“isolate link errors with repeated peer-transfer and NCCL tests under load” addresses collective communication stability was not included in component acceptance, but the current hardware or infrastructure issue is accelerator visibility mask does not match the intended test scope. The option “set device visibility for all GPUs under the approved local test” is more directly tied to the required outcome.

Useful for switch power state does not prove functioning data-plane interconnect, but it does not remove the specific condition: accelerator visibility mask does not match the intended test scope.

“check unique rank mapping and supported local collective configuration” addresses local rank and GPU device assignments are inconsistent at full cardinality, but the current hardware or infrastructure issue is accelerator visibility mask does not match the intended test scope. The option “set device visibility for all GPUs under the approved local test” is more directly tied to the required outcome.

“local NCCL throughput and all required peer links meet the HGX baseline” addresses required intra-node NVLink paths are not operational, but the current hardware or infrastructure issue is accelerator visibility mask does not match the intended test scope. The option “all intended device UUIDs appear in NCCL ranks and test summary” is more directly tied to the required outcome.

“defective path is identified and post-repair pairwise NCCL tests pass” addresses peer-specific NVLink route or GPU interconnect hardware may be defective, but the current hardware or infrastructure issue is accelerator visibility mask does not match the intended test scope. The option “all intended device UUIDs appear in NCCL ranks and test summary” is more directly tied to the required outcome.

“NCCL logs and link telemetry show expected interconnect usage” addresses collective library fallback masks a missing high-bandwidth GPU link, but the current hardware or infrastructure issue is accelerator visibility mask does not match the intended test scope. The option “all intended device UUIDs appear in NCCL ranks and test summary” is more directly tied to the required outcome.

“use the documented NCCL algbw and busbw definitions when assessing limits” addresses benchmark metrics were interpreted without understanding collective data movement, but the current hardware or infrastructure issue is collective communication stability was not included in component acceptance. The option “isolate link errors with repeated peer-transfer and NCCL tests under load” is more directly tied to the required outcome.

“pin qualified NCCL version and algorithm settings for baseline runs” addresses benchmark software and algorithm differences invalidate performance comparison, but the current hardware or infrastructure issue is collective communication stability was not included in component acceptance. The option “isolate link errors with repeated peer-transfer and NCCL tests under load” is more directly tied to the required outcome.

Useful for peer-specific NVLink route or GPU interconnect hardware may be defective, but it does not remove the specific condition: collective communication stability was not included in component acceptance.

“eight-rank NCCL completes all-reduce without hangs or duplicated devices” addresses local rank and GPU device assignments are inconsistent at full cardinality, but the current hardware or infrastructure issue is collective communication stability was not included in component acceptance. The option “multiple full test loops finish without GPU or NVLink faults” is more directly tied to the required outcome.

“reported figures are evaluated against the correct topology-adjusted acceptance baseline” addresses benchmark metrics were interpreted without understanding collective data movement, but the current hardware or infrastructure issue is collective communication stability was not included in component acceptance. The option “multiple full test loops finish without GPU or NVLink faults” is more directly tied to the required outcome.

“all intended device UUIDs appear in NCCL ranks and test summary” addresses accelerator visibility mask does not match the intended test scope, but the current hardware or infrastructure issue is collective communication stability was not included in component acceptance. The option “multiple full test loops finish without GPU or NVLink faults” is more directly tied to the required outcome.

“use the documented NCCL algbw and busbw definitions when assessing limits” addresses benchmark metrics were interpreted without understanding collective data movement, but the current hardware or infrastructure issue is benchmark software and algorithm differences invalidate performance comparison. The option “pin qualified NCCL version and algorithm settings for baseline runs” is more directly tied to the required outcome.

Useful for collective communication stability was not included in component acceptance, but it does not remove the specific condition: benchmark software and algorithm differences invalidate performance comparison.

“check unique rank mapping and supported local collective configuration” addresses local rank and GPU device assignments are inconsistent at full cardinality, but the current hardware or infrastructure issue is benchmark software and algorithm differences invalidate performance comparison. The option “pin qualified NCCL version and algorithm settings for baseline runs” is more directly tied to the required outcome.

“all intended device UUIDs appear in NCCL ranks and test summary” addresses accelerator visibility mask does not match the intended test scope, but the current hardware or infrastructure issue is benchmark software and algorithm differences invalidate performance comparison. The option “measured bandwidth comparison uses matching library and test parameters” is more directly tied to the required outcome.

“defective path is identified and post-repair pairwise NCCL tests pass” addresses peer-specific NVLink route or GPU interconnect hardware may be defective, but the current hardware or infrastructure issue is benchmark software and algorithm differences invalidate performance comparison. The option “measured bandwidth comparison uses matching library and test parameters” is more directly tied to the required outcome.

“NCCL logs and link telemetry show expected interconnect usage” addresses collective library fallback masks a missing high-bandwidth GPU link, but the current hardware or infrastructure issue is benchmark software and algorithm differences invalidate performance comparison. The option “measured bandwidth comparison uses matching library and test parameters” is more directly tied to the required outcome.

“isolate link errors with repeated peer-transfer and NCCL tests under load” addresses collective communication stability was not included in component acceptance, but the current hardware or infrastructure issue is peer-specific NVLink route or GPU interconnect hardware may be defective. The option “perform pairwise peer link isolation and check NVLink Switch error records” is more directly tied to the required outcome.

“check unique rank mapping and supported local collective configuration” addresses local rank and GPU device assignments are inconsistent at full cardinality, but the current hardware or infrastructure issue is peer-specific NVLink route or GPU interconnect hardware may be defective. The option “perform pairwise peer link isolation and check NVLink Switch error records” is more directly tied to the required outcome.

“pin qualified NCCL version and algorithm settings for baseline runs” addresses benchmark software and algorithm differences invalidate performance comparison, but the current hardware or infrastructure issue is peer-specific NVLink route or GPU interconnect hardware may be defective. The option “perform pairwise peer link isolation and check NVLink Switch error records” is more directly tied to the required outcome.

“trend corrected and uncorrected FEC errors against vendor thresholds” addresses rising correction rate can warn of shrinking physical signal margin, but the current hardware or infrastructure issue is link-up alone does not guarantee adequate cable signal integrity. The option “inspect pre-FEC bit errors corrected errors and signal levels under load” is more directly tied to the required outcome.

“measure receive levels and inspect connectors and bend radius” addresses connector contamination or bend loss has impaired channel margin, but the current hardware or infrastructure issue is link-up alone does not guarantee adequate cable signal integrity. The option “inspect pre-FEC bit errors corrected errors and signal levels under load” is more directly tied to the required outcome.

Useful for installed cable environment or routing is compromising signal quality, but it does not remove the specific condition: link-up alone does not guarantee adequate cable signal integrity.

“BER meets the stated threshold across the full-rate test window” addresses signal integrity must be validated at the actual operational rate, but the current hardware or infrastructure issue is link-up alone does not guarantee adequate cable signal integrity. The option “observed error counters remain inside the approved fabric quality thresholds” is more directly tied to the required outcome.

“link quality remains good during routine door and cable movements” addresses mechanical cable handling changes optical loss under normal installation conditions, but the current hardware or infrastructure issue is link-up alone does not guarantee adequate cable signal integrity. The option “observed error counters remain inside the approved fabric quality thresholds” is more directly tied to the required outcome.

“long-duration traffic test produces no incremental uncorrected link errors” addresses installed cable environment or routing is compromising signal quality, but the current hardware or infrastructure issue is link-up alone does not guarantee adequate cable signal integrity. The option “observed error counters remain inside the approved fabric quality thresholds” is more directly tied to the required outcome.

Useful for mechanical cable handling changes optical loss under normal installation conditions, but it does not remove the specific condition: connector contamination or bend loss has impaired channel margin.

“inspect pre-FEC bit errors corrected errors and signal levels under load” addresses link-up alone does not guarantee adequate cable signal integrity, but the current hardware or infrastructure issue is connector contamination or bend loss has impaired channel margin. The option “measure receive levels and inspect connectors and bend radius” is more directly tied to the required outcome.

“trend corrected and uncorrected FEC errors against vendor thresholds” addresses rising correction rate can warn of shrinking physical signal margin, but the current hardware or infrastructure issue is connector contamination or bend loss has impaired channel margin. The option “measure receive levels and inspect connectors and bend radius” is more directly tied to the required outcome.

“stress and time-series checks show acceptable margin without escalation” addresses rising correction rate can warn of shrinking physical signal margin, but the current hardware or infrastructure issue is connector contamination or bend loss has impaired channel margin. The option “optical readings and bit-error tests return to qualified channel margins” is more directly tied to the required outcome.

“link quality remains good during routine door and cable movements” addresses mechanical cable handling changes optical loss under normal installation conditions, but the current hardware or infrastructure issue is connector contamination or bend loss has impaired channel margin. The option “optical readings and bit-error tests return to qualified channel margins” is more directly tied to the required outcome.

“fabric counters stabilize and sustained communication reaches accepted bandwidth” addresses marginal physical link quality is producing retried or degraded data transmission, but the current hardware or infrastructure issue is connector contamination or bend loss has impaired channel margin. The option “optical readings and bit-error tests return to qualified channel margins” is more directly tied to the required outcome.