tirx-harness 0.1.2.post1

Release the latest main commit, a2347ec1d2aea4090329b043c808a033ebec3109, as 0.1.2.post1.

Changes

• Upgrade to TVM 0.27.0.post1 and require tirx-kernels>=0.1.2.post1, including the tensor-map encoding fixes for CUDA host exports.
• Document exporting CUDA source and running the resulting kernels with tvm-ffi, without installing TVM or the harness.
• Fix ptxas resource parsing so values stay scoped to the first kernel report.
• Add the Grouped GEMM optimization task and DeepGEMM benchmark wheels.
• Add the Python API reference and update the documentation site, examples, and worktree setup output.

Validation

The CUDA export workflow was tested on B200. FP16/BF16 GEMM and RMSNorm ran correctly in a separate consumer environment without TVM or tirx-kernels. Existing generated-code and native frontend tests passed, along with the strict documentation build.

The release workflow checks the source distribution and builds and tests Linux x86_64 and aarch64 wheels before publishing to PyPI.

Full changelog: v0.1.2...v0.1.2.post1

tirx-harness 0.1.2

Release tirx-harness 0.1.2 with stable kcoral and tirx-kernels dependencies.

• Install the harness dependency set directly from PyPI, including kcoral >= 0.1.0 and tirx-kernels >= 0.1.2.
• Provide Linux x86_64 and aarch64 wheels for Python 3.12 and 3.13.
• Include the pinned native build sources as a release asset for builds from this repository.

Install with python -m pip install tirx-harness==0.1.2.