karpathy/llama2.c

llama2.c calculation in FP16

Opened this issue · 0 comments

ztachip commented

Is it possible to replace all calculation in float32 with float16?
How would this impact accuracy? Is this attempted before?