CortexFoundation / tachikoma

neural network inference standard for zero-knowledge-proof systems

Home Page:https://zk-ml.xyz

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

tachikoma: neural network inference standard for zero-knowledge-proof systems


tachikoma defines how a neural network's inference process should be serialized into a graph of operator computational traces, each of which containing the input, expected output, relevant metadata (including parameters), and an identifier relating back to the original operator in TVM's intermediate representation.


We are actively working on consolidating the standards into a stable form and release relevant artifacts, as well as forming a committee and organizing regular meetings. If you are interested in this effort, please reach out!


in addition, tachikoma's TVM fork is useful for:

  • converting a floating-point neural network or a framework-prequantized model into an integer-only form
  • generating a computational trace binary respecting the tachikoma standard.
  • as a proof of concept, how tachikoma can be used in ZKP systems. We will be implementing a simple graph runtime on top of the tachikoma standard, as well as a circuit builder in ZEXE. The code will be available here: https://github.com/zk-ml/tachikoma-poc-runtime

About

neural network inference standard for zero-knowledge-proof systems

https://zk-ml.xyz

License:Apache License 2.0


Languages

Language:Python 57.2%Language:C++ 37.7%Language:C 1.0%Language:Rust 1.0%Language:Shell 0.8%Language:CMake 0.6%Language:Java 0.5%Language:Go 0.3%Language:TypeScript 0.2%Language:Objective-C++ 0.2%Language:Makefile 0.2%Language:Jinja 0.1%Language:Cython 0.1%Language:Objective-C 0.1%Language:Cuda 0.0%Language:JavaScript 0.0%Language:Assembly 0.0%Language:Groovy 0.0%Language:Batchfile 0.0%Language:HTML 0.0%Language:RenderScript 0.0%