Einblicke, Tutorials und Neuigkeiten von unserem Team
Dynamia has partnered with MetaX to complete full adaptation of MetaX C-series GPUs for the llm-d community, validating single-GPU, multi-GPU, and Prefill/Decode disaggregated inference deployments — the first Chinese GPU ever added to llm-d's officially supported accelerator list.
On the evening of September 6, Dynamia, HAMi, and CNCF co-hosted the HAMi Incubating Community Night in Shanghai to celebrate HAMi entering CNCF Incubating. The event launched the HAMi Ambassador Program and honored Maintainers, Approvers, and Contributors.
Five production talks unpacking AI compute efficiency. Recap of the HAMi Meetup Shanghai Incubating special event: remote GPUs and HAMi 2.10, the llm-d distributed inference control plane, SLO-driven inference on Chinese GPUs, Volcano + HAMi-core scheduling and isolation, and UCloud's AI platform engineering.
Inference replicas want tight binpacking, multi-GPU Pods want the same NUMA node for bandwidth, and latency-sensitive workloads want whole GPUs to themselves — three policies in one sentence, and you used to have to pick just one. In the third article of our HAMi 2.10 deep-dive series, we explain the filter-then-sort evaluation model and verify each step with a real selection path across four T4s.
GPUs have HAMi for hard memory isolation — but what about Ascend NPUs? In the second article of our HAMi 2.10 deep-dive series, we validate Volcano scheduling + HAMi-core soft slicing on a real Ascend 310P3: the container sees only its 8192 MiB slice, two Pods binpack onto the same physical card, and monitoring metrics report faithfully.
The biggest anxiety around GPU sharing is "we agreed to split it fifty-fifty — so how can you oversubscribe?" This is the first article in our HAMi 2.10 deep-dive series: with a reproducible test on GKE, we answer whether HAMi-core can truly pin each Pod's GPU memory to its quota after KAI Scheduler co-schedules two Pods onto the same NVIDIA T4. Requesting 3 GiB succeeds; accumulating 5 GiB fails with an immediate out-of-memory.
HAMi v2.10.0 is officially released! Flexible MIG creates and reclaims MIG instances on demand via NVML, AMD Instinct MI300X gains software vGPU support, the mutex policy and composable scheduling policy chains cover real-cluster needs, and the KAI Resource Isolator plus Volcano vNPU integration carry HAMi-core's isolation into a broader scheduling ecosystem.
On July 25, the CMB Digital Intelligence Innovation Meetup landed in Hangzhou. HAMi community Approver and Dynamia engineer Wang Jifei delivered a talk on "Making NPU Easier to Share." This article reviews the full evolution of HAMi vNPU from hard to soft partitioning, and its joint validation in production with China Merchants Bank.
On July 2, 2026, HAMi, the open-source project founded and long led by Dynamia, officially became a CNCF Incubating project. As the only open-source project today focused on heterogeneous GPU virtualization and efficient scheduling that has reached the CNCF Incubating stage, HAMi's milestone sends a clear signal: the AI infrastructure race is shifting from "who has more compute" to "who can manage compute more efficiently."
Get the latest technical articles, tutorials, and HAMi community updates.