cudaGraph_t graph;
cudaGraphCreate(&graph, 0);

cudaGraphNode_t kernelNodeA, kernelNodeB;
cudaKernelNodeParams paramsA = {};
paramsA.func = (void*)kernelA;
paramsA.gridDim = blocks;
paramsA.blockDim = threads;
paramsA.kernelParams = kernelAArgs;

cudaGraphAddKernelNode(&kernelNodeA, graph, nullptr, 0, &paramsA);
cudaGraphAddKernelNode(&kernelNodeB, graph, &kernelNodeA, 1, &paramsB);  // depends on A
