Kernel card

Supported operators

The callable functions and dimensional contracts are listed in the README. Weights and FP8 inputs use contiguous E4M3FN (N,K) storage; biases, gates, residuals and outputs are contiguous BF16; channel inverse-scale tensors are also BF16, matching the original FlashRT calibration contract. Scale floats follow the FlashRT static per-tensor calibration contract.

These are region megakernels, not arbitrary-shape linear layers. A mismatched dimension, dtype, layout, capacity or device raises an error. No silent eager fallback is present. SM110 uses standard FP8 MMA and explicit ordered launches to avoid the SM120 software-grid-barrier residency assumption. SM120 retains the original tuned path. Inference only.