홍석쓰 블로그

Jitendra lunch group meeting; source: (Israel / The Technion)Probability Theory 101What is the "convolution property of a PDF (probability density function)"?For two independent random variables, the PDF of their sum is the convolution of their individual PDFs. What’s the difference between "uncorrelation" and "independence" in expectation?x1​ and x2​ are uncorrelated if and only if the expectat..

What is KV caching? KV caching is specifically related to the auto-regressive approach of a transformer decoder. In a transformer decoder, it attends to the past and current tokens, but not to future tokens. At each time step, the transformer repeatedly calculates the attention scores between the query and the key, and computes the values by multiplying the scores with the previously computed va..