tritify
Deterministic AI inference across multiple machines
tritify is Pelloran’s open-source project for deterministic and reproducible AI inference.
The core idea is simple:
One model.
Multiple machines.
One inference process.
One reproducible result.
tritify lets a model be split across multiple machines while keeping execution deterministic and bit-verifiable.
Instead of treating small floating-point differences as unavoidable, tritify uses exact integer execution so the same model, input, and execution schedule can reproduce the same bits.
This is not a speed claim.
The goal is reproducibility, auditability, and verifiable distributed inference.
Open source under Apache 2.0.
Built with
