GPUStack uses one request model for GPUs, NPUs, MLUs, DCUs and PPUs. The resource name and the runtime isolation method depend on the manufacturer.
Accelerator Requests explains whole, shared, logically sliced and
physically partitioned requests. Device Discovery explains
how the per-node Devices record connects discovery to allocation.
Follow Heterogeneous Devices Walkthrough for end-to-end device allocation and Kueue queuing.
Check a node before installation with Preflight Checks. If you use hardware partitioning, follow the runbook for NVIDIA, T-Head or Hygon.
See also — GPU Instances (a container workspace using an accelerator) · Architecture (the scheduling chain)
Next → Accelerator Requests — choose a request shape.