Skip to content
NVIDIA CUDA-Q Onboarding

NVIDIA CUDA-Q Onboarding Skill

NVIDIA's official onboarding skill for CUDA-Q — installs the platform, writes your first quantum kernel, picks a GPU simulator and routes you to real QPU hardware.

GitHub Docs

Install

Install with the skills CLI

One command for Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Windsurf and dozens of other agents. Needs Node.js; the skill is fetched straight from NVIDIA/skills.

$ npx skills add NVIDIA/skills --skill cudaq-guide

Run it in a terminal. The CLI finds the agents installed on your machine, asks which ones to add the skill to, and whether to install it into the folder you ran it from or for all your projects.

Other ways to install

Copy the whole skill folder — SKILL.md and any scripts or references beside it — into ~/.claude/skills/cudaq-guide/, where Claude Code picks it up in every project. Only git is needed.

tmp=$(mktemp -d) && \
  git clone --depth 1 --branch main https://github.com/NVIDIA/skills.git "$tmp" && \
  mkdir -p ~/.claude/skills && \
  cp -R "$tmp"/skills/cudaq-guide ~/.claude/skills/cudaq-guide

No scripts or references next to it? Saving the SKILL.md alone as ~/.claude/skills/cudaq-guide/SKILL.md works too.

Using another agent? Its personal skills folder
Codex
~/.agents/skills/
Cursor
~/.cursor/skills/
GitHub Copilot
~/.copilot/skills/
Gemini CLI
~/.gemini/skills/
Windsurf
~/.codeium/windsurf/skills/
OpenCode
~/.config/opencode/skills/

About

  • Routes to install, test program, GPU simulation, QPU or applications from one command
  • Validates a fresh install with a Bell state kernel instead of assuming success
  • Picks between qpp-cpu, the nvidia target, mgpu memory pooling and mqpu circuit batching
  • Shows async circuit dispatch with sample_async / observe_async across virtual QPUs
  • Distributes large Hamiltonians with cudaq.parallel.thread or cudaq.parallel.mpi
  • States the restricted-Python kernel subset and platform limits up front

An official, NVIDIA-authored agent skill that turns a coding agent into a CUDA-Q guide. CUDA-Q is NVIDIA's hybrid quantum-classical programming platform, and the usual failure mode for newcomers is not the physics — it is picking the wrong backend, writing kernel code the compiler rejects, or burning hours on a multi-GPU option that solves a different problem than the one they have.

What it teaches the agent
  • A routing table instead of a wall of docs. Invoked as /cudaq-guide [topic], it jumps straight to install, first test program, GPU simulation, QPU hardware, built-in applications or parallelisation. With no argument it presents the menu and asks what you actually want.
  • Installation that ends in a verified result. The default path is pip install cudaq, validated by building a Bell state kernel and checking the histogram comes out near { 00:~500 11:~500 } rather than assuming the install worked.
  • Backend selection. qpp-cpu for CPU-only work, the nvidia target for a single GPU, and a clear rule for the two multi-GPU strategies that are easy to confuse: --target-option mgpu pools GPU memory when one circuit is too large for a single card, while mqpu maps a virtual QPU per GPU so many independent circuits run in parallel via sample_async / observe_async.
  • Hamiltonian batching, where adding execution=cudaq.parallel.thread (single node) or cudaq.parallel.mpi (multi-node) to cudaq.observe distributes terms across GPUs with no other code change.
  • A two-step QPU dialogue — technology first, then provider — before reading the matching hardware documentation, so credential setup is not guessed at.
Limitations it makes explicit

GPU simulation requires Linux on x86_64 or ARM64; macOS is CPU-only. The mgpu target needs MPI. Kernel bodies accept only a restricted Python subset, so NumPy and SciPy calls inside @cudaq.kernel will not compile. QPU access needs provider-specific credentials.

Requirements

Python 3.10+ for the Python path, C++20 and Linux or WSL for the C++ path, and the CUDA Toolkit for GPU-accelerated targets. An NVIDIA GPU is optional given the CPU simulator.

Installing it

Part of NVIDIA's agent-skills catalogue, mirrored daily from NVIDIA's product repositories. Install the whole set with npx skills add nvidia/skills, or just this one with npx skills add nvidia/skills --skill cudaq-guide --yes. Works with Claude Code, Codex, Cursor and any runtime following the Agent Skills standard. Apache-2.0.

Try it

Once installed, ask your agent something like:

"/cudaq-guide parallelize — I have 5,000 independent VQE circuits and four A100s. Should I use mgpu or mqpu, and what does the dispatch code look like?"

SKILL.md

The instructions your agent loads, as published by NVIDIA.

SKILL.md
name
cudaq-guide
title
Cuda Quantum
description
CUDA-Q onboarding guide for installation, test programs, GPU simulation, QPU hardware, and quantum applications.
version
1.0.1
author
CUDA-Q Team <cuda-quantum@nvidia.com>
tags
[cuda-quantum, quantum-computing, onboarding, getting-started, nvidia]
tools
[Read, Glob, Grep]
license
Apache-2.0
compatibility
Python 3.10+, C++ 20
metadata
author: "CUDA-Q Team <cuda-quantum@nvidia.com>" tags: - cuda-quantum - quantum-computing - onboarding - getting-started - nvidia languages: - python - c++ domain: "quantum"
CUDA-Q Getting Started Guide

You are a CUDA-Q expert assistant. Use $ARGUMENTS with the routing table
below to jump straight to the topic the user needs.

Purpose

Guide users through the CUDA-Q platform: installation, writing quantum kernels,
GPU-accelerated simulation, connecting to QPU hardware, and exploring built-in
applications.

Prerequisites
  • Python 3.10+ (for Python installation path)
  • CUDA Toolkit (for GPU-accelerated targets on Linux; not required on macOS)
  • NVIDIA GPU (optional; CPU-only simulation available via qpp-cpu)
  • For C++ path: Linux or WSL on Windows
  • For QPU access: provider-specific credentials and account
Instructions
  • Invoke with /cudaq-guide [argument]
  • If no argument is given, display the full onboarding menu and ask what

the user wants to explore

  • Pass an argument from the routing table below to jump directly to that topic
  • Read local CUDA-Q documentation files to answer questions accurately
References
SectionDoc file
Installdocs/sphinx/using/install/install.rst, docs/sphinx/using/quick_start.rst
Test Programdocs/sphinx/using/basics/kernel_intro.rst, docs/sphinx/using/basics/build_kernel.rst
GPU Simulationdocs/sphinx/using/backends/sims/svsims.rst, docs/sphinx/using/examples/multi_gpu_workflows.rst
QPUdocs/sphinx/using/backends/hardware.rst, docs/sphinx/using/backends/cloud.rst
Applicationsdocs/sphinx/using/applications.rst
Parallelizedocs/sphinx/using/examples/multi_gpu_workflows.rst
Routing by Argument
ArgumentAction
installWalk through installation (see Install section)
test-programBuild and run a Bell state kernel to verify CUDA-Q is working properly
gpu-simExplain GPU-accelerated simulation targets (see GPU Simulation section)
qpuExplain how to run on real QPU hardware (see QPU section)
applicationsShowcase what can be built with CUDA-Q (see Applications section)
parallelizeShow how to run circuits in parallel across multiple QPUs (see Parallelize section)
_(none)_Print the full menu below and ask what they'd like to explore

---

Full Menu (no argument)

Present this when invoked with no argument

CUDA-Q Getting Started

CUDA-Q is NVIDIA's unified quantum-classical programming model for CPUs, GPUs, and QPUs.
Supports Python and C++. Docs https://nvidia.github.io/cuda-quantum/

Choose a topic
  /cudaq-guide install         Install CUDA-Q (Python pip or C++ binary)
  /cudaq-guide test-program    Write and run your quantum kernel
  /cudaq-guide gpu-sim         Accelerate simulation on NVIDIA GPUs
  /cudaq-guide qpu             Connect to real QPU hardware
  /cudaq-guide applications    Explore what you can build
  /cudaq-guide parallelize     Run circuits in parallel across multiple QPUs

---

Install

Instructions

  • Default to Python installation unless the user explicitly mentions C++ or

the nvq++ compiler.

  • After installation, always guide the user through the validation step

(run the Bell state example and confirm output shows { 00:~500 11:~500 }).

  • Default to GPU-accelerated targets (nvidia) unless: the user is on

macOS/Apple Silicon, mentions no GPU available, or explicitly asks for
CPU-only simulation - in those cases use qpp-cpu.

  • Do not suggest cloud trial or Launchpad options unless the user has no

local environment or asks about cloud access.

Platform notes

  • Linux (x86_64, ARM64): full GPU support -

pip install cudaq + CUDA Toolkit

  • macOS (ARM64/Apple Silicon): CPU simulation only -

pip install cudaq (no CUDA Toolkit needed)

  • Windows: use WSL, then follow Linux instructions
  • C++ (no sudo):

bash install_cuda_quantum*.$(uname -m) --accept -- --installpath $HOME/.cudaq

  • Brev (cloud, no local setup): Log in at the NVIDIA Application Hub,

open a CUDA-Q workspace, then SSH in with the Brev CLI:

  brev open ${WORKSPACE_NAME}

CUDA-Q and the CUDA Toolkit are pre-installed.

---

Test Program

Key concepts to explain

  • @cudaq.kernel / __qpu__ marks a quantum kernel - compiled to Quake MLIR
  • cudaq.qvector(N) allocates N qubits in |0⟩
  • cudaq.sample() - kernel measures qubits; returns bitstring histogram

(SampleResult)

  • cudaq.run() - kernel returns a classical value; runs shots_count times

and returns a list of those return values

  • cudaq.observe() - computes expectation value ⟨H⟩ for a spin operator
  • cudaq.get_state() - returns the full statevector (simulator only)

Kernel restrictions

  • Only a restricted Python subset is valid inside a kernel - it compiles to

Quake MLIR, not regular Python.

  • NumPy and SciPy cannot be used inside a kernel. Use them outside the kernel

for classical pre/post-processing.

  • Kernels can call other kernels; the callee must also be a @cudaq.kernel.

For compiler internals (inspect module -> ast_bridge.py -> Quake MLIR ->
QIR -> JIT), route to /cudaq-compiler.

---

GPU Simulation

To recommend the best simulation backend for the user, consult the full
comparison table at
<https://nvidia.github.io/cuda-quantum/latest/using/backends/simulators.html>

Available GPU Targets
TargetDescriptionUse when
nvidia (default)Single-GPU state vector via cuStateVec (up to ~30 qubits)Default choice for most simulations on a single GPU
nvidia --target-option fp64Double-precision single GPUHigher numerical precision needed (e.g. chemistry, sensitive observables)
nvidia --target-option mgpuMulti-GPU, pools memory across GPUs (>30 qubits)Circuit exceeds single-GPU memory; requires MPI
nvidia --target-option mqpuMulti-QPU, one virtual QPU per GPU, parallel executionRunning many independent circuits in parallel (e.g. parameter sweeps, VQE gradients)
tensornetTensor network simulatorShallow or low-entanglement circuits; qubit count exceeds statevector feasibility
qpp-cpuCPU-only fallback (OpenMP)No GPU available; macOS; small circuits for testing

---

QPU

When the user invokes this section, do not dump all providers at once.
Instead, follow this two-step dialogue:

Step 1 - ask which technology they want

Which QPU technology are you targeting?
  1. Ion trap       (IonQ, Quantinuum)
  2. Superconducting (IQM, OQC, Anyon, TII, QCI)
  3. Neutral atom   (QuEra, Infleqtion, Pasqal)
  4. Cloud / multi-platform (AWS Braket, Scaleway)

Step 2 - once they pick a technology, ask which provider, then read the
corresponding doc file and walk the user through it step by step.

TechnologyProviderDoc file
Ion trapIonQdocs/sphinx/using/backends/hardware/iontrap.rst (IonQ section)
Ion trapQuantinuumdocs/sphinx/using/backends/hardware/iontrap.rst (Quantinuum section)
SuperconductingIQMdocs/sphinx/using/backends/hardware/superconducting.rst (IQM section)
SuperconductingOQCdocs/sphinx/using/backends/hardware/superconducting.rst (OQC section)
SuperconductingAnyondocs/sphinx/using/backends/hardware/superconducting.rst (Anyon section)
SuperconductingTIIdocs/sphinx/using/backends/hardware/superconducting.rst (TII section)
SuperconductingQCIdocs/sphinx/using/backends/hardware/superconducting.rst (QCI section)
Neutral atomInfleqtiondocs/sphinx/using/backends/hardware/neutralatom.rst (Infleqtion section)
Neutral atomQuEradocs/sphinx/using/backends/hardware/neutralatom.rst (QuEra section)
Neutral atomPasqaldocs/sphinx/using/backends/hardware/neutralatom.rst (Pasqal section)
CloudAWS Braketdocs/sphinx/using/backends/cloud/braket.rst
CloudScalewaydocs/sphinx/using/backends/cloud/scaleway.rst

After walking through the provider steps, always close with

  • Test locally first with emulate=True before submitting to real hardware.
  • Use cudaq.sample_async() / cudaq.observe_async() for non-blocking submission.
  • Handle provider credentials securely: export them as environment variables

in your shell session (or a local profile that is not committed to version
control) rather than hardcoding them in source or notebooks. Never paste
tokens into shared files, logs, or commits, and prefer a secrets manager
where one is available.

---

Applications

CUDA-Q ships with ready-to-run application notebooks

CategoryExamples
OptimizationQAOA, ADAPT-QAOA, MaxCut
ChemistryVQE, UCCSD, ADAPT-VQE
Error CorrectionSurface codes, QEC memory
AlgorithmsGrover's, Shor's, QFT, Deutsch-Jozsa, HHL
MLQuantum neural networks, kernel methods
SimulationHamiltonian dynamics, Trotter evolution
FinancePortfolio optimization, Monte Carlo

---

Parallelize

CUDA-Q supports two distinct multi-GPU parallelization strategies - pick based
on what you are trying to scale.

GoalStrategyTarget option
Single circuit too large for one GPUPool GPU memorynvidia --target-option mgpu
Many independent circuits at onceRun circuits in parallelnvidia --target-option mqpu
Large Hamiltonian expectation valueDistribute terms across GPUsmqpu + execution=cudaq.parallel.thread
Circuit batching with mqpu (sample_async / observe_async)

The mqpu option maps one virtual QPU to each GPU. Dispatch circuits
asynchronously with qpu_id to all GPUs simultaneously.

import cudaq

cudaq.set_target("nvidia", option="mqpu")
n_qpus = cudaq.get_platform().num_qpus()

futures = [
    cudaq.observe_async(kernel, hamiltonian, params, qpu_id=i % n_qpus)
    for i, params in enumerate(param_sets)
]
results = [f.get().expectation() for f in futures]
Hamiltonian batching

For a single kernel with a large Hamiltonian, add execution= to
cudaq.observe — no other code change needed.

# Single node, multiple GPUs
result = cudaq.observe(kernel, hamiltonian, *args,
                       execution=cudaq.parallel.thread)

# Multi-node via MPI
result = cudaq.observe(kernel, hamiltonian, *args,
                       execution=cudaq.parallel.mpi)

See the docs above for complete working examples of both patterns.

---

Examples
  • /cudaq-guide — print the onboarding menu and ask the user which topic to

explore.

  • /cudaq-guide install — walk through installation, defaulting to the Python

pip install cudaq path, then validate with the Bell state example.

  • /cudaq-guide test-program — build and run a Bell state kernel and confirm

the output shows roughly { 00:~500 11:~500 }.

  • /cudaq-guide gpu-sim — recommend a simulation backend (for example

nvidia for a single GPU, or nvidia --target-option mgpu for circuits
larger than one GPU's memory).

  • /cudaq-guide qpu — start the two-step QPU dialogue (technology, then

provider) and read the matching hardware doc.

  • /cudaq-guide parallelize — choose between mgpu (pool memory for one large

circuit) and mqpu (run many circuits in parallel).

---

Limitations
  • GPU simulation requires Linux (x86_64 or ARM64); macOS is CPU-only
  • Multi-GPU mgpu target requires MPI
  • Kernel code must use a restricted Python subset; NumPy/SciPy are not

allowed inside kernels

  • QPU access requires provider-specific credentials and accounts
Troubleshooting
  • Import error after pip install cudaq: Ensure Python 3.10+ and a

supported OS (Linux or macOS)

  • No GPU detected: Verify CUDA Toolkit is installed and nvidia-smi

shows your GPU; fall back to qpp-cpu

  • Kernel compile error: Check that only supported Python constructs are

used inside @cudaq.kernel

  • QPU submission fails: Confirm credentials are set as environment

variables per the provider docs

Install

Diagnoses wrong gradients in differentiable NVIDIA Warp programs by measuring first — comparing autodiff against finite differences on a shrunk reproduction before proposing any fix.

Trains, distils, quantises, evaluates, exports and runs inference for RT-DETR real-time object detection in NVIDIA TAO, routing training through AutoML hyperparameter search by default.

Browse all skills →