This probably runs
Three days later, CUDA. src/cuda/kernels/vecadd.cu is a whole program, not a library call. Ten thousand floats, host malloc, cudaMalloc, copy, vectorAdd at 256 threads per block, copy back, check within 1e-5, free. The kernel is C[i] = A[i] + B[i]. matmul.cu is an empty file with the right name. The commit message is the honest one: This probably runs.
They are not wired through Zig’s C FFI. They are not called from Tensor.add. They are the GPU half of the original bet, sitting next to the Zig, waiting for someone to link them. An a.out from compiling the sample sneaks into the tree and leaves a week later, along with *.out in gitignore. Not architecture. The kernels are.