Blog

July 6, 2026

The openFuyao v26.06 community release is officially launched! In AI inference acceleration scenarios, InferNex achieves another performance breakthrough, with TTFT reduced by 55% and average total throughput improved by 32%. The newly added weight dispatcher component significantly shortens model cold-start time. At the resource scheduling layer, KubeVirt enables co-management of virtual machines and containers on Kunpeng, and many-core orchestration is integrated to improve secure colocation density. For diverse computing power access, NPU DRA capabilities are further strengthened to support a richer range of device types. Additionally, a declarative component upgrade architecture and built-in automated compliance scanning tools are introduced, achieving dual improvements in O&M efficiency and security. This release evolves across multiple dimensions, helping users achieve leaps in both efficiency and stability in AI inference scenarios.

Experience openFuyao v26.06 now

Enhanced NPU DRA Capabilities ​

SIG-orchestration-engine npu-dra-plugin capability enhancement: Building on the capabilities of previous openFuyao versions, it now supports mounting of 310P and 910C cards and hard-partitioned device mounting. Current capabilities include: View Repository

• Supports resource discovery and reporting for Ascend 910B, 910C, and 310P series NPU devices.
• Supports device filtering through DeviceClass/CEL.
• Supports resource requests using ResourceClaim/ResourceClaimTemplate, enabling binding between business Pods and ResourceSlices.
• Supports device injection into containers through CDI.
• Supports 910B hard partitioning (fixed template vNPU) functionality, automatically matching templates based on requested memory size to partition a full card into multiple vNPU instances for different business Pods.

AI Inference Acceleration Capability Upgraded Again ​

SIG-ai-inference continues to enhance AI inference acceleration capabilities, improving inference performance across multiple dimensions including TTFT, total throughput, and weight distribution.

InferNex further optimizes large model deployment capabilities and inference performance
InferNex deployment mode has changed from Deployment to the more flexible and powerful LeaderWorkerSet (LWS) mode, with new multi-DP architecture deployment capabilities, supporting one-click deployment of the latest MOE large models such as MiniMax 2.7, DeepSeek v4, and GLM 5.2. Performance has been validated in multi-turn dialogue and fixed-length system prompt scenarios, achieving an average 55% reduction in first-token latency and 32% improvement in total throughput compared to baseline. View Repository

Table 1 InferNex Performance

• infernex-bridge: New KServe integration plugin, supporting launching InferNex inference services in NPU environments through KServe. View Repository:
• infernex-checker: Provides systematic environment validation tools for InferNex deployment, performing comprehensive checks on hardware, Kubernetes clusters, and business configurations before helm install, effectively identifying potential risks and improving deployment success rates. View Repository
• hermes-router: New routing strategy based on inference latency prediction and backend compute saturation, with core component K8s GIE upgraded to the latest version 1.5.0, new tokenizer module, comprehensively enhancing request-level scheduling capabilities. View Repository
• cache-indexer: Refactored the original Python project in Go, optimizing local HBM KVCache index maintenance for inference instances, adding Mooncake-based memory-level KVCache index awareness, migrating the tokenizer step to intelligent router hermes-router, supporting more accurate global request-level scheduling. View Repository
• elastic-scaler: The elastic-scaler framework further improves CRD semantics, implements key components such as MetricsManager and Context Builder, supports LWS-based elastic scaling for large-parameter LLMs, provides APA as the default algorithm, and supports user-developed custom scaling algorithms. View Repository
• eagle-eye: New weight distribution and Lingqu architecture-based supernode network dynamic metrics collection capabilities, covering 30+ dynamic performance metrics including link status, actual transmission rate, remaining bandwidth, and packet loss rate for node-side RDMA NICs and card-side RoCE NICs, with periodic collection through Prometheus and near-real-time push through NATS. View Repository

New Weight Dispatcher Component Significantly Shortens Model Cold-Start Time

• weight-dispatcher: New model weight distribution component, enabling simultaneous distribution of model weights from a single storage node to multiple compute nodes, improving model transfer speed in cold-start scenarios. In multi-node concurrent weight pull scenarios, end-to-end model transfer time is reduced by 20%. View Repository

Enhanced KubeVirt-Based VM-Container Co-Management ​

SIG-orchestration-engine builds on virtual machine lifecycle management capabilities (including create, delete, start, stop, and status synchronization) to implement advanced virtualization management capabilities on Kunpeng environments, including NIC hot-plug, macvlan, SR-IOV, SR-IOV live migration, and VM live migration. View Repository

• NIC Hot-Plug Capability: Add or remove NICs for running Pods by adding the annotation k8s.v1.cni.cncf.io/networks. The controller listens for Pod changes, obtains container information (container ID and Net NS), and calls the CNI interface to add NICs to containers. For KubeVirt community network hot-plug user guide, see KubeVirt Network Hot-Plug User Guide.
• macvlan: Bypasses traditional Linux bridges, allowing data packets to be sent and received directly through physical NICs, reducing latency and improving throughput. In KubeVirt environments, enables VMs to directly connect to the host's underlay network for near bare-metal network performance.
• SR-IOV: Supports SR-IOV (Single Root I/O Virtualization) specification in Kunpeng hardware environments, bypassing the OS kernel for direct communication, greatly reducing network latency and CPU overhead, achieving near bare-metal performance.
• SR-IOV Live Migration: VMs configured with SR-IOV passthrough NICs support online live migration.
• UI-Based VM Management: The openFuyao container platform management plane supports custom resource management. VMs managed through KubeVirt capabilities are displayed as virtualmachines.kubevirt.io resources on the custom resource page, supporting basic lifecycle management operations.

Many-Core Orchestrator Capability Integration ​

SIG-orchestration-engine Many-Core Orchestrator: For online/offline colocation scenarios on many-core servers, provides host interference metrics collection, node interference analysis, interference-aware scheduling, and Kata VM-level isolation capabilities. Without modifying application code, it reduces online business long-tail latency jitter and improves secure colocation density. Current capabilities include: View Repository

• Host Interference Metrics Collection and Analysis: Continuously collects underlying metrics such as cache hit rate, I/O latency, and system pressure on each node, analyzes them to generate interference levels for each node, and aggregates cluster-level interference reports for scheduling decisions and O&M visibility. Collection covers three interference dimensions: cache pollution, memory pressure, and I/O congestion. When some nodes lack pressure metrics collection capability, the system automatically switches to alternative evaluation methods.
• Interference-Aware Scheduling: Based on real-time collected and analyzed underlying node interference metrics, automatically avoids high-interference nodes and prioritizes low-interference nodes during scheduling, preventing online workloads from being assigned to nodes with resource contention, reducing latency jitter risks in colocation environments from the scheduling layer.
• Kata VM-Level Isolation: Automatically completes Kata runtime installation and registration. Users only need to specify the runtime type in Pods to enable VM-level isolation for offline workloads, physically cutting off cache and I/O interference from offline workloads to online workloads. Scheduling and isolation capabilities can be used independently or combined for dual protection.
• Observability: All components provide standardized monitoring metrics integration. Interference reports are exposed through custom resources, supporting automatic discovery by mainstream monitoring systems for continuous observation of node interference status and scheduling decisions.
• Reliability Guarantee: The system follows the "do not block business scheduling" principle. When analysis components are unavailable or interference data expires, scheduling automatically falls back to basic mode without affecting normal business deployment. Each node independently determines collection capability without affecting others.

Version Upgrade Adapted to Declarative Component Architecture ​

SIG-installation adapts to the declarative component upgrade architecture in v26.06, supporting automatic orchestration, validation, and execution of upgrades based on community release package component lists and dependencies, facilitating user-defined component upgrade handling and enabling version component release and upgrade path management. Details:

• Declarative Version Package Definition: Introduces three CRDs—ClusterVersion, ReleaseImage, and UpgradePath—to define the complete component set for each openFuyao version and upgrade paths between versions, providing unified version upgrade content and entry points through images.
• Dependency-Based Upgrade Capability: Implements a DAG-scheduled upgrade engine based on component dependencies, ensuring components are installed according to dependency definitions while maximizing parallel component upgrades to reduce time.

Compliance Scanning Tool ​

SIG-security-committeeCompliance Scanning Tool: Compliance Operator is an automated compliance scanning tool based on the Kubernetes Operator pattern, performing security compliance scans on Kubernetes clusters, compatible with CIS Benchmark and DISA STIG baselines. View Repository