llama2.c
llama2.c copied to clipboard
llama2.c calculation in FP16
Is it possible to replace all calculation in float32 with float16? How would this impact accuracy? Is this attempted before?