Light Unified Matrix Architecture
One programming model across photonic switching and compute
LUMA — the Light Unified Matrix Architecture — is the software layer above LightScale's photonic hardware. A single API spans the Scale1 switching fabric and Photon1 compute: developers write to LUMA, not to individual devices.
Topology, routing, and link assignment are exposed to software and reconfigured at packet cadence. The same code targets every generation, from Scale1 onward, with no per-device rewrite.
One fabric, every topology
Every parallelism strategy has a network shape it wants. Data parallelism moves gradients around rings, expert routing scatters tokens all-to-all, pipelines stream activations down chains. A fixed fabric picks one compromise at build time — and every workload pays for it on every step.
Hyperlane — topology on demand
Hyperlane exposes topology, routing, and link assignment to software, and reconfiguration happens at packet cadence — fast enough to reshape the fabric inside a training step, not just between jobs.
Each phase gets the network it wants: chains while activations stream, a ring while gradients reduce, all-to-all when experts fire — and inference traffic routes on shapes tuned for latency rather than a trained-in compromise. Less time blocked on communication is more time computing, so throughput per accelerator rises without touching the model.
Drops into your stack
LUMA ships as a library that slots cleanly into leading open-source frameworks — PyTorch first. It sits beneath the framework's collective and runtime layers, so existing training and inference code runs unchanged: no new operators, no rewrites, no changes to the model.
The fabric reshapes underneath while your code never notices — the same script that runs today runs on Hyperlane, just with the network out of the way.