Re:Estimating required GPU memory for serving LLMs
by brianjking on Sunday November 19, 2023 @01:08PM
(#38335646)
My issue is figuring out how to identify how many concurrent users you can support on average on a given GPU.
Understanding the vram to simply load the weights is easy enough. When you are allowing for something like content generation with varying lengths of input/output tokens, how do you even begin to identify the GPUs you need?
Wrote a blog post to demystify the process of GPU memory usage estimating.