v0.4.0

Added

• Every GPU of a claim gets its own HSA_CU_MASK entry and HIP_DEVICE_MEMORY_LIMIT_<index>, numbered in render node order, instead of all limits naming device 0 (#17).
• GPU health is reported to kubelet from gpu_busy_percent; a GPU that left the bus or does not answer is Unhealthy (#17).
• On a GPU without a dmem region, a static CDI createContainer hook (vgpu-hook-check) stops a memory slice whose image can not load the vGPU hook: musl, or glibc older than 2.34 (#17).

Changed

• The image is based on UBI 9.8 minimal (#18), and amd-hami-core is at its current main (#19).

Image: ghcr.io/project-hami/amd-dra-driver:0.4.0

v0.3.0

Added

• A brand attribute (Radeon or Instinct) from the product name (#15).

Fixed

• Topology values are range-checked before integer conversion, images and build tools are pinned by digest, fuzz tests were added, and go.etcd.io/etcd/client/pkg/v3 was upgraded (#14).

Image: ghcr.io/project-hami/amd-dra-driver:0.3.0

v0.2.0

Added

• The image bundles HAMi's vGPU hook from the amd-hami-core submodule and installs it on the node at --vgpu-hook-host-path.
• The uuid attribute and a productName for RDNA cards, from AMD SMI (#2).
• The VRAM of a memory slice is capped in the kernel through the pod's dmem.max (#3).
• The layout, CI, security policy and Scorecard of the other HAMi AMD repositories (#4), with README, STATUS.md and ONBOARDING.md (#1, #13).

Image: ghcr.io/project-hami/amd-dra-driver:0.2.0

v0.1.0

Added

• The GPUSharing feature gate: whole GPUs are published with consumable capacity, and each share gets its own CUs through HSA_CU_MASK and a memory limit for HAMi's vGPU hook.
• The image is built by CI and pushed to ghcr.io/project-hami/amd-dra-driver.

Image: ghcr.io/project-hami/amd-dra-driver:0.1.0