Insights, tutorials, and news from our team
This article takes the PR as an entry point, combined with community Issues and mailing records, to fully restore a 'HAMi × vLLM' landing path from deployment to verification, helping you quickly achieve multi-model deployment and resource reuse in Kubernetes.
This article takes the PR as an entry point, combined with community Issues and mailing records, to fully restore a 'HAMi × vLLM' landing path from deployment to verification, helping you quickly achieve multi-model deployment and resource reuse in Kubernetes.
In its core model training scenarios, a leading autonomous driving company utilizes multi-machine distributed training with scheduling frameworks like Ray and Volcano.
In the fast-growing field of AI education, PREP EDU (prepedu.com) is emerging as a focal point in Southeast Asia's EdTech sector.
Yesterday, we took a deep dive into how KAI-Scheduler achieves fractional GPU sharing. Thank you all for your attention and lively discussion! A reader pointed out a key technical detail that needed further clarification, and today we're going to break down that very issue.
Today, we're doing a technical deep dive to compare the implementation methods of KAI-Scheduler and HAMi, and to look ahead at the possibilities for future collaboration.
Koordinator v1.6 has been released, featuring deep collaboration with the CNCF Sandbox project HAMi to introduce strong GPU sharing isolation capabilities, providing a more efficient resource scheduling and isolation solution for AI training and inference scenarios.
I'll discuss this from several angles: What is HAMi, and how does it fundamentally differ from NVIDIA's solutions? Why doesn't NVIDIA do it? Is it a technical limitation or a business decision? What are HAMi's core competitiveness and long-term moat?
This article analyzes how Hygon DCU devices can be virtualized through HAMi to achieve unified management and scheduling.
Get the latest technical articles, tutorials, and HAMi community updates.