Skip to content

Releases

Swage v0.5.3

Choose a tag to compare

@abhiksark abhiksark released this 08 Nov 12:48
v0.5.3
6d1496d

Swage 0.5.3 updates build and CI dependencies and fixes release qualification
compatibility with pytest 9. Compiler and runtime behavior are unchanged.

  • Update the locked build and test tools, including scikit-build-core 1.1.1
    and pytest 9.2.0, and refresh artifact, wheel, SBOM, and CodeQL actions.
  • Keep the wheel reproducibility check aligned with the installed build
    backend and repository manifest.
  • Materialize native test parameters for compatibility with pytest 9.
  • Isolate installed-wheel qualification from unrelated pytest plugins in
    the shared CUDA environment.
  • Preserve supported dependency ranges during routine updates. Optional
    PyTorch support remains >=2.6,<3.

The signed-tag release workflow
passed all four CPython wheel lanes, source distribution and wheel
reproducibility, security and SBOM checks, and the actual installed release
wheel's correctness, cache, and performance gates. The installed CPython 3.13
wheel passed 2,001 GPU tests with no skips on NVIDIA RTX A6000 (sm_86).
Instrumented LLVM sanitizer checks
and trusted GPU qualification
also passed on the same commit.

Binary wheels target Linux x86-64 with glibc 2.28 or newer and regular
CPython 3.10 through 3.13. Swage remains pre-alpha. GPU qualification covers
RTX A6000 (sm_86); other admitted GPUs remain best-effort. The two opt-in
Compute Sanitizer racecheck tests were skipped.

Release assets include SHA-256 checksums, the SPDX SBOM, signed provenance
and SBOM attestation bundles, and retained CPU and GPU qualification evidence.

Install with PyTorch support:

python -m pip install "swage-compiler[pytorch]==0.5.3"
python -m swage.env --json

See the changes since v0.5.2
and support matrix.

Swage v0.5.2

Choose a tag to compare

@abhiksark abhiksark released this 07 Nov 12:50
v0.5.2
9ba54b7

Swage v0.5.2 packages the native compiler and runtime with the Python API.
Binary wheels target Linux x86-64 with glibc 2.28 or newer and regular
CPython 3.10 through 3.13. The previous v0.5.1 wheel contains only Python.

  • Fixed vector addition and multiplication run on the Native CPU or CUDA
    backend, with float32, float16, and both supported FP8 storage formats.
  • Public CUDA segmented reductions provide sum, maximum, minimum, and mean over
    float32 or float64 scalars and rows of features. Segmented softmax uses
    float32. Both calls support gradients, including second derivatives.
  • Ahead-of-time artifacts let the segmented calls run without loading the
    compiler. Artifact loading now rejects incompatible subgroup metadata
    before a launch instead of leaving segment outputs unwritten.
  • Environment diagnostics, typed fixed-kernel interfaces, persistent kernel
    caching, and installed-wheel checks accompany the native distribution.
  • Documentation states the limits of sum and mean accuracy after
    intermediate overflow. CI and documentation dependencies use the patched
    certifi and pytest releases.

Install the native package with PyTorch support:

python -m pip install "swage-compiler[pytorch]==0.5.2"
python -m swage.env --json

The release workflow
checked all four wheel ABIs on CPU and qualified the installed CPython 3.13
wheel on RTX A6000.
The assets include checksums, an SPDX SBOM, signed provenance and SBOM
attestations, and the retained CPU and GPU qualification evidence.

Swage remains pre-alpha. The Python kernel language admits the documented
fixed vector bodies; caller-written segment programs remain planned.
Persistent scheduling remains private and experimental. The supported
release GPU qualification device is RTX A6000 (sm_86). See the
support matrix and
changelog for the full compatibility boundary.

Swage v0.5.1

Choose a tag to compare

@abhiksark abhiksark released this 24 Aug 12:20
eac03c8

Swage v0.5.1 is a fix-forward release for the documentation and package presentation added after v0.5.0.

Highlights:

  • Adds the private M8 split-CTA correctness path for oversized canonical identity-sum segments.
  • Adds Read the Docs configuration with pinned MkDocs dependencies.
  • Adds the Swage logo and favicon to the README and hosted documentation.
  • Publishes the branded README as the PyPI project description.

The public execution boundary is unchanged. Public segmented syntax and segmented launch remain planned.

Install the Python package with:

python -m pip install swage-compiler==0.5.1

The wheel contains the importable swage package. Native mlir_swage bindings, emit_mlir(), and execution require a bindings-enabled source build.

Full changelog: v0.5.0...v0.5.1

Swage v0.5.0

Choose a tag to compare

@abhiksark abhiksark released this 24 Aug 08:41
2b90b7f

Swage v0.5.0 publishes the importable swage Python package and completes the internally qualified M1 through M7 compiler/runtime foundation.

Highlights:

  • Python AST-to-MLIR emission and strict build-tree native binding boundary.
  • Fixed vector-add CUDA launch on the current PyTorch stream.
  • Internal segmented reduction, ragged-softmax, host classification, and mixed warp/CTA qualification.
  • Pinned LLVM/MLIR 22.1.8 build and exact-target NVPTX lowering.

The PyPI wheel contains swage only. mlir_swage, emit_mlir(), CUDA execution, and internal segmented qualification require a bindings-enabled native build. This release does not add a public segmented launch API.

See CHANGELOG.md for the complete change history.