問題文
A user's training script indexes accelerators from zero and works when it gets the whole node, but fails when the job is allocated two of eight devices. The user asks whether the scheduler assigned the wrong devices. What is the accurate explanation?
選択肢
- The variable is only set inside the job's pre-processing and post-processing hooks.
- The scheduler allocated devices that are not adjacent on the interconnect, and a script that assumes adjacency cannot address them at all until the allocation is requested with a placement constraint.
- The devices allocated to the job are exposed with indices starting at zero from the job's point of view, so a script that assumes global indices sees the wrong numbering.
- The devices were allocated to another job first and the numbering reflects the leftover.