#HAMi

HAMi is a sharing and scheduling product path for accelerator sharing, virtualization, and scheduling behavior. It builds on vendor integrations such as NVIDIA GPU or Ascend NPU, depending on the product version and backend support.

#Use HAMi when

• multiple workloads need to share accelerator devices;
• users need accelerator sharing modes such as virtual GPU, virtual NPU, hard partitioning, or soft partitioning when supported by the target backend;
• the product documentation confirms support for the target hardware, runtime, and ACP version.

#HAMi backends

HAMi documentation owns the HAMi part of a combined deployment: HAMi installation, scheduler behavior, HAMi device plugins, resource keys, sharing semantics, monitoring integration, and troubleshooting.

HAMi backend pages should point to the vendor integration prerequisites they depend on:

• HAMi on NVIDIA GPU depends on the NVIDIA GPU base path for driver, runtime, toolkit, and NVIDIA-specific prerequisites.
• HAMi on Ascend NPU or Ascend vNPU depends on the Ascend base path for NPU nodes, driver lifecycle, runtime integration, and CDI prerequisites.

For accelerator allocation without HAMi sharing semantics, use the corresponding vendor integration:

• For NVIDIA GPU allocation without HAMi sharing, use the NVIDIA vendor integration.
• For Ascend NPU allocation without HAMi sharing, use the Ascend vendor integration.

#Related concepts

• Device Plugin
• DRA
• RuntimeClass and CDI
• Accelerator resource quota
• GPU accelerator family
• NPU accelerator family

#Product documentation

Use the HAMi product site for installation, compatibility, release notes, troubleshooting, and backend-specific behavior.