News › ai-infrastructure
By Zayden R., August 9, 2026
Kubernetes Dynamic Resource Allocation (DRA) is now generally available in v1.34, offering native GPU memory allocation. This development impacts HAMi's role in fractional GPU requests, though HAMi's enforcement features remain crucial.
Kubernetes has officially rolled out Dynamic Resource Allocation (DRA) as a generally available feature starting with version 1.34, a significant development for engineers managing GPU resources. This enhancement allows Kubernetes to natively handle fractional GPU memory allocations, a capability previously absent from its API vocabulary.
Before DRA, projects like HAMi had to innovate around Kubernetes' limitations. HAMi, which became an incubating project under the CNCF Technical Oversight Committee on July 15, 2026, created a sophisticated pipeline to manage fractional GPU requests. This included a mutating webhook, scheduler extender, and in-container enforcement mechanisms. The native Kubernetes API could only count whole devices, forcing HAMi to develop its own vocabulary for requests like '8,000 MiB and 10% of a GPU'.
With the arrival of DRA, Kubernetes v1.34 can now accept such fractional requests directly from the scheduler, eliminating the need for annotations and simplifying the process for developers. This change is enabled by default as of Kubernetes v1.35. However, DRA does not render HAMi obsolete. While DRA takes over the role of encoding fractional requests, it does not handle enforcement within the container, a critical function that HAMi continues to provide.
HAMi has adapted by focusing on its enforcement capabilities and rebuilding its encoding on top of DRA across three repositories. For engineers, this means that while DRA simplifies certain aspects of GPU management, HAMi remains essential for precise in-container resource control.
The practical impact of DRA is significant. Engineers should prepare to update their Kubernetes environments to version 1.34 or later to take advantage of these improvements. They should also consider how HAMi's continuing enforcement functionality can complement DRA's native capabilities for a comprehensive GPU management strategy.
The Linux Camp teaches these topics as hands-on labs on real virtual machines, verified as you type.