Inyigisho igereranya kongeramo GPU hamwe no guhagarika ububiko bwa LLM kugirango buhendutse
Urupapuro rushya rwa arXiv rugereranya tensor parallelism hamwe na KV-cache compression ya moderi nini yerekana ururimi runini ikora, ivuga ko compression yarihendutse inshuro 1.20 kugeza kuri 2.00 mugihe cyageragejwe, mugihe GPU yongeyeho aribwo buryo bwonyine bwageragejwe bwagabanije ubukererwe.