Currently I observe some inconsistency for fp8 and int4. - binary / unary / computation ops do not support either of them - Some ops (flatten, reshape etc) support int4 but not fp8 - Allocation ops (full, arange) support both (well it's because we cast in the end)