HAMi

Welcome Junis Lin to the HAMi community! He has been contributing to HAMi since April 2026, with improvements to the scheduler, DRA driver, and CI infrastructure, including bug fixes, feature enhancements, and technical discussions. Thanks for your contributions, and we look forward to continuing our collaboration! https://lnkd.in/gU5vjnpN #HAMi #OpenSource #Kubernetes #CloudNative

Later today at OSS Europe, I'm talking about the threat model for AI agents. I start from one assumption: every tool is hostile, so limits belong in the tool, not the agent. I'll go through deny-by-default scopes, human approval for destructive actions, and sandboxing from processes to microVMs. I'll also how HAMi fits into this and ask whether we really need GPU passthrough for microVMs. 📍 Forum Hall (Floor 2), 14:40 CEST Designing Permissioned AI Agents That Can Run Offline https://lnkd.in/dbDmyM7w #OSSummit #AIAgents #OpenSource

To support 2,000+ engineers running concurrent AI workloads, Sangfor Technologies implemented vGPU slicing to run 8x more models per GPU without adding extra hardware. By building a shared vGPU platform on Kubernetes with Volcano and HAMi, Sangfor enabled fine-grained resource slicing down to 1 percent compute and 256MB memory. This lets large models share a single card alongside smaller workloads under strict isolation. Overall GPU utilization increased by more than 3x, external model invocation costs dropped 50 percent, and platform stability stayed above 95 percent during peak traffic stress testing. Read the full case study: https://lnkd.in/e6utJBTQ #Kubernetes #AI

Thanks to the Cloud Native Computing Foundation (CNCF) CNAI Community for having us today! Mengxuan Li, dynamia.ai Co-founder & CTO and HAMi author, showed how HAMi gets more out of heterogeneous AI clusters: GPU sharing, isolation & topology-aware scheduling on K8s.

Happy Mid-Autumn Festival from the HAMi community! Wishing our contributors, users, partners, and friends around the world a joyful season of reunion, connection, and shared moments. #HAMi #OpenSource #Community #MidAutumnFestival

Your GPU really costs more than a car. Also your GPU: 12% busy. HAMi lets multiple Kubernetes workloads share one GPU, with per-pod limits on GPU memory and GPU compute. Same hardware, way less idle silicon. 60-second explainer, What's your cluster's real utilization? #HAMi #GPU #AIInfra #CloudNative

Want to see how HAMi manages heterogeneous GPU resources on Kubernetes? Join the CNCF AI Technical Community Group session on September 29 to learn how HAMi approaches topology-aware scheduling and GPU sharing across different accelerator vendors. A good opportunity to see HAMi in action and discuss practical challenges in heterogeneous AI infrastructure. Event details: https://lnkd.in/gU9YDQUj #HAMi #Kubernetes #AIInfrastructure #GPU #CNCF

New HAMi lab: serving models from a KitOps ModelKit on HAMi by Rudraksh Karpe and Shivay Lamba This hands-on guide shows how to package and pull a model as an OCI artifact, then serve it with SGLang or vLLM on HAMi-managed GPU shares. Try it here: https://lnkd.in/gc8gZtvJ #HAMi #KitOps #Kubernetes #AIInfrastructure #GPU

Boosting developer capacity by 15x on the exact same hardware sounds impossible until you rethink GPU scheduling. CETC Cloud needed to run production AI models and active R&D debugging on portable, 8-GPU devices. Instead of locking physical GPUs into dedicated silos, the technical team adopted Kubernetes and HAMi to virtualize and schedule heterogeneous GPUs on demand. The platform team expanded concurrent developer Pod capacity from 2 to 30 on the same device while maintaining stable performance for production text generation, embedding, and rerank pipelines. Explore how CETC Cloud built a dynamic resource foundation for edge AI: https://lnkd.in/eyFMhvTF Ready to optimize your AI infrastructure? Help build and refine GPU virtualization with Kubernetes and HAMi on GitHub: https://lnkd.in/gjem9Tnk #Kubernetes #AI

Welcome ADITYA RAUT to the HAMi community! Aditya has already made 15 merged contributions across the scheduler, device plugin, multi-vendor backends, and test coverage. Thanks for the solid work and continued contributions to HAMi. We’re glad to have you in the community and look forward to building more together. #HAMi #OpenSource #CloudNative #GPU #CNCF