New Features

MindIO ACP

MindIO ACP is open-sourced.

MindIO TFT

• MindIO TFT supports hot switchover upon subhealthy faults in the MindSpore scenario.
• MindIO TFT is open-sourced.

MindCluster ToolBox

• Adds PCIe full eyed diagram test and optimizes on-chip memory stress test duration and algorithms for A3 servers (A+X).
• Adds the 310P power consumption stress test.
• Adds the DSA stress test.

MindCluster Ascend FaultDiag

Adds fault events of A3 AI servers.

MindCluster Ascend Deployer

N/A

MindCluster components

• Process-level online recovery for UnifiedBus L1-L2 link faults is supported when operator re-execution is disabled.
• Traffic isolation for faulty NPU instances can be deployed based on AIBrix vLLM.
• NPU Exporter can output the SN.
• AIBrix vLLM-based serving instance-level rescheduling is supported.
• Based on the CRD definition of the AIBrix community, YAML files can be generated by one-click mode, and one-click configuration and delivery are supported.
• Based on the native CRD definition, YAML files can be generated by one-click mode, and one-click configuration and delivery are supported.
• SGLang OME deployment and instance-level rescheduling are supported.
• The reliability of UnifiedBus fault reporting is enhanced.
• An adaptation layer is added to Volcano to isolate the differences between different job controllers and support affinity scheduling in all podGroups that meet the format requirements.
• The scheduling resource usage is optimized. If scheduling is not complete, a job can be re-enqueued after a certain period of time.
• Pre-isolation processing is supported for public faults.
• NPU Exporter supports custom metrics.
• Multi-instance inference job scheduling is supported for A3 servers.
• A3 servers are compatible with accelerator-type of A2 servers.
• The ecosystem component compatibility is verified.
• A reference design for the inference job daemon is introduced.
• NPU fault detection and rectification are supported for appliances.
• Volcano scheduling supports StatefulSet.
• Hot switching under MindSpore is supported.
• The usability of quick training recovery is enhanced.