tirx-kernels v0.1.2.post1

Release the latest main commit, 8b6ed130a330e522a22b62eed2fa15ed487acec0, as 0.1.2.post1.

Changes

• Add txl.cu_tensor_map_encode_tiled, which emits TVM's typed tensor-map encoding operation while preserving the existing operand order.
• Migrate FP16/BF16 GEMM and DeepSeek-V4 MLA host preludes to support CUDA host source export with TVM 0.27.0.post1.

Fix: #3

Validation

• TIRx-lite API tests: 37 passed.
• Built the wheel and source distribution as 0.1.2.post1; isolated wheel validation passed for 105 kernels and 105 configurations.
• On B200 with TVM 0.27.0.post1, checked FP16 and BF16 GEMM at M=N=K=1024 through both the default host path and exported CUDA host source. The exported kernels passed correctness checks on a non-default CUDA stream in a consumer environment without TVM or tirx-kernels.

tirx-kernels v0.1.2

Changes

• Describe the library as curated high-performance TIRx GPU kernels and link to tirx-harness support (#2).
• Simplify the README installation instructions (#1).

Installation

pip install tirx-kernels==0.1.2

tirx-kernels v0.1.0

Changes

• Publish the stable 0.1.0 release with the kernel implementations from v0.1.0rc2.
• Update installation and release examples for 0.1.0 and remove an obsolete PyPI version warning.

Installation

python -m pip install tirx-kernels==0.1.0

GPU runtimes and the TIRx compiler are installed separately as described in the README.

The release workflow builds and validates the wheel and source distribution on Python 3.10 and 3.12 before uploading them to PyPI.

tirx-kernels v0.1.0rc2

Changes

• Remove the legacy curated_ prefix from eight kernel registry names and their benchmark configuration filenames; update documentation and tests.
• Remove six obsolete baseline rows for retired kernels and regenerate the baseline view. The remaining 232 measurements are unchanged.
• Check that pinned baseline entries resolve to registered kernels and configured workloads.

Registry names

Callers using these kernels must drop the curated_ prefix. The new names are:

• kda_forward_portfolio_multishape
• kda_backward_packed
• kda_decode_multishape
• alphamoe_fp8_blockscale_qwen3next
• msa_prefill_multishape
• msa_decode_multishape
• mla_dsv4_multishape
• vsa_multishape

Kernel computation code is unchanged. The package still contains 105 registered kernels and 105 benchmark configuration files.

Installation

python -m pip install tirx-kernels==0.1.0rc2

This is a release candidate. GPU runtimes and the TIRx compiler are installed separately as described in the README.

The release workflow builds and validates the wheel and source distribution on Python 3.10 and 3.12 before uploading them to PyPI.

tirx-kernels v0.1.0rc1

Changes

• Organize native kernels by workload and group upstream ports under tirx_kernels.ported.
• Update the packed KDA forward implementation and its benchmark measurements.

Installation

python -m pip install tirx-kernels==0.1.0rc1

This is a release candidate. The package includes Python kernel sources, benchmark configurations, reference resources, and licenses. GPU runtimes and the TIRx compiler are installed separately as described in the README.

The release workflow builds and validates the wheel and source distribution on Python 3.10 and 3.12 before uploading them to PyPI.