Add BI-150 vLLM image and ModelHub release workflow

2026-09-29 16:01:24 +08:00
commit 5718091425
5 changed files with 205 additions and 0 deletions
--- a/.dockerignore
+++ b/.dockerignore
@@ -0,0 +1,2 @@
+**
+!Dockerfile
--- a/.gitea/workflows/docker-build-push.yml
+++ b/.gitea/workflows/docker-build-push.yml
@@ -0,0 +1,132 @@
+name: Docker Build and Push
+
+on:
+  push:
+    tags:
+      - "v*"
+
+jobs:
+  docker:
+    runs-on: amd64-ubuntu-24.04
+
+    steps:
+      - name: Clone repository
+        run: |
+          git clone "${{ gitea.server_url }}/${{ gitea.repository }}.git" .
+          git checkout "${{ gitea.ref_name }}"
+
+      - name: Set image metadata
+        run: |
+          IMAGE_NAME="$(echo "${{ gitea.repository }}" | tr '[:upper:]' '[:lower:]' | tr '_' '-')"
+          IMAGE="${DOCKER_REGISTRY}/${DOCKER_USERNAME}/${IMAGE_NAME}:${{ gitea.ref_name }}"
+
+          echo "IMAGE_NAME=${IMAGE_NAME}" >> "$GITEA_ENV"
+          echo "IMAGE=${IMAGE}" >> "$GITEA_ENV"
+
+      - name: Load and Validate Task Info
+        run: |
+          set -a
+          . .gitea/workflows/task_info.env
+          set +a
+
+          for name in FRAMEWORK GPU_TYPE TASK_TYPE; do
+            eval "value=\${${name}:-}"
+            if [ "$name" = "FRAMEWORK" ] && [ -z "$value" ]; then
+              echo "${name} is empty in .gitea/workflows/task_info.env"
+              exit 1
+            fi
+
+            echo "${name}=${value}" >> "$GITEA_ENV"
+          done
+
+      - name: Validate Image Verify Metadata
+        run: |
+          if [ -z "${FIXED_TOKEN:-}" ]; then
+            echo "FIXED_TOKEN is not configured on runner"
+            exit 1
+          fi
+
+          if ! response="$(curl --silent --show-error --location --get 'https://modelhub.org.cn/adminApi/image-verify/validate' \
+            --header "Xc-Token: ${FIXED_TOKEN}" \
+            --data-urlencode "gpuType=${GPU_TYPE:-}" \
+            --data-urlencode "taskType=${TASK_TYPE:-}")"; then
+            echo "failed to call image verify validate API"
+            exit 1
+          fi
+
+          VALIDATE_RESPONSE="$response" python3 - <<'PY'
+          import json
+          import os
+          import sys
+
+          raw = os.environ.get("VALIDATE_RESPONSE", "")
+          try:
+              body = json.loads(raw)
+          except json.JSONDecodeError:
+              print("image verify validate API returned invalid JSON")
+              print(raw)
+              sys.exit(1)
+
+          if body.get("code") == 0 and body.get("data") is True:
+              print("image verify metadata validation passed")
+              sys.exit(0)
+
+          message = body.get("message") or "unknown error"
+          print(f"image verify metadata validation failed: {message}")
+          print(raw)
+          sys.exit(1)
+          PY
+
+      - name: Login to Docker Registry
+        run: |
+          echo "$DOCKER_PASSWORD" | docker login "$DOCKER_REGISTRY" \
+            -u "$DOCKER_USERNAME" \
+            --password-stdin
+
+      - name: Build Docker Image
+        run: |
+          docker build -t "$IMAGE" .
+
+      - name: Push Docker Image
+        run: |
+          for attempt in 1 2 3; do
+            echo "Starting docker push attempt ${attempt}/3 for ${IMAGE}"
+            docker push "$IMAGE" &
+            PUSH_PID=$!
+
+            while kill -0 "$PUSH_PID" 2>/dev/null; do
+              echo "docker push is still running at $(date -u '+%Y-%m-%dT%H:%M:%SZ')"
+              sleep 60
+            done
+
+            if wait "$PUSH_PID"; then
+              echo "docker push completed successfully"
+              exit 0
+            fi
+
+            echo "docker push failed on attempt ${attempt}/3"
+            sleep 30
+          done
+
+          echo "docker push failed after 3 attempts"
+          exit 1
+
+      - name: Notify Image Verify
+        run: |
+          if [ -z "${FIXED_TOKEN:-}" ]; then
+            echo "FIXED_TOKEN is not configured on runner"
+            exit 1
+          fi
+
+          curl --silent --show-error --fail-with-body --location --request POST 'https://modelhub.org.cn//adminApi/image-verify' \
+            --header "Xc-Token: ${FIXED_TOKEN}" \
+            --header 'Content-Type: application/json' \
+            --data-raw "{
+              \"framework\": \"${FRAMEWORK}\",
+              \"gpuType\": \"${GPU_TYPE}\",
+              \"imageUrl\": \"${IMAGE}\",
+              \"taskType\": \"${TASK_TYPE}\",
+              \"createBy\": \"${{ gitea.actor }}\",
+              \"repoUrl\": \"${{ gitea.server_url }}/${{ gitea.repository }}\",
+              \"tag\": \"${{ github.ref_name }}\"
+            }"
--- a/.gitea/workflows/task_info.env
+++ b/.gitea/workflows/task_info.env
@@ -0,0 +1,3 @@
+FRAMEWORK=vllm
+GPU_TYPE="Iluvatar_bi-150"
+TASK_TYPE=text-generation
--- a/4
+++ b/4
@@ -0,0 +1,4 @@
+FROM registry.iluvatar.com.cn:10443/customer/sz/vllm0.17.0-4.4.0-x86:v7
+
+# Preserve the vendor runtime used in the validated BI-150 deployment.
+CMD ["/bin/bash"]
--- a/README.md
+++ b/README.md
@@ -0,0 +1,64 @@
+# Step-3.5-Flash 天垓150部署
+
+基于用户已验证的天数镜像：
+`registry.iluvatar.com.cn:10443/customer/sz/vllm0.17.0-4.4.0-x86:v7`。
+使用 16 张卡，TP=8、PP=2。镜像不包含模型权重。
+
+## ModelHub 发布
+
+CI 配置沿用 https://dev.modelhub.org.cn/4pdadmin/cicd_demo。
+推送 `v*` 标签后自动构建、推送镜像并提交审核。
+GPU 分类采用 demo 的 `Iluvatar_bi-150`，由平台接口校验。
+Runner 使用 amd64-ubuntu-24.04，需能访问并拉取上述天数基础镜像。
+推送凭据和审核令牌由 runner 提供，不写入仓库。
+
+## 模型配置准备
+
+模型目录：`/mnt/disk0/stepfun-ai/Step-3.5-Flash`。
+在用户原部署中，`num_hidden_layers=45`，但以下数组各有48项，
+引发层数校验失败：`layer_types`、`rope_theta`、
+`partial_rotary_factors`、`swiglu_limits`、`swiglu_limits_shared`。
+
+原部署的处理方式是先备份 config.json，再将这些数组保留前45项。
+该操作修改宿主机挂载的模型配置，并未修改容器镜像。
+本镜像不会自动修改外部模型文件；部署时使用已验证的45项配置。
+如果换用其他版本权重，应先核对配置和权重结构，不应盲目截断。
+
+## 创建容器
+
+将 STEP_IMAGE 设置为 CI 日志中实际推送成功的镜像地址，然后执行：
+
+```bash
+docker run -dit -v /usr/src:/usr/src -v /lib/modules:/lib/modules -v /dev:/dev \
+  -v /home:/home -v /mnt/disk0:/mnt/disk0 \
+  --network=host --name=step3p5-flash --ipc=host --privileged --cap-add=ALL --pid=host \
+  "${STEP_IMAGE:?请先设置CI产出的镜像地址}" /bin/bash
+```
+
+## 启动服务
+
+```bash
+docker exec -it step3p5-flash bash -c '
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15
+vllm serve /mnt/disk0/stepfun-ai/Step-3.5-Flash \
+  --served-model-name step3p5-flash \
+  --host 0.0.0.0 --port 1237 \
+  --tensor-parallel-size 8 \
+  --pipeline-parallel-size 2 \
+  --disable-cascade-attn \
+  --trust-remote-code \
+  --gpu-memory-utilization 0.92 \
+  --max-model-len 8192'
+```
+
+原部署中 corex MoE 算子要求切分后的中间维度为32的倍数。
+1280/16=80不满足要求，1280/8=160满足要求，因此采用TP=8、PP=2。
+各流水线阶段的显存占用取决于实际层分配，不保证各卡相等。
+
+## 验证
+
+```bash
+curl -sS http://127.0.0.1:1237/v1/chat/completions \
+  -H 'Content-Type: application/json' \
+  -d '{"model":"step3p5-flash","messages":[{"role":"user","content":"请介绍一下人工智能"}],"temperature":0.7,"top_p":0.9,"max_tokens":1024}'
+```