Operate the DOCA UROM Service container on BlueField Arm for host-offloaded remote memory operations.
npx skills add https://github.com/NVIDIA/skills --skill doca-urom-svc
Please help me install this Agent Skill. Skill: doca-urom-svc Source: https://github.com/NVIDIA/skills/tree/main/skills/doca-urom-svc Command: npx skills add https://github.com/NVIDIA/skills --skill doca-urom-svc
Deploying and operating the DOCA UROM Service on BlueField DPUs involves a paired host-library/service version contract, UCX component and collective configuration, queue sizing, and Comch/RDMA access boundaries that are easy to misconfigure and hard to debug.
Ask your agent to deploy and configure the DOCA UROM Service container on a BlueField paired with hosts using the doca-urom library.
High-intent search queries and answers about installing and using this skill.
Pull the service container from NGC at the tag matching your paired host doca-urom library version, set the daemon's SERVICE_ARGS flags and UROM_PLUGIN_PATH, mount the plugins directory, and start it under the BlueField container runtime per the public Container Deployment Guide.
Deploy it when host nodes use the doca-urom library and host CPU cost of posting RDMA communication is the bottleneck. Do not deploy it when hosts do not use doca-urom, when the stack is neither MPI nor UCX, or when the BlueField generation cannot expose the intended collectives.
DOCA_ERROR_NOT_PERMITTED is a DOCA Comch pairing, RDMA permission, or host-side doca_dev access signal, not a service-side authorization rejection. The shipped UROM service binary has no standalone authorization list; verify the Comch endpoint pairing and RDMA permissions.
Stalled completions usually come from a saturated service-side queue, a stuck handler, or an unhealthy underlying RDMA substrate. Check container state and service logs first, then inspect RDMA substrate counters before changing any service configuration.
Yes. The host-side doca-urom library version and the DPU-side service container tag are a paired contract governed by the DOCA Compatibility Policy. A mismatch fails subtly as NOT_SUPPORTED errors or silent stalls rather than a loud refusal.
No. One service container owns the BlueField's UROM execution state; running two competing containers on the same BlueField is a configuration error. Redundancy is achieved with multiple BlueFields each running their own service instance.
air-gappedAutomate NVIDIA datacenter GPU bring-up on Ubuntu 24.04 with health validation.
ChegginAutomate multi-service deployments across Convex, Railway, and Vercel with rollback safeguards.
theslashdojoExpose local services through a remote host using reverse SSH tunneling.
NVIDIAConfigure and troubleshoot NVIDIA networking deployments on Kubernetes.
