CubeK Reduce

Implements a wide variety of reduction algorithms across multiple instruction sets and hardware targets for efficient tensor reduction.

Running Tests

Important Environment Variables

Test behavior is controlled by the shared CUBE_TEST_MODE env var (see cubek-test-utils).

• CUBE_TEST_MODE=Correct (default): numerical errors fail the test; compilation / hardware-incompatibility errors are accepted.
• CUBE_TEST_MODE=Strict: both numerical and compilation errors fail the test. Useful to surface tests that are silently skipped on your hardware.
• CUBE_TEST_MODE=PrintAll[:<filter>] / PrintFail[:<filter>]: print tensor elements; see cubek-test-utils docs.

Important Feature Flags

• heavy: enables the Cube reduction-routine strategy tests. They are part of the base suite on GPU backends (the cargo test-cuda and friends aliases pass heavy), but too slow to run on CPU.
• extended: implies heavy, and adds the full breadth of the reduce_dim matrix: both dtypes, both output-vectorization settings, both forced plane dims, every shape configuration and the whole k sweep. Meant for large-coverage runs (before a release, for instance), so the base suite keeps one point per collapsed axis and one shape per layout class. CI never compiles the extended tier while testing, so cargo clippy -p cubek-reduce --tests --features extended guards it separately.
• full: implies extended, and enables the benchmark catalogue.

Examples

# Light suite on CUDA
cargo test --features cubecl/cuda

# Base GPU suite: light matrix plus the Cube-strategy tests
cargo test --features cubecl/cuda,heavy

# Full `reduce_dim` matrix
cargo test --features cubecl/cuda,extended

# Fail on any silently-skipped tests
CUBE_TEST_MODE=Strict cargo test --features cubecl/cuda,extended