CubeK: high-performance multi-platform kernels in CubeCL
CubeK provides reusable compute kernels written in Rust with CubeCL. It powers core operations in Burn and can also be used directly in CubeCL applications, with specialized implementations for matrix multiplication, convolution, attention, reductions, and more.
| 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 |
| Attention backward | mma unit plane cube |
| Gated delta rule | cube walk |
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.
If you want to contribute new kernels, please read the GUIDE.md.
The full testing guide — suites, CUBE_TEST_MODE, failure-message format, and filter syntax — lives
in cubek-test-utils.