K3-Node brings multibackend GNNs to Keras 3

K3-Node targets a real gap: Keras provides backend portability, but graph learning still heavily favors PyTorch. The pitch is compelling, though its success will depend on true numerical and performance parity—not just matching function names.

• –PyG API compatibility could significantly lower migration costs for graph-learning teams.
• –Support for GraphMAE2, Graphormer, GraphGPS, GROVER, and Mole-BERT makes this more than a basic GCN/GAT port.
• –Keras already supports JAX, TensorFlow, and PyTorch backends, giving K3-Node a credible portability foundation.
• –Sparse kernels, samplers, 3D operators, and compiler behavior will be the real tests of the “100% parity” claim.
• –Independent benchmarks and production-grade releases will determine whether it can challenge PyG beyond demos.