NVIDIA/numba-cuda-mlir v0.4.0

v0.4.0

Repository: NVIDIA/numba-cuda-mlir

Tag: v0.4.0

Published: 2026-06-15T16:24:32Z

Prerelease: no

Release notes: This first update focuses on platform support, debugging, ecosystem enablement, performance, and broader CUDA Python compatibility.

Highlights

• Added Support for Windows, and integrated Windows tests into CI.
• Added CUDA-gdb CI workflow and debugging support validation.
• Added experimental third-party ecosystem coverage for nvmath-python, RAPIDS/cuDF, and numbast extension backends.
• Improved warm compile-time performance by redesigning extension registry refresh behavior, delivering an additional ~40% speedup over the previous implementation on our benchmark suite and reaching ~1.8x geomean speedup on warm compile-time vs. numba-cuda.

Platform and Tooling Support

• Introduced Windows CI coverage and related build fixes, including static CRT usage.
• Added CUDA-gdb workflow coverage to validate debugging behavior.
• Improved compatibility with newer libc++ versions.
• Removed the implicit nvjitlink dependency derived from cudatoolkit.

Performance and Compilation

• Replaced implicit context refresh with explicit initialization and version-tracked registries, reducing warm compile overhead.
• Optimized CUDA Array Interface launch caching.
• Avoided finalizing internal device callees during compilation.
• Added user-controlled handling for LTOIR linker optimization disabling instead of unconditionally disabling it.

CUDA Python Compatibility Improvements

• Added full array.view() support, including dtype bitwidth changes.
• Added support for vector types in local and shared memory.
• Added CUDA vector / scalar operations and vector-to-complex conversions.
• Added support for custom dtypes.
• Added complex constructor support, including complex32.
• Added support for complex CPointer getitem/setitem lowering.
• Added support for NamedTuple usage in kernels.
• Improved support for array slicing and shared-memory views.

Lowering and Type System Fixes

• Unified vector type handling by replacing VectorTypeStub with VectorType / VectorTypeClass.
• Introduced a value/storage data model to fix float16 and bool memory representation issues.
• Fixed lowering for defaults, tuples, dtype tokens, heterogeneous tuple assignment, optional values, and string constant folding.
• Fixed array.real / array.imag on shared-memory arrays preserving address space.
• Fixed VectorType to complex setitem behavior.
• Fixed to_numba_type handling for NumPy dtypes.

Ecosystem and Extension Support

• Enabled extension linkage in MLIR lowerings.
• Added Extension API documentation.
• Added Numbast MLIR source CI tests.
• Added experimental cuDF / RAPIDS third-party test coverage, including use of pylibcudf from the active conda environment.
• Prevented unintended invocation of the Numba-CUDA JIT and addressed resulting issues.

Documentation and Maintenance

• Updated reference documentation.
• Added PR documentation preview infrastructure.
• Fixed PyPI-hosted README links.
• Removed outdated conda install documentation.
• Removed legacy @intrinsic implementations.
• Removed dead NRT C++ code.
• Removed cudasim support.
• Removed unnecessary packaging dependency from numba_cuda._compat.

Bug Fixes

• Fixed ICE for raise-only kernels.
• Fixed shared-memory view behavior with None starts.
• Fixed array slicing issues.
• Fixed multiple lowering edge cases involving tuples, optionals, constants, and complex/vector interactions.
• Fixed cuDF CI and Numbast CI issues.