“verify embedded DPU OS boot and its supported management route independently” addresses host PCIe enumeration was treated as proof of DPU control-plane readiness, but the current hardware or infrastructure issue is dPU operating-system image compatibility and provenance were not validated. The option “stage and verify a supported signed BFB image for that DPU model” is more directly tied to the required outcome.

“set the supported DPU mode according to the approved architecture and vendor procedure” addresses blueField mode and host ownership assumptions are incompatible with intended offload, but the current hardware or infrastructure issue is dPU operating-system image compatibility and provenance were not validated. The option “stage and verify a supported signed BFB image for that DPU model” is more directly tied to the required outcome.

“restrict DPU management accounts and reachability independently from workload users” addresses dPU administrative trust boundary was not configured for a production cluster, but the current hardware or infrastructure issue is dPU operating-system image compatibility and provenance were not validated. The option “stage and verify a supported signed BFB image for that DPU model” is more directly tied to the required outcome.

“DPU management recovers and the host sees restored network functions” addresses embedded DPU lifecycle is independent of the host reboot action, but the current hardware or infrastructure issue is dPU operating-system image compatibility and provenance were not validated. The option “firmware install and first boot complete with the approved DPU OS identity” is more directly tied to the required outcome.

“unauthorized host processes cannot reach DPU admin endpoints and logged authorized access succeeds” addresses dPU administrative trust boundary was not configured for a production cluster, but the current hardware or infrastructure issue is dPU operating-system image compatibility and provenance were not validated. The option “firmware install and first boot complete with the approved DPU OS identity” is more directly tied to the required outcome.

“tested traffic increments the expected hardware counters with approved offload behavior” addresses software configuration presence was mistaken for actual offload operation, but the current hardware or infrastructure issue is dPU operating-system image compatibility and provenance were not validated. The option “firmware install and first boot complete with the approved DPU OS identity” is more directly tied to the required outcome.

“restrict DPU management accounts and reachability independently from workload users” addresses dPU administrative trust boundary was not configured for a production cluster, but the current hardware or infrastructure issue is software configuration presence was mistaken for actual offload operation. The option “compare DPU steering and packet counters with workload CPU and flow tests” is more directly tied to the required outcome.

“verify embedded DPU OS boot and its supported management route independently” addresses host PCIe enumeration was treated as proof of DPU control-plane readiness, but the current hardware or infrastructure issue is software configuration presence was mistaken for actual offload operation. The option “compare DPU steering and packet counters with workload CPU and flow tests” is more directly tied to the required outcome.

Useful for BlueField mode and host ownership assumptions are incompatible with intended offload, but it does not remove the specific condition: software configuration presence was mistaken for actual offload operation.

“expected host traffic exits the correct fabric port under packet capture” addresses dPU physical port topology and host-facing interface mapping are inconsistent, but the current hardware or infrastructure issue is software configuration presence was mistaken for actual offload operation. The option “tested traffic increments the expected hardware counters with approved offload behavior” is more directly tied to the required outcome.

“firmware install and first boot complete with the approved DPU OS identity” addresses dPU operating-system image compatibility and provenance were not validated, but the current hardware or infrastructure issue is software configuration presence was mistaken for actual offload operation. The option “tested traffic increments the expected hardware counters with approved offload behavior” is more directly tied to the required outcome.

“cross-tenant probe traffic is blocked while permitted offloaded flows succeed” addresses accelerator offload network rules violate tenant isolation requirements, but the current hardware or infrastructure issue is software configuration presence was mistaken for actual offload operation. The option “tested traffic increments the expected hardware counters with approved offload behavior” is more directly tied to the required outcome.

“verify embedded DPU OS boot and its supported management route independently” addresses host PCIe enumeration was treated as proof of DPU control-plane readiness, but the current hardware or infrastructure issue is dPU administrative trust boundary was not configured for a production cluster. The option “restrict DPU management accounts and reachability independently from workload users” is more directly tied to the required outcome.

Useful for software configuration presence was mistaken for actual offload operation, but it does not remove the specific condition: DPU administrative trust boundary was not configured for a production cluster.

“stage and verify a supported signed BFB image for that DPU model” addresses dPU operating-system image compatibility and provenance were not validated, but the current hardware or infrastructure issue is dPU administrative trust boundary was not configured for a production cluster. The option “restrict DPU management accounts and reachability independently from workload users” is more directly tied to the required outcome.

“authenticated DPU management session returns platform health and operating version” addresses host PCIe enumeration was treated as proof of DPU control-plane readiness, but the current hardware or infrastructure issue is dPU administrative trust boundary was not configured for a production cluster. The option “unauthorized host processes cannot reach DPU admin endpoints and logged authorized access succeeds” is more directly tied to the required outcome.

“firmware install and first boot complete with the approved DPU OS identity” addresses dPU operating-system image compatibility and provenance were not validated, but the current hardware or infrastructure issue is dPU administrative trust boundary was not configured for a production cluster. The option “unauthorized host processes cannot reach DPU admin endpoints and logged authorized access succeeds” is more directly tied to the required outcome.

“reported software and device firmware versions match the qualified stack” addresses dPU firmware and DOCA version compatibility was not established, but the current hardware or infrastructure issue is dPU administrative trust boundary was not configured for a production cluster. The option “unauthorized host processes cannot reach DPU admin endpoints and logged authorized access succeeds” is more directly tied to the required outcome.

“select a supported instance profile providing sufficient memory and compute resources” addresses requested MIG instance memory profile does not meet workload demand, but the current hardware or infrastructure issue is whole-device scheduling cannot provide requested MIG resource isolation. The option “enable supported MIG mode and create qualified GPU and compute instances” is more directly tied to the required outcome.

“drain workloads and follow the supported MIG reconfiguration procedure” addresses changing GPU partition mode has workload disruption and reset implications, but the current hardware or infrastructure issue is whole-device scheduling cannot provide requested MIG resource isolation. The option “enable supported MIG mode and create qualified GPU and compute instances” is more directly tied to the required outcome.

“benchmark candidate MIG sizes and choose the qualified performance profile” addresses compute instance capacity and workload demand are mismatched, but the current hardware or infrastructure issue is whole-device scheduling cannot provide requested MIG resource isolation. The option “enable supported MIG mode and create qualified GPU and compute instances” is more directly tied to the required outcome.

“scheduled processes see only their assigned MIG UUID and profile” addresses scheduler resource discovery does not match physical GPU partitioning, but the current hardware or infrastructure issue is whole-device scheduling cannot provide requested MIG resource isolation. The option “nvidia-smi displays intended MIG instances and tenant jobs run on assigned devices” is more directly tied to the required outcome.

“the job allocates its working set and completes within its assigned instance” addresses requested MIG instance memory profile does not meet workload demand, but the current hardware or infrastructure issue is whole-device scheduling cannot provide requested MIG resource isolation. The option “nvidia-smi displays intended MIG instances and tenant jobs run on assigned devices” is more directly tied to the required outcome.

“no jobs occupy the target GPU during mode change and new instances enumerate afterward” addresses changing GPU partition mode has workload disruption and reset implications, but the current hardware or infrastructure issue is whole-device scheduling cannot provide requested MIG resource isolation. The option “nvidia-smi displays intended MIG instances and tenant jobs run on assigned devices” is more directly tied to the required outcome.

“enable supported MIG mode and create qualified GPU and compute instances” addresses whole-device scheduling cannot provide requested MIG resource isolation, but the current hardware or infrastructure issue is requested MIG instance memory profile does not meet workload demand. The option “select a supported instance profile providing sufficient memory and compute resources” is more directly tied to the required outcome.

Useful for scheduler resource discovery does not match physical GPU partitioning, but it does not remove the specific condition: requested MIG instance memory profile does not meet workload demand.

“benchmark candidate MIG sizes and choose the qualified performance profile” addresses compute instance capacity and workload demand are mismatched, but the current hardware or infrastructure issue is requested MIG instance memory profile does not meet workload demand. The option “select a supported instance profile providing sufficient memory and compute resources” is more directly tied to the required outcome.

“measured latency and throughput meet service thresholds per assigned instance” addresses compute instance capacity and workload demand are mismatched, but the current hardware or infrastructure issue is requested MIG instance memory profile does not meet workload demand. The option “the job allocates its working set and completes within its assigned instance” is more directly tied to the required outcome.

“HPC allocation obtains the required visible memory and completes its workload” addresses workload shape conflicts with the partition strategy, but the current hardware or infrastructure issue is requested MIG instance memory profile does not meet workload demand. The option “the job allocates its working set and completes within its assigned instance” is more directly tied to the required outcome.

“instance inventories and cross-tenant resource tests show prescribed separation” addresses isolation requirements were not translated into appropriate MIG instance assignments, but the current hardware or infrastructure issue is requested MIG instance memory profile does not meet workload demand. The option “the job allocates its working set and completes within its assigned instance” is more directly tied to the required outcome.

“benchmark candidate MIG sizes and choose the qualified performance profile” addresses compute instance capacity and workload demand are mismatched, but the current hardware or infrastructure issue is changing GPU partition mode has workload disruption and reset implications. The option “drain workloads and follow the supported MIG reconfiguration procedure” is more directly tied to the required outcome.

“enable supported MIG mode and create qualified GPU and compute instances” addresses whole-device scheduling cannot provide requested MIG resource isolation, but the current hardware or infrastructure issue is changing GPU partition mode has workload disruption and reset implications. The option “drain workloads and follow the supported MIG reconfiguration procedure” is more directly tied to the required outcome.

Useful for MIG profile capabilities are device-generation and profile specific, but it does not remove the specific condition: changing GPU partition mode has workload disruption and reset implications.