Skip to main content

Dynamia ascend-device-plugin Chart 0.3.0

Release date: 2026-10-26

VersionPreviousCurrent
Chart0.2.20.3.0
appVersionv1.4.2v1.4.3

This minor Chart release delivers application v1.4.3 and commercial libvnpu v1.4.3. The application adds per-card resource control across multiple NPUs. The published Chart retains the force compute-policy default; its ENPU deployment configuration is completed by the Chart 0.3.1 patch.

Open-source baselines​

This release selectively integrates the following feature baselines while retaining Enterprise resource management and runtime extensions.

ComponentOpen-source baselineDelivery scope
Ascend Device Pluginab41becPer-card quotas and monitoring, shared virtual-device counts, and backend selection. The published Chart has the deployment limitations described below.
hami-vnpu-core98ec21f, with HostArgs support from 86b120ePer-device runtime state, spatial compute control, and additional memory-allocation and launch hooks.

HAMi Enterprise compatibility​

Feature or modeMinimum HAMi Enterprise versionCompanion requirements and upgrade guidance
Template-based hardware partitioningv2.9.0-rc0The existing scheduler can be retained.
hami-core static soft partitioning, including Enterprise compute policies, task priority, and Turbov2.10.0-r2Upgrade Enterprise first if using an earlier version, and complete license activation. Application v1.4.3 requires the companion runtime authorization provided by this release.
Automatic memory expansionv2.10.0-r2Requires an ordinary container explicitly using hami-core with one vNPU, expansion approval, and admission protection. The companion Chart requires Kubernetes 1.30 or later. Multi-NPU allocation does not extend automatic expansion to multiple NPUs.

The application includes an ENPU backend, but this Chart package lacks its runtime and configuration mounts. Use Chart 0.3.1 for the completed deployment configuration and its Enterprise companion requirements.

Enterprise extensions​

The following capabilities were delivered in earlier commercial releases and remain available:

  • Automatic memory expansion raises a single-vNPU container's memory limit only after HAMi Enterprise approval. It does not change the Pod resource request or AI Core quota.
  • NPU_CORE_UTILIZATION_POLICY=force limits the average compute duty cycle. Smaller NPU_TASK_PRIORITY values have higher priority at cooperative submission boundaries. Submitted work finishes naturally; fixed preemption latency is not guaranteed, and sustained high-priority demand can delay lower-priority work.
  • NPU_TURBO_MODE=true disables compute throttling and priority arbitration while retaining memory limits, accounting, and approved expansion. Use it only for trusted, compute-exclusive workloads.
  • Image pinning with image.digest and the vnpuMonitor metrics service, ServiceMonitor, and PrometheusRule configuration remain available.

Changes​

Chart changes​

  • [Upstream sync] Renames the bundled configuration field and Helm value from vnpuDeviceSplitCount to deviceSplitCount, retaining the default of 10. A positive node-level vDeviceCount takes precedence. Source: ascend-device-plugin #144.
  • [Enterprise feature] Changes the default image.tag from v1.4.2 to an empty value, so the image template follows appVersion. An explicit image.tag or image.digest still overrides that selection.

Application changes​

  • [Upstream sync] Supports memory and compute quotas per physical NPU across multi-card allocations. Each card receives its own quota rather than the sum across cards. A container's allocated cards must have equal effective memory and compute quotas; unequal quotas are rejected. Monitoring follows the allocated physical devices. Sources: ascend-device-plugin #147, hami-vnpu-core #44.
  • [Upstream sync] Adds spatial compute control and the dynamic, fixed, and overdraft scheduling policies. Spatial control sets cooperative Cube/Vector budgets through CANN; it does not create exclusive hardware partitions or guarantee proportional inference throughput. The published Chart still injects force by default. Sources: ascend-device-plugin #145, hami-vnpu-core #43, #44.
  • [Upstream sync] Extends quota accounting to cached/configured and runtime HBM allocation paths. Allocation identity and retained physical handles prevent duplicate charges and incorrect refunds; failed native allocations roll back their reservations. Source: hami-vnpu-core #46.
  • [Upstream sync] Accounts for ACL HostArgs eager launches through the same compute-admission path as other intercepted submissions. Source: hami-vnpu-core #41.
  • [Enterprise feature] Adds runtime authorization to hami-core workloads. Complete Enterprise license activation before creating workloads; recreate Pods after license renewal so they receive the updated authorization.

Upgrade guidance​

Prepare the companion and runtime​

  1. For hami-core, first upgrade HAMi Enterprise to v2.10.0-r2 if needed and complete license activation. Use UTC in the soft-partitioned workload containers.
  2. Record each workload's actual shared-memory paths and host mount mappings. Stop the old soft-partitioned workloads being migrated and confirm that all processes using their shared state have exited.
  3. Back up the affected state and use fresh local and global paths, or remove only the old files owned by those stopped workloads. The new runtime uses .device-<physical-device-ID> files for each card. Default local files are under /usr/local/hami-vnpu-core/containers/<Pod-UID>_<container-name>/ on the host; the new default global files are /usr/local/hami-shared-region/global_registry.device-<physical-device-ID>. Use the effective paths if mounts or HOOK_PATH are customized. Do not remove live mappings or mix old and new libraries on the same files.
  4. Upgrade the plugin, confirm that the matching libvnpu v1.4.3 library is installed on the nodes, and recreate the workloads. Rolling the plugin alone does not update the library and mounts already used by running containers.

Review configuration​

  • Rename a custom Helm value vnpuDeviceSplitCount to deviceSplitCount, preserving its value. With the default config.create=false, update the actual shared-device ConfigMap. The application accepts the old device-configuration field as a fallback; if both fields are present, deviceSplitCount wins. A node-level vDeviceCount still takes precedence. Restart the plugin after changing its configuration.
  • Check image.tag and image.digest. Keeping an explicit v1.4.2 tag or its digest prevents the new application from reaching the nodes; the target default image is ghcr.io/dynamia-ai/ascend-device-plugin:v1.4.3.
  • Review manually supplied NPU_MEMORY_LIMIT_MB, NPU_COMPUTE_LIMIT, and legacy quota variables. Limits now apply per card; when canonical and legacy aliases are both provided, their numeric values must agree. An omitted or zero -core request remains memory-only best effort and does not reserve or guarantee full compute capacity.
  • The Chart's hamiVnpuCore.coreUtilizationPolicy remains force. Use the 0.3.1 patch guidance when adopting the completed backend configuration or the default / spatial defaults.

Existing users of automatic memory expansion must retain its approval and admission configuration, explicit huawei.com/vnpu-mode: hami-core, and one allocated vNPU per ordinary container. Init containers and multi-vNPU containers are outside that expansion scope.

Values reference​

The parameter descriptions below are preserved from the README bundled with Chart 0.3.0. The original values.yaml from that package defines its defaults.

Expand values table
KeyTypeDefaultDescription
config.createboolfalseCreate the bundled device configuration ConfigMap for an installation without HAMi. Leave disabled when HAMi owns the device configuration.
config.deviceConfigMapNamestring"hami-scheduler-device"Name of the device configuration ConfigMap created when config.create=true.
config.existingDeviceConfigMapNamestring"hami-scheduler-device"HAMi-managed device configuration ConfigMap to reuse when config.create=false. The ConfigMap must be in the release namespace.
daemonSet.args[0]string"--config_file"
daemonSet.args[1]string"/device-config.yaml"
daemonSet.args[2]string"--v=4"
daemonSet.args[3]string"--node_config_file"
daemonSet.args[4]string"/node-config.yaml"
daemonSet.namestring"hami-ascend-device-plugin"Device plugin DaemonSet name.
deviceSplitCountint10Number of virtual devices advertised per physical Ascend NPU in soft-slicing mode when this chart creates the device configuration. A positive node-level vDeviceCount overrides it.
fullnameOverridestring""Override the fully qualified resource name.
hamiVnpuCore.coreUtilizationPolicystring"force"Default compute policy injected into hami-core containers that do not declare NPU_CORE_UTILIZATION_POLICY.
hamiVnpuCore.enabledboolfalseEnable hami-vnpu-core in the bundled device configuration when config.create=true. This mode requires the HAMi scheduler.
image.digeststring""Optional immutable OCI manifest digest. When set, it takes precedence over image.tag.
image.pullPolicystring"IfNotPresent"Kubernetes image pull policy.
image.repositorystring"ghcr.io/dynamia-ai/ascend-device-plugin"Container image repository.
image.tagstring""Container image tag. Defaults to the chart appVersion when empty.
nameOverridestring""Override the chart name used in resource names.
nodeConfigstring"nodes: []"Per-node hami-vnpu-core configuration written to the node ConfigMap.
nodeConfigMap.createbooltrueCreate the per-node configuration ConfigMap.
nodeConfigMap.namestring"hami-device-node-config"Per-node configuration ConfigMap name.
nodeSelector.ascendstring"on"Node label value used to schedule the device plugin.
rbac.namestring"hami-ascend"Name shared by the chart-managed RBAC resources.
resources.limits.cpustring"500m"
resources.limits.memorystring"500Mi"
resources.requests.cpustring"500m"
resources.requests.memorystring"500Mi"
runtimeClass.createbooltrueCreate the Ascend RuntimeClass.
runtimeClass.handlerstring"ascend"Container runtime handler used by the RuntimeClass.
runtimeClass.namestring"ascend"RuntimeClass resource name.
serviceAccount.createbooltrueCreate a ServiceAccount for the device plugin.
serviceAccount.namestring"hami-ascend"ServiceAccount name.
vnpuMonitor.enabledboolfalseExpose device plugin metrics and create optional Prometheus Operator resources.
vnpuMonitor.prometheusRule.createbooltrueCreate a PrometheusRule when vnpuMonitor.enabled=true.
vnpuMonitor.prometheusRule.groupLabels.vendorstring"ascend"
vnpuMonitor.prometheusRule.groupNamestring"ascend-vnpu"Recording-rule group name.
vnpuMonitor.prometheusRule.intervalstring"15s"Recording-rule evaluation interval.
vnpuMonitor.prometheusRule.labels.releasestring"prometheus"
vnpuMonitor.prometheusRule.labels.rolestring"recording-rules"
vnpuMonitor.prometheusRule.labels.vendorstring"ascend"
vnpuMonitor.prometheusRule.namestring"hami-ascend-vnpu-monitor"PrometheusRule resource name.
vnpuMonitor.prometheusRule.namespacestring"monitoring"Namespace for the PrometheusRule.
vnpuMonitor.service.namestring"hami-ascend-device-plugin-metrics"Headless Service that exposes the device plugin metrics port.
vnpuMonitor.serviceMonitor.createbooltrueCreate a ServiceMonitor when vnpuMonitor.enabled=true.
vnpuMonitor.serviceMonitor.intervalstring"15s"Prometheus scrape interval.
vnpuMonitor.serviceMonitor.labels.releasestring"prometheus"
vnpuMonitor.serviceMonitor.namestring"hami-ascend-vnpu-monitor"ServiceMonitor resource name.
vnpuMonitor.serviceMonitor.namespacestring"monitoring"Namespace for the ServiceMonitor.
vnpuMonitor.serviceMonitor.pathstring"/metrics"Metrics HTTP path.