/uchikoma

essentialist neural network transpiler for non-floating-point runtimes

Primary LanguagePython

uchikoma: essentialist neural network transpiler for non-floating-point runtimes

uchikoma is a transpiler from TVM's intermediate representation to programming languages targeting emergent runtimes that do not support floating point operations. We currently support/plan to support:

  • Circom for ZKP systems, which uses field arithmetic
  • Solidity for Ethereum Virtual Machines, which uses signed and unsigned integer arithmetic

The finalist submission for ETHSanFransico 2022 can be found under branch ethsf-2022.