ignis-autograd adds a differentiable tensor (Ignis::Tensor) and a reverse-mode autograd tape on top of the Ignis GPU foundation. Build computation graphs over GPU arrays and get exact gradients (verified against finite differences). The building block for neural-network training in pure Ruby on an NVIDIA GPU.

ignis-collective brings NCCL-style collective communication to Ruby: ring/tree all-reduce, and P2P / IPC (VMM) / host-staged / TCP transports, on the Ignis foundation. EXPERIMENTAL — the transports require multiple GPUs / nodes and cannot be exercised on a single GPU, so this is shipped separately from the verified single-GPU stack.

ignis-numerics is GPU numerical computing for Ruby on the Ignis foundation: dense/sparse linear algebra, FFT (cuFFT), random number generation (cuRAND), linear solvers & decompositions (cuSOLVER), and Einstein-notation tensor contraction (cuTENSOR), with an nvmath-style API (Ignis.fft, Ignis.solve, Ignis.zeros, …). The lineage NvRuby began from, revived as a standalone gem.