pub fn exec_of(device: &Device, weight: Format, activation: Format) -> ExecExpand description
2026-09-30: How device executes a node of weight reading activation. NVFP4 weights
under FP8 activations (W4A8 with group-16 E4M3 scales) have no single-instruction MMA on any
device unless native_mma says so (fp4_fp8_nvfp4_scaled); they take the exact E4M3 path.