Configure DOCA GPI channels so CUDA kernels drive RDMA directly from GPU memory.
npx skills add https://github.com/NVIDIA/skills --skill doca-gpi
Please help me install this Agent Skill. Skill: doca-gpi Source: https://github.com/NVIDIA/skills/tree/main/skills/doca-gpi Command: npx skills add https://github.com/NVIDIA/skills --skill doca-gpi
Developers building GPU-resident DOCA applications need to wire a GPU-Packet-Initiator context so a CUDA kernel can post RDMA work directly from GPU memory without host CPU mediation, but the doca_gpi object model, lifecycle ordering, and GPU-side handle handoff are easy to get wrong.
Ask your agent to help you bring up a DOCA GPI channel and hand the GPU-side handle to your CUDA kernel so it can post RDMA work directly from GPU memory.
High-intent search queries and answers about installing and using this skill.
Use the DOCA GPI library: create a doca_gpi instance on a doca_dev, configure it with doca_gpi_set_* attributes, start it, create a domain and channel, then call doca_gpi_gpu_channel_get to obtain the doca_gpu_gpi_channel* handle your CUDA kernel consumes.
doca-gpi is the lower-level channel surface where a CUDA kernel drives RDMA work directly via a GPU-side channel handle. doca-gpunetio is the higher-level Send/Receive-shaped Ethernet I/O surface for CUDA kernels; pick GPI when you need direct control over per-channel work submission.
You need DOCA SDK installed at /opt/mellanox/doca on Linux with a BlueField DPU or ConnectX NIC, an NVIDIA GPU with a compatible CUDA Toolkit, and pkg-config modules doca-gpi, doca-gpunetio, doca-dpa, and doca-verbs all reporting the same DOCA version.
This usually means the channel was not yet created on a configured, started GPI instance. Verify that all doca_gpi_set_* calls ran before doca_gpi_start, and that doca_gpi_channel_create completed before requesting the GPU handle.
No. Every doca_gpi_* symbol is annotated DOCA_EXPERIMENTAL with no DOCA_STABLE subset, so any call may change between releases. Pin your DOCA version and re-test the full workflow after every DOCA or CUDA Toolkit upgrade.
Avoid doca-gpi for the Send/Receive Ethernet-shaped GPU NetIO API (use doca-gpunetio), host-CPU-initiated RDMA queue lifecycles (use doca-rdma), DPA-side initiation (use doca-rdmi), or general CUDA programming questions covered by upstream CUDA documentation.
NVIDIAConfigure and troubleshoot NVIDIA networking deployments on Kubernetes.
yangwhaleInstall, configure, and troubleshoot DeepEP on NVIDIA GPU systems.
awfixers-stuffInspect ELF binaries for dependencies, symbols, sections, and debug info.
open-circleReviews PRs and source code changes in Git repositories for compliance with packaging and framework conventions.
