diff --git a/blog/kubecon-india-2026-recap/index.md b/blog/kubecon-india-2026-recap/index.md new file mode 100644 index 000000000..281c7c4f2 --- /dev/null +++ b/blog/kubecon-india-2026-recap/index.md @@ -0,0 +1,152 @@ +--- +title: "HAMi at KubeCon + CloudNativeCon India 2026: Bringing GPU Sharing to the Community" +date: "2026-06-29" +description: "A recap of HAMi at KubeCon + CloudNativeCon India 2026 in Mumbai, including the keynote demo, booth demonstrations, and community discussions around Kubernetes-native GPU sharing." +tags: ["KubeCon", "GPU", "Kubernetes", "AI", "India"] +authors: [hami_community] +--- + +Held on June 18-19, 2026, in Mumbai, India, [KubeCon + CloudNativeCon India 2026](https://events.linuxfoundation.org/kubecon-cloudnativecon-india/) brought together cloud native practitioners, platform engineers, AI infrastructure teams, and open source contributors from across the ecosystem. As AI emerged as one of the conference's defining themes, HAMi showcased how Kubernetes-native GPU sharing helps organizations maximize accelerator utilization while maintaining workload isolation and operational flexibility. + +From the opening keynote to live booth demonstrations and technical discussions with engineering teams, the event highlighted a growing industry focus: making expensive GPU infrastructure practical for multi-tenant AI workloads. + + + +## AI Infrastructure Takes Center Stage + +The conference opened with a strong focus on AI infrastructure. During the opening keynote, Chris Aniszczyk highlighted how cloud native technologies are becoming the foundation for the next generation of AI platforms. + +HAMi's presence at KubeCon India built directly on that message by demonstrating how Kubernetes can share GPU resources across multiple teams and inference workloads without assigning a full GPU or a separate cluster to every workload. + +## Keynote Demo: Multi-Tenant AI Inference on Kubernetes + +![Saiyam Pathak presenting the HAMi GPU sharing demo during the KubeCon India keynote](/img/kubecon-india-2026-recap/keynote-demo.jpg) + +During [Saiyam Pathak's keynote](https://kccncind2026.sched.com/event/2NgRQ/keynote-from-platforms-to-ai-factories-has-kubernetes-solved-it-saiyam-pathak-principal-developer-advocate-vcluster), HAMi powered a live demonstration showing GPU sharing in action. + +The architecture combined: + +- A MacBook running the Kubernetes control plane +- NVIDIA DGX Spark as the GPU worker +- vCluster for tenant isolation +- HAMi for fractional GPU allocation +- Multiple AI inference workloads running simultaneously on a shared GPU + +![Demo architecture showing two vCluster-backed teams sharing one DGX Spark GPU through HAMi](/img/kubecon-india-2026-recap/demo-architecture.png) + +To make the concepts relatable, the demo used a Mumbai-themed application featuring local trains, dabbawalas, and kaali peeli taxis. Images captured on a mobile application were routed through Kubernetes to AI models running on the shared GPU infrastructure. + +Rather than simply explaining GPU virtualization, the keynote demonstrated how multiple independent AI teams can share the same physical accelerator while maintaining isolated Kubernetes environments. + +The demo illustrated how HAMi and vCluster together support practical multi-tenant AI platforms that are resource efficient and cloud native. + +## Live Booth Demonstrations + +Throughout the Project Pavilion, visitors experienced two hands-on demonstrations highlighting real-world AI inference on Kubernetes. + +For readers who want to reproduce parts of the vLLM workflow, see the [HAMi vLLM lab](https://project-hami.io/tutorials/labs/hami-vllm). + +### Demo 1: Production AI Inference with vLLM and SGLang + +The first demonstration showcased production-style LLM serving using vLLM and SGLang on Kubernetes, with HAMi managing GPU allocation across heterogeneous hardware. The demo used two model delivery approaches: serving models directly from Hugging Face, and packaging models with KitOps ModelKits delivered through Kubernetes initContainers, showing flexible deployment strategies for AI platforms. + +The inference stack was deployed on: + +- NVIDIA H100 GPUs provided by Nebius +- NVIDIA A10 GPUs provided by Reza Jelveh + +Visitors could observe: + +- Production-style LLM serving +- Kubernetes-native deployment workflows +- Resource-aware GPU scheduling +- Efficient GPU sharing across inference workloads +- End-to-end request routing through modern inference engines + +By combining HAMi with popular inference frameworks such as vLLM and SGLang, the demo showed how organizations can deploy scalable AI services while improving GPU utilization. + +### Demo 2: Dynamic GPU Sharing + +The second demonstration focused on HAMi's GPU virtualization capabilities. + +Attendees explored how multiple AI workloads can coexist on the same physical GPU with explicit resource controls using: + +- Fractional GPU allocation +- Dynamic resource assignment +- Flexible scheduling policies +- GPU virtualization +- Kubernetes-native deployment + +Visitors also validated allocations using NVIDIA tooling while observing inference requests flowing through shared infrastructure in real time. + +Together, both demonstrations showcased how Kubernetes can evolve beyond simple GPU allocation toward more intelligent accelerator scheduling for modern AI platforms. + +## Conversations from the Booth + +Beyond the demonstrations, the booth became a hub for technical discussions with engineers, startup founders, platform teams, CTOs, and students exploring AI infrastructure. + +Although attendees represented different industries, many conversations centered around the same challenge: + +> How can organizations share expensive GPU resources with clearer isolation and more predictable workload placement? + +### HyperVerge: Memory Isolation for Multi-Tenant AI + +One of the most detailed discussions was with HyperVerge. + +Their engineering team described previous experiences with GPU sharing solutions that lacked clear GPU memory isolation controls between inference workloads. + +For organizations serving multiple customers on shared infrastructure, this presents significant operational challenges. + +The discussion focused on how HAMi approaches: + +- GPU memory isolation +- Intelligent scheduling +- Bin-packing strategies +- Spread scheduling +- Multi-tenant workload placement + +The ability to combine efficient GPU utilization with stronger workload isolation controls was particularly valuable for enterprise AI platforms. + +### Enterprise Scheduling Discussions + +Engineers from BlackRock were interested in HAMi's scheduling algorithms and how placement decisions are made across heterogeneous GPU workloads. + +Similarly, the InfraCloud team discussed how HAMi integrates into existing Kubernetes environments while comparing traditional Kubernetes Device Plugins, Dynamic Resource Allocation (DRA), and HAMi's scheduling model. + +These discussions reflected a broader trend across enterprise Kubernetes platforms: + +> Organizations are actively looking for smarter ways to eliminate GPU underutilization while supporting increasingly diverse AI workloads. + +## Community Behind the Booth + +![HAMi community contributors at the KubeCon India Project Pavilion booth](/img/kubecon-india-2026-recap/hami-booth.jpg) + +Open source projects succeed because of their communities. + +A huge thank you to everyone who helped make the HAMi booth possible: + +- [Shivam Kumar](https://github.com/maishivamhoo123) +- [Satyam Soni](https://github.com/satyampsoni) +- [Rudraksh Karpe](https://github.com/rudrakshkarpe) +- [Shivay Lamba](https://github.com/shivaylamba) +- [Reza Jelveh](https://github.com/fishman) + +Special thanks to [Jimmy Song](https://github.com/rootsongjc) for preparing booth materials and supporting the community throughout the event. + +## Looking Ahead + +The conversations throughout KubeCon India reinforced a clear message. + +Organizations are moving beyond the traditional "one GPU per team" model and are actively searching for Kubernetes-native solutions that enable efficient, scalable GPU sharing with practical isolation controls. + +As AI infrastructure continues to evolve, technologies such as HAMi will play an increasingly important role in helping organizations: + +- Increase GPU utilization +- Reduce infrastructure costs +- Support multi-tenant AI platforms +- Deploy production inference on Kubernetes +- Scale accelerator infrastructure more efficiently + +Thank you to everyone who visited the booth, attended the keynote, shared feedback, and contributed ideas. + +We're excited to continue building the future of GPU infrastructure for Kubernetes and we look forward to seeing you at the next KubeCon. diff --git a/i18n/zh/docusaurus-plugin-content-blog/kubecon-india-2026-recap/index.md b/i18n/zh/docusaurus-plugin-content-blog/kubecon-india-2026-recap/index.md new file mode 100644 index 000000000..64adc77fc --- /dev/null +++ b/i18n/zh/docusaurus-plugin-content-blog/kubecon-india-2026-recap/index.md @@ -0,0 +1,152 @@ +--- +title: "HAMi 亮相 KubeCon + CloudNativeCon India 2026：将 GPU 共享带给社区" +date: "2026-06-29" +description: "回顾 HAMi 在孟买 KubeCon + CloudNativeCon India 2026 的亮相，包括 Keynote 现场演示、展台 Demo，以及围绕 Kubernetes 原生 GPU 共享的社区交流。" +tags: ["KubeCon", "GPU", "Kubernetes", "AI", "India"] +authors: [hami_community] +--- + +2026 年 6 月 18-19 日，[KubeCon + CloudNativeCon India 2026](https://events.linuxfoundation.org/kubecon-cloudnativecon-india/) 在印度孟买举行，来自云原生生态的实践者、平台工程师、AI 基础设施团队和开源贡献者齐聚一堂。随着 AI 成为本届大会最重要的主题之一，HAMi 展示了 Kubernetes 原生 GPU 共享如何帮助组织提升加速器利用率，同时保持工作负载隔离和运维灵活性。 + +从开场 Keynote 到展台现场演示，再到与工程团队的技术讨论，本次活动凸显了一个越来越明确的行业关注点：让昂贵的 GPU 基础设施真正适用于多租户 AI 工作负载。 + + + +## AI 基础设施成为大会焦点 + +本届大会一开始就高度聚焦 AI 基础设施。在开场 Keynote 中，Chris Aniszczyk 强调，云原生技术正在成为下一代 AI 平台的基础。 + +HAMi 在 KubeCon India 的亮相正是对这一趋势的直接回应：通过现场演示，展示 Kubernetes 如何在多个团队和推理工作负载之间共享 GPU 资源，而不必为每个工作负载分配一整张 GPU 或一套独立集群。 + +## Keynote Demo：Kubernetes 上的多租户 AI 推理 + +![Saiyam Pathak 在 KubeCon India Keynote 中展示 HAMi GPU 共享演示](/img/kubecon-india-2026-recap/keynote-demo.jpg) + +在 [Saiyam Pathak 的 Keynote](https://kccncind2026.sched.com/event/2NgRQ/keynote-from-platforms-to-ai-factories-has-kubernetes-solved-it-saiyam-pathak-principal-developer-advocate-vcluster) 中，HAMi 支撑了一场展示 GPU 共享能力的现场演示。 + +演示架构包括： + +- 运行 Kubernetes 控制平面的 MacBook +- 作为 GPU 工作节点的 NVIDIA DGX Spark +- 用于租户隔离的 vCluster +- 用于细粒度 GPU 分配的 HAMi +- 多个 AI 推理工作负载同时运行在共享 GPU 上 + +![演示架构：两个基于 vCluster 的团队通过 HAMi 共享一张 DGX Spark GPU](/img/kubecon-india-2026-recap/demo-architecture-zh.png) + +为了让概念更直观，演示使用了带有孟买本地特色的应用场景，包括本地火车、dabbawala 送餐员和 kaali peeli 出租车。移动应用采集的图片会被路由到 Kubernetes，并交由运行在共享 GPU 基础设施上的 AI 模型完成推理。 + +这场 Keynote 并不只是解释 GPU 虚拟化概念，而是展示了多个独立 AI 团队如何共享同一张物理加速卡，同时保持彼此隔离的 Kubernetes 环境。 + +该演示说明，HAMi 与 vCluster 结合后，可以支撑资源高效、云原生的多租户 AI 平台。 + +## 展台现场演示 + +在 Project Pavilion，参观者体验了两场动手演示，展示 Kubernetes 上真实的 AI 推理场景。 + +如果你希望复现其中一部分 vLLM 工作流，可以参考 [HAMi vLLM 实验](https://project-hami.io/tutorials/labs/hami-vllm)。 + +### Demo 1：基于 vLLM 与 SGLang 的生产级 AI 推理 + +第一场演示展示了在 Kubernetes 上使用 vLLM 和 SGLang 进行生产级 LLM serving，并由 HAMi 管理异构硬件上的 GPU 分配。演示使用了两种模型交付方式：直接从 Hugging Face 加载模型，以及通过 Kubernetes initContainers 分发 KitOps ModelKits 打包的模型，从而展示 AI 平台中灵活的部署策略。 + +推理栈运行在： + +- Nebius 提供的 NVIDIA H100 GPU +- Reza Jelveh 提供的 NVIDIA A10 GPU + +参观者可以看到： + +- 生产级 LLM serving +- Kubernetes 原生部署工作流 +- 感知资源的 GPU 调度 +- 跨推理工作负载的高效 GPU 共享 +- 通过现代推理引擎完成端到端请求路由 + +通过将 HAMi 与 vLLM、SGLang 等流行推理框架结合，这场演示展示了组织如何在提升 GPU 利用率的同时部署可扩展的 AI 服务。 + +### Demo 2：动态 GPU 共享 + +第二场演示聚焦 HAMi 的 GPU 虚拟化能力。 + +参观者了解了多个 AI 工作负载如何通过显式资源控制共存于同一张物理 GPU 上，相关能力包括： + +- 分数 GPU 分配 +- 动态资源分配 +- 灵活调度策略 +- GPU 虚拟化 +- Kubernetes 原生部署 + +参观者还通过 NVIDIA 工具验证资源分配情况，并实时观察推理请求如何流经共享基础设施。 + +这两场演示共同展示了 Kubernetes 如何从简单的 GPU 分配，进一步演进到面向现代 AI 平台的智能加速器调度。 + +## 展台交流 + +除演示之外，HAMi 展台也成为技术交流的中心，吸引了工程师、创业者、平台团队、CTO 以及正在探索 AI 基础设施的学生。 + +虽然参会者来自不同领域，但许多讨论都围绕同一个问题展开： + +> 组织如何在更清晰的隔离和更可预测的工作负载放置基础上，共享昂贵的 GPU 资源？ + +### HyperVerge：面向多租户 AI 的显存隔离 + +其中一次深入交流来自 HyperVerge。 + +他们的工程团队分享了此前使用 GPU 共享方案的经历：一些方案缺少推理工作负载之间清晰的 GPU 显存隔离控制。 + +对于在共享基础设施上服务多个客户的组织来说，这会带来显著的运维挑战。 + +这次讨论聚焦 HAMi 如何处理： + +- GPU 显存隔离 +- 智能调度 +- Bin-packing 策略 +- Spread 调度 +- 多租户工作负载放置 + +对于企业级 AI 平台而言，将高效 GPU 利用率与更强的工作负载隔离控制结合起来，具有很高价值。 + +### 企业级调度讨论 + +来自 BlackRock 的工程师关注 HAMi 的调度算法，以及它如何在异构 GPU 工作负载之间做出放置决策。 + +类似地，InfraCloud 团队也讨论了 HAMi 如何集成到现有 Kubernetes 环境中，并对比了传统 Kubernetes Device Plugin、动态资源分配（DRA）和 HAMi 调度模型。 + +这些讨论反映了企业级 Kubernetes 平台中的一个更广泛趋势： + +> 组织正在积极寻找更智能的方式，以消除 GPU 利用率不足问题，同时支持日益多样化的 AI 工作负载。 + +## 展台背后的社区 + +![HAMi 社区贡献者在 KubeCon India Project Pavilion 展台](/img/kubecon-india-2026-recap/hami-booth.jpg) + +开源项目的成功离不开社区。 + +感谢所有帮助 HAMi 展台顺利运行的人： + +- [Shivam Kumar](https://github.com/maishivamhoo123) +- [Satyam Soni](https://github.com/satyampsoni) +- [Rudraksh Karpe](https://github.com/rudrakshkarpe) +- [Shivay Lamba](https://github.com/shivaylamba) +- [Reza Jelveh](https://github.com/fishman) + +特别感谢 [Jimmy Song](https://github.com/rootsongjc) 准备展台材料，并在整个活动期间支持社区。 + +## 展望 + +KubeCon India 期间的交流强化了一个清晰信号。 + +组织正在从传统的“每个团队一张 GPU”模式中走出来，积极寻找 Kubernetes 原生方案，以实现高效、可扩展、具备实际隔离控制的 GPU 共享。 + +随着 AI 基础设施持续演进，HAMi 这类技术将在帮助组织实现以下目标时发挥越来越重要的作用： + +- 提升 GPU 利用率 +- 降低基础设施成本 +- 支撑多租户 AI 平台 +- 在 Kubernetes 上部署生产级推理 +- 更高效地扩展加速器基础设施 + +感谢每一位来到展台、参加 Keynote、提供反馈并贡献想法的朋友。 + +我们很高兴继续构建 Kubernetes GPU 基础设施的未来，也期待在下一届 KubeCon 再见。 diff --git a/src/css/custom.css b/src/css/custom.css index 24dd42b04..d1a1ad269 100644 --- a/src/css/custom.css +++ b/src/css/custom.css @@ -375,7 +375,7 @@ body.hami-lightbox-open { z-index: 10000; background: rgba(6, 10, 16, 0.88); display: grid; - grid-template-rows: 1fr auto; + grid-template-rows: minmax(0, 1fr) auto; align-items: center; justify-items: center; padding: clamp(18px, 3vw, 40px); @@ -389,7 +389,7 @@ body.hami-lightbox-open { .hami-lightbox__image { max-width: min(96vw, 1600px); - max-height: min(86vh, 1100px); + max-height: min(82vh, 1100px); width: auto; height: auto; background: #ffffff; @@ -403,7 +403,7 @@ body.hami-lightbox-open { fit both dimensions (viewBox + preserveAspectRatio), so the whole diagram is always visible without internal scrolling. */ width: min(96vw, 1600px); - height: min(86vh, 1100px); + height: min(82vh, 1100px); display: flex; align-items: center; justify-content: center; @@ -447,6 +447,24 @@ html[data-theme="dark"] .hami-lightbox__svg { line-height: 1.45; } +.hami-lightbox__caption[hidden] { + display: none; +} + +.hami-lightbox__image[hidden], +.hami-lightbox__svg[hidden] { + display: none; +} + +.hami-lightbox[data-has-caption="false"] { + grid-template-rows: minmax(0, 1fr); +} + +.hami-lightbox[data-has-caption="false"] .hami-lightbox__image, +.hami-lightbox[data-has-caption="false"] .hami-lightbox__svg { + max-height: min(90vh, 1100px); +} + .hami-lightbox__close { position: absolute; top: 14px; diff --git a/src/theme/utils/useImageLightbox.js b/src/theme/utils/useImageLightbox.js index bc1c88448..46646c155 100644 --- a/src/theme/utils/useImageLightbox.js +++ b/src/theme/utils/useImageLightbox.js @@ -74,6 +74,14 @@ function ensureLightbox() { svgHost.hidden = true; svgHost.replaceChildren(); caption.textContent = ""; + root.dataset.hasCaption = "false"; + }; + + const setCaption = (captionText = "") => { + const normalizedCaption = captionText.trim(); + caption.textContent = normalizedCaption; + caption.hidden = !normalizedCaption; + root.dataset.hasCaption = normalizedCaption ? "true" : "false"; }; root.addEventListener("click", (event) => { @@ -93,13 +101,13 @@ function ensureLightbox() { } }); - root.__hamiLightboxOpen = ({ src, alt }) => { + root.__hamiLightboxOpen = ({ src, alt, caption: captionText }) => { lightboxImage.src = src; lightboxImage.alt = alt || ""; lightboxImage.hidden = false; svgHost.hidden = true; svgHost.replaceChildren(); - caption.textContent = alt || ""; + setCaption(captionText || alt || ""); root.hidden = false; document.body.classList.add("hami-lightbox-open"); }; @@ -118,7 +126,7 @@ function ensureLightbox() { svgHost.hidden = false; lightboxImage.hidden = true; lightboxImage.removeAttribute("src"); - caption.textContent = captionText || ""; + setCaption(captionText || ""); root.hidden = false; document.body.classList.add("hami-lightbox-open"); }; @@ -126,6 +134,18 @@ function ensureLightbox() { return root; } +function getImageCaption(image) { + const figureCaption = image.closest("figure")?.querySelector("figcaption")?.textContent; + + return ( + figureCaption || + image.getAttribute("aria-label") || + image.getAttribute("title") || + image.getAttribute("alt") || + "" + ); +} + function handleImageClick(event) { const target = event.target; if (!(target instanceof Element)) { @@ -166,9 +186,11 @@ function handleImageClick(event) { } const lightbox = ensureLightbox(); + const caption = getImageCaption(image); lightbox.__hamiLightboxOpen({ src: image.currentSrc || image.src, - alt: image.alt || "", + alt: image.alt || caption, + caption, }); } diff --git a/static/img/kubecon-india-2026-recap/demo-architecture-zh.png b/static/img/kubecon-india-2026-recap/demo-architecture-zh.png new file mode 100644 index 000000000..0a58763f7 Binary files /dev/null and b/static/img/kubecon-india-2026-recap/demo-architecture-zh.png differ diff --git a/static/img/kubecon-india-2026-recap/demo-architecture.png b/static/img/kubecon-india-2026-recap/demo-architecture.png new file mode 100644 index 000000000..4a8fbc1ed Binary files /dev/null and b/static/img/kubecon-india-2026-recap/demo-architecture.png differ diff --git a/static/img/kubecon-india-2026-recap/hami-booth.jpg b/static/img/kubecon-india-2026-recap/hami-booth.jpg new file mode 100644 index 000000000..ef76c9db7 Binary files /dev/null and b/static/img/kubecon-india-2026-recap/hami-booth.jpg differ diff --git a/static/img/kubecon-india-2026-recap/keynote-demo.jpg b/static/img/kubecon-india-2026-recap/keynote-demo.jpg new file mode 100644 index 000000000..b9364cc4c Binary files /dev/null and b/static/img/kubecon-india-2026-recap/keynote-demo.jpg differ