nvidia

Connects NemoClaw to a local inference server. Use when setting up Ollama, vLLM, TensorRT-LLM, NIM, or any OpenAI-compatible local model server with NemoClaw. Trigger keywords - nemoclaw local inference, ollama nemoclaw, vllm nemoclaw, local model server, openai compatible endpoint, switch nemoclaw inference model, change inference runtime, nemoclaw additional model, nemoclaw sub-agent model, openclaw sub-agent, agents.list, sessions_spawn, vlm-demo, nemoclaw tool calling, ollama tool calls, vllm tool-call-parser, raw json in tui, nemoclaw inference options, nemoclaw onboarding providers, nemoclaw inference routing.

Solve LP, MILP, QP (beta) with cuOpt Python API — linear/quadratic objectives, integer variables, scheduling, portfolio, least squares.

cuOpt REST server — what it does and how requests flow. Domain concepts; no deploy or client code.

Describes the NemoClaw plugin and blueprint architecture and how they orchestrate the OpenClaw sandbox. Use when looking up architecture, plugin structure, or blueprint design. Trigger keywords - nemoclaw architecture, nemoclaw plugin blueprint structure, nemoclaw vs openshell, which cli, nemoclaw cli, openshell cli, sandbox commands, nemoclaw cli commands, nemoclaw command reference, nemoclaw network policy, sandbox egress control operator approval, nemoclaw troubleshooting, nemoclaw debug sandbox issues.

Explains how OpenClaw, OpenShell, and NemoClaw form the ecosystem, NemoClaw's position in the stack, what NemoClaw adds beyond the community sandbox, and when to prefer NemoClaw versus integrating OpenShell and OpenClaw directly. Use when users ask about the relationship between OpenClaw, OpenShell, and NemoClaw, or when to use NemoClaw versus OpenShell. Trigger keywords - nemoclaw ecosystem, openclaw openshell, nemoclaw vs openshell, sandboxed openclaw, how nemoclaw works, nemoclaw sandbox lifecycle blueprint, nemoclaw overview, openclaw always-on assistants, nvidia openshell, nvidia nemotron, nemoclaw release notes, nemoclaw changelog.

Describes the agent skills shipped with NemoClaw and how to access them by cloning the repository. Use when users ask about AI agent support, coding assistant integration, or the .agents/skills/ directory. Trigger keywords - nemoclaw agent skills, ai coding assistant, cursor, claude code, copilot.

Explains how to run NemoClaw on a remote GPU instance, including the deprecated Brev compatibility path and the preferred installer plus onboard flow. Use when deploying NemoClaw to a remote VM, onboarding a Brev instance, or migrating away from the legacy `nemoclaw deploy` wrapper. Trigger keywords - deploy nemoclaw remote gpu, nemoclaw brev cloud deployment, nemoclaw plugins, openclaw plugins, install openclaw plugin, nemoclaw onboard from dockerfile, nemoclaw brev web ui, nemoclaw getting started, brev quickstart, nvidia nemotron agent, nemoclaw sandbox hardening, container security, docker capabilities, process limits.

Inspects sandbox health, traces agent behavior, and diagnoses problems. Use when monitoring a running sandbox, debugging agent issues, or checking sandbox logs. Trigger keywords - monitor nemoclaw sandbox, debug nemoclaw agent issues.

Explains operational tasks after the quickstart: listing sandboxes, status and health checks, logs, diagnostics, port forwards, multiple sandboxes, credential reset, rebuilds, network presets, upgrades, and uninstall. Trigger keywords - manage nemoclaw sandboxes, nemoclaw status, nemoclaw list, nemoclaw dashboard port, nemoclaw rebuild, nemoclaw upgrade sandboxes, nemoclaw uninstall, sandbox mutability, sandbox runtime configuration, sandbox rebuild, nemoclaw backup, nemoclaw restore, workspace backup, openshell sandbox download upload, nemoclaw messaging channels, nemoclaw telegram, nemoclaw discord, nemoclaw slack, nemoclaw wechat, nemoclaw whatsapp, openshell channel messaging, nemoclaw workspace files, soul.md, user.md, identity.md, agents.md, sandbox persistence.

After solving a non-trivial problem, detect generalizable learnings and propose skill updates. Always active — applies to every interaction.

Presents a risk framework for every configurable security control in NemoClaw. Use when evaluating security posture, reviewing sandbox security defaults, or assessing control trade-offs. Trigger keywords - nemoclaw security best practices, sandbox security controls risk framework, nemoclaw credential storage, openshell provider, api key security, openclaw security controls, nemoclaw security boundary, prompt injection, tool access control.

Hardware-offloaded erasure coding with DOCA Erasure Coding (Reed-Solomon, doca_ec) on a BlueField DPU, ConnectX NIC, or host: generating parity for data blocks, rebuilding lost blocks from the survivors, and refreshing parity after a change. Trigger on "one data block changed, how do I refresh parity without re-encoding", "a disk failed and 2 parity blocks are gone, can I rebuild", "RAID-6 resilience across 12 disks", "my doca_ec_task_create returns NOT_PERMITTED", or "is this N+K layout still recoverable", even without naming DOCA Erasure Coding. Covers bringing up a doca_ec context, the create / recover / update tasks, matrix type / N / K / block size, doca_ec_cap_* queries before sizing, doca_mmap src/dst permissions, and DOCA_ERROR_* from doca_ec_task_*. Refuse and route elsewhere for non-Reed-Solomon codes (fountain / LDPC / raptor), pure-replication designs, network FEC, or other DOCA accelerator libraries (SHA / Compress / AES-GCM / DMA).

Use this skill when the user is operating the DOCA Firefly Service container on BlueField — picking the four PTP configuration axes (role / profile / domain / interface), wiring the BlueField PHC + host follower + consumer workload pairing, deciding whether PTP-grade time is even needed (vs. chrony / NTP), or debugging a Firefly deployment where PTP isn't syncing or the host clock isn't following. Trigger even when the user does not explicitly mention "DOCA Firefly" or "PTP" — typical implicit phrasings include "container green but PTP never advances past LISTENING", "Firefly says synced but the host clock still drifts", "sync acquired but offset is tens of microseconds", "my Rivermax SMPTE workload needs PTP", or "is chrony good enough". Refuse and route elsewhere for installing DOCA, host-side chrony / ptp4l config bodies, PTP topology / boundary-clock design, building DOCA apps that read the disciplined PHC, or other DOCA services (DMS, Flow-Inspector, HBN) — those belong to other skills.

Build and debug DOCA Flow applications on supported NVIDIA NICs/DPUs: define match/action pipes, initialize ports and representors, choose forwarding targets, validate pipes before hardware programming, read counters, match the Flow version to the installed DOCA release, and diagnose Flow API errors. Trigger on DOCA packet steering, classifier, representor, rule-matching, hairpin, or 5-tuple-to-queue questions even when "DOCA Flow" is not named. Route plain DPDK `rte_flow`, kernel TC, OVS, BFB bring-up, and DPU OS installation elsewhere. DPU OS installation is destructive and always requires explicit confirmation.

Use this skill when the user is doing hands-on DOCA Flow DPA Provider work — exporting a `doca-flow` pipe or external resource (index-selector/memory) into BlueField DPA address space so a DPACC-built kernel can read counters, mutate hash-pipe entries, and update/read memory or index-selector resources inline with Flow. Covers per-port `doca_flow_dpa_ctx`, three queue types (general/resources-write/resources-read), the order-sensitive export handshake (`_export_prepare` → add entries → `_export` → `_get_device_addr`), and DPA-side device API. Trigger even when the user does not say "DOCA Flow DPA Provider" — implicit phrasings include "DPA kernel never sees entries in the exported pipe", "BAD_STATE from `_pipe_export`", "how do I disable a hash entry from a DPA kernel", "DPA memory read returns no value", or "DPA-side post keeps returning AGAIN". Refuse and route elsewhere for `doca-flow` pipe construction, generic host-side DPA (`doca-dpa`), or DPA-side kernel-writing — those belong to other skills.

Use this skill when the user is doing hands-on DOCA GPUNetIO programming — wiring a CUDA kernel on an NVIDIA GPU to a doca-eth queue via doca_gpu_eth_rxq / doca_gpu_eth_txq, standing up the per-CUDA-device doca_gpu context, designing the persistent CUDA kernel that drains the GPU-visible queue, running the dual capability check (DOCA cap-query plus cudaGetDeviceProperties), registering cudaMalloc pools via doca_buf_arr_create_*, or debugging DOCA_ERROR_* returns from the GPUNetIO API. Trigger even when the user does not explicitly mention "DOCA GPUNetIO" or "persistent kernel" — typical implicit phrasings include "CUDA kernel reading packets directly from the NIC", "GPU-initiated networking on BlueField", "DOCA_ERROR_DRIVER on doca_gpu_create", "nvidia_peermem not loaded", "kernel-per-packet is too slow", or "which GPU supports GPU-side packet I/O". Refuse and route elsewhere for general CUDA programming, DOCA Ethernet queue bring-up, DOCA DPA, or DOCA install — those belong to other skills.

Use this skill when the user is measuring GPU-kernel-initiated RDMA WRITE latency through doca-gpunetio — building and running the `gpunetio_ib_write_lat` client + server pair under `doca/tools/gpunetio_ib_write_lat/`, checking GPU-NIC pairing, reading the half-iter / full-iter / CUDA-side usec columns, characterizing median / p99 / jitter for a real-time control loop, picking GPUNetIO vs GPI vs CPU-initiated `perftest`, or weighing the latency-vs-batching trade-off. Trigger even without 'GPUNetIO' or 'ib_write_lat': 'GPU kernel RDMA latency benchmark', 'how fast can a CUDA kernel post a WRITE', 'p99 RDMA latency on H100 + ConnectX', 'kernel-launched WR tail latency', or 'compare GPU-init vs CPU-init perftest'. Route elsewhere for bandwidth runs (doca-gpunetio-ib-write-bw), the GPI surface (doca-gpi), library debugging (doca-gpunetio), or DOCA install.

Use this skill when the user needs to locate authoritative information about NVIDIA DOCA without access to the source tree — finding the right docs.nvidia.com page for a library/service/tool, identifying which DOCA libraries are installed and at what version, locating a sample on disk or its public GitHub source, decoding an on-disk path under /opt/mellanox/doca, or recovering from a 404'd or renamed doc URL. Trigger even when the user does not explicitly mention 'DOCA' or 'docs.nvidia.com' — typical implicit phrasings include 'where can I read about this library', 'which version do I have installed', 'where is the sample for X', 'this NVIDIA URL is broken what is the new one', 'what is in /opt/mellanox/doca', or 'where can I ask NVIDIA about this'. Refuse and route elsewhere for hands-on programming patterns, env prep and install verification, library API tutorials, or hardware/firmware mutation — those belong to doca-programming-guide, doca-setup, the per-library skills, and doca-hardware-safety.

Use this skill when the user needs SHA-1, SHA-256, or SHA-512 digest acceleration on BlueField-2 silicon via doca_sha (absent on BF-3) — selecting one-shot `doca_sha_task_hash` versus incremental `doca_sha_task_partial_hash` for chunked input, consulting `doca_sha_cap_*` for per-algorithm digest-size limits and input alignment, or decoding DOCA_ERROR_* statuses such as BAD_STATE on a resumed partial hash. Also load it for implied asks like "accelerate SHA-256 on the NIC", "streaming digest over chunks", "digest output size mismatch", or "is SHA-512 present on this card". Refuse and route to the proper owner for cryptographic-hash theory (collision resistance, SHA-3 selection), the AES-GCM / Compress / DMA libraries, or SDK install and BFB bring-up.

Use this skill when the operator is driving the DOCA Socket Relay to bridge a socket-oriented host application onto a BlueField DPU peer without rewriting it — picking the deployment shape (in-process, sidecar, or BlueField service container), configuring the host-side socket and the DPU-side forwarding endpoint, walking the bind → connect → round-trip → admit-fleet smoke, or diagnosing a stuck/silent relay. Trigger even when the user does not explicitly mention "DOCA Socket Relay" — typical implicit phrasings include "move my socket app onto the BlueField without rewriting it", "host app gets ECONNREFUSED on the relay", "relay accepts the connection but bytes never arrive on the DPU side", "first round-trip works, the rest hang", "bridge an AF_UNIX (UDS) socket to a DPU peer over Comch", or "I want a sidecar that forwards my socket to the BlueField". Refuse and route elsewhere for the comch programming API, line-rate raw packet I/O via doca-eth, and DOCA install/bring-up — those belong to other skills.

Use this skill when the user is invoking `doca_spcx_cc` (the host-side CLI under /opt/mellanox/doca/tools/) to load, parameterize, start, observe, or stop an SPC-X congestion-control algorithm on a BlueField with a DPA processor against a live RDMA / RoCE fabric, or picking SPCX vs the established `doca-pcc` surface. Trigger even when the user does not say "DOCA SPCX" or "doca_spcx_cc" — typical implicit phrasings include "I want to write a custom RTT-based CC algorithm for my RoCE fabric", "my SPCX session loaded but throughput / latency didn't change", "doca_pcc status shows Active but factory CC seems to still be in charge", "DOCA_PCC_PS_ERROR on start", "is the programmable-CC surface available on my install", or "DPA-side algorithm image won't load". Refuse and route elsewhere for DPA-side algorithm authoring detail, factory PCC firmware configuration, read-only PCC counter inspection, raw DPA cycle profiling, RDMA library programming, or general DOCA install — those belong to other skills.

Operate the DOCA UROM Service container on BlueField Arm for remote memory operations (puts, gets, atomics, collectives) enqueued by a paired host using `doca-urom`: pull the NGC image, choose the UCX component, size queues, configure Comch pairing, and align host and service versions. SECURITY: the service has no standalone access control; Comch pairing and RDMA permissions are the boundary. Pair only intended hosts, expose least-privilege memory regions, and verify both views before start. Trigger for slow UCX collectives, unexpected NOT_PERMITTED, or missing completions. Do not use for host application code, MPI/UCX integration design, or DOCA install.

Guide for selecting and configuring distributed training strategies in NeMo AutoModel, including FSDP2, Megatron FSDP, DDP, and parallelism settings.

Validate and use CUDA graph capture in Megatron Bridge, including local full-iteration graphs and Transformer Engine scoped graphs for attention, MLP, and MoE modules.

Assess a host and choose a Holoscan SDK 5.x installation method using release availability, CUDA 13 compatibility, and component needs.

Use when searching, extracting, ingesting, or querying a document collection with the NeMo Retriever 26.8.1 CLI, including local LanceDB indexes and deployed Retriever services. Use for PDFs, images, Office files, HTML, text, audio, and video; not for editing documents or web search.

Performance benchmarking for a deployed NVIDIA RAG Blueprint server: profiling pass + aiperf load test driven by a single YAML config. Not for accuracy / RAGAS scoring (use rag-eval) or for deploying / repairing services (use rag-blueprint).

Enable Jetson Thor 25G/10G/1G MGBE QSFP via kernel-DT overlay. Do NOT use for UPHY lane allocation or ODMDATA edits.

Use when you need to add, remove, edit, list, or change the boot default of an nvpmodel power mode on a Jetson/Tegra (Orin, Thor) target. Triggers: edit power mode, tune frequency caps.

Author a new Jetson target-platform profile (reference_devkit + optional custom_carrier) and update the active pointer. Use to create a target; not for switching existing profiles.

Techniques for reducing peak GPU memory in Megatron Bridge — expandable segments, PEFT + SP input re-gather, parallelism resizing, activation recompute, CPU offloading constraints, and common OOM fixes.

Use when a user needs to get started with PAIDF Auto-Labeling, plan a scenario, run or debug a shipped cookbook, author prompts or cookbooks, migrate a pipeline, or configure a stage. Confirm critical inputs (data path, output path, endpoints) and ask when any are missing. This is a router: read the matching reference instead of inventing a workflow.

Use when a user describes a custom PAIDF Orchestration pipeline — a specific ordered combination of stages such as augmentation only, auto-labeling only, detection+captioning only, or image attribute augmentation without full auto-labeling — that no existing DAG in airflow/dags/workflows/ covers, and asks for a new Kubernetes DAG. Also use to check that a generated or existing DAG's model/container/prompt choices match an external spec document (e.g. a PAIDF `launchable.md`).

Discover and run Holoscan Sensor Bridge example applications on a connected devkit. Filters available apps by the user's platform, HSB software version, board type, and sensors. Supports timed execution, failure analysis, code-edit suggestions, and iterative re-runs.

Choose the right MoE token dispatcher (`alltoall`, DeepEP, or HybridEP) for the hardware, EP degree, and optimization stage. Summarizes patterns from DSV3, Qwen3, Qwen3-Next, and VLM bring-up work.

Install and verify cuPyNumeric for Python — requirements, commands, verification. Source builds are out of scope.

The data-mover for TAO jobs — decides the storage tier (A pre-positioned mount with zero fetch / B volume-from-S3 / C ephemeral in-compute fetch), stages inputs (bulk + annotation-selective + archive extract + HF/NGC PTM), maps credentials to env, routes outputs 3-way with upload-excludes, and runs the compute-frame verify gate. A support skill other platform skills (docker, kubernetes, slurm, brev, virtualenv) call to get data to and from the compute container without the TAO SDK. Trigger phrases include "stage inputs", "mount the dataset", "upload TAO results", "download only referenced files", "resolve results_dir", "verify the container can read the data".

One-time session setup and orchestration map for the TAO skill bank. Run this first when the TAO skills were installed individually (e.g. from a public skills catalog) so the session gets the cross-skill discovery flow, credential checks, and host preflight that the bundled plugin hook would otherwise inject automatically. Trigger phrases include "set up TAO skills", "TAO session setup", "prepare TAO environment", "TAO getting started".

PointPillars for 3D object detection from LiDAR point clouds. Encodes point clouds into a pseudo-image via a pillar-based representation, then applies 2D detection — used in autonomous driving and robotics. Use when training, evaluating, exporting, pruning, retraining, or running inference for a TAO PointPillars model. Trigger phrases include "train PointPillars", "LiDAR 3D detection", "point-cloud object detection", "pillar-based 3D detector".

Run `tao-daft validate` to check NVIDIA TAO DAFT datasets for structure, schema, and cross-reference errors. Do not use for non-DAFT formats. Use when the user asks to validate a DAFT dataset, check DAFT schema, validate a TAO dataset format, or run `tao-daft validate`.

NVIDIA RAG Blueprint — deploy, configure, troubleshoot, and manage. Handles any RAG action: deploy, install, start, enable, disable, toggle, change, configure, troubleshoot, debug, fix, shutdown, stop, or tear down any RAG feature or service (Agentic RAG, VLM, guardrails, query rewriting, models, search, ingestion, observability, summarization, reasoning, and more).

Plan, configure, and chain repo-native Nemotron customization steps into single-step or multi-step pipelines: curation, translation, SFT/PEFT (AutoModel or Megatron-Bridge), pretraining/CPT, RL alignment (DPO/RLVR/GRPO/RLHF), BYOB/MCQ benchmarks, checkpoint conversion, ModelOpt optimization, env profiles, and evaluation of trained checkpoints or existing/hosted endpoints. Use when a request names a Nemotron step or workflow, or asks to clean, translate, train, fine-tune, align, convert, optimize, evaluate, or compose these into a pipeline. Do NOT use for frontend/dashboard/visualization work, generic ML advice, billing/access, or non-Nemotron coding tasks.

Operational guide for enabling Megatron FSDP in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.

Expand workflow HDF5 demonstrations with action jitter, optionally scoped to node segments. Use for synthetic variants; do not use to collect data, alter state directly, or generate new images.

Preflight and set up the root-level workflow runtime. Use for installation, missing component environments, or third-party failures; do not use for rollout validation.

Read and write large cuPyNumeric arrays to HDF5 with Legate's parallel, distributed HDF5 I/O (legate.io.hdf5: to_file, from_file, from_file_batched). Use when a developer needs to save a cuPyNumeric array to an .h5/.hdf5 file, load an HDF5 dataset into a distributed cuPyNumeric array, read a large HDF5 dataset in chunks, hand arrays to an HPC pipeline as a single file, or accelerate HDF5 disk I/O with GPUDirect Storage (GDS). Do not use it for Parquet/cuDF/raw-binary or other sharded/custom layouts (see the cupynumeric-parallel-data-load skill), Zarr or object-store/S3 output, .npz or pickled archives, plain h5py without cuPyNumeric, or pure array compute such as FFT, matmul, or reductions.

MoE expert-parallel communication overlap in Megatron Bridge. Covers dispatch/combine overlap, flex dispatcher backends, and expert wgrad scheduling.

Real-time stereo depth estimation using FastFoundationStereo (FFS), the distilled bp2 commercial variant of FoundationStereo. Predicts disparity maps from stereo image pairs with ~10× lower latency than full FoundationStereo. Use when training, evaluating, exporting, or running inference for a TAO FastFoundationStereo (FFS) model. Trigger phrases include "train fast stereo", "real-time stereo disparity", "FastFoundationStereo", "distilled stereo depth".

Representative, point-in-time MoE training playbooks by hardware and model family. Use them as candidate seeds, then revalidate the exact runtime, semantics, topology, and steady-state throughput.

Recommend and customize Megatron Bridge library and benchmark recipes for a user's model, GPU count, hardware, sequence length, and pretrain/SFT/PEFT goal. Use when selecting a starting recipe, comparing library and benchmark configs, resizing parallelism for a GPU allocation, or distinguishing convergence changes, semantics-preserving execution tuning, and benchmark-only shortcuts.

Docker conventions for running NVIDIA GPU container workloads — NGC authentication, --gpus flag, mount patterns, env-var passthrough, container inspection, data-root relocation for split-disk hosts, and common error modes. Use when another skill requires running an nvcr.io container or any docker run command on a GPU host. Trigger keywords — docker, docker run, nvcr.io, NGC, --gpus, nvidia-container-toolkit, container image, docker login, docker pull.

Diagnose and optimize OpenUSD scene performance

Integrate TileGym kernels into Hugging Face `transformers` models by replacing the library's submodule(s) and certain class(es)' implementations, and patching certain class(es)' init/forward/load weight methods prior to instantiating models. Used when the user requires integrating TileGym kernels into `transformers` models.

Convert existing Hugging Face Transformers Trainer or TRL SFTTrainer training code into an NVFLARE federated job using flare.patch(trainer), local validation, and job export; use when the user names Hugging Face or preliminary source inspection identifies one Hugging Face owner, and not for manual PyTorch loops, Lightning, inference-only pipelines, deployment, or experiment workflows.

Use when the user asks why a reported NVFLARE job failure signal occurred: the job failed, stalled, timed out, lost clients, ended with EXECUTION_EXCEPTION, or produced suspicious errors. Diagnose in simulation, POC, or production by collecting bounded evidence and mapping failure patterns to recovery actions.

Compute federated statistics over tabular data (count, sum, mean, stddev, var, histogram, quantile, noise-protected min/max) and image data (count, failure_count, pixel-intensity histogram) across NVFLARE sites via FedStatsRecipe — automatic and non-interactive from the dataset, feature names (header or supplied), and optionally a README or notes declaring which statistics to compute; do not use for model training conversion, hierarchical statistics, deployment, POC/production lifecycle, or failed-job diagnosis.

Route open-ended NVFLARE advice and only conversion requests whose preliminary source inspection reports unresolved or conflicting ownership; never load this skill merely to inspect a concrete conversion request before selecting its detected framework converter.

Internal NVFLARE conversion references and templates. Use only when another NVFLARE skill directs you to a shared workflow, policy, or asset.

Evaluate models, datasets, and agents with the NeMo Evaluator plugin. Use for metric selection, SDK checks, platform jobs, and result retrieval.

Create, refine, or fix NVIDIA voice agents (Cascaded or Omni) with Pipecat or LiveKit. Use when building, scaffolding, or iterating on a real-time voice-agent pipeline, including speech (ASR/TTS) customization and cloud or local deployment. Not for offline/batch speech-to-text, text-only chat or RAG, or generic Docker or CUDA work unrelated to a voice agent.

Base rules for end users calling NVIDIA cuOpt (routing/LP/MILP/QP/install/server). Not for cuOpt internals — use cuopt-developer for those.

Replay a workflow HDF5 episode through its original Scene. Use for visual trajectory and recording verification; do not use for policy evaluation or LeRobot data.

NVIDIA DeepStream SDK development with Python pyservicemaker API. Use when building video analytics pipelines, GStreamer-based video processing, TensorRT inference integration, object detection/tracking, or Kafka/message broker integration.

Used for generating synthetic CT volumes and masks with NV-Generate-CTMR rflow-ct. Not for production training data without review.

Used for finetuning NV-Generate-CTMR MR-brain diffusion UNet from a NIfTI datalist. Not for clinical or production data approval.

Used for finetuning the NV-Generate-CTMR MAISI VAE from CT/MRI NIfTI datalists. Not for clinical or production data approval.

Use when authoring or validating PAIDF augmentation YAML configs, or running remote Cosmos Transfer (including Cosmos3 WSM controls), Cosmos Predict, image-edit, or image-to-video inference.

Used for command-shape or live NV-Reason-CXR chest X-ray reasoning smoke tests. Not for diagnosis or clinical reporting.

Run the full DEFT smart-data-augmentation loop for NVIDIA TAO Grounding DINO object detection: zero-shot baseline inference, KPI analysis, per-class gap analysis, SigLIP embedding of weak images, unique-neighbor mining against a source pool, ODVG dataset staging, and retraining — repeated for a fixed number of iterations. Also prepares the source pool the loop mines from, as a separate run: Co-DETR pseudo-labeling, folding to the target classes, KITTI→COCO→ODVG conversion, and embedding. Use for prompts like "run the DEFT OD loop", "run smart data augmentation for grounding dino", "mine and retrain my detection model", "improve OD mAP with gap analysis and mining", "prep the source pool", or "pseudo-label my unlabeled images for mining"; do not use for standalone TAO training, one-off inference, or gap analysis alone.

Running a DOCA-linked binary directly on the hardware, with no container or kubelet: on host x86 with a BlueField NIC over PCIe, or bare-metal on BlueField Arm. Use this skill for launching and supervising it (direct, tmux, systemd), and when it "exits 1 right after launch", "systemd keeps restarting it", reports "no matching device on the BF", or the BlueField itself is unhealthy: "bfb-install exited 0 but DPU is dead", "ping 192.168.100.2 works but ssh fails", "host PFs aren't showing netdevs". Refuse service containers under kubelet (doca-container-deployment). Covers PCI/NUMA/CPU/IRQ binding, co-tenant isolation (cgroup-v2/netns/numactl), a seven-layer error taxonomy, platform lifecycle (BFB install, RShim/TMFIFO, host PF rebind, post-BFB recovery) and a six-state BlueField lifecycle classifier. Destructive firmware burn / mlxconfig set requires explicit confirmation via doca-hardware-safety; library APIs, env prep, and build use other skills.

Use this skill to deploy and operate a CollectX (clx) based DOCA telemetry collector on a host or BlueField — wiring providers / counters into the collector, running the collection daemon, and shaping its exporters (Prometheus pull, Fluent Bit push, NetFlow, file / IPC) so the metrics actually leave the box. Trigger even when the user never says CollectX or clx — implicit phrasings: {collector emits nothing downstream}, {add a provider to the clx collector}, {turn on the Prometheus endpoint}, {ship counters to Fluent Bit from the DPU}, {daemon starts but no schema rows appear}. This skill owns the CollectX collection mechanism plus the operator's own doca-telemetry / doca-telemetry-exporter usage; it ROUTES the productized DOCA Telemetry Service (DTS) to public docs (AGENTS.md Non-goal #7), the reader API to doca-telemetry, and the publisher API to doca-telemetry-exporter. Refuse to invent clx symbols, provider names, schema fields, flags, or config paths — describe the class and route to the live source.

Use this skill when the user is doing hands-on DOCA Device Emulation on a BlueField DPU — exposing a custom emulated PCIe device the host sees as a real peripheral while DPU-side code runs the backend, picking the sub-library (PCI Generic, virtio-net, virtio-fs), wiring the per-sub-library Core context plus doorbell / DMA primitives, querying `doca_devemu_*_cap_*`, or debugging DOCA_ERROR_* from a `doca_devemu_*` call. Trigger even when the user does not say "devemu" — typical implicit phrasings include "expose a custom PCIe device from BlueField to the host", "host should see a virtio NIC backed by my DPU code", "lspci does not show my emulated device", "device enumerated but no driver binds", "DPU sees nothing when host kicks the queue", or "virtio feature negotiation failed at bind". Refuse and route elsewhere for the packaged DOCA SNAP / Virtio-net Services, host-side virtio kernel drivers, backend body design, or standard BlueField NIC behavior — those belong to other skills.

Use this skill when the user is debugging any DOCA symptom — a build that won't compile, a link step that can't resolve a doca_* symbol, a runtime call returning DOCA_ERROR_*, a silent service or tool, or a stack trace / valgrind / core dump — and needs the layered ladder (install → version → build → link → runtime → program → driver), verbosity controls (--sdk-log-level, DOCA_LOG_LEVEL, the doca-{lib}-trace flavor), container-debug constraints, or how to capture state for a Developer Forum post. Trigger even when the user does not say "DOCA debug" — implicit phrasings include "undefined reference to doca_*", "how do I get more logs", "packets aren't reaching the wire", "doca_caps returned nothing", or "hugepages empty in the container". Refuse and route elsewhere for library-specific debug (Flow pipe trace, RDMA QP, Comch stats), env-class pkg-config or hugepages symptoms, the DOCA_ERROR_* taxonomy and lifecycle interpretation, and performance or incident-response work — those belong to other skills.

Used for running NV-Segment-CTMR on CT or MRI NIfTI volumes and recording label-map evidence. Not for clinical interpretation.

Use this skill for hands-on DOCA Ethernet packet-queue work on a BlueField DPU or ConnectX NIC — bringing up a `doca_eth_rxq` or `doca_eth_txq` on a port / representor / SF, picking among the four `enum doca_eth_rxq_type` values (`_REGULAR` / `_CYCLIC` / `_MANAGED_MEMPOOL` / `_SHARED_MEMPOOL`), sizing burst or scatter-gather length against the `_cap_*` queries, submitting `doca_eth_txq_task_send` / `_lso_send` (carrying packet `doca_buf`s — no `doca_eth_frame` struct exists), or debugging DOCA_ERROR_* from an Ethernet call. Trigger on implicit phrasings: "my RX queue is up but no packets arrive", "send-task returns AGAIN at line rate", "which queue type for fixed-MTU ingress", "device open fails without sudo", or "is L3 checksum offload available here". Refuse and route elsewhere for installing DOCA, flow-rule / steering programming, host↔DPU control messaging, or RDMA data movement.

PLAINTEXT-ONLY: the shipped `doca_flow_grpc` server uses `grpc::InsecureServerCredentials()` with NO TLS / mTLS / token-auth knob on the binary — transport security must come from external infrastructure (e.g. an mTLS proxy / sidecar) on a trusted segment. Use this skill when bringing up, configuring, hardening, or debugging `doca_flow_grpc` — the DOCA-shipped gRPC remote-control surface in front of `doca-flow` that lets non-C++ clients (Python, Go, Rust, Java) program Flow pipes and entries over RPC instead of linking `libdoca_flow.so` directly. Trigger even when the user doesn't say 'doca-flow-grpc-server' or 'gRPC' — e.g. 'program Flow rules from Python on another host', 'remotely configure pipes on the BlueField', 'client times out connecting to the Flow server', 'where is the .proto for Flow', 'UNAUTHENTICATED / FAILED_PRECONDITION on a Flow RPC'. Route elsewhere for the underlying doca-flow API, generic gRPC tooling (protoc, language bindings), or DOCA install / BFB bring-up.

Use this skill when the user is measuring the host or DPU-CPU control-plane rate of a DOCA Flow pipeline with doca_flow_perf — picking a JSON policy from configs/, choosing the DPDK or DOCA backend, running the single-iteration smoke then the iterative eval loop, interpreting per-iteration CPU cycles and num_pushed / num_failed, or capturing the four-tuple (DOCA version, BlueField/firmware, JSON policy, worker/queue/burst config) that makes a Kops/sec number defensible. Trigger even when the user does not explicitly mention "doca-flow-perf" — typical implicit phrasings include "how many rules per second can my BlueField insert", "5-tuple hairpin rule rate", "Kops/sec for steering", "flow-perf number does not match release notes", "DPDK vs DOCA benchmark", or "rule-install variance too high". Refuse and route elsewhere for optimizing a live Flow app (doca-flow-tune), the DPA-offloaded path (doca-flow-dpa-perf), dataplane throughput or latency, or library-internal pipe semantics — those belong to other skills.

Use this skill when the user is doing hands-on host-side DOCA PCC work to load a CUSTOM Programmable Congestion Control algorithm onto a BlueField DPU — creating per-port `doca_pcc` contexts, loading a `dpacc`-compiled `doca_pcc_app` onto the `doca_dev` for the RoCE-bearing port, parameterizing it, walking triple-axis capability discovery (DOCA cap-query + DPA-capable BlueField + firmware custom-PCC slot enabled), or debugging `DOCA_ERROR_*` from `doca_pcc_*`. Trigger even without explicit "DOCA PCC" phrasing — implicit forms include "loading my own congestion control onto a BF port", "DOCA_ERROR_NOT_PERMITTED on algorithm load", "DOCA_ERROR_DRIVER when I attach my custom algorithm", "my custom rate-update isn't affecting RoCE traffic", or "load succeeds but no on-wire change". Refuse and route elsewhere for DPA-side algorithm-body design, the `pcc_counters` CLI, default factory PCC in ConnectX firmware, or setting up the RDMA / RoCE traffic — those belong to other skills.

Use this skill when the user is doing hands-on DOCA RDMI (RDMA Initiator) programming — picking doca-rdmi vs doca-rdma for an accelerator-initiated one-sided RDMA flow, standing up a doca_rdmi_connection or doca_rdmi_poster, attaching a doca_dpa_completion or doca_verbs_cq before doca_ctx_start(), retrieving the DPA-side handle for a DPA kernel, auditing whether a doca_rdmi_* symbol is EXPERIMENTAL on this DOCA, or debugging DOCA_ERROR_* returns from RDMI calls. Trigger even when the user does not say "DOCA RDMI" or "initiator" — implicit phrasings include "my DPA kernel needs to post RDMA writes to a remote responder", "DPA kernel sees no completions", "function not found on doca_rdmi_* at link time", "DOCA_ERROR_BAD_STATE from completion attach", or "the DPA posted but the work request never arrived". Refuse and route elsewhere for two-sided or host-CPU RDMA, the DPA programming model, GPU-side RDMA initiation, or general RDMA/IB/RoCE concepts — those belong to other skills.

Use this skill when the user is dealing with the DOCA environment around their workload — verifying an install is healthy, preparing the build env (pkg-config, headers, LD_LIBRARY_PATH, hugepages, devlink, representors), debugging env-class failures, deciding container-vs-bare-metal deployment shape, or reaching a DOCA install from a host that doesn't have one yet via the NGC DOCA container Stage-1 fallback. Trigger even when the user does not explicitly mention "DOCA setup" — typical implicit phrasings include "I just got a BlueField, what now", "my code is built, how do I run it", "pkg-config can't find doca-flow", "no free 2048 kB hugepages", "representor X not found", "I'm on a Mac and want to learn DOCA". Refuse and route elsewhere for library API specifics (Flow pipes, RDMA queues), the modify-a-sample first-app workflow or DOCA_ERROR_* program-side debugging, and "where is X documented" knowledge-map questions — those belong to other skills.

Run TAO Data Services KPI analysis for object detection, comparing inference annotations against ground truth to compute per-class TP/FP/FN/TN, precision, recall, accuracy, and AP at a fixed IoU. Use when an object detection workflow needs per-class mAP reported after inference, or when the user asks to "run KPI analyze", "compute detection mAP", or "score my OD predictions against ground truth".

Use this skill to read DOCA hardware-counter events from a `doca_dev` through the per-domain Telemetry reader libraries: `doca_telemetry_pcc`, `_dpa`, `_diag`, `_adp_retx`, `_phy`, and `_pci`. It covers capability checks, context creation, startup, and per-domain reads or samples. Trigger for implicit requests such as "read PCC counters from my BlueField app", "sample DPA counter exports", or "expose PHY, PCI, or DIAG counters from this doca_dev". This is the counter-reader surface, not a NetFlow, IPFIX, or local-socket collector. Route publishing and export to `doca-telemetry-exporter`; route the deployed DOCA Telemetry Service to `doca-dts`, and collectors and plain stdout logging elsewhere.

Not for writing library code (name that library skill). Use this skill whenever another DOCA skill says "prefer the structured tool per doca-structured-tools-contract", or when the user wants a one-shot answer consolidating what multiple manual commands would produce — DOCA env / version / devices / capabilities / validate / host vs DPU state. Trigger even without the phrases "structured tool" or "doca-env --json" — typical phrasings include "is there one command that tells me everything about my DOCA install", "what version is X capability available since", "every PF/VF/SF visible on this BlueField with PCIe address", or "will this pipe pass validate before commit". Refuse and route elsewhere for general DOCA orientation, library API how-to, or install-from-scratch guidance — those belong to the per-library skill, doca-public-knowledge-map, or doca-setup.

Use this skill when the user is invoking `doca_telemetry_utils` on a host with DOCA installed — discovering the diagnostic-counter schema, translating counter names to binary Data IDs, validating per-device counter support before committing a DOCA Telemetry exporter config, or reverse-resolving a captured Data ID. Trigger even when the user does not explicitly mention "doca_telemetry_utils" or "Data ID" — typical implicit phrasings include "my exporter ships but the collector sees nothing", "this metric silently drops downstream", "which counters does this BlueField expose", "translate this 0x... back to a counter name", "what do node / pcie_index / depth mean here", or "is this counter supported on this device before I commit it". Refuse and route elsewhere for developer-side collector / exporter library programming, DTS deployment, or DOCA install / repair — those belong to doca-telemetry, doca-public-knowledge-map, and doca-setup.

Run the full DEFT AOI improvement loop for NVIDIA TAO VisualChangeNet / ChangeNet PCB inspection models: baseline evaluate, RCA, Cosmos AnomalyGen / AMP synthetic defects, k-NN mining, retraining, and deployment gating against a customer-defined primary metric and optional constraints. Use only when the request identifies an AOI / automated-optical-inspection, PCB-defect, VisualChangeNet, or ChangeNet workflow. Supports air-gapped/offline runs with pre-staged assets. Never infer AOI from generic iterative-improvement language. Do not use for CLIP / SigLIP image retrieval, attribute-labelled image data, standalone TAO training, one-off inference, generic anomaly generation, or RCA-only analysis.

Run container-backed AutoML / hyperparameter optimization (HPO) for NVIDIA TAO networks using AutoMLRunner. Handles algorithm selection (bayesian, hyperband, asha, bohb, llm, hybrid, autoresearch), WandB experiment tracking, job execution on any TAO SDK platform, result interpretation, and per-rec custom evaluation hooks. Use when the user mentions TAO AutoML, hyperparameter optimization, HPO, automl, automl_settings, AutoMLRunner, tao_automl, bayesian search, hyperband, ASHA, LLM-guided search, autoresearch, or wants to tune train/evaluate/inference/distill/prune/quantize for a TAO network. Model actions use the resolved image; venv training requires an explicit request. Platform-agnostic — runs on any SDK (Brev, SLURM, Kubernetes, Docker). Do not use generic "keep improving" language alone to override a matching domain-specific DEFT workflow; attribute-labelled CLIP / SigLIP image-retrieval loops belong to tao-run-deft-pas unless HPO is explicit.

Run the canonical NVIDIA AOI three-phase training pipeline — Phase 1 AutoML baseline (HPO), Phase 2 DEFT loop (RCA → SDG → mining → plain-train retrain), Phase 3 AutoML refinement on the DEFT-augmented dataset. Use when the user asks to "run the AOI workflow", "fine-tune my PCB AOI model end-to-end", "improve my AOI ChangeNet model", or "AOI workflow with AutoML" request — route here instead of tao-run-deft-aoi directly unless the user explicitly asks for the DEFT loop ONLY (e.g. "run JUST the DEFT loop", "skip AutoML, only DEFT"). Also handles the same three-phase pattern for non-AOI DEFT applications — AutoML baseline then DEFT loop warm-started from AutoML's winning HPs then post-DEFT AutoML refinement on the iteration-augmented dataset. Trigger phrases include "run the AOI workflow", "AOI end-to-end", "AutoML + DEFT", "AutoML then DEFT", "tune hyperparameters then DEFT", "DEFT with AutoML at both ends", "warm-start DEFT", "improve my AOI model".

Standard single-step train/eval/export workflow for any TAO model. Use when training a TAO model on a dataset without iterative data augmentation, AutoML, or DEFT loops. Trigger phrases include "single train run", "train then evaluate then export", "plain TAO training", "normal training", "no AutoML", "skip the loop". Routes through the per-model SKILL.md for action specifics and through `tao-launch-workflow` for platform/credentials/dataset intake.

Installs NemoClaw, launches a sandbox, and runs the first agent prompt. Use when onboarding, installing, or launching a NemoClaw sandbox for the first time. Trigger keywords - nemoclaw quickstart, install nemoclaw openclaw sandbox, nemohermes quickstart, hermes agent nemoclaw, run hermes openshell sandbox, nemoclaw prerequisites, nemoclaw supported platforms, nemoclaw hardware software, nemoclaw windows wsl2 setup, nemoclaw install windows docker desktop.

Use when you need to rebuild the BSP overlay — DT, OOT modules, or kernel — from changes under bsp_sources/. Triggers: build bsp, rebuild dtb, rebuild kernel.

Enable MIPI/GMSL camera sensors on a Jetson Thor or Orin custom carrier by rendering a kernel-DT overlay from the in-tree sensor DTSI. Do NOT use for UPHY lane allocation or ODMDATA edits.

Use to lock/cap Jetson CPU/GPU/EMC clocks, toggle EMC/CPU DVFS, or change cpufreq governors by editing BPMP DTB and nvpower.sh pre-flash. Do NOT use for live tuning or nvpmodel edits.

Configure Jetson UPHY lane allocation (uphy0/uphy1-config) on Orin/Thor custom carriers. Do NOT use for pinmux or PCIe-only edits.

Enable/disable Jetson USB2/USB3 SS ports via kernel-DT overlay. Do NOT use for UPHY lane allocation or ODMDATA edits.

Bootstrap a custom carrier board by forking carrier files and scaffolding a DT overlay from the reference devkit. Use after jetson-init-source; not for module-level or kernel-DTB changes.

Read-only Jetson health snapshot for identity, memory, GPU, thermal, power, storage, services, and top processes.

Download NVIDIA Jetson Linux BSP artifacts (BSP tarball, sample rootfs, public_sources, x-tools, guides) for the active target. Use for Auto Setup; not for extraction or profile edits.

Use to flash a promoted BSP image to a Jetson DUT in RCM mode via flash.sh or l4t_initrd_flash.sh. Do NOT use for BSP customization, image promotion, or carrier derivation.

Build a per-target knowledge-base markdown next to the active profile by walking the BSP root and source tree. Use after init-image / init-source; not for editing profile fields.

Plan and apply safe Jetson headless-mode changes to reclaim GUI and daemon memory.

Pick the serving stack and per-runtime memory flags (vLLM, SGLang, llama.cpp, TensorRT Edge-LLM) for an LLM/VLM workload on any NVIDIA Jetson.

Extract Jetson Linux + sample-rootfs tarballs and run apply_binaries.sh for the active target, then record bsp_image in the profile. Use after jetson-init-target; not for source-tree setup.

Bind pre-downloaded Jetson reference docs (developer guide, design guide, pinmux, schematics) into the active profile documents block. Use after staging docs on disk; not for downloading.

Stand up vLLM or SGLang serving on Jetson, using upstream vLLM on Thor and Orin JetPack 7.2+, and NVIDIA-AI-IOT vLLM on older Orin.

Measure Jetson DRAM/NvMap usage and verify before/after memory reclamation with live audit data.

Pick Jetson-compatible containers, vLLM runtime images, and Jetson AI Lab PyPI indexes; maps Orin SM 8.7 vs Thor SM 11.0 and JetPack-specific package choices.

Use when you need to print Jetson BSP info (L4T version, board configs, rootfs state) from a Linux_for_Tegra root on the host PC. This is an example skill.

Use when you need to print Jetson device info (module model, L4T version, kernel, OS version, current power mode) from a running Jetson target. This is an example skill.

Use to promote overlay files and built artifacts into the staged BSP image. Do NOT use to flash or build. Triggers: promote bsp image.

Entry skill for Jetson / IGX BSP customization. Asks one core click-to-select setup questionnaire and passes prefilled answers to downstream setup skills.

Switch the active Jetson target-platform pointer to an existing profile YAML. Use before customize/build/flash to change target; not for authoring profiles — use jetson-init-target instead.

Add EAGLE-3 or draft-model speculative decoding to a Jetson vLLM server when TPOT is the bottleneck.

Use this skill when an application owns tool or LLM/provider call sites and needs to wrap them with NeMo Relay scopes and managed execution APIs for lifecycle events, middleware, or guardrails.

Use this skill when concurrent requests, async tasks, threads, workers, goroutines, or agents need independent NeMo Relay scope stacks and correct ancestry propagation.

Use this skill when baseline NeMo Relay instrumentation exists and the user wants to configure or evaluate adaptive plugin behavior, including telemetry, state, adaptive_hints, tool_parallelism, ACG, hint consumption, or measured rollout.

Flash the FPGA on an HSB board connected to an NVIDIA devkit. Supports HSB Lattice boards (FPGA versions 2407, 2412, 2507, 2510) and Leopard Imaging VB1940 "all-in-one" cameras (FPGA versions 2507, 2510). Uses release-specific YAML manifests and board-type-specific program commands. Lattice and VB1940 commands must never be mixed.

Use to call the VIOS REST API (sensor list, timelines, clip extraction, snapshots, add/delete sensors and streams). Not for VLM inference or search.

Deploy and operate the RTVI-CV-3D stack (also known as MV3DT, Multi-View 3D Tracking, or RTVI-CV-MV3DT) — per-camera DeepStream perception plus BEV Fusion over multiple calibrated cameras. Applies to requests such as "deploy RTVI-CV-3D", "deploy rtvi-cv-3d", "deploy MV3DT", "deploy multi-view 3D tracking", "deploy rtvi-cv-mv3dt", "enable multi-camera tracking", "enable multi camera tracking", "set up multi-camera tracking", "multi-camera tracking", "run RTVI-CV-3D on my videos", "run MV3DT on my videos", "run RTVI-CV-3D / MV3DT on RTSP", "run on the sample dataset", "set up 3D tracking", or provides a 4-camera warehouse video/RTSP set. Routes between sample-data, custom-videos, and custom-RTSP flows; auto-chains to `vss-generate-video-calibration` when calibration data is missing. Not for the full warehouse blueprint with agents / LLM / VLM (use `vss-deploy-profile`) or 2D single-camera detection (use `vss-deploy-detection-tracking-2d`).

Stage 1 of Clinical ASR Flywheel. Use when bootstrapping a cycle: NVCF+MW disclosure, NVIDIA_API_KEY check, deps install, TTS+ASR smoke test.

Used for smoke or dataset finetuning of NV-Segment-CT VISTA3D on CT NIfTI labels. Not for clinical validation.

Documentation conventions for NeMo-RL. Covers docs/index.md updates and docstring format. Do NOT use for: bug fixes, test fixes, dependency bumps, refactoring, CI/CD changes, performance tuning, or any task that does not involve writing or updating documentation.

Operational guide for enabling hierarchical context parallelism in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.

Generates BYO custom safety policies for NVIDIA Nemotron content-safety guardrails — Nemotron-Content-Safety-Reasoning-4B (text) and multimodal Nemotron-3-Content-Safety. Produces a Markdown policy, JSON taxonomy, and drop-in inference prompts. Maps rough words or an existing policy to V2 categories, adding custom categories or topic-following rules.

Configure NeMo AutoModel job launches for interactive runs, Slurm clusters, and SkyPilot cloud execution.

Stage 3 of Clinical ASR Flywheel. Score a NeMo manifest, produce the five-section KER leaderboard (by-ipa_source diagnostic). Not for ASR auth (/riva-asr).

Validate and use CPU offloading in Megatron Bridge, including layer-level activation offloading and fractional optimizer state offloading with HybridDeviceOptimizer.

Use when the user wants to create a dataset, generate synthetic data, or build a data generation pipeline. The argument is the description of the dataset to generate.

Exports sanitized metadata, parameters, reproducibility details, quality metrics, and optional review artifacts from Medical AI inference runs or evidence packs to MLflow. Use after inference, including NV-Generate runs; not for live training tracking, model registration, or clinical use.

Use to ask the VSS agent's video_understanding tool a fresh visual question about a recorded clip. Not for prior tool output, search hits, or metadata-answerable questions.

Linting and formatting for Megatron-LM. Covers running autoformat.sh, tools (ruff, black, isort, pylint, mypy), and code style rules.

Select, validate, patch, and deploy existing NVIDIA Dynamo Kubernetes recipes. Use for model/backend/GPU/deployment-mode recipe bring-up; use router-starter for router-only mode work and troubleshoot for broken deployments.

Use when the user wants to set up, scale, validate, or harden NVIDIA physical AI infrastructure for synthetic data generation workflows across local MicroK8s or Azure AKS, including Kubernetes clusters, inference endpoint deployment, OSMO deployment, workload submission readiness, and infrastructure failure recovery. Trigger keywords: physical ai infrastructure, resilient scaling, SDG infrastructure, microk8s, azure aks, NVCF deployment, NIM Operator, OSMO deploy, workflow scaling. Don't trigger for: OSMO log summarization or workload-only operations unless infrastructure setup, scaling, validation, or recovery is requested.

Start or patch Dynamo router modes and run router endpoint smoke checks. Use for round-robin, KV-aware, least-loaded, or device-aware routing setup; use recipe-runner for recipe deployment and troubleshoot for failure diagnosis.

Autonomous NeMo-RL research agent workflow for directed hypothesis testing and open-ended discovery. Guides agents through the full experiment lifecycle: understanding recipes and environments, wiring RL or NeMo-gym runs, launching reproducible baselines and iterations, analyzing results, preserving human oversight, and using git plus TSV logs as the research ledger. Do NOT use for: bug fixes, code review, documentation, refactoring, dependency updates, or single-file changes.

Build deterministic forecast scripts with Earth2Studio (model, data source, IO, inference). Do NOT use for ensemble, diagnostics, data-only fetch, or install.

Used for converting one CT DICOM series folder to a HU NIfTI volume with affine evidence. Not for multi-frame DICOM or clinical use.

Cosmos-Embed1 video-text embedding for text-to-video retrieval, video-to-video search, semantic deduplication, and fine-tuning. Use when the user asks to "fine-tune Cosmos-Embed1", "run cosmos-embed inference", "export Cosmos-Embed1", "embed videos", or "search videos with text".

Use this skill when reading video-analytics metrics, incidents, alerts, and sensor data via the VA-MCP server (port 9901). Not for live VLM or incident-range narrative reports.

Install Holoscan SDK 5.x CUDA 13 Python wheels in an isolated environment, choosing the full stack or version-matched components.

Use when running video data augmentation and auto-labeling workflows on OSMO: flow selection, preflight, submit-time interpolation, monitoring, and output retrieval. Trigger keywords: video data augmentation, data enrichment, auto labeling, VDA demo, OSMO workflow, pseudo labeling.

Add a new cuTile GPU kernel operator to TileGym. Covers dispatch registration in ops.py, cuTile backend implementation, __init__.py exports, test creation, and benchmark in tests/benchmark. Use when adding, creating, or implementing a new cuTile operator/kernel in TileGym, or when asking how to register a new cuTile op.

Operational guide for enabling TP, DP, and PP communication overlap in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.

Long-context MoE training guidance for Megatron Bridge. Covers CP sizing, selective recompute, dispatcher choices, and practical patterns from DSV3, Qwen3, and Qwen3-Next long-context experiments.

Official NVIDIA-authored guidance for navigating PhysicsNeMo — pick the model, datapipe, or example for a SciML/AI4Science task (surrogates, forecasting, downscaling, physics-informed, inverse, generative). Points at existing files via live repo search; never writes code. Do NOT use for installation or environment setup, training-loop or other code authoring/scaffolding, contributor/CI/packaging questions, repo-specific questions in physicsnemo-sym/-cfd/-curator, or general (non-physics) ML/PyTorch.

Manage durable working-session memory for coding agents. Use when a user asks to preserve or recover agent context across disconnects, VS Code restarts, long-running work, handoffs, or any session where important state should be written periodically under the repo's session directory. Do NOT use for: simple questions, short tasks, one-off commands, linting, or code review.

Validate that a Dynamo deployment's NIXL/UCX/NCCL interconnect is ready for disaggregated serving over RDMA/NVLink. Use after recipe-runner brings a deployment up (especially disagg/multi-node) to confirm the KV transport is correct; use troubleshoot for diagnosing already-failed pods.

Practical guidance for training MoE VLMs in Megatron Bridge. Compares FSDP and 3D-parallel approaches, using rounded lessons from Qwen3-VL, Qwen3-Next, and other multimodal experiments.

Stage 2 of the Clinical ASR Flywheel. Use when curating clinical terms, tagging IPA, and synthesizing a NeMo manifest. NOT for scoring (use /digital-health-clinical-asr-eval).

Diagnose failed or unhealthy Dynamo deployments. Use when pods, model-cache jobs, PVCs, workers, frontend/router health, endpoints, or benchmark jobs fail; use recipe-runner/router-starter before this for normal bring-up.

Used for header-only preflight of one DICOM series folder before conversion or inference. Not for de-identification or clinical clearance.

Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.

Convert single-node scripts to multi-node Slurm sbatch jobs and debug common multi-node failures. Covers srun-native vs uv run torch.distributed approaches, container setup, NCCL timeouts, OOM sizing for MoE models, and interactive allocation.

Use as the top-level router for Omniverse Realtime Viewer USD app requests and focused viewer reference documents.

Use to run top-level VSS fusion search on archived video, or to ingest video files / RTSP streams for search. Not for ad-hoc Q&A or live captioning.

How to launch distributed Megatron-LM training jobs on a SLURM cluster. Covers a minimal sbatch skeleton, environment-variable setup for torch.distributed.run, CUDA_DEVICE_MAX_CONNECTIONS rules across hardware and parallelism modes, container conventions, monitoring, and per-rank failure diagnosis.

Use for VSS alert workflows — real-time monitoring, Alert-Bridge subscriptions, Slack notifications, incident queries, camera onboarding. Not for non-alert analytics.

Guide installing Earth2Studio via uv or pip, selecting model extras, and configuring the environment. Do NOT use for writing inference code, choosing models, or PhysicsNeMo questions.

Use to select, configure, deploy, verify, debug, or tear down a VSS profile (base, search, lvs, warehouse, edge). Not for standalone microservices — use the vss-deploy-* skill.

Find Earth2Studio models, data sources, and examples for a weather/climate use case. Do NOT use for writing inference code, downloading data, or installation.

Use to run AutoMagicCalib on local MP4s, RTSP, or the bundled sample dataset, and to deploy vss-auto-calibration when needed. Not for non-AMC calibration or runtime analytics.

Stage 4 of the Clinical ASR Flywheel. Use when priority KER is above 0.3 to run stock NeMo SFT on Parakeet TDT v2 and offline cycle N+1 re-eval. NOT for generic word boosting (use /finetune-asr).

Use to deploy the vss-behavior-analytics service standalone (entrypoint, config-source, optional calibration). Not for the full warehouse deploy.

Playbook for launching, monitoring, stopping, and debugging NeMo-RL recipes on a Kubernetes cluster via the nrl-k8s CLI. Covers ephemeral vs long-lived RayCluster modes, iterating on runs, and debugging hung or failed training jobs.

Use to deploy, run, debug, or tear down the RTVI-CV 2D detection / tracking microservice and call its REST API. Not for VLM, embedding, or analytics — use the matching vss-* skill.

Clone the latest NVIDIA Holoscan Sensor Bridge repo, ask which supported devkit is being used, configure the host per platform, build the correct demo container, run it, and verify HSB connectivity by pinging 192.168.0.2. Use for Holoscan Sensor Bridge setup, build, container launch, and first-connectivity bring-up.

Build and verify Holoscan SDK 5.x from an explicit source ref with the repository run wrapper; use for Early Access or SDK development.

Test system for Megatron-LM. Covers test layout, recipe YAML structure, adding and running unit and functional tests, golden values, marker filters, and CI parity.

Operational guide for choosing and combining parallelism strategies in Megatron Bridge, including sizing rules, hardware topology mapping, and combined parallelism configuration.

CUDA-Q onboarding guide for installation, test programs, GPU simulation, QPU hardware, and quantum applications.

Handle Holoscan SDK 5.x Conda installation requests: explain 5.0 package unavailability, check later releases, and select a supported installation method.

Execute QA test plans on Holoscan Sensor Bridge hardware. Reads a user-provided test document, filters tests by the user's setup, determines which tests can run automatically, executes them with pass/fail evaluation, and produces a structured test results report.

Fetch weather/climate data via Earth2Studio data sources for specific variables and times. Do NOT use for inference pipelines, model discovery, or installation.

Coordinate the end-to-end CAD/source-asset to SimReady workflow. Use for broad requests such as CAD to SimReady, source asset to simulation-ready USD, or prop packaging that require conversion, material/physics assignment, SimReady conformance, validation, and optional package creation; deploy or verify Content Agents services first when property assignment is enabled; route single-stage work through nested references.

Run `tao-daft convert` to convert NVIDIA TAO DAFT datasets between supported formats. Do not use for non-DAFT data. Use when the user asks to convert a DAFT dataset, change DAFT format, change a TAO dataset format, or run `tao-daft convert`.

Answer what the TAO Skill Bank plugin can do by generating the response from packaged application, data, model, AutoML, and platform manifests. Use when the user asks "what can TAO Skill Bank do", "list TAO models", "which TAO workflows are available", or "what supports AutoML".

Start, query, and stop a network-specific TAO inference microservice ({network_arch}-inference-microservice) by delegating container execution to the appropriate platform skill. Handles container image resolution, job-payload JSON construction, and the service registry. Use when the user wants to run inference on a TAO model checkpoint using a microservice container, deploy a TAO inference endpoint, or stop a running inference container.

Brev managed GPU instances with Docker support. Use when running TAO training, evaluation, or inference on Brev GPU instances, managing Brev deployments, or dispatching TAO jobs through the Brev CLI. Trigger phrases include "run on Brev", "Brev GPU instance", "submit job to Brev", "Brev CLI deployment".

Remote SLURM GPU cluster execution over SSH with sbatch/srun, Pyxis/Enroot containers, and Lustre-backed results. Use when running TAO training/eval/inference jobs on an on-prem or DGX SLURM cluster. Trigger phrases include "run on SLURM", "submit sbatch", "DGX SLURM cluster", "Pyxis/Enroot container", "Lustre dataset".

Local or remote Docker execution for TAO SDK job containers using a Docker daemon with NVIDIA GPU runtime. Use when running TAO jobs on the current machine, a directly attached Docker host, or a remote GPU box exposed through DOCKER_HOST. Trigger phrases include "run locally", "local Docker", "remote Docker", "use my GPU", "run on my machine", "host Docker daemon".

Kubernetes execution platform — submits TAO container jobs as k8s Jobs with NVIDIA GPU scheduling; single-pod for one node, Indexed Jobs for multi-node distributed training. Use when running on EKS / GKE / AKS / on-prem clusters with the NVIDIA GPU Operator installed, or when integrating TAO into an existing k8s-native ML platform.

BEVFusion for multi-sensor 3D object detection. Fuses LiDAR point clouds and camera images in bird's-eye-view (BEV) space, used in autonomous driving for robust 3D perception. Use when training, evaluating, or running inference for a TAO BEVFusion model. Trigger phrases include "train BEVFusion", "LiDAR + camera fusion", "BEV 3D detection", "multi-sensor 3D perception".

TAO Execution SDK for submitting and monitoring GPU training jobs on supported platforms (Brev, SLURM, local Docker, Kubernetes). Use when the user wants to run TAO jobs through the SDK, get job tracking, S3 I/O wrapping, multi-node distributed training, or platform-specific features that docker-run can't provide. Trigger phrases include "use the TAO SDK", "call tao_sdk", "AutoMLRunner", "ActionWorkflow", "Job handles", "S3 I/O wrapping", "TAO platform run".

NVPanoptix3D for panoptic 3D scene reconstruction from posed RGB images. Produces 3D panoptic segmentation (semantic, instance, and panoptic masks) with occupancy completion. Built on a VGGT backbone with a Mask2Former-style head and 3D frustum reconstruction. Use when training, evaluating, exporting, or running inference for a TAO NVPanoptix3D model. Trigger phrases include "train NVPanoptix3D", "panoptic 3D reconstruction", "3D scene segmentation", "occupancy completion".

Expert cuTile programming assistant. Write high-performance GPU kernels using cuTile's tile-based programming model with proper validation and optimization. Supports deep agent orchestration for complex multi-kernel tasks.

Use when adding, modifying, optimizing, or debugging CuTile autotuning code. Trigger signals: `exhaustive_search` / `replace_hints` / `hints_fn` / `cuda.tile.tune` in code, `autotune` in filenames, or correctness/performance issues in autotuned CuTile kernels. Covers: tune-once/cache/launch pattern, per-architecture configs (sm80–sm120), parameter space design (tile sizes, occupancy, num_ctas), and 7 common pitfalls with solutions.

Use this skill to check or change this machine's NVIDIA display and GPU settings, such as resolution, refresh rate, V-Sync, G-SYNC, brightness, color, and GPU performance.

NVDINOv2 for self-supervised visual representation learning. Trains vision transformers via self-distillation (teacher-student) without labels and produces general-purpose visual features. Use when training, exporting, or running inference for a TAO NVDINOv2 backbone. Trigger phrases include "train NVDINOv2", "self-supervised ViT pretraining", "DINOv2 backbone", "visual representation learning".

Calibrate a new dataset from live RTSP camera streams via the AutoMagicCalib REST API. Use when the user provides RTSP URLs or asks to calibrate live cameras; VIOS records clips, AMC ingests them, then runs calibration.

Port a published computer vision paper's official code and training recipe onto a customer's own dataset, or diagnose why such a transfer produced bad numbers. Use this whenever someone wants to reproduce a CV paper, run a paper's repo on their own images, fine-tune a published detection/segmentation/classification/keypoint model on customer data, adapt a training recipe to a new dataset, or figure out why a fine-tuned vision model scores well on validation but fails in production. Also use for post-mortems on any failed or disappointing CV training run, and whenever a user mentions mAP that looks too good, a model that "worked in training but not in deployment", or transferring hyperparameters from a paper to their own data. Trigger even if the user only says "train a model on my dataset" and a published architecture or repo is involved.

Use when you need to add, remove, edit, list, or change the boot default of an nvfancontrol fan profile on a Jetson/Tegra (Orin, Thor) target. Triggers: edit fan profile, tune fan curve.

Per-controller PCIe enable / disable / lanes / link-speed for a Jetson Thor or Orin custom carrier via ODMDATA + kernel-DT overlay. Do NOT use for UPHY lane allocation or endpoint-mode bring-up.

Per-pin SFIO / direction / initial-state configurator for a Jetson Orin or Thor custom carrier from the pinmux XLSM. Do NOT use for kernel-DT overlay or ODMDATA edits.

Set up the BSP source workspace: Linux_for_Tegra overlay tracker, bsp_sources, Crosstool-NG toolchain. Use after jetson-init-image; not for fetching inputs.

Benchmark Jetson LLM/VLM serving performance across vLLM, llama.cpp, and Ollama with structured JSON output.

Reclaim DRAM by disabling unused subsystems across MB1 BCT, MB2 BCT, kernel reserved-memory, and SWIOTLB. Use for headless or no-camera Jetson deployments; not for CPU/GPU frequency tuning.

Run and validate an end-to-end Mission Control showcase with a locally installed Isaac Sim launched in its GUI window, driven through the isaac-sim-remote Python server, with Nova Carter SIL. Use for demos, showcase replays, Mission Control driving a simulated robot, or diagnosing the integrated small-warehouse scenario. Detect existing Isaac Sim installations without modifying them, automatically select a usable runtime without prompting whenever compatibility can be confirmed, and delegate requested installation or version changes to isaac-sim-installation. Defaults to a canonical Isaac 6.1 warehouse and deterministic circular route when the user does not specify another scenario.

Resiliency features in Megatron Bridge including fault tolerance, straggler detection, in-process restart, preemption, and re-run state machine.

Modify, build, test, debug, and contribute to NVIDIA cuOpt (C++/CUDA, Python, server, CI). Use for solver internals, PRs, DCO, and code conventions.

NOTE: your protein sequence and the retrieved MSA alignment are transmitted to external NVIDIA-hosted APIs (health.api.nvidia.com) on every call. Use local NIM containers for confidential or proprietary sequences. Run a complete protein structure prediction pipeline using NVIDIA BioNeMo NIMs: search for MSA alignments with MSA-Search (ColabFold), then predict the structure with OpenFold3 using the retrieved alignments. Use this skill whenever the user wants to predict a protein structure with maximum accuracy using MSA context, run the full AlphaFold3-style pipeline, generate MSA-informed structure predictions, or improve structure prediction accuracy by providing evolutionary information. Triggers on: MSA structure prediction pipeline, structure prediction pipeline, MSA-informed prediction, OpenFold3, ColabFold MSA, AlphaFold3 pipeline, protein structure, homology search, a3m alignment, UniRef30, NIM microservice. This pipeline chains MSA-Search and OpenFold3.

Use when writing or debugging nvMolKit Python code for GPU-accelerated RDKit fingerprints, similarity, conformers, clustering, and molecular searches.

Mod or remaster a game with RTX Remix - open and edit projects, swap textures and models. Connect to the Remix Toolkit App via MCP. Not for non-Remix game interaction.

NV-Tesseract Forecasting — transformer-based multivariate time series forecasting with DARR (context-enhanced kNN retrieval), interpretability, and fine-tuning. Use when the user asks to "forecast with NV-Tesseract", "run forecasting inference", "use perform_forecasting", "DARR mode", "context-enhanced forecasting", "lag horizon attribution", "interpretability", "fine-tune forecasting", "fine-tune forecasting with automl", "hyper-parameter optimization with forecasting", or or mentions "nv-tesseract-forecasting", "moment_head_512_6hr", or "run8_best_model_cr".

NV-Tesseract AD Diffusion — diffusion-based anomaly detection and fine-tuning for multivariate time series. Use when the user asks to "fine-tune NV-Tesseract", "run AD diffusion inference", "detect anomalies with diffusion", "time series anomaly detection", "finetune ad-diffusion", "use perform_anomaly_analysis_with_diffusion", "automl ad-diffusion", "hyperparameter search ad-diffusion", "hyperparameter optimization" or mentions "curriculum_medium.yaml", "final_model.pth", "nv-tesseract-ad-diffusion", "ad_diffusion", or "TSDiffuser_Generic".

Run TAO Data Services TMM nearest-neighbor mining from embedding parquet files. Use when a workflow needs to mine source samples closest to target samples.

Run TAO Data Services image embedding to turn a parquet of image filepaths into an embedding parquet using CLIP, SigLIP, or a TAO checkpoint. Use when a workflow needs embeddings before nearest-neighbor or unique-neighbor mining, or when the user asks to "embed images", "compute image embeddings", or "generate SigLIP embeddings".

Run TAO Data Services TMM unique-neighbor matching mining from embedding parquet files for object detection workflows. Use when an object detection workflow needs to mine a bijectively-assigned set of unique source images closest to target samples. Use global allocation when mining without class constraints. Use class_stratified when rare classes are specified.

Run the disk-backed DEFT AOI improvement loop for NVIDIA Cosmos Reason 3 / Cosmos3 models, using Nano by default and Edge or Super when explicitly requested: evaluate the base model on Proxy and frozen Benchmark splits, mine real image pairs from Proxy gaps, assemble a per-iteration Train JSON from selected Mining samples, train with cosmos-rl LoRA SFT, and repeat through the selected platform's submit/status/logs/cancel contract. This migration supports bare labels only: the assistant response must be exactly OK or NG. Use for "run Cosmos3 DEFT AOI", "CR3 AOI loop", or "improve Cosmos3 PCB inspection with bare OK/NG"; do not use for rich/reasoning annotation, one-off Cosmos training, or generic anomaly generation.

Run the mining-based DEFT improvement workflow for ITS Cosmos-Reason binary video questions, focused on the non-reasoning classification/evaluation path. Use when the user asks for a DEFT CR ITS mining workflow, traffic-camera Cosmos Reason improvement loop, collision-identification workflow with data mining, or iterative Cosmos-RL refinement driven by gap analysis.

Use when operating PAIDF Curation and Retrieval or NVIDIA Cosmos Curator pipelines (split, filter, caption, embed, dedup, shard, image annotate) or PAIDF Data Mining nearest-neighbor matching on Curator embeddings. Activate for Make or CLI pipeline config, GPU run preflight, FFmpeg sidecar, SAM3 keys, or Curator-to-TAO handoff. Do not use for generic ETL, vector-database RAG, model training, orchestration, or embeddings outside Cosmos Curator and PAIDF Data Mining.

Multi-step video annotation pipeline that turns raw videos into Chain-of-Thought training data — multi-level captions, structured descriptions, and QA pairs (MCQ, binary, open-ended) with reasoning traces, via VLM/LLM distillation. Use when the user wants to "create video training data", "generate video QA datasets", "build CoT reasoning traces from videos", "auto-label videos", or run the video_reasoning_annotation pipeline. Triggers include "video annotation", "video CoT", "video QA", "chain-of-thought", "video captioning pipeline", "video distillation".

Converts cuTile Python GPU kernels (@ct.kernel) to cuTile.jl Julia equivalents. Handles kernel syntax translation, 0-indexed to 1-indexed conversion, broadcasting differences, memory layout (row-major to column-major), type system mapping, and launch API differences. Use when converting, porting, or translating cuTile Python kernels to Julia cuTile.jl, or debugging/optimizing existing Julia cuTile translations.

OCRNet for scene text recognition. Recognizes text content from cropped text-region images and supports CTC and attention-based decoders. Use when training, evaluating, exporting, pruning, quantizing, retraining, or running inference for a TAO OCRNet model. Trigger phrases include "train OCRNet", "scene text recognition", "OCR cropped text", "CTC / attention text decoder".

Use this skill for hands-on DOCA Arg Parser CLI work on a shipped sample or new DOCA-using app — adding / removing / renaming flags; wiring `doca_argp_init` → register params → `doca_argp_start` → `doca_argp_destroy` in order; picking a parameter type from the full public enum (`DOCA_ARGP_TYPE_STRING`, `_INT`, `_BOOLEAN`, `_DEVICE`, `_DEVICE_REP`, `_DOUBLE` — six values, not three); preserving the built-in DOCA flags (`-h` `-v` `-l` `--sdk-log-level` `--log-filter` `-j`/`--json`, NOT `--json-config`); `--device` is per-program; or debugging `DOCA_ERROR_BAD_STATE` / `INVALID_VALUE` / `NOT_SUPPORTED` / `IO_FAILED` from `doca_argp_*`. Trigger on implicit phrasings: "add a custom flag to a DOCA sample", "should I use getopt here", "BAD_STATE registering a new param", "my JSON config key is rejected", or "my sample's --json is ignored". Refuse and route elsewhere for variadic-flag / subcommand / shell-completion features, DOCA Core context, or DOCA Log internals.

Guide for onboarding new model architectures into NeMo AutoModel, including architecture discovery, implementation patterns, registration, and validation.

Use this skill when choosing or running NeMo Relay installation for an agent runtime, application language, framework integration, or plugin bundle, or checking a runtime's native Relay readiness. Do not use it for first-run instrumentation or export configuration.

The contract home for TAO's SDK-free execution pipeline — authoritative JSON Schemas for the four typed artifacts (spec-bundle, job-record, results_dir layout, best_rec) plus the fixed job-status vocabulary and the nested-not-dotted spec rule. Use when authoring or validating a spec-bundle before submit, writing or reading a .tao/jobs job-record, resolving where results land, or consuming AutoML's best_rec.json. Trigger phrases include "validate the spec bundle", "job record schema", "status vocabulary", "results_dir layout", "best_rec schema".

Raw verbs inside DOCA Core (doca-verbs): direct control of QP, CQ, PD, SRQ and address-handle objects through doca_verbs_context and doca_verbs_qp_create / doca_verbs_qp_modify / doca_verbs_cq_create, porting libibverbs code into DOCA Core, or a QP attribute, opcode or WR flag the doca-rdma task API does not expose. Trigger on "raw QP attribute the task API does not expose", "porting libibverbs code", "IO_FAILED on WR submit", "QP state transition rejected", or "attach a congestion-control group", even without naming doca-verbs. Refuse and route elsewhere for ordinary doca-rdma / doca-eth / doca-rmax workloads, DOCA install, Core internals, and general libibverbs theory.

Use this skill when integrating NVIDIA NeMo Fabric into a consumer application, service, evaluation harness, or platform through the typed Python SDK — translating the consumer's own application, job, or deployment config into an in-memory FabricConfig, choosing the single-invocation convenience API or an explicitly started runtime, validating with plan and doctor, and consuming normalized results, artifacts, and telemetry.

Convert existing PyTorch Lightning training code into an NVFLARE federated job using the Lightning Client API patch, local validation, and job export; use only when the request names federated/NVFLARE conversion or asks multiple sites to train collaboratively while keeping each site's data local, and either names PyTorch Lightning or preliminary source inspection identifies one Lightning owner; do not use for non-federated Lightning work such as DDP, profiling, inference serving, or training-loop changes, nor for plain PyTorch, TensorFlow/Keras, other frameworks, deployment, POC/production lifecycle, or experiment workflows.

Running an in-bundle DOCA service as a container on a BlueField (Argus, DMS, Firefly, or UROM service): kubelet standalone watching a static-pod manifests directory, YAML pod-spec drop, kubelet status / ENTRYPOINT logs / per-service liveness. Trigger on "how do I run my built service on the BlueField?", "where do I drop the pod-spec YAML?", "pod stuck in Pending / ImagePullBackOff / CrashLoopBackOff", "container Running but service isn't ready", "pod restart-loops after edit", or "DMS and Firefly together", even without naming container deployment. A DOCA-linked binary launched directly on the host or Arm belongs to doca-bare-metal-deployment. Covers smoke-before-bulk and the layered error taxonomy (pod-spec, scheduling, image pull, runtime, mount, network, version, host). Refuse and route elsewhere for per-service config schemas, DOCA install, library-API questions, external NVIDIA services (BlueMan, HBN, SNAP, Virtio-net), or full Kubernetes-cluster ops.

RDMA data movement on NVIDIA BlueField/ConnectX NICs (RoCE/InfiniBand) with DOCA RDMA, instead of hand-written libibverbs/rdma_cm. Use when the user moves data between processes or hosts over the NIC, or hits "one-sided read returns permission denied", "completions never arrive after submit", "connection callback never fires", atomic compare-and-swap over RoCE, or DOCA_ERROR_FULL on the send queue under burst, even without naming DOCA RDMA. Covers bringing up an RDMA context on a doca_dev, connection methods (RDMA CM, bridge/OOB, or gRPC exchange of doca_rdma_export()), the eleven task types (Send/Receive/Send-Imm, Read/Write/Write-Imm, Atomic CmpSwap/FetchAdd, Get/Set/Add Remote Sync Event), matching mmap + RDMA permissions, queue and connection sizing, doca_rdma_cap_* queries, and DOCA_ERROR_* from RDMA calls. Raw QP/verbs control belongs to doca-verbs. Refuse and route elsewhere for general RDMA / ibverbs theory (queue pairs, MRs, RoCE vs IB), installing DOCA itself, or non-RDMA DOCA libraries.

Install Holoscan SDK 5.x native C++ Debian packages on Ubuntu 24.04, choosing the full SDK or matching modular packages.

Operate NVIDIA DOCA Management Service on a BlueField, Arm/x86 host, or Kubernetes pod: from DOCA 3.5 the local `dms-cli` (one command per YANG path, no server), and the `dmsd` + `dmspe` service: choose deployment and authentication, configure `-allowed_users` and `dmsgroup`, use gNMI Get/Set, run supported gNOI workflows, and debug frontend/backend failures. Trigger even without "DMS" for "manage a remote BlueField over gRPC", "gNOI reboot from orchestrator", or fleet-management requests. SAFETY: reboot, OS install, factory-reset, and managed-file deletion are destructive and require target-bound explicit confirmation; never invoke them speculatively. Route installation and library/API build questions elsewhere, and route telemetry streaming and aggregation to doca-dts (DOCA Telemetry Service).

Run `doca_bench` (DOCA 2.7.0 or newer) to measure throughput, bulk latency, precision latency, or maximum bandwidth for RDMA, Compress, AES-GCM, SHA, DMA, EC, Ethernet, Comch, or GPUNetIO on a host or BlueField Arm. Use it to discover enabled benchmark libraries, capture a reproducible command/version/device/environment baseline, compare stable runs against a declared tolerance, or diagnose configuration, device-binding, workload-precondition, and measurement failures. Trigger for requests such as measuring BlueField compression speed, NIC RDMA throughput, crypto latency, or a pre-upgrade baseline. Do not use for application end-to-end timing, custom benchmark code, DOCA installation, or binary patches.

Use this skill when the user builds an NVMe-over-Fabrics storage target on BlueField with DOCA STA — instantiating doca_sta, declaring doca_sta_subsystem NQNs and namespaces over local NVMe disks via doca_sta_be, validating platform support with doca_sta_cap_is_supported, sizing I/O queue depth, or decoding DOCA_ERROR_* statuses such as IO_FAILED on NVMe reads. Also load it for implied asks like "Identify Controller times out over RoCE", "offload the nvmf target", "NVMe-oF Connect never completes", or "16 I/O queues at depth 1024 — supported?". Refuse and route to the proper owner for raw RDMA movement, packet I/O, steering rules, initiator-side NVMe, or SDK install.

DOCA version handling: detect the installed release and validate the four-way match across pkg-config doca-common, applications/VERSION, doca_caps --version, and bfver/mlnx-release on BlueField; reason about NGC tags; look up whether a capability exists on this release; diagnose build-vs-runtime drift. Use this skill alongside doca-debug when a symptom smells of version skew: "built but nothing on the wire", "undefined reference to a symbol the docs claim exists", "DOCA_ERROR_NOT_SUPPORTED at runtime", "what does latest mean for this tag", or "is my LTS still supported". Once all four sources match, the component's build / run workflow belongs to its own library, service or tool skill. Refuse install / package choice (doca-setup), per-library API questions (matching library skill), DOCA_ERROR_* taxonomy (doca-programming-guide), and the general debug ladder (doca-debug).

Use this skill when the user is doing hands-on DOCA Management programming against BlueField / ConnectX devices — standing up a management or representor context (doca_mgmt_dev_ctx / doca_mgmt_dev_rep_ctx), querying device caps (data-direct, caps-general), toggling congestion-control global status, modifying diagnostics-data, setting ICM quotas, or issuing a raw firmware command via doca_mgmt_raw_cmd with the right scope (CONFIGURATION / DEBUG_READ_ONLY / DEBUG_WRITE / DEBUG_WRITE_FULL). Trigger even when the user does not say "DOCA Management" — typical implicit phrasings include "fleet tool that walks every BlueField and reads device state", "toggle data-direct on a VF", "set an ICM quota per representor", "send a raw firmware command from C", "DOCA_ERROR_IO_FAILED from raw_cmd", or "fwctl ioctl is failing". Refuse and route elsewhere for mlxconfig direct operation, BFB / firmware reflash, streaming telemetry, doca_caps CLI snapshots, or DOCA install itself — those belong to other skills.

Host-side DOCA UROM library (doca_urom_*): offload an HPC / UCX / MPI application's remote memory operations (puts, gets, atomics, collectives) from the host CPU onto a BlueField DPU. Trigger on "MPI all-reduce burning host CPU", "push UCX traffic onto the BlueField", "first doca_urom call returns NOT_PERMITTED", or "host library and DPU service look out of sync", even without naming DOCA UROM. An offload needs both halves: this host library and the UROM Service on the DPU (doca-urom-svc), which owns deploying that container. Covers creating UROM Service contexts (doca_urom_service_*) and Worker contexts (doca_urom_worker_*) that run plugins on the DPU, discovering plugins via doca_urom_service_get_plugins_list, progressing completions, and DOCA_ERROR_* from doca_urom_* calls. Route elsewhere for UROM Service deployment on the DPU side, MPI / UCX collective algorithm design, and RDMA / RoCE / IB substrate bring-up.

Use this skill when the user is invoking the DOCA PCC Counters tool — the `pcc_counters.sh` bash script under the DOCA tools directory — to arm and read the fixed firmware/hardware PCC (Programmable Congestion Control) diagnostic counters (CNP, RTT, WRED-drop, etc.) on a ConnectX / BlueField device via mst + the mlx5 debugfs `diag_cnt` interface. The script takes two positional args — `set | query` and an mst device path — with no `--help` or subcommands. Trigger even without "pcc_counters.sh" or "PCC counters": "how do I read the CNP / RTT / WRED-drop counters", "PCC counter stuck at zero", "the script says Bad Device", or "is congestion control dropping packets on this port?". Route elsewhere for writing a custom PCC algorithm (doca-pcc), factory firmware PCC config, DOCA install, or fleet-wide CC tuning.

Convert existing plain or manual PyTorch training code into an NVFLARE federated job using Client API model exchange, local validation, and job export; use when the user names plain PyTorch or preliminary source inspection identifies one plain-PyTorch owner, and not for Lightning, other frameworks, deployment, POC/production lifecycle, or experiment workflows.

Build typed third-party NeMo Fabric adapters

Routes NVIDIA Nemotron Speech (Riva) NIM tasks — deploys, runs, and tests ASR, TTS, and NMT NIMs on build.nvidia.com or self-hosted.

Use this skill when the user is doing hands-on DOCA AES-GCM work on a BlueField DPU or ConnectX NIC — configuring `doca_aes_gcm_task_encrypt` / `_task_decrypt`, querying `doca_aes_gcm_cap_*` for per-key-type (only `DOCA_AES_GCM_KEY_128` / `_256` — AES-192 not supported) and per-task support, sizing plaintext against the max-buf cap, setting source / destination mmap permissions, validating with a NIST GCMVS or RFC 5288 vector, or debugging DOCA_ERROR_* including the security-critical tag-verification-failed outcome on decrypt. Trigger even when the user does not explicitly mention "DOCA AES-GCM" or "AEAD" — typical implicit phrasings: "decrypt completion IO_FAILED", "auth tag isn't verifying", "NOT_PERMITTED on my encrypt buffer", "is AES-192-GCM on this BlueField" (no), or "encrypted record came back tampered". Refuse and route elsewhere for non-GCM AES modes (CBC / CTR / XTS — CPU OpenSSL), key management (KMS / HSM / rotation), SHA (doca-sha), or general AEAD background.

WARNING: guides IRREVERSIBLE BlueField-4 hardware operations (PLDM firmware burns, ISO reflashes, power cycles, BMC factory resets); every mutating step is governed by doca-hardware-safety (window, rollback). BlueField-4 only: BlueField-3 bring-up is doca-bf3-deployment. Use this skill for BlueField-4 (BF4) day-1 platform bring-up from the BMC: installing the BlueField/DOCA bundle ISO onto the DPU (Grace, the Arm complex) over UEFI HTTP Boot, PXE, or Redfish Virtual Media; the PLDM platform-firmware update via the Redfish UpdateService and pldmtool; and a Grace Ubuntu image with optional cloud-init. Trigger on BlueField-4/BF4 bring-up phrasings even without "BF4": {bring up my new BlueField-4}, {the BlueField ISO will not boot over HTTP from the BMC}, {attach BF4 virtual media via Redfish}, {BF4 firmware Task stuck at Running}, {prepare a BlueField-4 host: platform support, version alignment, safest first smoke test}. Route application launch and library APIs to other skills.

Use this skill when the user is doing hands-on DOCA Comch work on a host + BlueField pair — bringing up host ↔ DPU PCIe control-plane messaging, picking server (DPU) vs client (host) roles, choosing slow-path send-task / recv-callback vs fast-path producer / consumer, querying max-msg-size or max-clients capabilities, registering connection callbacks, or debugging DOCA_ERROR_* returns from the Comch API. Trigger even when the user does not explicitly mention "DOCA Comch" or "Comm Channel" (renamed in DOCA 2.5) — typical implicit phrasings include "send a control message from host to BlueField over PCIe", "DPU can't see the host representor", "DOCA_ERROR_NOT_PERMITTED on server_create", "DOCA_ERROR_AGAIN on task_send submit", "connect callback never fires", or "stream bulk data from a host driver to a DPU agent". Refuse and route elsewhere for installing DOCA itself, BFB / firmware bring-up, non-Comch DOCA libraries, or deploying Comch apps at scale — those belong to other skills.

Use this skill when the user is doing hands-on DOCA DMA programming — bringing up a doca_dma context, configuring the single doca_dma_task_memcpy task type, sizing buffers via the doca_dma_cap_task_memcpy_* queries, setting LOCAL_READ_ONLY / LOCAL_READ_WRITE permissions on source / destination doca_mmap regions (plus doca_mmap_export_* for cross-peer copies), driving the progress engine, or debugging DOCA_ERROR_* returns. Trigger even when the user does not explicitly mention "DOCA DMA" or "doca_mmap" — typical implicit phrasings include "memcpy host buffer to BlueField without using the CPU", "offload a bulk copy to the DPU", "copy returns NOT_PERMITTED on first submit", "buffer too big for one DMA task", "task submitted but no completion", or "scatter-gather copy between two memory regions". Refuse and route elsewhere for cross-network copies (DOCA RDMA), producer/consumer messaging (DOCA Comch), DOCA Core / progress-engine internals, or DOCA install — those belong to other skills.

Use this skill when the user runs doca_dpa_hl_tracer to capture/decode DPA-side traces at the programming-events layer (kernel entry/exit, sync points, comm primitive calls, RDMA WR submission, completion drain) — picking TRACE vs CRIT, tuning the JSON config (file-size limits + file_size_limit_policy, thread priorities/cores), decoding against the matching DPA-side ELF, or diagnosing empty/noisy captures. Trigger even when the user does not explicitly mention "DOCA DPA tracer" or "high-level tracer" — typical implicit phrasings include "DPA kernel returns wrong result but host completions look clean", "kernel-entry to first-comm latency is huge", "RDMA WR to drain gap on the DPA", "trace file truncated mid-run", "TRACE doubled my DPA latency", or "tracer wrote a file but parser shows zero events". Refuse and route elsewhere for writing DPA kernels, DPA-Comms/DPA-Verbs programming, raw per-cycle DPA profiling, host-side doca-dpa debugging, or production DPA telemetry — those belong to other skills.

Use this skill when the user is invoking doca_flow_dpa_perf on DPA-capable hardware (ConnectX-7 minimum supported, ConnectX-8 recommended, or BlueField-3) to measure rule update / disable rates on the DPA-offloaded DOCA Flow path — picking the active / passive device split, choosing workload-shape axes (burst, queue, completion threshold, workers, hash pipe algo, PSL tables), or reading Kops/sec iteration stats and the optional self-test. Trigger even when the user does not explicitly mention "doca_flow_dpa_perf" or "DPA Provider" — typical implicit phrasings include "how fast can the DPA program path-selector entries", "baseline rule-update rate on ConnectX-8", "tool reports zero ops on my BlueField", "self-test sentinel never shows on tcpdump", or "is my BlueField-2 DPA-capable". Refuse and route elsewhere for the host / DPU-CPU Flow path (doca-flow-perf), Flow pipeline tuning (doca-flow-tune), writing doca-flow / doca-dpa applications, or DOCA install — those belong to other skills.

Use this skill when the user is tuning a live or captured `doca-flow` pipeline with `doca_flow_tune` — snapshotting pipe / counter / KPI state, picking a tuning axis (rule placement, resource hints / table sizing, HW-offload mode) and a matching measurement (rule-install rate, lookup latency, hardware-counter delta), running offline or online (read-only or state-changing) modes, reading the dumper CSV / analyze JSON / visualize mermaid, or applying a recommendation back into the Flow program. Trigger even when the user does not explicitly mention "doca_flow_tune" — typical implicit phrasings include "Flow rule-install rate is low on BlueField", "table sizing looks wrong for this pipe", "tune visualize step is empty", "before/after counters don't move", or "which doca-flow knob does this recommendation hit". Refuse and route elsewhere for measuring baseline numbers (doca-flow-perf, doca-flow-dpa-perf), writing the doca-flow application, DOCA install, or streaming Flow telemetry — those belong to other skills.

Use this skill for hands-on DOCA GPI programming — wiring a GPU-Packet-Initiator context so a CUDA kernel drives RDMA queues directly from GPU memory without host CPU mediation. Covers picking GPI vs doca-gpunetio, the doca_gpi / domain / channel object model, the GPU-side handle handoff (doca_gpu_gpi_channel*), attaching GPU memory to a GPI domain, the domain and channel attribute objects, and debugging DOCA_ERROR_* from doca_gpi_* calls. Trigger even when the user does not explicitly mention "DOCA GPI" — implicit phrasings include "my CUDA kernel needs to post RDMA directly from GPU memory", "DOCA_ERROR_* from doca_gpi_gpu_channel_get", "how do I hand a GPU handle to my CUDA kernel", "how many channels can a GPI domain hold", or "GPU kernel driving RDMA without the host CPU on the path". Refuse and route elsewhere for the doca-gpunetio Send/Receive surface, the doca-rdma queue lifecycle, DPA-side initiation (doca-rdmi), or the CUDA programming model — those belong to other skills.

Use this skill when the user is building, running, or interpreting the doca/tools/gpunetio_ib_write_bw client+server benchmark — a CUDA kernel on the client posts RDMA WRITE work requests through the doca-gpunetio device-side surface to measure sustained GPU-driven WRITE bandwidth on a GPU+IB-device pair. Trigger even when the user does not explicitly mention "doca-gpunetio-ib-write-bw" or "GPUNetIO" — typical implicit phrasings include "measure WRITE BW when the GPU posts the WRs", "BW swings between runs on the same flags", "is the NIC saturated or am I CPU-bound on the CUDA kernel", "meson compile fails for the GPUNetIO bw tool", "nvidia_peermem isn't picking up my GPU buffer", or "GPU-initiated WRITE throughput vs CPU-initiated perftest". Refuse and route elsewhere for general doca-gpunetio library work, DOCA install, the GPU-initiated WRITE latency analog, the CPU-initiated upstream perftest, or application-level end-to-end throughput — those belong to other skills.

Use this skill whenever the agent is about to recommend or apply a change that touches DPU / NIC hardware state on a live system — mlxconfig firmware-parameter write, NIC firmware burn, BFB reflash, NIC ↔ DPU mode flip, SR-IOV or device-emulation slot enable, kernel boot-parameter change (IOMMU, hugepages, VFIO), PCIe rebind / rescan / link-state flip, BlueField cold reboot, or remediation of a BlueField thermal shutdown. Wrap the change in pre-flight inventory, OOB reachability, a maintenance window, replica rehearsal, and rollback. Trigger on implicit phrasing: "flip BlueField mode over SSH", "enable SR-IOV and reboot", "burned firmware but mlxconfig shows old value", "BlueField thermal shutdown". Not for read-only inventory or version queries (doca-version, doca-caps). Refuse orientation (doca-public-knowledge-map), install or env debug (doca-setup), and program-side debug (doca-debug, doca-programming-guide).

Use this skill when the user is doing hands-on deployment, tuning, or evaluation of the DOCA-shipped Zero-Touch RoCE RTT-based Congestion Control (ZTR RTTCC) reference algorithm on a BlueField-3 DPA — wiring `doca_pcc_dev_ztr_rttcc_algo` into the shipped DOCA PCC sample, picking a variant (vanilla / PM / RX-rate / multipath / window-probeless) at DPACC build time, tuning host-set parameters, or diagnosing `DOCA_PCC_DEV_STATUS_FAIL` from the algorithm. Trigger even when the user does not say 'DOCA PCC' or 'ZTR RTTCC' — typical implicit phrasings: 'my RoCE-v2 flows aren't being throttled', 'PCC sample isn't dispatching to my algo', 'how do I pick the multipath PCC variant', 'set-params returns fail', 'algorithm loaded but counters are flat', or 'do I need a custom CC algorithm on BF3'. Refuse and route elsewhere for writing a custom PCC algorithm from scratch, read-only PCC counter inspection, the host-side `doca-pcc` lifecycle, or firmware-only pre-Programmable PCC — those belong to other skills.

DOCA Rivermax (doca_rmax) receive streams for timing-precise media over IP (SMPTE ST 2110 video / audio, market data, scientific feeds) on a target BlueField DPU (DOCA Rivermax apps run only there), once the Rivermax libraries and license are installed; getting them is doca-setup / doca-public-knowledge-map territory. Trigger on "ST 2110 receive isn't getting frames", "sub-microsecond jitter on my media receiver", "NOT_SUPPORTED from doca_rmax_init", "no recv events after stream start", or "license check failing on a media receiver", even without naming Rivermax. Covers doca_rmax_in_stream sessions, the SDK + license precondition check, doca_rmax_get_*_supported capability queries, pairing with doca-eth queues and doca-flow steering, and DOCA_ERROR_* from Rivermax calls. Refuse and route elsewhere for programming the underlying queue (doca-eth), steering rules (doca-flow), PTP / clock sync (doca-firefly), or best-effort packet I/O.

Use this skill when wiring the DOCA SHA Offload Engine (an OpenSSL ENGINE) into an existing OpenSSL pipeline to offload one-shot SHA-1, SHA-256, or SHA-512 (EVP_Digest) onto DOCA SHA hardware without rewriting against doca-sha. Covers engine load mechanics (`openssl engine dynamic`, `set_pci_addr` ctrl, `-engine_impl`), the SHA-224 negative test that proves offload engaged, the message-size window where offload beats CPU SHA, and engine-vs-library selection. Trigger even when the user does not say "DOCA SHA Offload Engine" or "OpenSSL ENGINE" — typical implicit phrasings: "speed up openssl SHA on BlueField", "offload SHA without code changes", "is openssl using the accelerator or falling back to software", "prove DOCA SHA actually ran", "openssl dgst hashed but I'm not sure it was offloaded". Refuse and route elsewhere for new SHA pipelines (use doca-sha), MD5 / SHA-3 / SHA-224 / HMAC-SHA offload, incremental hashing via chained `EVP_DigestUpdate`, or OpenSSL PROVIDER authoring.

Changing DOCA versions safely: upgrading or downgrading the host DOCA packages, refreshing the BlueField BFB, bumping the NGC DOCA container tag, or rolling back, with the discipline detect -> report -> ASK -> only then a guided upgrade. Trigger on "is there a newer DOCA", "I want the newest DOCA features, can I just apt upgrade in place", "should I move to the next release", "the upgrade stopped midway and dpkg is in a broken state", "my component is being deprecated, what now", or "roll me back", even without the word upgrade. Detect what is installed, find the newer release, report the gap, then STOP and ask for explicit confirmation; never upgrade automatically. Do not use it for a runtime error on an unchanged install (doca-debug). Version detection alone belongs to doca-version, first-time install to doca-setup, any hardware / firmware / reboot step to doca-hardware-safety, and public-docs or sunset routing to doca-public-knowledge-map.

Deploy object detectors into DeepStream

Use when planning, debugging, tuning, evaluating, exporting, or deploying public Nemotron `embed`/`rerank` retrieval recipes.

Run Megatron-LM (MLM) and Megatron Bridge training with mock or real data. Covers correlation testing, available recipes, and multi-GPU examples.

Use this skill when adding NeMo Relay typed wrappers, domain types, or provider codecs while preserving JSON middleware semantics and caller-visible behavior.

Guides human users' AI agents to the NemoClaw docs MCP server and canonical Fern documentation in Markdown form. Use when users ask how to install, configure, operate, troubleshoot, secure, or learn NemoClaw with an AI coding assistant. Trigger keywords - nemoclaw docs, use nemoclaw with ai agent, nemoclaw mcp docs, nemoclaw install help, nemoclaw quickstart, nemoclaw markdown docs, llms.txt, agent skills.

LP, MILP, QP — concepts, problem-text parsing, and formulation patterns (parameters, constraints, decisions, objective). Concepts only; no API.

Host setup for TAO GPU backends. Checks and, after user approval, installs minimum-compatible NVIDIA driver, CUDA Toolkit, and NVIDIA Container Toolkit versions for Docker/local-Docker and Kubernetes GPU worker hosts. TAO-wide defaults can be overridden by the selected model's runtime profile. The `--check-only` path works on any Linux distribution; `--install` automates debian-family (Ubuntu/Debian/Pop!_OS/Mint/Zorin/Raspbian), rhel-family (Fedora/RHEL/Rocky/AlmaLinux), and suse-family (openSUSE/SLES) hosts, and prints actionable manual-install steps for everything else. Use when the user asks to "set up an NVIDIA GPU host", "check TAO Docker GPU runtime", or prepare a Kubernetes GPU worker for TAO.

Used for generating synthetic brain MRI volumes with NV-Generate-CTMR rflow-mr-brain. Not for production training data.

Used for running NV-Segment-CT VISTA3D on CT NIfTI volumes and recording label-map evidence.

Used for generating synthetic body MRI volumes with NV-Generate-CTMR rflow-mr. Not for paired masks or production training data.

LP, MILP, and QP (beta) with cuOpt — CLI only (MPS/LP/QPS files, cuopt_cli). Use when the user is solving LP, MILP, or QP from an MPS or LP file via command line.

Use this skill when producing a VSS analysis report — Mode A per-clip VLM, Mode B incident-range via video-analytics. Not for standalone video summarization, real-time alerts or ad-hoc Q&A.

Create and modify NeMo AutoModel training and evaluation recipes, including YAML structure, builders, and execution flow.

LP, MILP, and QP (beta) with cuOpt — C API only. Use when the user is embedding LP, MILP, or QP in C/C++.

Brev instance operating guidance for NeMo-RL agents working in /home/ubuntu/RL with limited workspace disk, a larger /ephemeral volume, and optional /home/ubuntu/RL/.env secrets. Use when running nemo-rl-auto-research campaigns, experiments, training jobs, model or dataset downloads, shared cache-heavy commands, log-producing runs, checkpoint generation, W&B or Hugging Face authenticated workflows, or any workflow that may create large files on Brev.

Split a PR into multiple PRs to reduce the number of required CODEOWNERS reviewer groups.

Use to deploy the vss-video-analytics-api REST service standalone (config-source, data-log bind, Elasticsearch, optional Kafka). Not for full warehouse deploy.

Use when the user wants to create a dataset, generate synthetic data, or build a data generation pipeline.

Investigate a failing GitHub Actions run or job and create a GitHub issue for the failure.

Install and verify the Holoscan SDK 5.x CUDA 13 development container after checking release availability and host compatibility.

Extract false-positive and false-negative gaps from VLM binary-classification-question (BCQ, yes/no) predictions. Use when the user asks to "analyze VLM BCQ gaps", "extract VLM false positives and false negatives", or identify failure cases from a predictions JSON for DEFT root-cause analysis on a binary-classification VLM workflow.

Shared launch intake for any TAO workflow or action. Use when the user wants to run TAO AutoML, train, evaluate, infer, export, generate TensorRT engines, or launch DEFT/workflow jobs on an execution platform.

Integrate a HuggingFace Computer Vision model into the NVIDIA TAO Toolkit ecosystem (tao-core config, tao-pytorch trainer, tao-deploy TensorRT pipeline). Use when the user asks to "integrate a HuggingFace model into TAO", "add an HF model to TAO Toolkit", "wire a HuggingFace ViT/DETR/ SegFormer into tao-pytorch", "build a TAO trainer + deploy pipeline for an HF CV model", or pastes a HuggingFace model URL/ID and wants it turned into a TAO model. Covers the full 7-phase loop: prerequisites check, HuggingFace inspection and validation, codebase exploration, tao-core configuration and native trainer implementation, ONNX export plus TensorRT deploy integration, packaging and L0 testing, container-based end-to-end validation, and (conditional) accuracy/latency tuning. Supports classification, object detection, semantic / instance / panoptic segmentation, zero-shot detection, and depth estimation.

Evidence-gated workflow for MoE performance optimization in Megatron Bridge. Covers measurement contracts, the Three Walls framework, parallel folding, profiling, matched A/B tuning, and final validation.

Use this skill when the user is deploying or operating the DOCA Argus Service — the packaged BlueField-side runtime-security container that watches the BlueField and attached host for suspicious activity, integrity violations, and operational anomalies, and forwards findings to a SIEM (Splunk / ELK / Sentinel / syslog). Covers the four-axis config (detection policy, forwarding, sampling, host coverage), running the NGC container on BlueField Arm, and wiring the forwarder. Trigger even without "DOCA Argus" by name — typical implicit phrasings: "container green but no findings arrive", "false-positive flood in Splunk", or "runtime security on a fleet of BlueField-3s". Refuse and route elsewhere for installing DOCA, SIEM-side ingest stanzas, pre-baked detection-rule packs, and metrics observability (DOCA Telemetry). Argus is NVIDIA's currently- promoted runtime-security framework, superseding the older App Shield library; name it first for new runtime-security work.

Use this skill when the user is writing their first DOCA app or asking a library-agnostic programming question — picking a shipped sample to copy and modify, wiring pkg-config + meson (or FFI from Rust / Go / Python against the public C ABI), walking the Core lifecycle (create → configure → start → connect → use → stop → destroy), or decoding a DOCA_ERROR_*. Trigger on implicit phrasing: "write my first DOCA program", "meson line for doca_rdma_*", "DOCA_ERROR_BAD_STATE on first call", "call DOCA from Rust", "built clean but nothing on the wire". Refuse install / hugepages / pkg-config (doca-setup), docs or version lookup (doca-public-knowledge-map), and library-internal API work (matching library skill).

Use this skill when the user wants to invoke the read-only doca_caps CLI to ask what DOCA sees on this host — listing DOCA devices and PCIe addresses, listing representor devices, asking which DOCA libraries are available on the current OS, checking per-device per-library capabilities, scoping output to a specific PCIe address, or capturing a side-effect-free capability snapshot for debug or install smoke-test. Trigger even when the user does not explicitly mention "doca_caps" or "capabilities print tool" — typical implicit phrasings include "what does DOCA actually see on this box", "is my BlueField PF visible to DOCA", "is Flow available on my RHEL host", "enumerate VF representors for pf0", "doca_caps: command not found", or "empty output for RDMA, is the tool broken". Refuse and route elsewhere for DOCA installation (doca-setup owns it; never walk the steps), library-internal capability matrices (Flow pipe creation, RDMA verbs features), streaming telemetry / DTS, or modifying the shipped binary.

Use this skill whenever the user is doing hands-on DOCA programming on a BlueField DPU or ConnectX NIC and needs the foundation primitives every per-library context rests on — walking the doca_ctx lifecycle, discovering doca_dev / doca_devinfo and gating on doca_*_cap_* before trusting a feature, wiring doca_mmap / doca_buf_inventory / doca_buf for zero-copy I/O across libraries, driving doca_pe for completions, or DOCA Log's two-tier (--sdk-log-level vs app-side) model. Trigger even when the user does not say "DOCA Common" — typical implicit phrasings include "my tasks submit but nothing completes", "DOCA_ERROR_BAD_STATE from doca_ctx_start", "--sdk-log-level does nothing for my DOCA_LOG_DBG lines", "share a buf between doca_dma and doca_rdma", or "crashes far from the offending line". Refuse and route elsewhere for per-library questions in isolation (load doca-flow / doca-rdma / doca-eth alongside), installing DOCA (doca-setup), or doc lookup (doca-public-knowledge-map).

Use this skill for BlueField-3 (BF3) day-1 platform bring-up via the classic RShim/BFB path: pushing a BlueField bundle (BFB) to the DPU over RShim with bfb-install from the host, the host-to-DPU TMFIFO management channel (tmfifo_net0, the 192.168.100.x convention), RShim daemon state and console-over-rshim, DPU mode selection (DPU/embedded-function vs separated-host/NIC mode) via mlxconfig, post-BFB recovery, a six-state BlueField-state classifier, and verifying the install (cat /etc/mlnx-release plus version checks). Trigger even when the user does not say "BF3" — typical phrasings include {push a BFB to my BlueField-3}, {bfb-install exited 0 but the DPU never came back}, {ping 192.168.100.2 works but ssh fails}, or {is DOCA on the host or the Arm side?}. BFB reflash, mlxconfig set, mode changes, and firmware burns are destructive: require explicit target-bound confirmation and load doca-hardware-safety. App launch, container deploy, env install, and the BF4 BMC-Redfish path route elsewhere.

Use only to generate or update a governance skill card for a specified existing agent skill directory. Do not use for explaining, listing, comparing, or discussing skill capabilities.

Load a sharded, on-disk dataset (sharded .npy, Parquet/Arrow, raw binary, sharded HDF5, custom layouts) into a distributed cuPyNumeric ndarray via a manual partition + leaf @task launch with CPU/OMP/GPU variants. Use when no single-call loader fits, including when per-shard row counts differ across files. Prefer cupynumeric.load or legate.io.hdf5.from_file when they apply.

Pre-migration readiness assessor for porting NumPy to cuPyNumeric. Use BEFORE substantial porting work begins when the user asks whether code will scale on GPU, whether they should migrate to cuPyNumeric, which NumPy patterns transfer cleanly, what must be refactored before porting, or mentions pre-port assessment, scaling analysis, or refactor planning. Inspect the user's source code, look up NumPy usage, cross-reference the cuPyNumeric API support manifest, and distinguish distributed-scaling-friendly patterns from blockers such as unsupported APIs, scalar synchronization, host round-trips, Python/object-heavy control flow, shape/data-dependent branching, and in-place mutation hazards. Produce a verdict of READY, LIGHT REFACTOR, SIGNIFICANT REFACTOR, or NOT RECOMMENDED, with concrete refactor pointers.

Router for NVIDIA NuRec/NRE: USDZ rendering, NCore conversion, 3DGS, gRPC sensor sim, carline adaptation, PhysicalAI HF datasets. Do NOT use for SimReady or infra setup.

Iteratively optimize cuTile kernel performance through systematic profiling, bottleneck analysis, IR comparison, and targeted tuning. Covers tile sizes, occupancy, autotune configs, TMA, latency hints, persistent scheduling, num_ctas, flush_to_zero, and IR-level debugging. Use when asked to "optimize cutile kernel", "improve kernel perf", "tune cutile performance", "make kernel faster", or iteratively benchmark and refine a cuTile GPU kernel in the TileGym project.

Validate and use selective and full activation recompute in Megatron Bridge to reduce GPU memory usage at the cost of extra compute. Use for activation memory OOMs or regressions involving recompute_granularity, recompute_num_layers, recompute_modules, recompute_method, selective recompute, full recompute, or activation checkpointing.

Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV/Parquet I/O, nullable semantics, and multi-GPU DataFrame workloads.

Used for extracting selected metadata from one DICOM file and flagging standard-tag PHI presence. Not for anonymization or clinical use.

Vehicle routing (VRP, TSP, PDP) with cuOpt — Python API only. Use when the user is building or solving routing in Python.

Use this skill when deploying, operating, or integrating the VSS 3.2 GA RT-Embed Video Embedding microservice. Covers Docker Compose bring-up, GPU and storage prerequisites, the `/v1` REST API (file uploads, text and video embeddings, live RTSP streams, health and metrics), Redis/Kafka/OTel integration, common failure modes, and teardown.

Use this skill when deploying standalone RT-VLM dense captioning or calling its REST API (uploads, captions, streams, chat-completions, Kafka). Not for VSS profile deploy or video-search ingestion.

Use to summarize a recorded video via the LVS summarization microservice (HITL-gated) with a VLM fallback. Not for report generation or live RTSP captioning.

Use when the user wants to orchestrate defect image generation, run associated setup, or handle outputs on OSMO. The Day 0 path handles cold-start with USD-to-ROI, image-edit augmentation, and AnomalyGen to create initial PCBA datasets. The Day 1 path performs inference and labeling on real images. This skill helps with first-time asset setup, creation of finetuning checkpoints, and configuring deployment. Trigger keywords: defect image generation, dig workflow, dig pipeline, defect image detection workflow, aoi pipeline, aoi anomalygen, usd2roi anomalygen, day 0 pcba, day 1 pcba, day 1 real-photo alignment, day 1 manual roi, metal surface anomaly, glass defect, anomalygen finetune, setup_pcb, setup_metal, setup_glass, setup_pretrained, dig setup, dig datasets, dig pretrained checkpoint, dig image-edit endpoint.

Write and migrate code with DALI's imperative API

cuOpt REST server — start server, endpoints, Python/curl client examples. Use when the user is deploying or calling the REST API.

Validate and use packed sequences and long-context training in Megatron-Bridge, including offline LLM packing, collate-time VLM packing, Energon online packing, and CP constraints.

Validate and use MoE expert-parallel communication overlap in Megatron-Bridge, including overlap_moe_expert_parallel_comm, delay_wgrad_compute, and flex dispatcher backends such as DeepEP and HybridEP.

Install cuOpt for Python, C, or server via pip, conda, or Docker; verify the install. For building cuOpt from source, see cuopt-developer.

Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.

Four-step image referring-expression pipeline: turns images plus KITTI bounding-box labels into region descriptions, scene captions, grounded referring expressions, and (optionally) verified expressions via VLM distillation. Use when the user wants to generate referring-expression annotations from images with KITTI labels, build region descriptions, produce grouped grounding phrases tied to bboxes, run a double-check verification pass on grounding expressions, auto-label traffic / scene images for referring datasets, or run the image_referring_expression pipeline. Triggers include 'referring expression', 'region description', 'KITTI labels', 'spatial relationship annotation', 'auto-label image referring expression', 'image_referring_expression'.

Scaffold a standalone DeepStream CV microservice that publishes detection metadata (and optional segmentation/frame-mask payloads) to Kafka mdx-raw for VSS Search and Alerts profiles.

Run DiffDock molecular docking via NVIDIA NIM to predict small-molecule binding poses against protein targets. Use for DiffDock, molecular docking, ligand docking, blind docking, SMILES or SDF ligands, ranked poses, confidence scores, hosted NVIDIA API, or local Docker deployment.

Generate novel drug-like molecules using the GenMol NIM microservice. Use for de novo generation, scaffold decoration, motif extension, lead optimization, SAFE notation, QED or LogP ranking, hosted NVIDIA API calls, or local Docker deployment. GenMol takes SAFE notation in the smiles field, not ordinary SMILES.

Generate and analyze DNA sequences using NVIDIA's Evo 2 BioNeMo NIM microservice. Use for Evo2/Evo 2, DNA generation, genomic sequence generation, hosted generation, local Docker deployment, local forward passes, layer outputs, logits, sampled probabilities, and BioNeMo NIM workflows.

Use this skill for MolMIM, NVIDIA's BioNeMo NIM microservice for small-molecule latent-space generation and optimization. Invoke for MolMIM, molecular embeddings, hidden states, latent decoding, sampling around a seed SMILES, CMA-ES guided molecule generation, QED or plogP optimization, hosted NVIDIA API calls, or local Docker deployment.

Run ProteinMPNN inverse folding via NVIDIA NIM to design protein sequences for a target backbone. Sends user-provided PDB files and design parameters to NVIDIA's hosted API, authenticated with an environment API key, or to a user-selected local NIM. Use for sequence design, backbone redesign, fixed chains and residues, omit_AAs, sampling temperature, soluble model, local Docker, and multi-FASTA output.

Generate multiple sequence alignments (MSAs) for protein sequences using the ColabFold MSA-Search NIM. Use for homolog search, UniRef30/ColabFold env searches, A3M or FASTA alignments, paired MSA search for complexes, PDB70 structural templates, hosted NVIDIA API calls, or local Docker deployment. For local deployment, download the databases in parallel with aria2c and launch via NIM_MODEL_NAME (the recommended default fast path, ~14 min vs over 80 min for the built-in downloader); a plain docker run uses the slow built-in downloader.

Use this skill for OpenFold3, NVIDIA's BioNeMo NIM microservice for biomolecular structure prediction. Invoke whenever the user mentions OpenFold3 or needs protein, protein-ligand, protein-DNA/RNA, or multi-chain complex prediction with the hosted NVIDIA API or local Docker NIM. Covers endpoint choice, auth, request payloads, output artifacts, confidence scores, and local container setup.

Train, evaluate, and export Workflow RL policies

Use this skill when the operator is authoring, building, loading, or debugging a custom doca-bench plug-in — a versioned shared library with DOCA_EXPERIMENTAL-marked C entry points that doca-bench loads to measure a workload class its built-in modes do not cover, with doca_bench_cuda as the shipped reference exemplar. Trigger even when the user does not say "doca-bench-extension" or "doca_bench_cuda" — typical implicit phrasings include "no built-in doca-bench mode fits my workload", "how do I benchmark a CUDA GPUNetIO RX/TX kernel", "doca-bench cannot find or load my custom .so", "extension exported symbols do not match what the parent expects", "soversion mismatch after a DOCA upgrade", or "my GPU kernel hangs because stop_flag was never set". Refuse and route elsewhere for questions about which built-in doca-bench mode to pick, DOCA GPUNetIO programming semantics, CUDA toolkit installation, or contributor work on in-tree extensions — those belong to other skills.

Hardware compression offload with DOCA Compress on a BlueField DPU, ConnectX NIC, or host: DEFLATE compress and decompress, plus LZ4 decompress (stream and block; LZ4 encode is NOT supported), instead of CPU zlib / zstd. Trigger on "offload this gzip", "decompress incoming network data on the NIC", "compress task returns INVALID_VALUE on alloc_init", "submitted a task but no completion arrives", or "decompress LZ4 on the BlueField", even without naming DOCA Compress. Covers enabling compress-deflate, decompress-deflate, decompress-lz4-stream and decompress-lz4-block tasks on a doca_compress context, sizing source / destination doca_buf against the per-task cap query, mmap permissions, offload vs CPU decisions, a round-trip smoke, and DOCA_ERROR_* from Compress calls. Refuse and route elsewhere for zstd / Snappy / brotli, LZ4 encode (a CPU LZ4 library), pure mmap-to-mmap copies (doca-dma), or DOCA Core lifecycle internals.

Use this skill when the user has an existing DPDK application and is adding DOCA capabilities in-place — most commonly DOCA Flow hardware steering — without rewriting the data-plane in DOCA-native form: binding a DPDK port id to a `doca_dev` (`doca_dpdk_port_probe` / `doca_dpdk_port_as_dev`), converting `rte_mbuf` ↔ `doca_buf`, querying `doca_dpdk_cap_is_rep_port_supported`, or debugging `DOCA_ERROR_*` from a bridge call. Trigger even without "DOCA DPDK Bridge": "how do I add DOCA Flow to my DPDK app", "make a DPDK port visible to DOCA", "the bridge loads but every operation returns errors", "pkg-config --exists doca-dpdk-bridge fails", or "DOCA_ERROR_NOT_FOUND on port registration". Route elsewhere for fresh DOCA-native packet I/O (doca-eth), flow-rule programming (doca-flow), DOCA or DPDK install (doca-setup), or RDMA data movement (doca-rdma).