PodGroup

ring-controller.atlas

标识atlas的Pod

• ascend-910
• ascend-{xxx}b

Ascend Device Plugin、HCCL Controller、Volcano

fault-scheduling

任务故障重调度开关

grace、force、off

Volcano、Resilience Controller

elastic-scheduling

任务弹性调度开关

on

Resilience Controller、Volcano

fault-retry-times

任务发生业务面故障可以重调度的次数

0-100

Volcano、Ascend Operator

tor-affinity

交换机亲和性策略

• normal-schema
• large-model-schema
• null

Volcano

npu-310-strategy

标记推理服务器（插Atlas 300I 推理卡）调度策略

• card
• chip

Volcano