download models

app.py

@@ -1,7 +1,21 @@
+import os
+import subprocess
 import gradio as gr
 from transformers import pipeline
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
+# 下载模型
+base_dir = "/root/.cache/huggingface/hub"
+if not os.path.isdir(base_dir):
+    os.makedirs(base_dir)
+
+cmd_list = ["cd", "base_dir", "&&", "git clone", "https://huggingface.co/gpt2", "models"]
+cmd_str = "".join(cmd)
+print("cmd_str:", cmd_str)
+ret, out = subprocess.getstatusoutput(cmd_str)
+print("ret:", ret)
+print("out:", out)
+
 tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path="/root/.cache/huggingface/hub/models")
 model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path="/root/.cache/huggingface/hub/models")
 generator = pipeline('text-generation', model=model, tokenizer=tokenizer)

deploy/knative-serving.yaml

@@ -0,0 +1,76 @@
+apiVersion: serving.knative.dev/v1
+kind: Service
+metadata:
+  name: custom-ksvc-gpt2
+  namespace: kubeflow-admin-cpaas-io
+spec:
+  template:
+    metadata:
+      annotations:
+        # 表示自动扩缩容使用 kpa 类型，也是 Knative 默认类型
+        autoscaling.knative.dev/class: "kpa.autoscaling.knative.dev"
+        # 表示平均每个 pod 支持 10 个并发
+        autoscaling.knative.dev/target: "10"
+        # 表示 Knative service 最小 pods 数为 1
+        autoscaling.knative.dev/minScale: "1"
+        # 最大 pods 数为 5
+        autoscaling.knative.dev/maxScale: "5"
+    spec:
+      containerConcurrency: 1
+      initContainers:
+      - args:
+        # 这里指定用户 space repo 的地址
+        # https://huggingface.co/gpt2
+        # https://huggingface.co/spaces/lanzhiwang/first-gradio
+        # https://gitee.com/lanzhiwang/first-gradio.git
+        - git+lfs://gitee.com/lanzhiwang/first-gradio
+        - /mnt/app
+        # 这个是固定的镜像，用于拉取用户 space 代码
+        image: 10.0.200.210:11443/mlops/storage-initializer:v0.8.0-acp-0914
+        imagePullPolicy: IfNotPresent
+        name: storage-initializer
+        # env:
+        #   - name: HTTPS_PROXY
+        #     value: http://192.168.144.12:7890
+        resources:
+          limits:
+            cpu: "1"
+            memory: 2Gi
+          requests:
+            cpu: 100m
+            memory: 100Mi
+        volumeMounts:
+        - mountPath: /mnt/app
+          name: space-code-location
+      containers:
+      - env:
+        - name: SPACE_DIR
+          value: /mnt/app
+        - name: APP_NAME
+          value: first-gradio
+        - name: PIP_MIRROR
+          value: https://pypi.tuna.tsinghua.edu.cn/simple
+        - name: APP_FILE
+          value: app.py
+        # 镜像确定方式参考上文
+        image: dev.local/gradio-runtime:v3.4
+        command:
+        - bash
+        - /home/user/app/endpoint.sh
+        imagePullPolicy: IfNotPresent
+        ports:
+          - containerPort: 7860
+        name: space-container
+        volumeMounts:
+        - mountPath: /root/.cache/huggingface/hub
+          name: models-location
+        - mountPath: /mnt/app
+          name: space-code-location
+      timeoutSeconds: 600
+      volumes:
+      - name: space-code-location
+        persistentVolumeClaim:
+          claimName: custom-ksvc-gpt2-app
+      - name: models-location
+        persistentVolumeClaim:
+          claimName: custom-ksvc-gpt2-models

deploy/nodeport.yaml

@@ -0,0 +1,13 @@
+apiVersion: v1
+kind: Service
+metadata:
+  name: custom-ksvc-gpt2-nodeport
+  namespace: kubeflow-admin-cpaas-io
+spec:
+  ports:
+    - port: 7860
+      targetPort: 7860
+      protocol: TCP
+  selector:
+    serving.knative.dev/revisionUID: 8037843b-8ffa-4686-a33f-36134dfaafb7
+  type: NodePort

deploy/pvc.yaml

@@ -0,0 +1,27 @@
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: custom-ksvc-gpt2-app
+  namespace: kubeflow-admin-cpaas-io
+spec:
+  accessModes:
+  - ReadWriteOnce
+  resources:
+    requests:
+      storage: 2Gi
+  storageClassName: sc-topolvm
+  volumeMode: Filesystem
+---
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: custom-ksvc-gpt2-models
+  namespace: kubeflow-admin-cpaas-io
+spec:
+  accessModes:
+  - ReadWriteOnce
+  resources:
+    requests:
+      storage: 10Gi
+  storageClassName: sc-topolvm
+  volumeMode: Filesystem