diff --git a/blog/hami-at-kubecon-eu-2026/index.md b/blog/hami-at-kubecon-eu-2026/index.md index acdec16d..a378e3ea 100644 --- a/blog/hami-at-kubecon-eu-2026/index.md +++ b/blog/hami-at-kubecon-eu-2026/index.md @@ -1,141 +1,141 @@ --- -title: "HAMi 即将亮相 KubeCon Europe 2026：构建 Kubernetes 中的 GPU 资源层" +title: "HAMi at KubeCon Europe 2026: Building the GPU Resource Layer in Kubernetes" date: "2026-03-19" -description: "HAMi 将在 KubeCon Europe 2026 的多项活动中亮相，包括 Project Pavilion 展台、技术分享、主舞台 Demo 等。作为 CNCF Sandbox 项目，HAMi 关注的 GPU 虚拟化、共享与调度问题，正在与 Kubernetes 生态中的 AI 基础设施议题发生越来越直接的交汇。" +description: "HAMi will be featured in multiple activities at KubeCon Europe 2026, including Project Pavilion booth, technical sessions, main stage demo, and post-conference AI events. As a CNCF Sandbox project, HAMi focuses on GPU virtualization, sharing, and scheduling, which is increasingly intersecting with AI infrastructure topics in the Kubernetes ecosystem." tags: ["KubeCon", "GPU", "Kubernetes", "AI"] authors: [hami_community] --- -下周，HAMi 将在 [KubeCon + CloudNativeCon Europe 2026](https://events.linuxfoundation.org/kubecon-cloudnativecon-europe/) 的多项活动中亮相，包括 Project Pavilion 展台、技术分享、主舞台 Demo，以及会后 AI 相关活动。 +Next week, HAMi will be featured in multiple activities at [KubeCon + CloudNativeCon Europe 2026](https://events.linuxfoundation.org/kubecon-cloudnativecon-europe/), including Project Pavilion booth, technical sessions, main stage demo, and post-conference AI-related events. -作为 CNCF Sandbox 项目，HAMi 关注的 GPU 虚拟化、共享与调度问题，正在与 Kubernetes 生态中的 AI 基础设施议题发生越来越直接的交汇。KubeCon + CloudNativeCon Europe 2026 将于 3 月 23 日至 26 日在阿姆斯特丹举行，其中 3 月 23 日为 pre-event programming，3 月 24 日至 26 日为主会期。 +As a CNCF Sandbox project, HAMi focuses on GPU virtualization, sharing, and scheduling, which is increasingly intersecting with AI infrastructure topics in the Kubernetes ecosystem. KubeCon + CloudNativeCon Europe 2026 will be held in Amsterdam from March 23-26, with March 23 as pre-event programming and March 24-26 as the main conference. -![KubeCon EU 2026 吸引了 13,000 名参会者](/img/blog-hami-at-kubecon-eu-2026-kubecon.png) +![KubeCon EU 2026 attracts 13,000 attendees](/img/blog-hami-at-kubecon-eu-2026-kubecon.png) -## 为什么这次 KubeCon 值得关注？ +## Why This KubeCon Matters -如果把过去几年云原生社区的讨论放在一起看，会发现一个越来越清晰的变化：AI 正在从应用层问题，进入 Kubernetes 的资源层、调度层与控制层。 +Looking at the cloud native community discussions over the past few years, a clear trend emerges: AI is moving from the application layer into Kubernetes' resource layer, scheduling layer, and control layer. -围绕 GPU 的讨论也不再停留在"设备可见性"或"驱动可用性"，而是进一步延伸到共享、切分、利用率、多租户隔离，以及 AI workload 的调度语义等问题。 +The discussions around GPUs are no longer limited to "device visibility" or "driver availability," but have extended to sharing, partitioning, utilization, multi-tenant isolation, and AI workload scheduling semantics. -KubeCon Europe 2026 的官方议程中，keynote、AI 相关 session、Project Pavilion 与 co-located events 都体现出这一趋势。 +The official agenda of KubeCon Europe 2026 reflects this trend across keynotes, AI-related sessions, Project Pavilion, and co-located events. -在这个背景下，HAMi 所对应的问题空间也变得更加明确：不是简单地"让 Kubernetes 能识别 GPU"，而是让 GPU 进一步成为一种可以被抽象、被共享、被调度的资源层能力。 +In this context, HAMi's problem space becomes clearer: it's not simply about "making Kubernetes recognize GPUs," but making GPUs a resource layer capability that can be abstracted, shared, and scheduled. -这也是为什么本次 KubeCon 对 HAMi 社区而言，不只是一次项目展示，更是一次和更大范围云原生生态对话的机会。 +This is why this KubeCon is more than just a project showcase for the HAMi community—it's an opportunity to engage with the broader cloud native ecosystem. -## 在 KubeCon 现场，如何找到 HAMi？ +## Finding HAMi at KubeCon -![欢迎来到 HAMi 展台](/img/blog-hami-at-kubecon-eu-2026-booth.png) +![Welcome to HAMi Booth](/img/blog-hami-at-kubecon-eu-2026-booth.png) -HAMi 将在 Project Pavilion 设置展台，方便与社区成员、用户和维护者进行现场交流。 +HAMi will have a booth at Project Pavilion for in-person exchanges with community members, users, and maintainers. -- **Booth**：**P-13B** -- **时间**： - - **3 月 24 日（周二）15:10–19:00** - - **3 月 26 日（周四）12:30–14:00** +- **Booth**: **P-13B** +- **Times**: + - **March 24 (Tuesday) 15:10–19:00** + - **March 26 (Thursday) 12:30–14:00** -如果你会到现场，欢迎来到 HAMi Booth，一起交流这些话题： +If you're attending, stop by the HAMi Booth to discuss: -- Kubernetes 中的 GPU 虚拟化与共享 -- AI workload 的资源调度与利用率优化 -- 多租户 GPU 资源管理 -- HAMi 与 [Volcano](https://volcano.sh/)、[Kueue](https://kueue.sigs.k8s.io/)、[vLLM](https://github.com/vllm-project/vllm) 等生态项目的协同 +- GPU virtualization and sharing in Kubernetes +- Resource scheduling and utilization optimization for AI workloads +- Multi-tenant GPU resource management +- HAMi's integration with ecosystem projects like [Volcano](https://volcano.sh/), [Kueue](https://kueue.sigs.k8s.io/), [vLLM](https://github.com/vllm-project/vllm), and others -Project Pavilion 是 KubeCon 主展区中的项目展示区域，面向社区项目、维护者与开发者交流。 +Project Pavilion is the project showcase area within the main KubeCon exhibition, designed for community projects, maintainers, and developers to connect. -## HAMi @ KubeCon Europe 2026 活动一览 +## HAMi @ KubeCon Europe 2026 Event Overview ### 1. Opening Keynote -- **时间**：3 月 24 日 09:00–09:35 -- **地点**：Hall 12 -- **讲者**：Jonathan Bryce (Linux Foundation Executive Director) & Chris Aniszczyk (CNCF CTO) -- **议程**：[Keynote: Welcome + Opening Remarks](https://kccnceu2026.sched.com/event/2CtKk/keynote-welcome-+-opening-remarks-jonathan-bryce-executive-director-cloud-and-infrastructure-linux-foundation-chris-aniszczyk-cto-cloud-and-infrastructure-linux-foundation?iframe=no) +- **Time**: March 24, 09:00–09:35 +- **Location**: Hall 12 +- **Speakers**: Jonathan Bryce (Linux Foundation Executive Director) & Chris Aniszczyk (CNCF CTO) +- **Agenda**: [Keynote: Welcome + Opening Remarks](https://kccnceu2026.sched.com/event/2CtKk/keynote-welcome-+-opening-remarks-jonathan-bryce-executive-director-cloud-and-infrastructure-linux-foundation-chris-aniszczyk-cto-cloud-and-infrastructure-linux-foundation?iframe=no) -本次开幕 keynote 由 Linux Foundation 与 CNCF 相关负责人带来。 +This opening keynote will be delivered by leadership from Linux Foundation and CNCF. -对于关注 AI 基础设施方向的社区成员来说，keynote 本身就是一个观察窗口：云原生主叙事是否正在吸纳更多 AI、GPU 与资源管理相关议题。 +For community members focused on AI infrastructure, the keynote serves as an observation window: Is the main cloud native narrative embracing more AI, GPU, and resource management topics? -### 2. HAMi 技术分享（Lightning Talks） +### 2. HAMi Technical Sessions (Lightning Talks) #### GPU Sharing in Kubernetes -- **时间**：3 月 23 日 17:15–17:25 -- **地点**：Hall 7 · Room B -- **讲者**：张潇（「Dynamia 密瓜智能」CEO，HAMi Maintainer） -- **议程**：[K8s Issue #52757: Sharing GPUs Among Multiple Containers](https://colocatedeventseu2026.sched.com/event/2DY9v/cllightning-talk-k8s-issue-?iframe=yes&w=100%&sidebar=yes&bg=no#52757-sharing-gpus-among-multiple-containers-xiao-zhang-dynamiaai) +- **Time**: March 23, 17:15–17:25 +- **Location**: Hall 7 · Room B +- **Speaker**: Xiao Zhang (CEO, Dynamia, HAMi Maintainer) +- **Agenda**: [K8s Issue #52757: Sharing GPUs Among Multiple Containers](https://colocatedeventseu2026.sched.com/event/2DY9v/cllightning-talk-k8s-issue-?iframe=yes&w=100%&sidebar=yes&bg=no#52757-sharing-gpus-among-multiple-containers-xiao-zhang-dynamiaai) -这场 lightning talk 会从 Kubernetes 社区长期存在的 GPU 共享问题出发，讨论多容器共享 GPU 的背景、挑战与相关实现路径。 +This lightning talk will address Kubernetes' long-standing GPU sharing problem, discussing the background, challenges, and implementation paths for multi-container GPU sharing. -#### HAMi 项目技术解读 +#### HAMi Project Technical Deep Dive -- **时间**：3 月 23 日 14:43–14:48 -- **地点**：Elicium 2 -- **讲者**：李孟轩（「Dynamia 密瓜智能」CTO，HAMi Maintainer） -- **议程**：[HAMi: Dynamic, Smart, Stable GPU-Sharing Middleware in Kubernetes](https://kccnceu2026.sched.com/event/2EFyZ/project-lightning-talk-hami-dynamic-smart-stable-gpu-sharing-middleware-in-kubernetes-mengxuan-li-maintainer?iframe=yes&w=100%&sidebar=yes&bg=no) +- **Time**: March 23, 14:43–14:48 +- **Location**: Elicium 2 +- **Speaker**: Mengxuan Li (CTO, Dynamia, HAMi Maintainer) +- **Agenda**: [HAMi: Dynamic, Smart, Stable GPU-Sharing Middleware in Kubernetes](https://kccnceu2026.sched.com/event/2EFyZ/project-lightning-talk-hami-dynamic-smart-stable-gpu-sharing-middleware-in-kubernetes-mengxuan-li-maintainer?iframe=yes&w=100%&sidebar=yes&bg=no) -这场分享会聚焦 HAMi 的核心架构与能力，包括 GPU 虚拟化、共享与调度机制，以及项目在稳定性与生产可用性上的设计思路。 +This session will focus on HAMi's core architecture and capabilities, including GPU virtualization, sharing and scheduling mechanisms, and the project's design philosophy around stability and production readiness. ### 3. Maintainer Summit -- **时间**：3 月 22 日 -- **信息**：[Maintainer Summit](https://events.linuxfoundation.org/kubecon-cloudnativecon-europe/features-add-ons/maintainer-summit/) +- **Time**: March 22 +- **Information**: [Maintainer Summit](https://events.linuxfoundation.org/kubecon-cloudnativecon-europe/features-add-ons/maintainer-summit/) -HAMi 也将参与 KubeCon 期间的 Maintainer Summit，并围绕 **Insights on AI Workloads** 与维护者群体展开交流。 +HAMi will also participate in the KubeCon Maintainer Summit, engaging with maintainers around **Insights on AI Workloads**. -Maintainer Summit 是 KubeCon 主会前一天举行的维护者活动，聚焦上游协作、SIG/WG 话题与项目间讨论。 +The Maintainer Summit is a maintainer-focused event held the day before the main conference, focusing on upstream collaboration, SIG/WG topics, and inter-project discussions. -对 HAMi 来说，这也是一个把 GPU 资源管理与 AI workload 问题带入更广泛维护者语境的重要场景。Maintainer Summit 确认于 3 月 22 日在 RAI Amsterdam 举行。 +For HAMi, this is an important venue to bring GPU resource management and AI workload topics into a broader maintainer context. The Maintainer Summit is confirmed for March 22 at RAI Amsterdam. ### 4. Poster Session -- **时间**：3 月 25 日 13:15–14:15 -- **地点**：Hall 1–5 · Gouda Zone · Poster Pavilion -- **讲者**：Satyam Soni (Devtron) & Rudraksh Karpe (ZS Associates) -- **议程**：[Kubernetes as the Universal GPU Control Plane for AI Workloads](https://kccnceu2026.sched.com/event/2CW0q/poster-session-kubernetes-as-the-universal-gpu-control-plane-for-ai-workloads-satyam-soni-devtronai-rudraksh-karpe-zs-associates-inc?iframe=yes&w=100%&sidebar=yes&bg=no) +- **Time**: March 25, 13:15–14:15 +- **Location**: Hall 1–5 · Gouda Zone · Poster Pavilion +- **Speakers**: Satyam Soni (Devtron) & Rudraksh Karpe (ZS Associates) +- **Agenda**: [Kubernetes as the Universal GPU Control Plane for AI Workloads](https://kccnceu2026.sched.com/event/2CW0q/poster-session-kubernetes-as-the-universal-gpu-control-plane-for-ai-workloads-satyam-soni-devtronai-rudraksh-karpe-zs-associates-inc?iframe=yes&w=100%&sidebar=yes&bg=no) -这场 poster session 从更生态化的角度讨论 Kubernetes 作为 GPU control plane 的可能性，这一方向与 HAMi 长期关注的问题高度相关。 +This poster session discusses the potential of Kubernetes as a GPU control plane from an ecosystem perspective—a direction highly relevant to HAMi's long-term focus. -### 5. 主舞台 Demo +### 5. Main Stage Demo -- **时间**：3 月 26 日 10:03–10:18 -- **地点**：Hall 12 -- **讲者**：李孟轩（「Dynamia 密瓜智能」CTO，HAMi Maintainer），Reza Jelveh（「Dynamia 密瓜智能」Head of Global Market & Solution Engineer） +- **Time**: March 26, 10:03–10:18 +- **Location**: Hall 12 +- **Speakers**: Mengxuan Li (CTO, Dynamia, HAMi Maintainer), Reza Jelveh (Head of Global Market & Solution Engineer, Dynamia) -KubeCon 期间的主舞台 Demo 将展示 GPU 共享与调度在 Kubernetes 中的实际运行方式。相比常规 PPT 分享，这类 Demo 更适合直观理解从资源抽象到系统落地的完整链路。 +The main stage demo during KubeCon will showcase GPU sharing and scheduling in action within Kubernetes. Compared to traditional slide presentations, this demo provides a more intuitive understanding of the complete chain from resource abstraction to system implementation. ### 6. AI Native Summit -- **时间**：3 月 27 日 09:00–16:00 -- **地点**：Van der Valk Hotel Amsterdam – Zuidas -- **议程**：[AI Native Summit Hosted by ETSI ISG NFV](https://kccnceu2026.sched.com/event/2HKYM/ai-native-summit-hosted-by-etsi-isg-nfv-separate-registration-required?iframe=no) +- **Time**: March 27, 09:00–16:00 +- **Location**: Van der Valk Hotel Amsterdam – Zuidas +- **Agenda**: [AI Native Summit Hosted by ETSI ISG NFV](https://kccnceu2026.sched.com/event/2HKYM/ai-native-summit-hosted-by-etsi-isg-nfv-separate-registration-required?iframe=no) -在主会结束后，AI Native Summit 也值得关注。该活动更适合从系统层面讨论 AI 基础设施中的资源层、控制层，以及 Kubernetes 在其中的角色。 +After the main conference, the AI Native Summit is also worth attention. This event is better suited for system-level discussions of the resource layer and control layer in AI infrastructure, and Kubernetes' role within them. -## 除了 HAMi，还可以关注哪些议题？ +## Beyond HAMi: Other Topics to Follow -如果你会参加本届 KubeCon，除了 HAMi 相关活动，也建议重点关注以下方向： +If you're attending this KubeCon, in addition to HAMi-related activities, we recommend focusing on these areas: - Device Management / DRA - AI workload scheduling - GPU observability -- inference platform 与 AI reference stack -- GPU 共享与资源抽象 +- Inference platforms and AI reference stacks +- GPU sharing and resource abstraction -这些议题虽然分散在不同会场，但共同指向一个问题：**Kubernetes 如何在 AI 时代具备更强的资源管理与调度能力。** +While these topics are spread across different venues, they all point to one question: **How can Kubernetes gain stronger resource management and scheduling capabilities in the AI era?** -## 社区动态与后续内容 +## Community Updates and Follow-up Content -大会期间，HAMi 社区也会持续整理和发布相关内容，包括技术分享要点、现场展示以及对 AI 基础设施趋势的观察。 +During the conference, the HAMi community will continue to curate and publish related content, including technical session highlights, on-site demonstrations, and observations on AI infrastructure trends. -欢迎关注： +Stay connected: -- [HAMi GitHub 仓库](https://github.com/project-hami/hami) -- [HAMi 社区官网](https://project-hami.io) +- [HAMi GitHub Repository](https://github.com/project-hami/hami) +- [HAMi Community Website](https://project-hami.io) -如果你也会在阿姆斯特丹，欢迎来 Project Pavilion 找到我们。 +If you'll be in Amsterdam, come find us at Project Pavilion. -`📍 HAMi Booth：P-13B` +`📍 HAMi Booth: P-13B` diff --git a/blog/kubecon-eu-2026-recap/index.md b/blog/kubecon-eu-2026-recap/index.md new file mode 100644 index 00000000..c7d28f01 --- /dev/null +++ b/blog/kubecon-eu-2026-recap/index.md @@ -0,0 +1,228 @@ +--- +title: "KubeCon EU 2026 Recap: HAMi From Project Pavilion to Main Stage Keynote Demo" +date: "2026-03-31" +description: "KubeCon EU 2026 has wrapped up in Amsterdam, sending a clear signal: cloud native is evolving from an application runtime platform into the foundation for AI infrastructure. As a CNCF Sandbox project, HAMi made a landmark appearance across the Maintainer Summit, technical sessions, Project Pavilion, and the main stage Keynote Demo." +tags: ["KubeCon", "GPU", "Kubernetes", "AI"] +authors: [hami_community] +--- + +The recently concluded **KubeCon + CloudNativeCon Europe 2026** sent an increasingly clear signal to the industry: + +**Cloud native is rapidly evolving from an "application runtime platform" into the operational foundation for AI infrastructure.** + + + +In Amsterdam, discussions around Kubernetes, GPUs, inference serving, Agentic AI, and heterogeneous compute scheduling have moved beyond concepts into concrete engineering practice, community collaboration, and infrastructure paradigm evolution. + +As a CNCF Sandbox project, HAMi made a landmark appearance at this year's conference, spanning the Maintainer Summit, Lightning Talks, Project Pavilion, and the main stage Keynote Demo. + +![Mengxuan Li and Reza Jelveh at the KubeCon Keynote Live Demo](/img/kubecon-eu-2026-recap/keynote-live-demo.jpg) + +## Kubernetes Is Entering the AI Infra Era + +If Kubernetes previously focused on container orchestration, microservice governance, and cloud native application delivery, the questions dominating this KubeCon were quite different: + +- How can AI workloads run more efficiently on Kubernetes? +- How can GPUs be shared, partitioned, scheduled, and isolated? +- How can LLM serving and underlying resource management work in concert? +- How can heterogeneous compute be unified into the cloud native scheduling system? + +These questions point to a more fundamental shift: + +> **Kubernetes is moving from "orchestrating applications" to "orchestrating compute."** + +This is exactly where HAMi operates. + +## Maintainer Summit: GPU Scheduling Enters Core Community Discussions + +At the pre-conference **Maintainer Summit**, HAMi Maintainer Mengxuan Li shared HAMi's insights on AI workloads. + +![HAMi Maintainer Mengxuan Li sharing AI Workloads insights at Maintainer Summit](/img/kubecon-eu-2026-recap/cto-maintainer-summit.png) + +The team also participated in closed-door CNCF meetings, engaging in in-depth discussions with CNCF TOC Chair Karena Angell, Red Hat, and vLLM community members Brian Stevens and Robert Shaw. + +![Discussions on GPU Sharing with CNCF TOC, Red Hat, and vLLM community](/img/kubecon-eu-2026-recap/cncf-toc-redhat-vllm.png) + +This discussion was particularly representative, as it didn't stay at the level of "how to build features for a project," but addressed a bigger question: + +> **When LLM serving, GPU resource management, and Kubernetes begin converging in real production environments, what new abstractions does the infrastructure layer need?** + +During the exchange, the direction HAMi is pushing drew noticeable attention. There is a growing realization that GPUs can no longer be treated as simple devices — they are becoming an infrastructure resource layer that can be scheduled, shared, and governed. + +This is also why the collaboration between HAMi and projects like vLLM is becoming increasingly natural. At this event, both sides began exploring joint content collaboration and technical exchange, indicating that the AI Infra ecosystem is accelerating from "standalone projects" to "composable collaboration." + +Additionally, HAMi is currently applying for CNCF Incubation and participated in discussions as a representative project during the TAG workshop. + +![TAG Workshop discussing CNCF project governance](/img/kubecon-eu-2026-recap/tag-workshop.jpg) + + + +## Two Technical Sessions: From Community Problems to Engineering Solutions + +### Xiao Zhang: K8s Issue #52757 — Sharing GPUs Among Multiple Containers + +This issue ([#52757](https://github.com/kubernetes/kubernetes/issues/52757)) is not new — it's a long-standing "unsolved problem" in the Kubernetes community. + +With the explosion of AI workloads, this problem has been amplified: + +- Inference serving requires more granular GPU usage +- Multi-tenant environments demand resource sharing +- AI workload patterns mean GPUs are no longer suitable for exclusive allocation + +This is why what appears to be a low-level problem has become one of the core issues in AI infrastructure. + +![Xiao Zhang presenting HAMi at the KubeCon Cloud Native AI forum](/img/kubecon-eu-2026-recap/zhangxiao-gpu-sharing.png) + +HAMi Maintainer Xiao Zhang's talk started from a classic, long-standing problem in the Kubernetes community: **How can multiple containers share a GPU?** + +While this question seems specific, it actually points to a challenge the entire AI infrastructure ecosystem faces. Once you enter inference, batch processing, online serving, and multi-tenant mixed scenarios, GPUs can no longer be simply allocated in an "exclusive whole-card" manner. + +The significance of this talk lies in putting HAMi's solution back into the original context of the Kubernetes community: not building an isolated solution from scratch, but addressing a long-standing upstream problem that hasn't been fully resolved. + +### Mengxuan Li: Dynamic, Smart, Stable GPU-Sharing Middleware in Kubernetes + +HAMi Maintainer Mengxuan Li's talk focused on HAMi's core architecture and capabilities, systematically covering: + +- GPU virtualization +- GPU sharing and scheduling mechanisms +- Stability and production readiness design +- Approaches to AI workload resource management in Kubernetes + +![Mengxuan Li presenting HAMi at KubeCon](/img/kubecon-eu-2026-recap/limengxuan-hami-talk.png) + +This wasn't just a project feature introduction — it was answering a more practical question: + +> **Before Kubernetes natively solves the GPU sharing problem, how can enterprises actually run AI workloads — stably and efficiently?** + +## Project Pavilion: Face-to-Face Global Community Exchange + +Beyond the speaking sessions, HAMi also had a booth at the KubeCon EU 2026 **Project Pavilion**. + +![A steady stream of visitors at the HAMi booth](/img/kubecon-eu-2026-recap/booth-crowd.jpg) + +Over the course of the event, the booth became a hub for intensive exchanges. Visitors included: + +- Overseas developers and contributors +- Enterprise users and platform teams +- University and research institution staff +- Cloud providers and GPU ecosystem professionals +- Community members interested in AI infra, heterogeneous compute, and Kubernetes GPU scheduling + +We also connected with more community contributors on-site, including contributors from India — Rudraksh Karpe and Shivay Lamba. + +![Indian contributors Rudraksh Karpe (center) and Shivay Lamba (right)](/img/kubecon-eu-2026-recap/indian-contributors.png) + +In the Poster Session, community contributors created a diagram illustrating "Kubernetes as the Universal GPU Control Plane." + +![Kubernetes as the Universal GPU Control Plane](/img/kubecon-eu-2026-recap/k8s-gpu-control-plane.jpg) + +The value of these exchanges goes beyond "increasing visibility" — it helps validate something: + +> **GPU scheduling, resource sharing, and heterogeneous compute management have become a real global demand, not a niche problem for any single market.** + +## Keynote Demo: HAMi on the KubeCon Main Stage + +![KubeCon Keynote hosted by Linux Foundation CEO Jonathan and CNCF CTO Chris](/img/kubecon-eu-2026-recap/keynote-hosts.png) + +If the talks and booth represented "recognition within professional circles," the most iconic moment of this KubeCon was undoubtedly: + +> **HAMi became a Chinese open source project to take the KubeCon EU 2026 main stage Keynote, completing a live Demo.** + +This was the most critical moment of the entire conference. + +During the main keynote, HAMi Maintainer **Mengxuan Li** and **Reza Jelveh** delivered a live Demo showcasing multi-workload GPU scheduling on Kubernetes. + +![Mengxuan Li and Reza during the live Demo](/img/kubecon-eu-2026-recap/limengxuan-reza-demo.jpg) + +The Demo used two typical AI workloads: YOLO inference serving and Qwen3-8B large model inference. In the traditional Kubernetes scheduling model, these two types of tasks would typically require exclusive GPU access. Under HAMi's scheduling model, GPUs are decomposed into "compute + memory" resource units that can be shared on-demand by multiple Pods. + +In the live demonstration, multiple YOLO instances were scheduled to run on the same GPU, while the Qwen3-8B model was co-located with other workloads on the same GPU through a binpack strategy. Different types of AI workloads coexisted on the same GPU while maintaining resource isolation and controllable scheduling. + +What this Demo presented was not just an improvement in GPU utilization, but more importantly, a new infrastructure capability: GPUs transitioning from "devices" to "schedulable resources," with Kubernetes gaining the foundational ability to manage AI workloads. + +### Why Does This Matter? + +**First, AI infrastructure topics have entered the KubeCon main narrative.** + +In the past, the KubeCon main stage focused primarily on Kubernetes itself and foundational platform capabilities. This time, a GPU resource management project like HAMi entering the main keynote demo signals that **how AI workloads run on Kubernetes has become a question the cloud native community must address head-on.** + +**Second, GPU scheduling is no longer a "niche topic."** + +GPU sharing, virtualization, resource isolation, and heterogeneous scheduling were previously confined to specialized circles. Now, they've evolved from "specialized domain problems" to "common infrastructure problems." In TOC discussions and community exchanges, multiple projects (including vLLM-related practices) have begun to directly depend on underlying GPU scheduling capabilities. + +**Third, this is the result of accumulated HAMi community effort.** + +An open source project making it to the KubeCon main stage isn't just about "having a feature to demo." Behind it is the alignment of technical direction with industry trends, community value being recognized, and the project's position in the ecosystem becoming clearer. + +This keynote demo also served as a positioning confirmation: + +> **HAMi is evolving from a GPU sharing tool into an important component of the AI compute resource layer on Kubernetes.** + +### AI Native Summit + +Following the main KubeCon conference, the co-located AI Native Summit was also held. + +Compared to the main KubeCon, the AI Native Summit's discussions focused more directly on one question: **AI workload operational efficiency is becoming the new infrastructure bottleneck.** + +In this context, GPU virtualization and scheduling are no longer internal Kubernetes optimizations — they are key factors directly impacting model serving costs, response times, and system throughput. + +Reza Jelveh presented "HAMi: Heterogeneous GPU Virtualization and Scheduling for AI-Native Infrastructure on Kubernetes." + +![Reza presenting HAMi at AI Native Summit](/img/kubecon-eu-2026-recap/reza-ai-native-summit.png) + +Reza also participated in a panel discussion titled "AI Native Technology." + +![Reza participating in the AI Native Technology panel discussion](/img/kubecon-eu-2026-recap/reza-panel-discussion.png) + +The AI Native Summit brought together technical experts from cloud native, AI infrastructure, and the telecom industry for in-depth discussions on the evolution of AI-native architectures. The conference focused on how infrastructure can evolve from traditional service-oriented, request-response patterns to a new generation of platforms designed for inference, conversation, and autonomous decision-making — covering key topics such as AI gateways, inference scheduling, multi-model routing, and multi-tenant isolation. + +## A Notable Detail: HAMi Enters the Broader Cloud Native Context + +Beyond the live demo and talks, there was another important external signal from this conference: HAMi was mentioned as a **representative case in the expanded Cloud Native Landscape** during the main stage keynote. + +![HAMi highlighted as a Cloud Native Landscape expansion project during Keynote](/img/kubecon-eu-2026-recap/landscape-mention.jpg) + +This indicates that HAMi's significance extends beyond "a project doing GPU scheduling" — it's being viewed as a representative of next-generation infrastructure problems within the broader cloud native evolution. + +The cloud native community is realizing: + +- The existing resource model built around CPU / memory / network / storage isn't enough +- The AI era demands new resource abstractions +- GPUs, inference, heterogeneous devices, and workload governance are becoming key infrastructure topics for the next phase + +HAMi sits precisely at this inflection point, offering a clear, pragmatic, and implementable engineering path. + +## Key Takeaways + +Looking back at KubeCon, several things stand out for the community: + +### 1. Global Community Focus on AI Infra Is Rapidly Increasing + +People are no longer satisfied with just discussing models and applications. They're asking: + +- How does it run at the infrastructure level? +- How are resources scheduled? +- How do we improve efficiency? +- How do we ensure system stability? + +### 2. The Kubernetes-AI Convergence Is Entering Deep Waters + +The question is no longer "can it run?" but: can it run efficiently, at scale, and stably in production environments? + +### 3. HAMi's Positioning Is Becoming Clearer + +HAMi is no longer just "a project that does GPU sharing." It's gradually forming its unique positioning: + +> **The GPU resource layer and heterogeneous compute scheduling capability for Kubernetes.** + +## Conclusion + +KubeCon EU 2026 has reinforced our conviction: **cloud native won't be replaced by AI — it will be redefined by AI.** + +From booth exchanges to technical sessions to the main stage demo, HAMi's presence at this conference was more than just an event appearance — it was a signal: + +> **Cloud native infrastructure around GPUs, inference, and heterogeneous compute is entering a new phase.** + +If you're also interested in AI infrastructure, GPU virtualization, and the evolution of Kubernetes in the AI era, we invite you to join the HAMi community and help drive the next steps in this space. + +![HAMi community members and contributors group photo at KubeCon](/img/kubecon-eu-2026-recap/team-photo.jpg) diff --git a/i18n/en/docusaurus-plugin-content-blog/hami-at-kubecon-eu-2026/index.md b/i18n/en/docusaurus-plugin-content-blog/hami-at-kubecon-eu-2026/index.md deleted file mode 100644 index a378e3ea..00000000 --- a/i18n/en/docusaurus-plugin-content-blog/hami-at-kubecon-eu-2026/index.md +++ /dev/null @@ -1,141 +0,0 @@ ---- -title: "HAMi at KubeCon Europe 2026: Building the GPU Resource Layer in Kubernetes" -date: "2026-03-19" -description: "HAMi will be featured in multiple activities at KubeCon Europe 2026, including Project Pavilion booth, technical sessions, main stage demo, and post-conference AI events. As a CNCF Sandbox project, HAMi focuses on GPU virtualization, sharing, and scheduling, which is increasingly intersecting with AI infrastructure topics in the Kubernetes ecosystem." -tags: ["KubeCon", "GPU", "Kubernetes", "AI"] -authors: [hami_community] ---- - -Next week, HAMi will be featured in multiple activities at [KubeCon + CloudNativeCon Europe 2026](https://events.linuxfoundation.org/kubecon-cloudnativecon-europe/), including Project Pavilion booth, technical sessions, main stage demo, and post-conference AI-related events. - -As a CNCF Sandbox project, HAMi focuses on GPU virtualization, sharing, and scheduling, which is increasingly intersecting with AI infrastructure topics in the Kubernetes ecosystem. KubeCon + CloudNativeCon Europe 2026 will be held in Amsterdam from March 23-26, with March 23 as pre-event programming and March 24-26 as the main conference. - - - -![KubeCon EU 2026 attracts 13,000 attendees](/img/blog-hami-at-kubecon-eu-2026-kubecon.png) - -## Why This KubeCon Matters - -Looking at the cloud native community discussions over the past few years, a clear trend emerges: AI is moving from the application layer into Kubernetes' resource layer, scheduling layer, and control layer. - -The discussions around GPUs are no longer limited to "device visibility" or "driver availability," but have extended to sharing, partitioning, utilization, multi-tenant isolation, and AI workload scheduling semantics. - -The official agenda of KubeCon Europe 2026 reflects this trend across keynotes, AI-related sessions, Project Pavilion, and co-located events. - -In this context, HAMi's problem space becomes clearer: it's not simply about "making Kubernetes recognize GPUs," but making GPUs a resource layer capability that can be abstracted, shared, and scheduled. - -This is why this KubeCon is more than just a project showcase for the HAMi community—it's an opportunity to engage with the broader cloud native ecosystem. - -## Finding HAMi at KubeCon - -![Welcome to HAMi Booth](/img/blog-hami-at-kubecon-eu-2026-booth.png) - -HAMi will have a booth at Project Pavilion for in-person exchanges with community members, users, and maintainers. - -- **Booth**: **P-13B** -- **Times**: - - **March 24 (Tuesday) 15:10–19:00** - - **March 26 (Thursday) 12:30–14:00** - -If you're attending, stop by the HAMi Booth to discuss: - -- GPU virtualization and sharing in Kubernetes -- Resource scheduling and utilization optimization for AI workloads -- Multi-tenant GPU resource management -- HAMi's integration with ecosystem projects like [Volcano](https://volcano.sh/), [Kueue](https://kueue.sigs.k8s.io/), [vLLM](https://github.com/vllm-project/vllm), and others - -Project Pavilion is the project showcase area within the main KubeCon exhibition, designed for community projects, maintainers, and developers to connect. - -## HAMi @ KubeCon Europe 2026 Event Overview - -### 1. Opening Keynote - -- **Time**: March 24, 09:00–09:35 -- **Location**: Hall 12 -- **Speakers**: Jonathan Bryce (Linux Foundation Executive Director) & Chris Aniszczyk (CNCF CTO) -- **Agenda**: [Keynote: Welcome + Opening Remarks](https://kccnceu2026.sched.com/event/2CtKk/keynote-welcome-+-opening-remarks-jonathan-bryce-executive-director-cloud-and-infrastructure-linux-foundation-chris-aniszczyk-cto-cloud-and-infrastructure-linux-foundation?iframe=no) - -This opening keynote will be delivered by leadership from Linux Foundation and CNCF. - -For community members focused on AI infrastructure, the keynote serves as an observation window: Is the main cloud native narrative embracing more AI, GPU, and resource management topics? - -### 2. HAMi Technical Sessions (Lightning Talks) - -#### GPU Sharing in Kubernetes - -- **Time**: March 23, 17:15–17:25 -- **Location**: Hall 7 · Room B -- **Speaker**: Xiao Zhang (CEO, Dynamia, HAMi Maintainer) -- **Agenda**: [K8s Issue #52757: Sharing GPUs Among Multiple Containers](https://colocatedeventseu2026.sched.com/event/2DY9v/cllightning-talk-k8s-issue-?iframe=yes&w=100%&sidebar=yes&bg=no#52757-sharing-gpus-among-multiple-containers-xiao-zhang-dynamiaai) - -This lightning talk will address Kubernetes' long-standing GPU sharing problem, discussing the background, challenges, and implementation paths for multi-container GPU sharing. - -#### HAMi Project Technical Deep Dive - -- **Time**: March 23, 14:43–14:48 -- **Location**: Elicium 2 -- **Speaker**: Mengxuan Li (CTO, Dynamia, HAMi Maintainer) -- **Agenda**: [HAMi: Dynamic, Smart, Stable GPU-Sharing Middleware in Kubernetes](https://kccnceu2026.sched.com/event/2EFyZ/project-lightning-talk-hami-dynamic-smart-stable-gpu-sharing-middleware-in-kubernetes-mengxuan-li-maintainer?iframe=yes&w=100%&sidebar=yes&bg=no) - -This session will focus on HAMi's core architecture and capabilities, including GPU virtualization, sharing and scheduling mechanisms, and the project's design philosophy around stability and production readiness. - -### 3. Maintainer Summit - -- **Time**: March 22 -- **Information**: [Maintainer Summit](https://events.linuxfoundation.org/kubecon-cloudnativecon-europe/features-add-ons/maintainer-summit/) - -HAMi will also participate in the KubeCon Maintainer Summit, engaging with maintainers around **Insights on AI Workloads**. - -The Maintainer Summit is a maintainer-focused event held the day before the main conference, focusing on upstream collaboration, SIG/WG topics, and inter-project discussions. - -For HAMi, this is an important venue to bring GPU resource management and AI workload topics into a broader maintainer context. The Maintainer Summit is confirmed for March 22 at RAI Amsterdam. - -### 4. Poster Session - -- **Time**: March 25, 13:15–14:15 -- **Location**: Hall 1–5 · Gouda Zone · Poster Pavilion -- **Speakers**: Satyam Soni (Devtron) & Rudraksh Karpe (ZS Associates) -- **Agenda**: [Kubernetes as the Universal GPU Control Plane for AI Workloads](https://kccnceu2026.sched.com/event/2CW0q/poster-session-kubernetes-as-the-universal-gpu-control-plane-for-ai-workloads-satyam-soni-devtronai-rudraksh-karpe-zs-associates-inc?iframe=yes&w=100%&sidebar=yes&bg=no) - -This poster session discusses the potential of Kubernetes as a GPU control plane from an ecosystem perspective—a direction highly relevant to HAMi's long-term focus. - -### 5. Main Stage Demo - -- **Time**: March 26, 10:03–10:18 -- **Location**: Hall 12 -- **Speakers**: Mengxuan Li (CTO, Dynamia, HAMi Maintainer), Reza Jelveh (Head of Global Market & Solution Engineer, Dynamia) - -The main stage demo during KubeCon will showcase GPU sharing and scheduling in action within Kubernetes. Compared to traditional slide presentations, this demo provides a more intuitive understanding of the complete chain from resource abstraction to system implementation. - -### 6. AI Native Summit - -- **Time**: March 27, 09:00–16:00 -- **Location**: Van der Valk Hotel Amsterdam – Zuidas -- **Agenda**: [AI Native Summit Hosted by ETSI ISG NFV](https://kccnceu2026.sched.com/event/2HKYM/ai-native-summit-hosted-by-etsi-isg-nfv-separate-registration-required?iframe=no) - -After the main conference, the AI Native Summit is also worth attention. This event is better suited for system-level discussions of the resource layer and control layer in AI infrastructure, and Kubernetes' role within them. - -## Beyond HAMi: Other Topics to Follow - -If you're attending this KubeCon, in addition to HAMi-related activities, we recommend focusing on these areas: - -- Device Management / DRA -- AI workload scheduling -- GPU observability -- Inference platforms and AI reference stacks -- GPU sharing and resource abstraction - -While these topics are spread across different venues, they all point to one question: **How can Kubernetes gain stronger resource management and scheduling capabilities in the AI era?** - -## Community Updates and Follow-up Content - -During the conference, the HAMi community will continue to curate and publish related content, including technical session highlights, on-site demonstrations, and observations on AI infrastructure trends. - -Stay connected: - -- [HAMi GitHub Repository](https://github.com/project-hami/hami) -- [HAMi Community Website](https://project-hami.io) - -If you'll be in Amsterdam, come find us at Project Pavilion. - -`📍 HAMi Booth: P-13B` diff --git a/i18n/zh/docusaurus-plugin-content-blog/hami-at-kubecon-eu-2026/index.md b/i18n/zh/docusaurus-plugin-content-blog/hami-at-kubecon-eu-2026/index.md new file mode 100644 index 00000000..acdec16d --- /dev/null +++ b/i18n/zh/docusaurus-plugin-content-blog/hami-at-kubecon-eu-2026/index.md @@ -0,0 +1,141 @@ +--- +title: "HAMi 即将亮相 KubeCon Europe 2026：构建 Kubernetes 中的 GPU 资源层" +date: "2026-03-19" +description: "HAMi 将在 KubeCon Europe 2026 的多项活动中亮相，包括 Project Pavilion 展台、技术分享、主舞台 Demo 等。作为 CNCF Sandbox 项目，HAMi 关注的 GPU 虚拟化、共享与调度问题，正在与 Kubernetes 生态中的 AI 基础设施议题发生越来越直接的交汇。" +tags: ["KubeCon", "GPU", "Kubernetes", "AI"] +authors: [hami_community] +--- + +下周，HAMi 将在 [KubeCon + CloudNativeCon Europe 2026](https://events.linuxfoundation.org/kubecon-cloudnativecon-europe/) 的多项活动中亮相，包括 Project Pavilion 展台、技术分享、主舞台 Demo，以及会后 AI 相关活动。 + +作为 CNCF Sandbox 项目，HAMi 关注的 GPU 虚拟化、共享与调度问题，正在与 Kubernetes 生态中的 AI 基础设施议题发生越来越直接的交汇。KubeCon + CloudNativeCon Europe 2026 将于 3 月 23 日至 26 日在阿姆斯特丹举行，其中 3 月 23 日为 pre-event programming，3 月 24 日至 26 日为主会期。 + + + +![KubeCon EU 2026 吸引了 13,000 名参会者](/img/blog-hami-at-kubecon-eu-2026-kubecon.png) + +## 为什么这次 KubeCon 值得关注？ + +如果把过去几年云原生社区的讨论放在一起看，会发现一个越来越清晰的变化：AI 正在从应用层问题，进入 Kubernetes 的资源层、调度层与控制层。 + +围绕 GPU 的讨论也不再停留在"设备可见性"或"驱动可用性"，而是进一步延伸到共享、切分、利用率、多租户隔离，以及 AI workload 的调度语义等问题。 + +KubeCon Europe 2026 的官方议程中，keynote、AI 相关 session、Project Pavilion 与 co-located events 都体现出这一趋势。 + +在这个背景下，HAMi 所对应的问题空间也变得更加明确：不是简单地"让 Kubernetes 能识别 GPU"，而是让 GPU 进一步成为一种可以被抽象、被共享、被调度的资源层能力。 + +这也是为什么本次 KubeCon 对 HAMi 社区而言，不只是一次项目展示，更是一次和更大范围云原生生态对话的机会。 + +## 在 KubeCon 现场，如何找到 HAMi？ + +![欢迎来到 HAMi 展台](/img/blog-hami-at-kubecon-eu-2026-booth.png) + +HAMi 将在 Project Pavilion 设置展台，方便与社区成员、用户和维护者进行现场交流。 + +- **Booth**：**P-13B** +- **时间**： + - **3 月 24 日（周二）15:10–19:00** + - **3 月 26 日（周四）12:30–14:00** + +如果你会到现场，欢迎来到 HAMi Booth，一起交流这些话题： + +- Kubernetes 中的 GPU 虚拟化与共享 +- AI workload 的资源调度与利用率优化 +- 多租户 GPU 资源管理 +- HAMi 与 [Volcano](https://volcano.sh/)、[Kueue](https://kueue.sigs.k8s.io/)、[vLLM](https://github.com/vllm-project/vllm) 等生态项目的协同 + +Project Pavilion 是 KubeCon 主展区中的项目展示区域，面向社区项目、维护者与开发者交流。 + +## HAMi @ KubeCon Europe 2026 活动一览 + +### 1. Opening Keynote + +- **时间**：3 月 24 日 09:00–09:35 +- **地点**：Hall 12 +- **讲者**：Jonathan Bryce (Linux Foundation Executive Director) & Chris Aniszczyk (CNCF CTO) +- **议程**：[Keynote: Welcome + Opening Remarks](https://kccnceu2026.sched.com/event/2CtKk/keynote-welcome-+-opening-remarks-jonathan-bryce-executive-director-cloud-and-infrastructure-linux-foundation-chris-aniszczyk-cto-cloud-and-infrastructure-linux-foundation?iframe=no) + +本次开幕 keynote 由 Linux Foundation 与 CNCF 相关负责人带来。 + +对于关注 AI 基础设施方向的社区成员来说，keynote 本身就是一个观察窗口：云原生主叙事是否正在吸纳更多 AI、GPU 与资源管理相关议题。 + +### 2. HAMi 技术分享（Lightning Talks） + +#### GPU Sharing in Kubernetes + +- **时间**：3 月 23 日 17:15–17:25 +- **地点**：Hall 7 · Room B +- **讲者**：张潇（「Dynamia 密瓜智能」CEO，HAMi Maintainer） +- **议程**：[K8s Issue #52757: Sharing GPUs Among Multiple Containers](https://colocatedeventseu2026.sched.com/event/2DY9v/cllightning-talk-k8s-issue-?iframe=yes&w=100%&sidebar=yes&bg=no#52757-sharing-gpus-among-multiple-containers-xiao-zhang-dynamiaai) + +这场 lightning talk 会从 Kubernetes 社区长期存在的 GPU 共享问题出发，讨论多容器共享 GPU 的背景、挑战与相关实现路径。 + +#### HAMi 项目技术解读 + +- **时间**：3 月 23 日 14:43–14:48 +- **地点**：Elicium 2 +- **讲者**：李孟轩（「Dynamia 密瓜智能」CTO，HAMi Maintainer） +- **议程**：[HAMi: Dynamic, Smart, Stable GPU-Sharing Middleware in Kubernetes](https://kccnceu2026.sched.com/event/2EFyZ/project-lightning-talk-hami-dynamic-smart-stable-gpu-sharing-middleware-in-kubernetes-mengxuan-li-maintainer?iframe=yes&w=100%&sidebar=yes&bg=no) + +这场分享会聚焦 HAMi 的核心架构与能力，包括 GPU 虚拟化、共享与调度机制，以及项目在稳定性与生产可用性上的设计思路。 + +### 3. Maintainer Summit + +- **时间**：3 月 22 日 +- **信息**：[Maintainer Summit](https://events.linuxfoundation.org/kubecon-cloudnativecon-europe/features-add-ons/maintainer-summit/) + +HAMi 也将参与 KubeCon 期间的 Maintainer Summit，并围绕 **Insights on AI Workloads** 与维护者群体展开交流。 + +Maintainer Summit 是 KubeCon 主会前一天举行的维护者活动，聚焦上游协作、SIG/WG 话题与项目间讨论。 + +对 HAMi 来说，这也是一个把 GPU 资源管理与 AI workload 问题带入更广泛维护者语境的重要场景。Maintainer Summit 确认于 3 月 22 日在 RAI Amsterdam 举行。 + +### 4. Poster Session + +- **时间**：3 月 25 日 13:15–14:15 +- **地点**：Hall 1–5 · Gouda Zone · Poster Pavilion +- **讲者**：Satyam Soni (Devtron) & Rudraksh Karpe (ZS Associates) +- **议程**：[Kubernetes as the Universal GPU Control Plane for AI Workloads](https://kccnceu2026.sched.com/event/2CW0q/poster-session-kubernetes-as-the-universal-gpu-control-plane-for-ai-workloads-satyam-soni-devtronai-rudraksh-karpe-zs-associates-inc?iframe=yes&w=100%&sidebar=yes&bg=no) + +这场 poster session 从更生态化的角度讨论 Kubernetes 作为 GPU control plane 的可能性，这一方向与 HAMi 长期关注的问题高度相关。 + +### 5. 主舞台 Demo + +- **时间**：3 月 26 日 10:03–10:18 +- **地点**：Hall 12 +- **讲者**：李孟轩（「Dynamia 密瓜智能」CTO，HAMi Maintainer），Reza Jelveh（「Dynamia 密瓜智能」Head of Global Market & Solution Engineer） + +KubeCon 期间的主舞台 Demo 将展示 GPU 共享与调度在 Kubernetes 中的实际运行方式。相比常规 PPT 分享，这类 Demo 更适合直观理解从资源抽象到系统落地的完整链路。 + +### 6. AI Native Summit + +- **时间**：3 月 27 日 09:00–16:00 +- **地点**：Van der Valk Hotel Amsterdam – Zuidas +- **议程**：[AI Native Summit Hosted by ETSI ISG NFV](https://kccnceu2026.sched.com/event/2HKYM/ai-native-summit-hosted-by-etsi-isg-nfv-separate-registration-required?iframe=no) + +在主会结束后，AI Native Summit 也值得关注。该活动更适合从系统层面讨论 AI 基础设施中的资源层、控制层，以及 Kubernetes 在其中的角色。 + +## 除了 HAMi，还可以关注哪些议题？ + +如果你会参加本届 KubeCon，除了 HAMi 相关活动，也建议重点关注以下方向： + +- Device Management / DRA +- AI workload scheduling +- GPU observability +- inference platform 与 AI reference stack +- GPU 共享与资源抽象 + +这些议题虽然分散在不同会场，但共同指向一个问题：**Kubernetes 如何在 AI 时代具备更强的资源管理与调度能力。** + +## 社区动态与后续内容 + +大会期间，HAMi 社区也会持续整理和发布相关内容，包括技术分享要点、现场展示以及对 AI 基础设施趋势的观察。 + +欢迎关注： + +- [HAMi GitHub 仓库](https://github.com/project-hami/hami) +- [HAMi 社区官网](https://project-hami.io) + +如果你也会在阿姆斯特丹，欢迎来 Project Pavilion 找到我们。 + +`📍 HAMi Booth：P-13B` diff --git a/i18n/zh/docusaurus-plugin-content-blog/kubecon-eu-2026-recap/index.md b/i18n/zh/docusaurus-plugin-content-blog/kubecon-eu-2026-recap/index.md new file mode 100644 index 00000000..9568ee96 --- /dev/null +++ b/i18n/zh/docusaurus-plugin-content-blog/kubecon-eu-2026-recap/index.md @@ -0,0 +1,228 @@ +--- +title: "KubeCon EU 2026 回顾：HAMi 从展台到主论坛 Keynote Demo" +date: "2026-03-31" +description: "KubeCon EU 2026 已于阿姆斯特丹落幕。本届大会释放出明确信号：云原生正在从应用运行平台演进为 AI 基础设施底座。HAMi 作为 CNCF Sandbox 项目，在 Maintainer Summit、技术分享、Project Pavilion 及主论坛 Keynote Demo 中完成了一次标志性亮相。" +tags: ["KubeCon", "GPU", "Kubernetes", "AI"] +authors: [hami_community] +--- + +刚刚结束的 **KubeCon + CloudNativeCon Europe 2026**，释放出一个越来越明确的行业信号： + +**云原生正在快速从“应用运行平台”演进为 AI 基础设施的运行底座。** + +在阿姆斯特丹，围绕 Kubernetes、GPU、推理服务、Agentic AI 和异构算力调度的讨论，已经不再停留在概念层面，而是进入到更具体的工程实践、社区协作与基础设施范式演进阶段。 + + + +作为 CNCF Sandbox 项目，HAMi 在本届大会上完成了从 Maintainer Summit、Lightning Talk、Project Pavilion 到主论坛 Keynote Demo 的一系列亮相。 + +![李孟轩和 Reza Jelveh 在 KubeCon Keynote Live Demo](/img/kubecon-eu-2026-recap/keynote-live-demo.jpg) + +## Kubernetes 正在进入 AI Infra 阶段 + +如果说过去 Kubernetes 主要解决的是容器编排、微服务治理和云原生应用交付，那么在这届 KubeCon 上，更受关注的问题已经变成了： + +- AI workload 如何更高效地运行在 Kubernetes 上？ +- GPU 如何被共享、切分、调度和隔离？ +- LLM serving 与底层资源管理如何协同？ +- 异构算力如何被统一纳入云原生调度体系？ + +这些问题背后对应的是一个更本质的变化： + +> **Kubernetes 正在从"编排应用"走向"编排算力"。** + +这也正是 HAMi 所处的位置。 + +## Maintainer Summit：GPU 调度进入更核心的社区讨论 + +在大会前的 **Maintainer Summit** 上，HAMi Maintainer 李孟轩分享了 HAMi 对 AI 工作负载的见解。 + +![HAMi Maintainer 李孟轩在 Maintainer Summit 上分享 AI Workloads 洞察](/img/kubecon-eu-2026-recap/cto-maintainer-summit.png) + +随后团队参与了 CNCF 闭门会议，与 CNCF TOC 主席 Karena Angell、Red Hat 以及 vLLM 社区成员 Brian Stevens、Robert Shaw 等进行了深入交流。 + +![与 CNCF TOC、Red Hat、vLLM 社区分享交流 GPU Sharing](/img/kubecon-eu-2026-recap/cncf-toc-redhat-vllm.png) + +这次讨论很有代表性，因为它并不停留在"某个项目怎么做功能"，而是在讨论一个更大的问题： + +> **当 LLM serving、GPU 资源管理和 Kubernetes 在真实生产环境中开始汇合时，基础设施层需要什么样的新抽象？** + +现场交流中，HAMi 所推动的方向引起了明显关注。大家越来越意识到，GPU 已经不能只被看作一个简单设备，而正在变成一种可以被调度、共享、治理的基础设施资源层。 + +这也是为什么 HAMi 与 vLLM 等项目之间的协同开始变得越来越自然。在本次活动中，双方已经开始探讨后续的联合内容合作与技术交流，这也说明 AI Infra 生态正在加速从"单点项目"走向"组合式协作"。 + +另外 HAMi 项目也正在申请 CNCF 孵化，在 TAG workshop 中作为代表项目参与了讨论。 + +![TAG Workshop 讨论 CNCF 的项目治理](/img/kubecon-eu-2026-recap/tag-workshop.jpg) + + + +## 两场技术分享：从社区问题到工程实现 + +### 张潇：K8s Issue #52757 — Sharing GPUs Among Multiple Containers + +这个问题（[#52757](https://github.com/kubernetes/kubernetes/issues/52757)）并不是一个新问题，而是在 Kubernetes 社区中存在多年的"未被彻底解决的问题"。 + +随着 AI workload 的爆发，这个问题被重新放大： + +- 推理服务需要更细粒度的 GPU 使用方式 +- 多租户环境要求资源共享 +- AI workload 的形态决定了 GPU 不再适合独占 + +这也是为什么，这个看似底层的问题，开始成为 AI 基础设施的核心问题之一。 + +![张潇在 KubeCon 的 Cloud Native AI 论坛上分享 HAMi](/img/kubecon-eu-2026-recap/zhangxiao-gpu-sharing.png) + +HAMi Maintainer 张潇的分享从 Kubernetes 社区长期存在的一个经典问题出发：**多个容器如何共享 GPU？** + +这个问题看似具体，但实际上指向的是整个 AI 基础设施生态共同面临的难题。因为一旦进入推理、批处理、在线服务和多租户混合场景，GPU 就不再适合以"整卡独占"的方式被简单分配。 + +这场分享的重要性，在于它把 HAMi 所解决的问题放回到了 Kubernetes 社区的原始语境中：不是另起炉灶做一个孤立方案，而是在回应一个长期存在、尚未被彻底解决的 upstream 问题。 + +### 李孟轩：Dynamic, Smart, Stable GPU-Sharing Middleware in Kubernetes + +HAMi Maintainer 李孟轩的分享聚焦 HAMi 的核心架构与能力，系统介绍了： + +- GPU 虚拟化 +- GPU 共享与调度机制 +- 稳定性与生产可用性设计 +- 在 Kubernetes 中实现 AI workload 资源管理的思路 + +![李孟轩在 KubeCon 上分享 HAMi](/img/kubecon-eu-2026-recap/limengxuan-hami-talk.png) + +这并不只是介绍一个项目功能，而是在回答一个更实际的问题： + +> **在 Kubernetes 尚未原生解决 GPU 共享问题的前提下，企业如何真正把 AI workload 跑起来，并跑得更稳、更高效？** + +## Project Pavilion：面对面的全球社区交流 + +除了议题分享，HAMi 还在 KubeCon EU 2026 的 **Project Pavilion** 设有展台。 + +![前往 HAMi 展台交流的人络绎不绝](/img/kubecon-eu-2026-recap/booth-crowd.jpg) + +在这几天里，展台成为了非常密集的交流现场。前来交流的人群覆盖了： + +- 海外开发者与贡献者 +- 企业用户与平台团队 +- 高校、研究机构人员 +- 云厂商及 GPU 生态相关从业者 +- 对 AI infra、异构算力和 Kubernetes GPU 调度感兴趣的社区成员 + +在现场我们也结交了更多社区贡献者，包括来自印度的贡献者 Rudraksh Karpe 和 Shivay Lamba。 + +![来自印度的贡献者 Rudraksh Karpe（中间）和 Shivay Lamba（右）](/img/kubecon-eu-2026-recap/indian-contributors.png) + +在 Poster Session 中社区贡献者制作了 Kubernetes as the universal GPU control plane 的示意图。 + +![Kubernetes as the universal GPU control plane](/img/kubecon-eu-2026-recap/k8s-gpu-control-plane.jpg) + +这类交流的价值，不只是"增加曝光"，而是帮助验证一件事： + +> **GPU 调度、资源共享和异构算力管理，已经成为全球范围内的真实需求，而不是某个局部市场的特殊问题。** + +## Keynote Demo：HAMi 登上 KubeCon 主论坛舞台 + +![KubeCon Keynote — Linux Foundation CEO Jonathan 和 CNCF CTO Chris 联合主持](/img/kubecon-eu-2026-recap/keynote-hosts.png) + +如果说演讲和展台代表的是"专业圈层中的认可"，那么这次 KubeCon 最具标志性的时刻，无疑是： + +> **HAMi 作为中国开源项目登上了 KubeCon EU 2026 的主论坛 Keynote，并完成了现场 Demo 展示。** + +这是本次大会最关键、也最值得被强调的一环。 + +在主论坛环节中，HAMi Maintainer **李孟轩** 与 **Reza Jelveh** 进行了现场 Demo，展示了基于 Kubernetes 的多 workload GPU 调度场景。 + +![李孟轩和 Reza 现场 Demo](/img/kubecon-eu-2026-recap/limengxuan-reza-demo.jpg) + +Demo 以两个典型 AI workload 为例：一类是 YOLO 推理服务，另一类是 Qwen3-8B 大模型推理任务。在传统 Kubernetes 调度模型中，这两类任务通常需要独占 GPU 运行，而在 HAMi 的调度模型下，GPU 被拆分为"算力 + 显存"的资源单元，可以被多个 Pod 按需共享。 + +在实际演示中，多个 YOLO 实例被调度到同一张 GPU 上运行，而 Qwen3-8B 模型则通过 binpack 策略与其他 workload 共同部署在同一 GPU 上。不同类型的 AI workload 在同一 GPU 上共存，同时保持资源隔离与调度可控。 + +这个 Demo 所呈现的，并不仅仅是 GPU 利用率的提升，更重要的是一个新的基础设施能力：GPU 从"设备"转变为"可调度资源"，而 Kubernetes 正在具备管理 AI workload 的基础能力。 + +### 为什么这件事有意义？ + +**第一，AI 基础设施议题已经进入 KubeCon 的主叙事。** + +过去 KubeCon 主论坛更多聚焦 Kubernetes 本身和基础平台能力。而这次，HAMi 这样的 GPU 资源管理项目能够进入主论坛 demo，说明 **AI workload 如何运行在 Kubernetes 上，已经成为云原生社区必须正面回答的问题。** + +**第二，GPU 调度不再只是"边缘话题"。** + +GPU 共享、虚拟化、资源隔离、异构调度这些问题，过去往往更多存在于专业小圈子中。但现在，它们已经从"专门领域问题"变成"基础设施共同问题"。在 TOC 讨论和社区交流中，多个项目（包括 vLLM 相关实践）已经开始直接依赖底层 GPU 调度能力。 + +**第三，这是 HAMi 社区共同积累的结果。** + +一个开源项目能够走到 KubeCon 主舞台，不会只是因为"有个功能能演示"。它背后是技术方向与行业趋势对齐、社区价值被看见、项目在生态中的位置变得更清晰。 + +这次 keynote demo 也是一次定位确认： + +> **HAMi 正在从 GPU sharing 工具，走向 Kubernetes 上 AI 算力资源层的重要组成部分。** + +### AI Native Summit + +在 KubeCon 大会后还举办了同场活动 AI Native Summit。 + +与 KubeCon 主会场相比，AI Native Summit 的讨论更加直接聚焦在一个问题上：**AI workload 的运行效率，正在成为新的基础设施瓶颈。** + +在这个语境下，GPU virtualization 和调度问题，不再是 Kubernetes 内部优化，而是直接影响模型服务成本、响应时间和系统吞吐能力的关键因素。 + +Reza Jelveh 分享了《HAMi: Heterogeneous GPU Virtualization and Scheduling for AI-Native Infrastructure on Kubernetes》。 + +![Reza 在 AI Native Summit 上分享 HAMi](/img/kubecon-eu-2026-recap/reza-ai-native-summit.png) + +Reza 还参与了主题为《AI Native Technology》的圆桌讨论。 + +![Reza 参与 AI Native Technology 圆桌讨论](/img/kubecon-eu-2026-recap/reza-panel-discussion.png) + +本次 AI Native Summit 汇聚了来自云原生、AI 基础设施及电信行业的技术专家，围绕 AI 原生架构的演进展开深入探讨。会议重点关注在大模型和 Agent 驱动背景下，基础设施如何从传统的服务化、请求响应模式，演进为面向推理、会话和自治决策的新一代平台体系，涵盖 AI 网关、推理调度、多模型路由以及多租户隔离等关键议题。 + +## 一个值得注意的细节：HAMi 进入更大的云原生语境 + +除了现场 demo 和分享之外，这次大会还有一个很重要的外部信号：在主舞台分享中，HAMi 也被作为 **Cloud Native Landscape 扩展背景下的代表性案例** 被提及。 + +![HAMi 作为 Cloud Native Landscape 的扩展项目在 Keynote 中被重点提及](/img/kubecon-eu-2026-recap/landscape-mention.jpg) + +这说明 HAMi 的意义已经不只是"某个项目在做 GPU 调度"，而是在更大的云原生演进语境里，被看作新一代基础设施问题的代表。 + +云原生社区正在意识到： + +- 过去那套围绕 CPU / 内存 / 网络 / 存储构建的资源模型还不够 +- AI 时代需要新的资源抽象 +- GPU、推理、异构设备与工作负载治理，正在成为下一阶段的重要基础设施议题 + +而 HAMi 正是在这个转折点上，提供了一种清晰、务实、可落地的工程路径。 + +## 本次参会的核心收获 + +回看这次 KubeCon，有几件事值得社区关注： + +### 1. 全球社区对 AI Infra 的关注正在快速升温 + +大家已经不再满足于讨论模型和应用本身，而是在追问： + +- 底层怎么跑？ +- 资源怎么调？ +- 效率怎么提升？ +- 系统怎么稳定？ + +### 2. Kubernetes 与 AI 的结合正在进入深水区 + +现在的问题已经不再是"能不能跑"，而是：能不能高效地跑、能不能大规模地跑、能不能在生产环境稳定地跑。 + +### 3. HAMi 的定位越来越明确 + +HAMi 不再只是"做 GPU 共享的一个项目"，而是在逐步形成自己的独特定位： + +> **面向 Kubernetes 的 GPU 资源层与异构算力调度能力。** + +## 结语 + +KubeCon EU 2026 让我们更加确信：**云原生不会被 AI 替代，反而会因为 AI 被重新定义。** + +从展台交流，到议题分享，再到主论坛 demo，HAMi 在这次大会上的亮相，更像是一个信号： + +> **围绕 GPU、推理与异构算力的云原生基础设施，正在进入新的阶段。** + +如果你同样关注 AI 基础设施、GPU 虚拟化以及 Kubernetes 在 AI 时代的演进，欢迎加入 HAMi 社区，与我们一起推动这一领域的下一步发展。 + +![HAMi 社区成员与贡献者在 KubeCon 会场合影](/img/kubecon-eu-2026-recap/team-photo.jpg) diff --git a/static/img/kubecon-eu-2026-recap/booth-crowd.jpg b/static/img/kubecon-eu-2026-recap/booth-crowd.jpg new file mode 100644 index 00000000..c2e24452 Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/booth-crowd.jpg differ diff --git a/static/img/kubecon-eu-2026-recap/cncf-toc-redhat-vllm.png b/static/img/kubecon-eu-2026-recap/cncf-toc-redhat-vllm.png new file mode 100644 index 00000000..b365dd6c Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/cncf-toc-redhat-vllm.png differ diff --git a/static/img/kubecon-eu-2026-recap/cto-maintainer-summit.png b/static/img/kubecon-eu-2026-recap/cto-maintainer-summit.png new file mode 100644 index 00000000..713513f1 Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/cto-maintainer-summit.png differ diff --git a/static/img/kubecon-eu-2026-recap/indian-contributors.png b/static/img/kubecon-eu-2026-recap/indian-contributors.png new file mode 100644 index 00000000..f1bc6ce5 Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/indian-contributors.png differ diff --git a/static/img/kubecon-eu-2026-recap/k8s-gpu-control-plane.jpg b/static/img/kubecon-eu-2026-recap/k8s-gpu-control-plane.jpg new file mode 100644 index 00000000..4ea14a66 Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/k8s-gpu-control-plane.jpg differ diff --git a/static/img/kubecon-eu-2026-recap/keynote-hosts.png b/static/img/kubecon-eu-2026-recap/keynote-hosts.png new file mode 100644 index 00000000..20bac1ba Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/keynote-hosts.png differ diff --git a/static/img/kubecon-eu-2026-recap/keynote-live-demo.jpg b/static/img/kubecon-eu-2026-recap/keynote-live-demo.jpg new file mode 100644 index 00000000..55adb18b Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/keynote-live-demo.jpg differ diff --git a/static/img/kubecon-eu-2026-recap/keynote-stage.jpg b/static/img/kubecon-eu-2026-recap/keynote-stage.jpg new file mode 100644 index 00000000..9d112bf5 Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/keynote-stage.jpg differ diff --git a/static/img/kubecon-eu-2026-recap/landscape-mention.jpg b/static/img/kubecon-eu-2026-recap/landscape-mention.jpg new file mode 100644 index 00000000..7013416a Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/landscape-mention.jpg differ diff --git a/static/img/kubecon-eu-2026-recap/limengxuan-hami-talk.png b/static/img/kubecon-eu-2026-recap/limengxuan-hami-talk.png new file mode 100644 index 00000000..e0a56793 Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/limengxuan-hami-talk.png differ diff --git a/static/img/kubecon-eu-2026-recap/limengxuan-reza-demo.jpg b/static/img/kubecon-eu-2026-recap/limengxuan-reza-demo.jpg new file mode 100644 index 00000000..3df53689 Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/limengxuan-reza-demo.jpg differ diff --git a/static/img/kubecon-eu-2026-recap/reza-ai-native-summit.png b/static/img/kubecon-eu-2026-recap/reza-ai-native-summit.png new file mode 100644 index 00000000..6c8a6147 Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/reza-ai-native-summit.png differ diff --git a/static/img/kubecon-eu-2026-recap/reza-panel-discussion.png b/static/img/kubecon-eu-2026-recap/reza-panel-discussion.png new file mode 100644 index 00000000..647bf20e Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/reza-panel-discussion.png differ diff --git a/static/img/kubecon-eu-2026-recap/tag-workshop.jpg b/static/img/kubecon-eu-2026-recap/tag-workshop.jpg new file mode 100644 index 00000000..9b30773c Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/tag-workshop.jpg differ diff --git a/static/img/kubecon-eu-2026-recap/team-photo.jpg b/static/img/kubecon-eu-2026-recap/team-photo.jpg new file mode 100644 index 00000000..b9590b47 Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/team-photo.jpg differ diff --git a/static/img/kubecon-eu-2026-recap/zhangxiao-gpu-sharing.png b/static/img/kubecon-eu-2026-recap/zhangxiao-gpu-sharing.png new file mode 100644 index 00000000..5bc5130f Binary files /dev/null and b/static/img/kubecon-eu-2026-recap/zhangxiao-gpu-sharing.png differ