pelloran-inc/tritify is a Rust engine that runs inference for language models and image segmentation models using integer arithmetic. The latest release is v0.0.1. The provided material does not separately list what changed in this version. The README describes a design in which the hash of run results also matches across devices under the same configuration. The license is Apache-2.0.
The scope covered by this engine includes forward passes for language models, mixture-of-experts models (MoE), and SAM 2.1 image paths. Floating-point arithmetic can yield different result bits on each device because of the order of additions or rounding differences in platform-specific math libraries. tritify computes inference paths with integer arithmetic to reduce such differences and aims to make the output bytes identical across runs and devices under the same configuration. According to the README, each operation is compared against the floating-point reference path, and if the difference exceeds the allowed tolerance, the check does not pass. Workspace tests and command-line examples are also covered.
This description concerns the design and verification methods presented by the repository. The provided information contains no results comparing performance or accuracy against other inference engines. It can therefore be viewed as a project aiming at reproducibility of outputs through integer arithmetic, but the speed or accuracy benefits in real workloads cannot be judged from this material alone.