Release Notes

Highlights

New Table Creation Workflow

The new workflow significantly simplifies connecting and adding new tables to Kumo after the connector is set up.

Workflow condensed to a single step
Table types (i.e., fact/dimension tables) are automatically assigned

PyTorch 2.3 and torch.compile Efficiency Improvements

Kumo now leverages the latest version of PyTorch. This enables model compilation for faster GNNs through iust-in-time (JIT)-compilation into optimized kernels. In particular, this means:

Kumo GNNs come with less Python overheads and utilize less GPU read/writes
Subsequent model operations are fused into dynamic and single GPU kernels
20% to 30% training time reductions with model compilation enabled

Other Updates

EMR on EKS in default mode for IAC2
Documentation updates - restructured for clarity and added more examples