Show HN: GT: experimental multiplexed distributed tensor framework (github.com)

GT (multiplexing tensor framework) is an experimental Python tensor runtime (pip install git+https://github.com/bwasti/gt.git) that rethinks distributed ML execution by separating eager, GPU-agnostic client math from a dynamic, asynchronous dispatcher+worker back end. Clients emit pure functional "instructions" which a single dispatcher rewrites into GPU-aware tasks and streams to N workers (one per GPU). The system uses ZeroMQ DEALER/ROUTER for high-throughput, batched messaging, supports IPC for localhost, and workers can JIT-compile "hot" paths annotated by the dispatcher. Users control sharding, pipeline stages, and compile directives via named signals and a YAML config; annotations are optional so the same code can run locally or distributed. Technically significant features include a PyTorch-like API with tape-based autograd at the client layer, support for data/model/pipeline parallelism, replicated parameters, and per-layer sharding strategies. GT provides observability (instruction logs, timeline visualizer, htop-style real-time worker monitor and trace capture) to debug idle workers, bottlenecks, and communication patterns. Implications for the AI/ML community: a more asynchronous, OS-inspired execution model could improve GPU utilization and flexible placement for large models, while offering tooling to visualize distributed execution—though it remains experimental and incurs serialization/transport overhead compared with in-process frameworks.