NVIDIA nvmath-python v1.0 is now generally available, providing Python bindings to CUDA-X math libraries (cuFFT, cuBLASLt, cuDSS, cuSPARSE, cuTENSOR, cuBLASMp, and more). It supports CPU, single-GPU, and multi-GPU/multi-node execution, works with NumPy, CuPy, and PyTorch arrays, and automatically infers execution space from input tensors. Key features include stateless (functional) and stateful (class-based) APIs, where stateful APIs amortize planning and autotuning costs across repeated executions. The library supports composite fused operations for higher arithmetic intensity, custom FFT callbacks compiled via numba-cuda, and device APIs callable from within GPU kernels. A universal sparse tensor (UST) enables user-defined sparse formats via a DSL. Installation is flexible via pip, conda, uv, or pixi with optional dependency selection.
