knowhere
knowhere copied to clipboard
Minimize the error when converting float32 to bf16/fp16 by using rounding rather than truncation