A school-district lab runs single-node NCCL all-reduce. What participation result is required for acceptance?
Select an answer to reveal the explanation.
Short Explanation
All-reduce means all the GPUs—if one ghosted the school-lab test, hunt it down before you stamp acceptance.
Full Explanation
School-district lab acceptance for single-node NCCL all-reduce requires confirming every GPU in the node participates; any silent absence must be investigated. Collectives that quietly omit a device indicate visibility, driver, cgroup, MIG partitioning, or topology faults that will later break multi-GPU training and inference for lab users who expect the full device count.
Accepting if any single GPU reports a result fails by concept—partial participation is not all-reduce success for a multi-GPU node and leaves defective or invisible devices undiscovered. Excluding half the GPUs deliberately to shorten the test fails because abbreviated participation hides the defects acceptance exists to find and falsifies the node’s usable GPU count in inventory. Running NCCL with zero GPUs selected fails as a non-test that cannot establish interconnect health or device participation at all.
Exam caveat: Domain 4 NCCL acceptance hinges on full device participation in collectives, not “some GPU returned a number” or intentionally reduced world sizes for convenience.
Operational check: configure the all-reduce to target all visible GPUs, assert world size equals the expected GPU count from school-lab inventory, fail the node on mismatch, and triage missing devices with nvidia-smi, topology dumps, and runtime visibility checks before lab pool admission.