Files
composable_kernel/include
Anton Gorenko 335f80033b Impove precision of CShuffle with scales or multi D
Two new template parameters are introduced:
 * CShuffleDataType allows to apply multiple Ds before downcasting to
 ODataType (prevents unexpected precision loss and/or overflow);
 * CComputeDataType to use scales with int32 AccDataType (int8 gemms);
2026-06-16 11:41:20 +05:00
..