How does DCHelix connect a GPU error to the job and the customer?
Per-GPU telemetry from NVIDIA DCGM or the Intel tools is joined to the scheduler's job metadata. An Xid on one GPU names the node, the job it stalls, every GPU in that allocation, the customer who owns it, the GPU time lost per hour and the work to redo since the last checkpoint.