HAMi Enterprise 2.9.0-rc0

HAMi Enterprise v2.9.0-rc0 adds AMD memory and compute-share requests, an Ascend soft-partition scheduling interface, and Biren device support. It also changes monitoring metric names, the device-configuration structure, and traffic routing for scheduler leader and follower instances. Existing deployments should carefully review monitoring rules, custom device configuration, and companion device plugins.

Version information

Item	Value
Enterprise release	HAMi Enterprise v2.9.0-rc0
Release date	2026-05-15
Enterprise upgrade baseline	v2.8.0-rc1
Open-source baseline	HAMi v2.9.0
Helm Chart	hami, version 2.9.0-rc0
Chart appVersion	2.9.0-rc0
Default Enterprise component image tag	v2.9.0-rc0

Software changes

Changes compared with Enterprise v2.8.0-rc1

• Device scheduling: AMD switches from recognizing full cards based on node capacity to reading device-plugin registration information and adds memory- and compute-share requests. Ascend adds scheduling and runtime configuration interfaces for soft partitioning. Biren adds scheduling support for full-card and die modes, plus UUID selection. Each device type requires a separately deployed matching device plugin; the main Chart does not include these plugins.
• Monitoring interfaces: Common scheduler and vGPU monitor metrics now use the hami_ prefix, and label names are updated. A compatibility switch for old metrics is added and disabled by default; it does not cover the three memory-detail metrics. See “Monitoring rule migration” for the complete mapping.
• GPU memory auto-expansion: This existing feature now has an explicit enablement condition. It processes expansion feedback only when CUDA_MEMORY_AUTO_ADJUST: "true" is declared directly on a workload container.
• Scheduler leader/follower routing: Readiness probes can now succeed on both leader and follower instances. The Service selects the leader using hami.io/scheduler-role: leader. Existing leader election remains available, but readiness no longer indicates that an instance is the leader.
• Allocation and recovery: Fixes index preservation in multi-container device annotations, CDI file updates after MIG reconfiguration, and some state handling for terminating Pods and device-registration recovery. Also fixes scoring for percentage-based memory requests and handling of zero capacity.

Enterprise extensions compared with open-source HAMi v2.9.0

Capability	Enterprise behavior and requirements
Enterprise license management	Continues to manage licenses by device instance or device-model count and expose expiration status. Confirm license support for each device backend against the delivery plan.
NVIDIA container allocation and memory auto-expansion	Continues to pass device, memory, and compute allocations to containers through the Downward API and to process memory-expansion requests and allocation updates. Expansion requires a compatible Enterprise runtime, is limited by allocatable memory recorded by the scheduler, and applies only to regular workload containers.
AMD share scheduling	Extends the open-source baseline's full-card handling with registered-device information, memory and compute requests, model filtering, and allocation-information coordination. Requires the updated device-registration protocol.
Alibaba PPU scheduling integration	Continues PPU resource requests, device selection, and namespace quota checks. Requires a separate companion PPU plugin and runtime. See “AMD and PPU integration” for the limitations of the default Chart resource mapping in this release.
AWS Neuron resource accounting	Continues Enterprise accounting semantics for core counts and memory shares, which differ from the open-source baseline. When migrating from open source, verify existing quotas, resource statistics, and device-type mappings.

Except for AMD share scheduling, the capabilities above were already provided in earlier Enterprise releases. The NVIDIA Downward API allocation path also already existed; useDownward and postNvidiaSmi remain enabled in the default device configuration.

Upstream changes included in this release

The following are the major upstream changes included in this release. Follow this page's upgrade notes for configuration and Enterprise usage boundaries.

Change	Impact on usage and integration	Upstream PR
Biren device scheduling	Adds scheduling integration for full-card and die modes and UUID selection; requires the vendor device plugin.	#1645
Webhook label filtering	Limits admission processing by namespace and Pod labels while retaining existing exclusion conditions.	#1653
Monitoring metric normalization	Renames metrics and labels and provides compatibility output for some old metrics.	#1644
Ascend soft partitioning and runtime interface	Adds HAMi-core mode, compute requests, RuntimeClass, and environment-variable handling. Runtime libraries and the device plugin must be provided separately.	#1771, #1804, #1812, #1837, #1738
Preload configuration for a custom library directory	Uses devicePlugin.libPath for preload files instead of continuing to reference a fixed default directory.	#1714
Resource usage by terminating Pods	Retains device-usage records while a Pod is being deleted to avoid releasing resources too early from the scheduler cache.	#1719
Zero-capacity scoring protection	Handles devices or nodes with zero capacity to prevent invalid scores.	#1820
Device-registration recovery	Fixes handshake-time parsing and allows a deleted device to register again after the cooldown period.	#1816, #1843
MIG and CDI integration	Updates and checks CDI files after MIG configuration changes in the applicable allocation path; fixes the A100 80 GB MIG profile.	#1826, #1847

Chart changes

• Adds legacyMetrics, defaulting to false. When set to true, the scheduler and vGPU monitor both append the supported legacy metrics to their output.
• Adds scheduler.admissionWebhook.namespaceSelector and objectSelector, supporting matchLabels and matchExpressions. Custom conditions apply together with existing conditions; namespaces or Pods labeled hami.io/webhook: ignore are still excluded.
• A non-empty component-level image.tag for the scheduler, NVIDIA Device Plugin, or vGPU monitor takes precedence over global.imageTag; an empty component-level tag falls back to the global tag. The settings are scheduler.extender.image.tag, devicePlugin.image.tag, and devicePlugin.monitor.image.tag.
• Consolidates scheduling-configuration templates for older and newer Kubernetes versions. For Kubernetes 1.22 and later, the scheduling ConfigMap name no longer includes -newversion; Chart-managed mounts are updated accordingly. Update deployments that reference the old ConfigMap name directly.
• Sets the NVIDIA Device Plugin DaemonSet to hostNetwork: false. Configurations that collect monitoring data directly from the host network address must use the monitor Service endpoint instead and review related network policies. Existing Service and ServiceMonitor interfaces remain available.
• Adds devices.ascend.hamiVnpuCore (default false) and devices.ascend.runtimeClassName (default empty), and adds Ascend compute-resource and 910C device configuration. devices.ascend.enabled remains disabled by default.
• scheduler.overwriteEnv is now used to generate NVIDIA device configuration and is also used for the new Ascend configuration. It remains "false" by default. Enabling it affects the device environment variables visible to containers that did not request the corresponding device; check existing environment-variable conventions before enabling it.
• Updates the optional hami-dra subchart from 0.1.0 to 0.2.0. dra.enabled remains disabled by default. Enabling it switches to the DRA deployment path and does not install the existing scheduler extender. DRA driver images have independent version settings and are not determined by the main application tag.

Upgrade notes

Package and companion components

Obtain hami-2.9.0-rc0.tgz, the Enterprise images, and the plugins and runtimes for the devices in use through the agreed delivery channel. The Chart remains named hami. Keep the existing Helm release name and namespace during the upgrade, and merge existing custom values with the new version's configuration.

If component-level image tags were pinned to an older version, changing only global.imageTag will not override them. Update the three component-level tags or clear them to use the global tag.

Monitoring rule migration

Before upgrading, update dashboards, alerts, recording rules, and label filters. During migration, set legacyMetrics: true to emit both new metrics and supported legacy metrics; disable legacy output after migration. Deployments that start binaries directly must separately set --legacy-metrics=true for the scheduler and vGPU monitor.

Old scheduler metric	New metric
GPUDeviceMemoryLimit	hami_gpu_memory_limit_bytes
GPUDeviceCoreLimit	hami_gpu_core_limit_ratio
GPUDeviceMemoryAllocated	hami_gpu_memory_allocated_bytes
GPUDeviceSharedNum	hami_gpu_shared_count
GPUDeviceCoreAllocated	hami_gpu_core_allocated_ratio
nodeGPUOverview	hami_node_gpu_overview
nodeGPUMemoryPercentage	hami_node_gpu_memory_allocated_ratio
nodeGPUMigInstance	hami_node_gpu_mig_instance_info
vGPUMemoryAllocated	hami_vgpu_memory_allocated_bytes
vGPUCoreAllocated	hami_vgpu_core_allocated_ratio
QuotaUsed	hami_resource_quota_used

Old monitor metric	New metric	Is the old name retained by the compatibility switch?
HostGPUMemoryUsage	hami_host_gpu_memory_used_bytes	Yes
HostCoreUtilization	hami_host_gpu_utilization_ratio	Yes
vGPU_device_memory_usage_in_bytes	hami_vgpu_memory_used_bytes	Yes
vGPU_device_memory_limit_in_bytes	hami_vgpu_memory_limit_bytes	Yes
Device_memory_desc_of_container	hami_container_device_memory_bytes	Yes
Device_utilization_desc_of_container	hami_container_device_utilization_ratio	Yes
Device_last_kernel_of_container	hami_container_last_kernel_elapsed_seconds	Yes
MigInfo	hami_mig_device_info	Yes
vGPU_device_memory_context_size_bytes	hami_vgpu_memory_context_bytes	No
vGPU_device_memory_module_size_bytes	hami_vgpu_memory_module_bytes	No
vGPU_device_memory_buffer_size_bytes	hami_vgpu_memory_buffer_bytes	No

New metrics also use the following label names. Update query filters, grouping, and joins accordingly.

Old label	New label
nodeid, nodename	node
deviceuuid, deviceidx, devicetype	device_uuid, device_index, device_type
devicecores, sharedcontainers, devicememorylimit	device_cores, shared_containers, device_memory_limit
podnamespace, quotanamespace	namespace
podname, ctrname, containeridx	pod, container, container_index
quotaName, migname	quota_name, mig_name
vdeviceid, instanceid	vdevice_index, instance_id
context, module, data in memory descriptions	context_size, module_size, buffer_size

The compatibility output for GPUDeviceSharedNum also adds the devicetype label. The Enterprise metric vGPULicenseRemains retains its existing name and labels.

Do not change thresholds uniformly based only on the _ratio suffix. hami_host_gpu_utilization_ratio still reports utilization from 0 to 100. Compute-limit and allocation metrics retain the device backend's values; a regular NVIDIA full card uses 100. hami_node_gpu_memory_allocated_ratio is allocated memory divided by total memory. For memory breakdowns, use the separate hami_vgpu_memory_*_bytes metrics instead of numeric labels in the memory-description metric.

GPU memory auto-expansion

For regular workload containers that use this feature, declare the following environment variable directly in the workload template and apply it through a normal rolling update or container recreation:

env:
  - name: CUDA_MEMORY_AUTO_ADJUST
    value: "true"

This release checks for a directly declared string value in the Pod. Providing the variable indirectly through valueFrom or envFrom does not meet the enablement condition. Expansion is constrained by allocatable memory; it does not expand nodes automatically and does not guarantee that CUDA out-of-memory errors will be eliminated.

Ascend configuration migration and soft-partitioning requirements

Device configuration generated by the Chart adopts the new structure automatically. If you provide files/device-config.yaml in the Chart, move the old vnpus list to vnpus.configs and preserve the existing chip configurations. The custom file replaces the entire default device configuration; entries are not merged individually.

vnpus:
  hamiVnpuCore: false
  configs: []  # Move the complete chip configurations from the old vnpus list here

Soft partitioning requires the companion Ascend device plugin, runtime, and HAMi-vNPU-core. Set devices.ascend.hamiVnpuCore: true to declare that nodes support this mode by default. The hami-vnpu-core node annotation takes precedence; only the string "true" enables it. Workloads must also set the Pod annotation huawei.com/vnpu-mode: hami-core and request the corresponding huawei.com/<device-type>-core compute resource; for example, a value of 40 means 40%. Nodes using soft partitioning reject some memory requests made in the legacy vNPU mode. In mixed deployments, use node annotations to distinguish modes and avoid switching nodes that still run legacy-mode workloads.

devices.ascend.runtimeClassName is added only when the Pod does not already specify a RuntimeClass. The container must provide bash, the companion runtime libraries, and /hami-vnpu-core/limiter. An existing postStart hook is not automatically combined with the limiter startup command; arrange to start /hami-vnpu-core/limiter in the background from the existing hook. These scheduling settings do not replace installation of the companion device plugin and runtime.

AMD and PPU integration

The AMD plugin must report hami.io/node-amd-register; setting Node capacity alone is no longer sufficient for this release to recognize devices. The memory resource name changes from amd.com/gpu-memory to amd.com/gpumem. Platforms that integrate allocation information must also update the device-type identifier from AMDGPU to amd.

In container resources.limits, amd.com/gpumem is measured in MiB, and amd.com/gpucores: 40 requests 40% of the total compute capacity reported by the device. If memory is omitted or has a non-positive value, the full-card memory is reserved; if compute is also omitted, full-card compute is reserved as well.

To request an AMD compute share, add amd.com/gpucores to Helm values because it is not included in the default list:

devices:
  amd:
    customresources:
      - amd.com/gpu
      - amd.com/gpumem
      - amd.com/gpucores

The default scheduler managedResources in this Chart does not include the previous PPU mapping, and devices.alibaba.enabled: true does not add those entries. Existing PPU deployments must add and retain alibaba.com/ppu, alibaba.com/ppu-memory, and alibaba.com/gpu-core, with ignoredByScheduler: true for each, under extenders[].managedResources in the Chart or custom scheduler configuration, then replace the existing configuration. Do not overwrite an existing PPU resource mapping with this release's default configuration.

Health-check difference from the open-source baseline

This release does not include open-source HAMi v2.9.0 #1879. For devices such as Ascend that use the shared health check, a registration handshake that remains in Requesting for more than 60 seconds may cause the device cache to be cleared and affect scheduling for new tasks, even if Kubernetes still reports allocatable resources for the device. When diagnosing this, check both device-plugin registration and scheduler device state; Node resource counts alone do not confirm successful registration.

Init-container limitations

This release improves init-container request parsing and allocation-index handling, but GPU admission injection, the Enterprise Downward API path, and monitoring are still incomplete. Workloads with GPU init containers should not be considered fully supported. GPU memory auto-expansion also applies only to regular workload containers.

See the Chart and appVersion mapping or the software center.

Values reference

The following parameter descriptions are taken from the README bundled with the Chart 2.9.0-rc0 release package and are preserved as published. Use this version's values.yaml as the source of default configuration.

Expand values table

Global Configuration

Parameter	Description	Default Value
global.imageRegistry	Global Docker image registry	""
global.imagePullSecrets	Global Docker image pull secrets	[]
global.imageTag	Image tag	"v2.8.0"
global.gpuHookPath	GPU Hook path	/usr/local
global.labels	Global labels	{}
global.annotations	Global annotations	{}
global.managedNodeSelectorEnable	Whether to enable managed node selector	false
global.managedNodeSelector.usage	Managed node selector usage	"gpu"
nameOverride	Name override	""
fullnameOverride	Full name override	""
namespaceOverride	Namespace override	""

Resource Name Configuration

NVIDIA GPU Resources

Parameter	Description	Default Value
resourceName	GPU resource name	"nvidia.com/gpu"
resourceMem	GPU memory resource name	"nvidia.com/gpumem"
resourceMemPercentage	GPU memory percentage resource name	"nvidia.com/gpumem-percentage"
resourceCores	GPU core resource name	"nvidia.com/gpucores"
resourcePriority	GPU priority resource name	"nvidia.com/priority"

Cambricon MLU Resources

Parameter	Description	Default Value
mluResourceName	MLU resource name	"cambricon.com/vmlu"
mluResourceMem	MLU memory resource name	"cambricon.com/mlu.smlu.vmemory"
mluResourceCores	MLU core resource name	"cambricon.com/mlu.smlu.vcore"

Hygon DCU Resources

Parameter	Description	Default Value
dcuResourceName	DCU resource name	"hygon.com/dcunum"
dcuResourceMem	DCU memory resource name	"hygon.com/dcumem"
dcuResourceCores	DCU core resource name	"hygon.com/dcucores"

Metax GPU Resources

Parameter	Description	Default Value
metaxResourceName	GPU resource name	"metax-tech.com/sgpu"
metaxResourceCore	GPU core resource name	"metax-tech.com/vcore"
metaxResourceMem	GPU memory resource name	"metax-tech.com/vmemory"
metaxsGPUTopologyAware	GPU topology awareness	"false"

Enflame GCU Resources

Parameter	Description	Default Value
enflameResourceNameVGCU	vGCU resource name	"enflame.com/vgcu"
enflameResourceNameVGCUPercentage	vGCU percentage resource name	"enflame.com/vgcu-percentage"

Kunlunxin XPU Resources

Parameter	Description	Default Value
kunlunResourceName	XPU resource name	"kunlunxin.com/xpu"

Scheduler Configuration

Parameter	Description	Default Value
schedulerName	Scheduler name	"hami-scheduler"
scheduler.nodeName	Define node name, scheduler will schedule to this node	""
scheduler.overwriteEnv	Whether to overwrite environment variables	"false"
scheduler.defaultSchedulerPolicy.nodeSchedulerPolicy	Node scheduler policy	binpack
scheduler.defaultSchedulerPolicy.gpuSchedulerPolicy	GPU scheduler policy	spread
scheduler.metricsBindAddress	Metrics bind address	":9395"
scheduler.forceOverwriteDefaultScheduler	Whether to force overwrite default scheduler	true
scheduler.livenessProbe	Whether to enable liveness probe	false
scheduler.leaderElect	Whether to enable leader election	true
scheduler.replicas	Number of replicas	1

Kube Scheduler Configuration

Parameter	Description	Default Value
scheduler.kubeScheduler.enabled	Whether to run kube-scheduler container in scheduler pod	true
scheduler.kubeScheduler.image.registry	Kube scheduler image registry	"registry.cn-hangzhou.aliyuncs.com"
scheduler.kubeScheduler.image.repository	Kube scheduler image repository	"google_containers/kube-scheduler"
scheduler.kubeScheduler.image.tag	Kube scheduler image tag	""
scheduler.kubeScheduler.image.pullPolicy	Kube scheduler image pull policy	IfNotPresent
scheduler.kubeScheduler.image.pullSecrets	Kube scheduler image pull secrets	[]
scheduler.kubeScheduler.extraNewArgs	Extra new arguments	["--config=/config/config.yaml", "-v=4"]
scheduler.kubeScheduler.extraArgs	Extra arguments	["--policy-config-file=/config/config.json", "-v=4"]

Extender Configuration

Parameter	Description	Default Value
scheduler.extender.image.registry	Scheduler extender image registry	"docker.io"
scheduler.extender.image.repository	Scheduler extender image repository	"projecthami/hami"
scheduler.extender.image.tag	Scheduler extender image tag	""
scheduler.extender.image.pullPolicy	Scheduler extender image pull policy	IfNotPresent
scheduler.extender.image.pullSecrets	Scheduler extender image pull secrets	[]
scheduler.extender.extraArgs	Scheduler extender extra arguments	["--debug", "-v=4"]

Admission Webhook Configuration

Parameter	Description	Default Value
scheduler.admissionWebhook.enabled	Whether to enable admission webhook	true
scheduler.admissionWebhook.customURL.enabled	Whether to enable custom URL	false
scheduler.admissionWebhook.customURL.host	Custom URL host	127.0.0.1
scheduler.admissionWebhook.customURL.port	Custom URL port	31998
scheduler.admissionWebhook.customURL.path	Custom URL path	/webhook
scheduler.admissionWebhook.reinvocationPolicy	Reinvocation policy	Never
scheduler.admissionWebhook.failurePolicy	Failure policy	Ignore

TLS Certificate Configuration

Parameter	Description	Default Value
scheduler.certManager.enabled	Whether to use cert-manager to generate self-signed certificates	false
scheduler.patch.enabled	Whether to use kube-webhook-certgen to generate self-signed certificates	true
scheduler.patch.image.registry	Certgen image registry	"docker.io"
scheduler.patch.image.repository	Certgen image repository	"jettech/kube-webhook-certgen"
scheduler.patch.image.tag	Certgen image tag	"v1.5.2"
scheduler.patch.image.pullPolicy	Certgen image pull policy	IfNotPresent
scheduler.patch.imageNew.registry	New certgen image registry	"docker.io"
scheduler.patch.imageNew.repository	New certgen image repository	"liangjw/kube-webhook-certgen"
scheduler.patch.imageNew.tag	New certgen image tag	"v1.1.1"

Scheduler Service Configuration

Parameter	Description	Default Value
scheduler.service.type	Service type	NodePort
scheduler.service.httpPort	HTTP port	443
scheduler.service.schedulerPort	Scheduler NodePort	31998
scheduler.service.monitorPort	Monitor port	31993
scheduler.service.monitorTargetPort	Monitor target port	9395

Device Plugin Configuration

Parameter	Description	Default Value
devicePlugin.image.registry	Device plugin image registry	"docker.io"
devicePlugin.image.repository	Device plugin image repository	"projecthami/hami"
devicePlugin.image.tag	Device plugin image tag	""
devicePlugin.image.pullPolicy	Device plugin image pull policy	IfNotPresent
devicePlugin.image.pullSecrets	Device plugin image pull secrets	[]

Monitor Configuration

Parameter	Description	Default Value
devicePlugin.monitor.image.registry	Monitor image registry	"docker.io"
devicePlugin.monitor.image.repository	Monitor image repository	"projecthami/hami"
devicePlugin.monitor.image.tag	Monitor image tag	""
devicePlugin.monitor.image.pullPolicy	Monitor image pull policy	IfNotPresent
devicePlugin.monitor.image.pullSecrets	Monitor image pull secrets	[]
devicePlugin.monitor.ctrPath	Container path	/usr/local/vgpu/containers
devicePlugin.monitor.extraArgs	Monitor extra arguments	["-v=4"]
devicePlugin.monitor.extraEnvs	Monitor extra environments	{}

Device Plugin Other Configuration

Parameter	Description	Default Value
devicePlugin.deviceSplitCount	Integer type, default value: 10. Maximum number of tasks assigned to a single GPU device	10
devicePlugin.deviceMemoryScaling	Device memory scaling ratio	1
devicePlugin.deviceCoreScaling	Device core scaling ratio	1
devicePlugin.runtimeClassName	Runtime class name	""
devicePlugin.createRuntimeClass	Whether to create runtime class	false
devicePlugin.migStrategy	String type, "none" means ignore MIG functionality, "mixed" means allocate MIG devices through independent resources	"none"
devicePlugin.disablecorelimit	String type, "true" means disable core limit, "false" means enable core limit	"false"
devicePlugin.passDeviceSpecsEnabled	Whether to enable passing device specs	false
devicePlugin.extraArgs	Device plugin extra arguments	["-v=4"]
devicePlugin.nodeConfiguration.config	Node configuration for device plugin by json	An example of default configuration.
devicePlugin.nodeConfiguration.externalConfigName	Node configuration for device plugin by external congimap	""
devicePlugin.extraEnvs	Device plugin extra environments	{}

Device Plugin Service Configuration

Parameter	Description	Default Value
devicePlugin.service.type	Service type	NodePort
devicePlugin.service.httpPort	HTTP port	31992

Device Plugin Deployment Configuration

Parameter	Description	Default Value
devicePlugin.pluginPath	Plugin path	/var/lib/kubelet/device-plugins
devicePlugin.libPath	Library path	/usr/local/vgpu
devicePlugin.nvidiaNodeSelector	NVIDIA node selector	{"gpu": "on"}
devicePlugin.updateStrategy.type	Update strategy type	RollingUpdate
devicePlugin.updateStrategy.rollingUpdate.maxUnavailable	Maximum unavailable count	1

Device Configuration

AWS Neuron

Parameter	Description	Default Value
devices.awsneuron.customresources	Custom resources	["aws.amazon.com/neuron", "aws.amazon.com/neuroncore"]

Kunlunxin

Parameter	Description	Default Value
devices.kunlun.enabled	Whether to enable	true
devices.kunlun.customresources	Custom resources	["kunlunxin.com/xpu"]

Mthreads

Parameter	Description	Default Value
devices.mthreads.enabled	Whether to enable	true
devices.mthreads.customresources	Custom resources	["mthreads.com/vgpu"]

NVIDIA

Parameter	Description	Default Value
devices.nvidia.gpuCorePolicy	GPU core policy	default
devices.nvidia.libCudaLogLevel	CUDA library log level	1

Huawei Ascend

Parameter	Description	Default Value
devices.ascend.enabled	Whether to enable	false
devices.ascend.image	Image	""
devices.ascend.imagePullPolicy	Image pull policy	IfNotPresent
devices.ascend.extraArgs	Extra arguments	[]
devices.ascend.nodeSelector	Node selector	{"ascend": "on"}
devices.ascend.tolerations	Tolerations	[]
devices.ascend.customresources	Custom resources	["huawei.com/Ascend910A", "huawei.com/Ascend910A-memory", ...]

Iluvatar

Parameter	Description	Default Value
devices.iluvatar.enabled	Whether to enable	false
devices.iluvatar.customresources	Custom resources	["iluvatar.ai/BI-V150-vgpu", "iluvatar.ai/BI-V150.vMem","iluvatar.ai/BI-V150.vCore", ...]