← ClaudeAtlas

cudaq-gpu-kernelslisted

NVIDIA CUDA-Q: kernels in Python (@cudaq.kernel) and C++ (__qpu__), sample/observe/get_state, spin operators, targets (nvidia, nvidia-mgpu, nvidia-mqpu, tensornet, density-matrix-cpu, hardware backends), async and multi-QPU batching, noise models, performance and memory limits. Load when writing CUDA-Q kernels or scaling circuit simulation on GPUs.
rustcircle/c101 · ★ 0 · API & Backend · score 65
Install: claude install-skill rustcircle/c101
# CUDA-Q GPU kernels CUDA-Q compiles typed quantum kernels and runs them on interchangeable targets: CPU and GPU statevector, multi-GPU pooled memory, one virtual QPU per GPU, tensor networks, density matrices and hardware. Pick the target from memory math, batch work across QPUs asynchronously, and check conventions (bit order, exp_pauli sign) once per project. Facts marked verified were checked on CUDA-Q 0.16 (CPU build); GPU targets exist only in CUDA-enabled Linux builds, so confirm them with `cudaq.get_targets()` on the machine. Simulator choice beyond CUDA-Q: `quantum-circuit-simulators`. ## 1. Kernel model ```python import cudaq @cudaq.kernel def ghz(n: int): q = cudaq.qvector(n) h(q[0]) for i in range(n - 1): x.ctrl(q[i], q[i + 1]) mz(q) @cudaq.kernel def ansatz(theta: float): # every argument annotated: int, float, bool, list[float], cudaq.pauli_word, ... q = cudaq.qvector(4) x(q[0]) x(q[1]) exp_pauli(theta, q, "XXXY") ``` - Gates: `h x y z s t rx ry rz r1 u3 swap`, controlled `x.ctrl(c, t)`, adjoint `t.adj(q)`, measurement `mz mx my`, Pauli exponentials `exp_pauli(theta, qubits, word)`; kernels can call other kernels and use `cudaq.control` / `cudaq.adjoint`. - Control flow: `for` over `range`, `if` on mid-circuit measurement results, kernel arguments. - Restrictions: kernels are a statically compiled Python subset. No general Python objects, external libraries, exceptions or untyped containers inside;