Alauda Build of HAMi

Choose a HAMi Deployment Scenario

HAMi adds sharing, virtualization, and scheduling capabilities to supported vendor accelerator stacks. Use this page when workloads need HAMi-managed device allocation.

HAMi has the following capability and product scope:

  • The HAMi community provides extensible backends for GPUs, NPUs, and other vendor accelerators, including accelerators from multiple domestic vendors.
  • The Alauda product integrations documented here focus on NVIDIA GPUs and Huawei Ascend NPUs.
  • The available accelerator models and allocation modes depend on the installed Alauda product combination. Confirm the current scope in Compatibility.
Direct allocation without HAMi

If a target node pool needs only exclusive whole-card allocation and does not need HAMi scheduling, slicing, or resource management, use the vendor direct-allocation path instead.

For Ascend NPU, keep the NPU Operator native Ascend Device Plugin enabled for that node pool and do not enable HAMi Ascend Device Plugin on the same devices. Continue with Request Ascend NPU resources directly.

TOC

Choose a HAMi scenarioRelated tasks

Choose a HAMi scenario

Choose the scenario that matches the accelerator backend and the allocation behavior required by the workload:

HAMi scenarioChoose whenKey workload behaviorContinue
HAMi on NVIDIA GPUNVIDIA workloads need HAMi scheduling, whole-GPU allocation, memory or compute sharing, or HAMi-managed dynamic MIG.Use hami-scheduler and HAMi NVIDIA resource keys such as nvidia.com/gpualloc, with optional memory or compute resources.HAMi on NVIDIA GPU
HAMi on Ascend NPU (whole-card)Ascend workloads need exclusive whole-card allocation but should still use HAMi scheduling and resource management.Request only the model-specific count resource and set runtimeClassName: ascend. Whole-card requests do not depend on the hard/soft node mode.HAMi on Ascend NPU
HAMi on Ascend vNPUAscend workloads need fixed-template hard slicing or hami-core soft slicing.Hard slicing uses hami-vnpu-core=false and requests count plus memory. Soft slicing uses hami-vnpu-core=true, adds huawei.com/vnpu-mode: hami-core, and requests count plus memory and optional core percentage. Both paths use runtimeClassName: ascend.HAMi on Ascend vNPU
One Device Plugin owner per device pool

Only one Device Plugin can expose and allocate the same physical devices in a node pool.

For HAMi-managed Ascend scenarios, keep the NPU Operator Driver, OCI Runtime, ascend RuntimeClass, and any required Exporter components, but disable its native Ascend Device Plugin. HAMi Ascend Device Plugin becomes the allocation owner for whole-card, hard-slice, and soft-slice workloads. Follow Switch the Ascend Device Plugin Owner before creating HAMi-managed workloads.