1. 镜像构建

 git clone https://github.com/hiyouga/LLaMA-Factory.git  -b v0.7.0
 cd LLaMA-Factory
 docker build -f ./Dockerfile -t llama-factory:v0.7.0 .

2.数据准备

• 模型

  为了节省时间，提前下载模型

  root@ndoe:/data/models# tree -L 2 
  .
  ├── BAAI
  │   └── bge-reranker-large
  ├── databricks
  │   ├── dbrx-instruct
  ├── hpcai-tech
  │   └── Open-Sora
  ├── meta-llama
  │   ├── Meta-Llama-3-70B-Instruct
  │   └── Meta-Llama-3-8B
  ├── mistralai
  │   └── Mixtral-8x7B-Instruct-v0.1
  ├── moka-ai
  │   └── m3e-large
  └── xai-org
      └── grok-1
  

• 数据集

  使用官方提供的测试用数据集
  https://github.com/hiyouga/LLaMA-Factory/blob/main/data/alpaca_data_zh_51k.json

  root@node:/data/llama-factory# tree -L 2 
  .
  ├── data
  │   ├── alpaca_data_zh_51k.json
  │   └── dataset_info.json
  └── output
  

• 数据集配置

  {
    "alpaca_data_zh_51k.json": {
      "file_name":"alpaca_data_zh_51k.json",
      "file_sha1":"2ba9827122c158dc256668d42bd1bcb8bc6b786e"
    }
  }
  

3.部署

• 使用k8s

  apiVersion: extensions/v1beta1
  kind: Ingress
  metadata:
    name: llamaf
    namespace: llm
  spec:
    ingressClassName: nginx
    rules:
    - host: llamaf.xxx.cn
      http:
        paths:
        - backend:
            serviceName: llamaf
            servicePort: 7860
          path: /
          pathType: Prefix
  ---
  apiVersion: v1
  kind: Service
  metadata:
    labels:
      app.kubernetes.io/name: llamaf
    name: llamaf
    namespace: llm
  spec:
    ports:
    - name: llamaf
      port: 7860
      protocol: TCP
      targetPort: 7860
    selector:
      app.kubernetes.io/name: llamaf
    sessionAffinity: None
    type: ClusterIP
  ---
  apiVersion: apps/v1
  kind: Deployment
  metadata:
    name: llamaf
    namespace: llm
  spec:
    replicas: 1
    selector:
      matchLabels:
        app.kubernetes.io/name: llamaf
    template:
      metadata:
        labels:
          app.kubernetes.io/name: llamaf
      spec:
        containers:
        - name: llamaf
          image: llama-factory:v0.7.0
          imagePullPolicy: IfNotPresent
          env:
          - name: TZ
            value: Asia/Shanghai
          resources:
            limits:
              cpu: "8"
              memory: 32Gi
              tencent.com/vcuda-core: "100"
              tencent.com/vcuda-memory: "320"
            requests:
              cpu: "8"
              memory: 16Gi
              tencent.com/vcuda-core: "100"
              tencent.com/vcuda-memory: "320"
          volumeMounts:
          - name: localtime
            mountPath: /etc/localtime
          - name: shm
            mountPath: /dev/shm
          - name: model
            mountPath: /root/.cache/huggingface
          - name: data
            mountPath: /app/data
          - name: output
            mountPath: /app/output
        volumes:
        - name: shm
          emptyDir:
            medium: Memory
            sizeLimit: 16G
        - name: localtime
          hostPath:
            path: /etc/localtime
        - name: model
          hostPath:
            path: /data/models
            type: DirectoryOrCreate
        - name: data
          hostPath:
            path: /data/llama-factory/data
            type: DirectoryOrCreate
        - name: output
          hostPath:
            path: /data/llama-factory/output
            type: DirectoryOrCreate
  

4.微调

微调配置

5.详解

TODO: