Guides host-side DOCA UROM library workflows for offloading remote memory operations to BlueField DPUs.
npx skills add https://github.com/NVIDIA/skills --skill doca-urom
Please help me install this Agent Skill. Skill: doca-urom Source: https://github.com/NVIDIA/skills/tree/main/skills/doca-urom Command: npx skills add https://github.com/NVIDIA/skills --skill doca-urom
Developers building HPC, UCX, or MPI applications often struggle to offload remote memory operations (puts, gets, atomics, collectives) from the host CPU to a BlueField DPU using the DOCA UROM library, facing confusing errors like DOCA_ERROR_NOT_PERMITTED when the DPU-side UROM Service is missing or version-mismatched.
Ask your agent to help you offload MPI remote memory operations to the BlueField DPU using the doca-urom library, starting with verifying the DPU-side UROM Service is running.
High-intent search queries and answers about installing and using this skill.
Create a doca_urom_service context bound to the BlueField's doca_dev, start it, discover plugins with doca_urom_service_get_plugins_list, then create Worker contexts and enqueue Command tasks. The DPU-side UROM Service must be deployed and running before the first enqueue.
Use doca-urom when host CPU communication overhead is the bottleneck, such as HPC collectives or dense MPI patterns. For simple point-to-point RDMA where the host CPU is not loaded, doca-rdma directly is the right tool and UROM's service contract adds unnecessary overhead.
This error means either the process cannot open the target doca_dev, or the DPU-side UROM Service is not deployed, running, or at a compatible version. The API error alone cannot distinguish these causes, so verify device access and service state separately.
Call doca_urom_service_get_plugins_list on a started Service; UROM has no doca_urom_cap_* capability family, so the discovered plugin list is the authoritative capability surface. Exact operation symbols must be read from installed headers and shipped samples, not assumed.
No, UROM rides on top of the RDMA substrate and only offloads the posting of RDMA work to the DPU. A failing RDMA fabric surfaces as DOCA_ERROR_IO_FAILED at the UROM API, so substrate health must be verified first.
It covers only host-side library usage and excludes DPU-side UROM Service deployment, MPI/UCX collective algorithm design, RDMA substrate bring-up, and DOCA installation. It ships no sample code, routing users to the verified samples under /opt/mellanox/doca/samples/doca_urom/ instead.
NVIDIAConfigure and troubleshoot NVIDIA networking deployments on Kubernetes.
AscendAutomate HCCL performance testing across Ascend NPU clusters with MPI.
yo-stevenProfile and optimize TensorRT-LLM host CPU overhead with line_profiler.
outfitter-devGuide the lifecycle of developing, validating, and distributing Claude Code plugins.
