Decentralised AI Training and Inference with BlockTrain

BlockTrain partitions a transformer into independently trainable blocks, each optimized on a local objective and composed at inference, eliminating the need for full-model optimizer state on any single worker. On byte-level WikiText, a six-worker run achieves cross-entropy 1.385 versus a centralized reference at 1.359, a gap of roughly 0.026 CE. The protocol targets decentralized training across commodity hardware using HTTP/TCP transport.