app.py

import streamlit as st

def number_field(label, **args):
  c1, c2 = st.columns([2, 4])
  c1.write(label)
  
  return c2.number_input('', **args)

TFLOPS = 312e12
GB_S = 1935e9

st.header("Transformer parameters")
col1, col2 = st.columns([2, 4])

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*1*d*3*d
mha_bytes = 2*bs*1*d + 2*3*d*d + 2*bs*1*3*d
mha_int = mha_flop/mha_bytes
mha_time = (mha_flop/TFLOPS + mha_bytes/GB_S)*1000

mha_flop = round(mha_flop/1e9, 2)
mha_bytes = round(mha_bytes/1e6, 2)


st.subheader("Multi-Head Attention")
c1, c2 = st.columns([2, 3])
c1.write("GFLOP:")
c2.write(str(mha_flop))
c1.write("MB: ")
c2.write(str(mha_bytes))
c1.write("Arithm. intensity:")
c2.write(str(mha_int))
c1.write("Time (ms):")
c2.write(str(mha_time))


mqa_flop = 2*bs*1*d*(1+2/h)*d
mqa_bytes = 2*bs*1*d + 2*(2/h)*d*d + 2*bs*1*(2/h)*d
mqa_intensity = mqa_flop/mqa_bytes
mqa_time = (mqa_flop/TFLOPS + mqa_bytes/GB_S)*1000

mqa_flop = round(mqa_flop/1e9, 2)
mqa_bytes = round(mqa_bytes/1e6, 2)

st.subheader("Multi-Query Attention")
c1, c2 = st.columns([2, 3])
c1.write("GFLOP:")
c2.write(str(mqa_flop))
c1.write("MB:")
c2.write(str(mqa_bytes))
c1.write("Arithm. intensity:")
c2.write(str(mqa_intensity))
c1.write("Time (ms):")
c2.write(str(mqa_time))

st.header('Attention')
mha_flop = 2*bs*h*(d/h)*n
mha_bytes = 2*bs*h*(d/h) + 2*bs*h*n*(d/h) + 2*bs*h*n
mha_int = mha_flop/mha_bytes
mha_time = (mha_flop/TFLOPS + mha_bytes/GB_S)*1000

mha_flop = round(mha_flop/1e9, 2)
mha_bytes = round(mha_bytes/1e6, 2)


st.subheader("Multi-Head Attention")
c1, c2 = st.columns([2, 3])
c1.write("GFLOP:")
c2.write(str(mha_flop))
c1.write("MB: ")
c2.write(str(mha_bytes))
c1.write("Arithm. intensity:")
c2.write(str(mha_int))
c1.write("Time (ms):")
c2.write(str(mha_time))

mqa_flop = 2*bs*h*(d/h)*n
mqa_bytes = 2*bs*h*(d/h) + 2*bs*n*(d/h) + 2*bs*h*n
mqa_intensity = mqa_flop/mqa_bytes
mqa_time = (mqa_flop/TFLOPS + mqa_bytes/GB_S)*1000

mqa_flop = round(mqa_flop/1e9, 2)
mqa_bytes = round(mqa_bytes/1e6, 2)

st.subheader("Multi-Query Attention")
c1, c2 = st.columns([2, 3])
c1.write("GFLOP:")
c2.write(str(mqa_flop))
c1.write("MB:")
c2.write(str(mqa_bytes))
c1.write("Arithm. intensity:")
c2.write(str(mqa_intensity))
c1.write("Time (ms):")
c2.write(str(mqa_time))