CuPy

CuPy: The high-performance, GPU-accelerated array library for Python, offering a near-perfect drop-in replacement for NumPy and SciPy.

CuPy delivers GPU acceleration for Python array computing: it is a NumPy and SciPy-compatible library, acting as a drop-in replacement for existing code. Simply swap `numpy` for `cupy` to leverage NVIDIA CUDA or AMD ROCm platforms. This library utilizes optimized CUDA Toolkit components (e.g., cuBLAS, cuFFT) to maximize GPU architecture performance, often providing speedups exceeding 100X on specific operations. It also supports advanced features: custom CUDA kernels, memory pooling, and multi-GPU environments via `cupyx.distributed`.

https://cupy.dev/