tirx-harness

Inspection, numerical simulation, and native correctness checks for TIRx kernels.

python -m pip install tirx-harness==0.1.0rc0

The release supports Python 3.12 and 3.13 on Linux x86_64 and aarch64 with glibc 2.28 or newer. Wheels include the compiled Rust/TVM frontend and the NumSim engine sources. Compatible Apache TVM and TVM FFI packages are installed as dependencies. CUDA is not needed to import the package or run CPU simulation.

The package exposes:

• tirx_harness.synccheck and tirx_harness.racecheck: native correctness checks
• tirx_harness.numsim: transpilation and deterministic numerical simulation
• tirx_harness.dump_kernel: CUDA, PTX, cubin, and SASS inspection

NumSim compiles each kernel's native artifact on first use. Install Rust 1.89 or newer, Cargo, and a C linker before running simulation or checkers. Source installation also requires these tools; installing a prebuilt wheel does not. GPU compilation and source dumping additionally require the CUDA toolkit and a suitable GPU environment.

Build from source

From a checkout, install the inner package rather than the root dependency bundle:

python -m pip install /absolute/path/to/TIRx-harness/tirx_harness

Build isolation installs the pinned TVM FFI build dependency automatically. The locked native bindings and their license notices are included in the source distribution; no private GitHub access is required. Cargo downloads its locked public dependencies during compilation.

The compiler baseline is apache-tvm==0.27.0. The full GPU test suite also requires the boolean BitwiseNot code-generation fix. Use a compatible TVM build containing that fix for those GPU tests.

See the documentation and source repository for examples.