import streamlit as st def number_field(label, columns=None, **input_params): c1, c2 = st.beta_columns(columns or [1, 4]) # Display field name with some alignment c1.markdown("##") c1.markdown(label) # Sets a default key parameter to avoid duplicate key errors input_params.setdefault("key", label) # Forward text input parameters return c2.number_input("", **input_params) def key_value(key, value, columns=None): c1, c2 = st.beta_columns(columns or [2, 3]) # Display field name with some alignment c1.markdown("##") c1.markdown(key) c2.markdown("##") c2.markdown(value) st.header("Transformer parameters") bs = number_field('Batch size: ', value=10) h = number_field('Num heads: ', value=16) d = number_field('Dimension: ', value=768) n = number_field('Seq length: ', value=1024) st.header('Query, Key, Value projection') mha_flop = 2*bs*n*d*3*d mha_bytes = 2*bs*n*d + 2*3*d*d + 2*bs*n*3*d st.subheader("Multi-query Attention") key_value("FLOP: ", str(mha_flop)) key_value("bytes: ", str(mha_bytes)) key_value("Arithm. intensity:", str(mha_flop/mha_bytes)) mqa_flop = 2*bs*n*d*(1+2/h)*d mqa_bytes = 2*bs*n*d + 2*(2/h)*d*d + 2*bs*n*(2/h)*d st.subheader("Multi-query Attention") key_value("FLOP: ", str(mqa_flop)) key_value("bytes: ", str(mqa_bytes)) key_value("Arithm. intensity:", str(mqa_flop/mqa_bytes)) st.header('Attention')