TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

StreamingLLM: tiny tweak to KV LRU improves long conversations

91 pointsby lucasluitjesover 1 year ago

4 comments

TrueDualityover 1 year ago
There was a really interesting post a while ago about adjusting the softmax function to allow attention heads to not make a choice (<a href="https:&#x2F;&#x2F;www.evanmiller.org&#x2F;attention-is-off-by-one.html" rel="nofollow">https:&#x2F;&#x2F;www.evanmiller.org&#x2F;attention-is-off-by-one.html</a>). It seems like that might remove the need for these attention sinks entirely. I keep meaning to go in and perform tests on this but boy time gets away from you...
评论 #39359975 未加载
评论 #39359798 未加载
popinman322over 1 year ago
Previous discussion, on a link to the implementation: <a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=37740932">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=37740932</a>
Translationautover 1 year ago
This seems only to work cause large GPTs have redundant, undercomplex attentions. See this issue in BertViz about attention in Llama: <a href="https:&#x2F;&#x2F;github.com&#x2F;jessevig&#x2F;bertviz&#x2F;issues&#x2F;128">https:&#x2F;&#x2F;github.com&#x2F;jessevig&#x2F;bertviz&#x2F;issues&#x2F;128</a>
gremlinsincover 1 year ago
I wonder if it could make sense to maybe have break away bots, where at 10k tokens a new one launches with the first 2k, and the last 1k and a table of contents such that when you go back to something you&#x27;re handed off to a model where that data is stronger reinforced or something like that. Sort of like mixture of experts but they&#x27;re only an expert about individual snippets of a long conversational thread.
评论 #39358282 未加载
评论 #39357638 未加载