IDEAS — Improved Decoding and Equivalence Automated testing at Scale

Dependencies

To install the Python and Rust toolchain dependencies, run make install.

Tip

This will install [email protected] and [email protected] for the current user.

C Project structure

The current toolkit assumes the following structure for the C projects that will be translated:

📦IDEAS
 ┣ 📂src/ideas  # Core library
 ┗ 📂examples   # Project folders go here
   ┣ 📂project_name # A single C project folder with an arbitrary name
   ┃ ┣ 📂test_case
   ┃ ┃ ┣ 📂include
   ┃ ┃ ┃ ┗ 📄lib.h
   ┃ ┃ ┣ 📂src
   ┃ ┃ ┃ ┗ 📄main.c
   ┃ ┃ ┗ 📄CMakeLists.txt # Must be correct and complete
   ┃ ┗ 📂test_vectors
   ┃   ┣ 📄test1.json
   ┃   ┗ 📄test2.json
   ┗ 📂other_project_name

To run translation on a C project folder, it must be placed in the top-level examples folder.

Basic usage

To start a local vLLM server, run:

make serve

By default, make serve without specifying VLLM_ARGS is optimized for parallelized eight-way inference, so will require eight available devices.

To run LLM-based translation and attempt to build all examples and save results in a newly created translation.{git-hash} sub-folder in each folder, run:

make examples/build

To run agentic code repair on existing translations for all examples, run:

make examples/repair

To run all tests (if available) on existing translations for all examples, run:

make examples/test

To control the output directory and translation hyper-parameters, set:

export TRANSLATION_DIR="custom-name"
export TRANSLATE_ARGS="hyperparam=value"

To delete the current set of translated examples, run:

make examples/clean

For translating a specific example using Rust compiler feedback, run:

make examples/path/to/project/build # Or repair/test/clean for other functionality

Direct usage

To directly call the Python translation module on a specific C translation unit and enable debugging, run:

uv run python -m ideas.translate filename=/path/to/file.i