yaml file:

apiVersion: v1
kind: Service
metadata:
  name: wasmedge-llama2-7b-chat-svc
  namespace: llm
spec:
  selector:
    app: wasmedge-llama2-7b-chat
  type: ClusterIP
  ports:
    - name: http
      protocol: TCP
      port: 80
      targetPort: 8080

Apply the changes:

kca deploy.yaml
kca svc.yaml

wasm-llm-portainer-application

Next, set up both the Kong service and route:

wasm-llm-kong-service-setup

wasm-llm-kong-route-setup

In Action

Chnage to using the new path in the Langchain4j application:

private static final String LOCAL_AI_URL = "http://api.local/wasmedge-llama2-7b-chat/v1";

After running the Spring Boot application, observe the result from the retrieve API:

wasm-llm-postman-retrieve-api

And that concludes our setup! Feel free to drop any questions or comments below.