Machine Learning zk-VM Compiler

Fork of zkonduit/ezkl, hijacked for a Hackathon project, see Devfolio submission and this presentation.

Setup

Install Powdr:

git clone git@github.com:powdr-labs/powdr.git
cd powdr
cargo install --path powdr_cli --features halo2

Compiling a VM

Go inside a directory with a network.onnx & input.json, for example:

• examples/onnx/2l_relu_fc
• examples/onnx/4l_relu_conv_fc

Run the following EZKL steps to prepare the network:

cargo run gen-settings -M network.onnx --param-visibility public --input-visibility private --output-visibility private --logrows 16
cargo run compile-circuit -M network.onnx -S settings.json --compiled-circuit network.ezkl
cargo run gen-witness -M network.ezkl -D input.json

Run the generate-program command added by this repository:

cargo run generate-program -M network.ezkl --witness witness.json

This will generate 2 files:

• vm.asm: The model-agnostic Powdr VM with the model-specific assembly program inserted
• memory.csv: The content of the memory, specific to the input (from input.json)

These can be passed to Powdr to compile the VM and make a prove of the program:

powdr pil vm.asm -f --witness-values memory.csv --prove-with estark

(or without the --prove-with estark flag if you just want to check everything works)

For a better comparison with EZKL, you can also create the proof with the Halo2 backend:

powdr pil vm.asm -f --witness-values memory.csv --prove-with halo2 --field bn254

For comparison, to benchmark EZKL:

cargo run -r get-srs --logrows=16 --srs-path=16.srs
cargo run -r setup -M network.ezkl --srs-path=16.srs
time cargo run -r prove -M network.ezkl --witness witness.json --pk-path=pk.key --proof-path=model.proof --srs-path=16.srs