Project-HAMi/HAMi

Health v2 · maintenance only

Adoption not part of the score

4438 stars · 794 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

HAMi (Heterogeneous AI Computing Virtualization Middleware) is a CNCF Incubating, Kubernetes-native GPU virtualization and scheduling middleware that lets multiple workloads share GPUs and other AI accelerators with hard memory/compute isolation. It works via device plugins, a mutating webhook, and a custom scheduler, supporting NVIDIA, Ascend, Cambricon, Hygon, and other vendors without application code changes.

Use cases

• share one GPU across multiple Kubernetes pods
• slice GPU memory for AI inference workloads
• schedule heterogeneous accelerators like NPU and MLU in Kubernetes
• prevent one pod from OOMing other tasks on a shared GPU
• improve GPU utilization in an AI cluster
• allocate fractional GPU resources (e.g. 1/2 or 1/4 of a card) to jobs
• monitor GPU memory and core usage across vendors

When to choose

• you run Kubernetes with expensive GPUs that are underutilized by exclusive allocation
• you need VRAM/compute isolation for shared GPUs without MIG-capable hardware
• you manage mixed-vendor accelerators (NVIDIA, Ascend, Cambricon, Hygon, etc.) and want one scheduling workflow
• you want device-aware scheduling policies like binpack, spread, or topology-aware placement

When to avoid

• you need hardware-level isolation guarantees such as NVIDIA MIG on A100/H100
• your cluster is not Kubernetes-based
• you only run a single workload per GPU and don't need sharing
• you require vendor-official support rather than a community middleware layer

Facets

service · maturity active

container-orchestration scheduling monitoring gpu-computing cloud-computing gpu-computing infrastructure-as-code machine-learning go self-hosted gpu-virtualization vgpu device-plugin gpu-sharing heterogeneous-accelerators cncf ai-infrastructure gpu-scheduling nvidia ascend cambricon containers kubernetes linux docker

Member repositories

For agents

markdown · JSON · MCP: product_card(name="Project-HAMi/HAMi")