openKylin FlagOS SIG 成立：给多元 AI 芯片搭一套统一软件栈

# /etc/flagos/accelerator-topology.yaml
# 描述本节点的加速器资源配置，FlagOS 调度插件会读取此文件
node_name: compute-node-01

accelerators:
  - type: gpu
    vendor: vendor-a
    model: xa100
    count: 4
    driver_version: "2.3.1"
    firmware_version: "1.0.8"
    memory_mb: 32768
    device_path: /dev/vendor-a-accel

  - type: npu
    vendor: vendor-b
    model: nb200
    count: 2
    driver_version: "5.0.0"
    firmware_version: "3.2.1"
    memory_mb: 65536
    device_path: /dev/vendor-b-npu

runtime_defaults:
  compute_backend: flagos-rt   # FlagOS 统一运行时
  compiler_backend: flagos-cc  # FlagOS 统一编译器

#!/usr/bin/env bash
# flagos-accel-check.sh — 检查本节点加速器驱动与设备节点状态
set -euo pipefail

TOPOLOGY_FILE="/etc/flagos/accelerator-topology.yaml"

if [[ ! -f "$TOPOLOGY_FILE" ]]; then
  echo "错误：拓扑文件 $TOPOLOGY_FILE 不存在，请先配置"
  exit 1
fi

echo "=== FlagOS 加速器状态检查 ==="
echo "节点：$(hostname)"
echo ""

# 从 YAML 中提取 device_path 字段（依赖 yq 工具）
if command -v yq &>/dev/null; then
  DEVICES=$(yq '.accelerators[].device_path' "$TOPOLOGY_FILE")
  for dev in $DEVICES; do
    if [[ -c "$dev" ]]; then
      echo "✅ $dev — 设备节点存在（字符设备）"
    elif [[ -e "$dev" ]]; then
      echo "⚠️  $dev — 节点存在但类型异常"
    else
      echo "❌ $dev — 设备节点缺失，驱动可能未加载"
    fi
  done
else
  echo "提示：安装 yq 可自动解析 YAML（pip install yq 或下载二进制）"
  echo "手动检查：ls /dev/vendor-* 确认设备节点"
  ls /dev/vendor-* 2>/dev/null || echo "未发现 /dev/vendor-* 设备节点"
fi

echo ""
echo "=== 内核模块加载情况 ==="
lsmod | grep -E 'accel|npu|gpu' || echo "无加速器相关内核模块"

# 安装 yq（可选，用于 YAML 解析）
pip install yq      # 或从 https://github.com/mikefarah/yq 下载 Go 版本

# 运行检查
chmod +x flagos-accel-check.sh
./flagos-accel-check.sh

# flagos-device-plugin-deploy.yaml
# 在 Kubernetes 集群中为每种加速器部署 DevicePlugin DaemonSet
apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: flagos-accel-device-plugin
  namespace: flagos-system
spec:
  selector:
    matchLabels:
      app: flagos-device-plugin
  template:
    metadata:
      labels:
        app: flagos-device-plugin
    spec:
      containers:
        - name: device-plugin
          image: flagos/device-plugin:0.1.0   # FlagOS 官方镜像，后续发布
          securityContext:
            privileged: true
          volumeMounts:
            - name: device-plugin
              mountPath: /var/lib/kubelet/device-plugins
            - name: dev
              mountPath: /dev
            - name: config
              mountPath: /etc/flagos
      volumes:
        - name: device-plugin
          hostPath: /var/lib/kubelet/device-plugins
        - name: dev
          hostPath: /dev
        - name: config
          hostPath: /etc/flagos

# 创建命名空间并部署
kubectl create namespace flagos-system
kubectl apply -f flagos-device-plugin-deploy.yaml

# 验证设备插件注册
kubectl get nodes -o wide
kubectl describe node compute-node-01 | grep -A5 Capacity