I don't think the tests in https://github.com/JuliaDiff/SparseDiffTools.jl/blob/master/test/test_gpu_ad.jl verify correctness of the Jacobian, it's only about whether or not the code runs. Since #302 breaks GPU tests, it means the code path with the wrong indexing was taken by GPUs, so GPU Jacobians have probably been wrong all along.