Please enable JavaScript.
Coggle requires JavaScript to display documents.
LLM Systems Seminar (CS7670): Week03a, vLLM (paper, hotcrp, lottery) -…
LLM Systems Seminar (CS7670): Week03a, vLLM (paper, hotcrp, lottery)
1. LLM serving basics
-
-
-
-
decoding strategies
motivation
common decoding strategy
-
- top-p decoding (Nucleus sampling)
-
-
-
-
-
2. Debate question
vLLM is strongly inspired by operating system paging, yet it has had a very large impact in LLM serving (vLLM citations: 3079, memory ballooning, citations: 2203).
How should we evaluate such a paper that adopts an existing idea from another domain and applies it successfully in a new context?
Pro side: Adapting established ideas to a new setting demonstrates the value of cross-domain thinking and can generate high practical impact.
Con side: Reusing established ideas limits originality and risks being seen more as engineering application than genuine research innovation.