(26.02.2024, 09:08)Daniel Wrote: This is how it meant to work. You need to convert it back to float on master device.
mb.setfloat16precision(2) sets how float16 is converted to integer. By default it's 2 decimals (0.01 precision). Change 2 to 0 to convert to integer without any decimals.
Hi! Yes, it works. Not a big problem to convert back to float on master device, but just to clarify. Is it possible to make pure float register (either it can be little or big endian)?