What it does DGL-KE trains, evaluates, and runs inference on knowledge graph embeddings — dense vector representations of entities and relations in graphs. It wraps six established models (TransE, TransR, RESCAL, DistMult, ComplEx, RotatE) in a command-line interface that works on single machines, GPUs, or distributed clusters.

The interesting bit The speed claims are specific and large: on a graph with 86M nodes and 338M edges, it finishes in 100 minutes on 8 GPUs or 30 minutes on a 4-machine CPU cluster, reportedly 2–5× faster than competing tools. The README pins these numbers to exact hardware and rival systems, which is rarer than you’d expect.

Key highlights

• Built on Deep Graph Library (DGL); PyTorch backend
• Distributed training via dglke_dist_train, single-machine via dglke_train
• Includes evaluation (dglke_eval) and two inference modes: link prediction and embedding similarity
• Ships with benchmark comparisons against GraphVite (FB15k) and PyTorch-BigGraph (Freebase)
• Published at SIGIR 2020 with a full optimization paper

Caveats

• The project appears in maintenance mode; the README now redirects TransE/DistMult/RotatE users to the newer GraphStorm project
• Installation still uses sudo pip3, which may surprise modern Python developers

Verdict Worth a look if you’re embedding billion-scale knowledge graphs and need reproducible speed benchmarks. Skip it if you only need standard small-graph KGE — GraphStorm or plain PyTorch-BigGraph may be more actively maintained.