GPU Sharing: Design#
conceptstub
Trade-offs, decision records, and when to choose which option.
Planned content
- Decision matrix: workload type → sharing mechanism
- Why sliced GPUs do not compose into NCCL communicators
- Per-node configuration via labels