PHP extension enables native NVIDIA CUDA GPU tensor operations

A new native PHP extension, php-gpu-tensors, has been developed to enable NVIDIA CUDA-accelerated numerical workloads and machine-learning data pipelines directly within PHP. This extension allows for GPU tensor operations without the requirement of a Python runtime.

The extension provides features such as broadcasting, matrix multiplication, reductions, and Python-style slicing. It also supports the compilation of custom CUDA C++ kernels at runtime using NVRTC. To address performance bottlenecks in small training steps, the extension includes an opt-in kernel-fusion mode. This mode merges elementwise operations into generated CUDA kernels to reduce per-operation overhead and eliminate intermediate buffers.