Module compute 

2026-10-02: The compute unit a kernel runs on: tensor cores (with the MMA atom and the operand format it issues), CUDA cores, or memory (data movement, element-wise work and reductions). Declared per family in KERNEL_FAMILIES.toml (compute, mma), per point where a family’s points differ, and per kernel where one family’s kernels differ; read by the tensor-core policy (hardware::tc_policy) and shown on every planned group.

Owner: metrale-circuit (venn). Invariants:

• Nothing defaults: every family states its unit; tensor_core names its atom, the other units name none.
• A kernel’s unit is its kernel_compute entry, else its family’s. A family whose points state different units names every kernel’s unit, so no kernel’s unit is inferred.

Structs

FamilyCompute

2026-10-02: A family’s units: its own, and the kernels that differ from it.

Enums

ComputeUnit

2026-10-02: Where a kernel’s arithmetic runs.

Functions

group_unit

2026-10-02: The unit a plan group runs its node on, from its kernels’ units: tensor cores when any kernel issues MMAs (the others quantize, stage or reduce around it), else CUDA cores when any kernel computes, else memory. None for a group without kernels.