跳到主要内容

Resource Pool Manager Chart 0.0.7

Chart:0.0.7 · appVersion:0.0.3 · RPM Release:rpm-0.0.3 · 发布时间:2026-11-10(UTC)

软件变更​

Ascend 库存与准入​

RPM 0.0.3 在上报设备容量、选择资源池和验证分配时使用 Ascend 运行模式。RPM 从现有调度器设备 ConfigMap 读取 vnpus.hamiVnpuMode,并根据节点和单个设备上报的模式匹配 Pod 请求。只有支持所请求模式的设备才会计入候选资源池的可用容量。

对于 hami-core 设备,RPM 将旧版插件上报的物理核数归一为百分比容量 100,并保留大于 100 的已上报百分比容量。模板设备不将物理 AI 核数作为百分比容量上报。ENPU 设备的核心容量为 100,Pod 必须显式设置 huawei.com/vnpu-mode: enpu 注解。ENPU 准入要求每个容器恰好请求一个 Ascend 设备。

NVIDIA Downward API SM 限额​

对于启用 useDownward 的 NVIDIA 工作负载,RPM 要求 CUDA_DEVICE_SM_LIMIT 通过 Downward API 引用当前容器对应的 HAMi 核心限额注解。缺少该显式引用时,RPM 会补充引用,防止通过 envFrom 导入的值覆盖运行时 SM 限额。此行为同时适用于普通容器和 init 容器,不依赖两个 Webhook 的执行顺序。

旧版调度器可能额外注入 CUDA_DEVICE_CORE_LIMIT。RPM 接受引用同一注解的旧变量,同时仍要求 CUDA_DEVICE_SM_LIMIT。RPM 保留已有的显式条目,并在最终准入时拒绝错误引用或字面量值。

Chart 变更​

  • appVersion 从 0.0.2 升至 0.0.3;默认镜像从 ghcr.io/dynamia-ai/resource-pool-manager:v0.0.2 更新为 ghcr.io/dynamia-ai/resource-pool-manager:v0.0.3。
  • Chart 模板及其余 Values 默认值与 Chart 0.0.6 相同,无需新增 Values 配置键。
  • 单独发布的资源池 CRD Chart 仍为 hami-enterprise-crds 0.4.2,其 appVersion 为 0.4.0。本次发布不涉及 CRD schema 迁移。

升级说明​

同时升级 Manager Chart 和 RPM 镜像。安装配置覆盖了 image.tag 时,将其显式设为 v0.0.3;保留旧覆盖值会继续使用旧镜像。

Ascend 安装保持 devices.ascend.enabled: true,并核对现有调度器设备 ConfigMap、节点及设备注册信息和工作负载模式一致。vnpus.hamiVnpuMode 接受 template、hami-core 或 enpu。未设置该字段时,旧配置 vnpus.hamiVnpuCore: true 选择 hami-core,否则默认为 template。ENPU 工作负载需显式设置 Pod 模式,每个容器请求一个 Ascend 设备。RPM 的设备开关不会安装或升级 Ascend 设备插件及运行库,使用前需准备支持所选模式的配套组件。模式前置条件见 Ascend 设备文档。

使用 Downward API 限额的 NVIDIA 安装需检查工作负载模板中显式设置的 CUDA_DEVICE_SM_LIMIT 和 CUDA_DEVICE_CORE_LIMIT。移除应用提供的字面量覆盖,或将引用修正为当前容器对应的 HAMi 核心限额注解。错误的显式条目可能导致升级后新建或重建的 Pod 被拒绝。

Chart 继续要求 Kubernetes 1.30 或更新版本、单独安装的资源池 CRD、HAMi extender 调度模式,以及通过 hami.existingConfigMap 指定现有调度器设备 ConfigMap。修改调度器设备配置后,使用完整的 manager values 文件升级 RPM,以重新生成 Webhook 规则和 manager 配置。安装操作见 HAMi Enterprise 安装文档中的 RPM 章节。

Values 说明​

以下参数说明取自 Chart 0.0.7 发布包随附的 README,保留原文。默认配置以该版本的 values.yaml 为准。

展开参数表
KeyTypeDefaultDescription
config.createbooltrueInitialize the runtime ConfigMap if absent. It is retained outside Helm ownership for external policy updates. For render-and-apply GitOps, precreate the ConfigMap and set this to false.
config.namestring""Runtime ConfigMap name; defaults to <workload name>-config. With config.create=false it must already exist in the manager namespace.
config.poolSchedulerPolicystring"binpack"Initial pool selection policy. Later changes belong in the runtime ConfigMap, not Helm values.
devices.ascend.enabledboolfalseInclude Ascend profiles from the scheduler device ConfigMap in resource-pool admission.
devices.hygon.enabledboolfalseInclude Hygon HCU resources from the scheduler device ConfigMap in resource-pool admission.
devices.nvidia.enabledbooltrueInclude NVIDIA resources from the scheduler device ConfigMap in resource-pool admission.
fullnameOverridestring""Override the workload name. Names are limited to 43 characters to leave room for resource suffixes.
hami.existingConfigMapstring""Required name of the scheduler device ConfigMap in this release namespace. Helm reads data.device-config.yaml from the cluster and the manager mounts the same ConfigMap. Inline configuration and overrides are not supported. Upgrade the manager after scheduler device configuration changes.
hami.schedulerNamestring"hami-scheduler"Scheduler name used by the existing HAMi installation. The scheduler must use HAMi extender admission and scheduling, not DRA.
healthProbeBindAddressstring":8081"Health server listen address.
image.pullPolicystring"IfNotPresent"Resource-pool-manager image pull policy.
image.repositorystring"ghcr.io/dynamia-ai/resource-pool-manager"Resource-pool-manager image repository, including registry.
image.tagstring"v0.0.3"Required published image tag. Independent of this chart's appVersion; update explicitly when releasing a new manager image.
imagePullSecretslist[]Registry Secret names used by the manager and certificate Jobs.
leaderElectbooltrueElect a leader for inventory and status reconcilers.
metricsBindAddressstring":8080"Metrics server listen address.
nameOverridestring""Override the chart name used in names and labels.
nodeSelectorobject{}Node selector for the manager.
podAnnotationsobject{}Additional annotations for manager Pods.
replicasint1Number of manager replicas. Use leader election when running multiple replicas.
resourcesobject{}Resource requests and limits for the manager.
timeoutSecondsint10Admission webhook timeout in seconds.
tolerationslist[]Tolerations for the manager.
webhook.tls.certManagerboolfalseUse an installed cert-manager controller for certificates and CA injection. When false, certgen hook Jobs manage webhook certificates.
webhook.tls.certgen.image.pullPolicystring"IfNotPresent"Certificate generator image pull policy.
webhook.tls.certgen.image.repositorystring"registry.k8s.io/ingress-nginx/kube-webhook-certgen"Certificate generator image repository, including registry.
webhook.tls.certgen.image.tagstring"v1.6.9"Certificate generator image tag.
webhook.tls.certgen.nodeSelectorobject{}Node selector for certificate Jobs.
webhook.tls.certgen.podAnnotationsobject{}Additional annotations for certificate Job Pods.
webhook.tls.certgen.priorityClassNamestring""Priority class for certificate Jobs.
webhook.tls.certgen.runAsUserint65532Non-root UID for certificate Jobs.
webhook.tls.certgen.tolerationslist[]Tolerations for certificate Jobs.
webhookPortint9443HTTPS port in the manager container.