Ethan @ethansmith2000.com · Nov 28

it's crazy to me that RoPE's issue with BF16 wasn't noticed earlier. For a reasonable N of 2048, these are the computed frequencies prior to cos(x) & sin(x) for fp32 above and bf16 below. Given how short the period is of simple trig functions, this difference is catastrophic for large values.

8 likes 1 replies

?

Replies

Birchlabs · Nov 28

your experiment is showing just the computation of rotation tables. these tables can be re-used, so this only needs to be done once and can be done in higher-precision.