KERN■Ship the model as a program

A small runtime verifies a typed manifest, loads hash-pinned GPU kernels, and executes the program the artifact declares.

Quick start

Understand the artifact

One typed contract

Buffers, state, calls, launch arguments, and serving protocol live in one verified manifest.

Kernels from anywhere

Load cubins produced by existing frameworks or handwritten code. Every module is selected by SHA-256.

Model-agnostic runtime

The executor knows how to verify and run a program. Model-specific structure remains in the artifact.

Evidence for a swap

kern test compares a candidate manifest with a reference through structural, numerical, fuzz, and timing evidence.

Start with the boundary

A deployable model has three parts:

text

manifest.json     typed buffers, state, programs, and serving protocol
kernels/          compiled device modules pinned by digest
weights           the model's own checkpoint; the manifest says which tensor lands where

The runtime refuses an invalid manifest before using it. When it loads an artifact, it also checks that the supplied modules and their launch ABI match what the manifest declares.

Read the overview →