Numba-CUDA-MLIR support#

Numbast can generate bindings for both the default Numba-CUDA backend and the experimental numba-cuda-mlir backend. The CUDA C++ parsing workflow is the same for both backends; the differences are the Python import namespace, runtime dependency, and static binding configuration.

Backend differences#

Area	Numba-CUDA	numba-cuda-mlir
Dynamic Numbast imports	Import binding helpers from top-level `numbast`.	Import binding helpers from `numbast.experimental.mlir`.
Static binding generation	Use the default static generator configuration.	Add `MLIR Backend: true` to the static binding config.
Generated module target	Generated bindings target the `numba-cuda` runtime.	Generated bindings target the `numba-cuda-mlir` runtime.
Dependency setup	Install Numbast with the default `numba-cuda` dependency.	For dynamic binding generation, install Numbast with the `mlir` extra so `numba-cuda-mlir` is available.

Dynamic binding generation#

If you intend to use dynamic binding generation with numba-cuda-mlir, install Numbast with the mlir extra:

pip install "numbast[mlir]"

For dynamic binding generation, change imports from top-level numbast to numbast.experimental.mlir. The Numbast helper names remain the same.

Before:

from numbast import bind_cxx_function, bind_cxx_struct, MemoryShimWriter

After:

from numbast.experimental.mlir import (
    bind_cxx_function,
    bind_cxx_struct,
    MemoryShimWriter,
)

Use numba_cuda_mlir types, models, and CUDA entry points with the generated bindings.

from numba_cuda_mlir import cuda, types
from numba_cuda_mlir.models import PrimitiveModel

Static binding generation#

For static binding generation, add the MLIR Backend entry to the same config file you use with the default Numbast CLI.

Entry Point: /path/to/library/header.hpp
File List:
  - /path/to/library/header.hpp
GPU Arch: ["sm_80"]

MLIR Backend: true

Then run the standard Numbast static binding generator:

python -m numbast --cfg-path config.yml --output-dir ./output

With MLIR Backend: true, the CLI routes generation through numbast.experimental.mlir and writes a module for numba-cuda-mlir.

Notes#

The MLIR backend is experimental and is not re-exported from top-level numbast.
The environment that generates or imports MLIR-backed bindings must provide numba_cuda_mlir.
Backend-specific static config keys that mention MLIR require MLIR Backend: true.