---

CubeK: high-performance multi-platform kernels in CubeCL

Algorithms

Algorithms	Variants
Random	bernoulli normal uniform
Quantization	symmetric per-block per-tensor q2 q4 q8 fp4
Reduction	mean sum prod max min arg[max|min] per-cube per-plane
Matmul	mma unit tma multi-stage specialization ordered multi-rows
Convolution	mma unit tma multi-stage im2col
Attention	mma unit multi-rows

Runtime configuration

CubeK loads runtime settings from cubek.toml or CubeK.toml in the current directory or a parent directory. Missing settings use defaults.

To disable automatic F32-to-TF32 conversion for matrix multiplication:

[matmul]
allow_tf32 = false

allow_tf32 defaults to true. Disabling it preserves F32 operands during automatic selection and may reduce performance.

Configuration is loaded on first use. Restart the application after editing the file; no recompilation is needed.

Contributing

If you want to contribute new kernels, please read the GUIDE.md.

Running tests

The full testing guide — suites, CUBE_TEST_MODE, failure-message format, and filter syntax — lives in cubek-test-utils.