Ascend Device Plugin

English | 中文

Introduction

This Ascend device plugin is implemented for NPU-Slicing for HAMi and volcano. It supports three modes:

1. Template-based Hard Slicing (vNPU)

Memory slicing is supported based on virtualization template. For detailed information, check template

2. Soft Slicing with Runtime Interception (hami-vnpu-core)

The hami-core backend uses libvnpu.so for memory accounting and compute sharing. It defaults to cooperative CANN Cube/Vector budgets (spatial), with explicit dynamic, fixed, and overdraft time-scheduling policies available. See the HAMi configuration guide and hami-vnpu-core.

Note: hami-vnpu-core currently only supports ARM platforms.

3. Soft Slicing with ubs-virt-enpu (vCANN-RT)

The enpu backend provides compute and memory quotas through vCANN-RT, with optional memory oversubscription through a compatible enpu-manager/runtime. It supports A2/910B and A3/910C; the single-DIE mode prerequisite applies only to A3/910C. An existing ConfigMap entry for 910A does not imply ENPU support.

Before enabling ENPU, follow the official vCANN-RT configuration guide for the hardware, driver/CANN, runtime assets and container prerequisites. Use the HAMi ENPU guide for integration settings and the complete ordinary soft-slicing and mem-swap Pod examples. Mem-swap requires a version that includes that feature; annotations alone do not enable it in an older release.

Deployment & Usage

Prerequisites, deployment steps and usage examples differ depending on which scheduler you use:

• HAMi
• Volcano

Compile

update submodule:

git submodule update --init --recursive

make all

Build image

docker buildx build -t $IMAGE_NAME .

Build an ENPU-enabled image

CI and plain docker build reuse the public ENPU release 1.0.0 / CANN 9.1.0 runtime image pinned by digest in the Dockerfile, verify its assets, and bundle them alongside hami-vnpu-core. CI builds the AMD64 and ARM64 plugin images on native runners; only main and version-tag pushes publish a combined manifest. Prepare the existing hami-vnpu-core assets before building locally. To rebuild ENPU from source instead, run make docker-enpu VERSION=<image-tag> on an ARM64 Linux host with Docker, Git and Python 3. No NPU or host driver installation is required for compilation. See release build prerequisites and options.

License