gpumetropolis 0.7: from a fast sampler to a full joint model
Between 0.4 and 0.7 the package grew an engine and a destination. The engine is the gradient: reverse-mode automatic differentiation of the compiled log-density, which turns on a Metropolis-adjusted Langevin path and a conjugate exact fast path. The destination is the joint distribution: automatic marginal selection plus a copula layer, both halves of Sklar's theorem, with a new log-gamma operation carried identically across CPU, CUDA and Vulkan.
Read the post →