SpecMemo prunes speculative-decoding token trees and pre-allocates memory to run multi-turn LLM chatbots on GPUs with as little as 8GB, retaining about 96% of throughput while cutting buffer memory by 65%.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
SpecMemo: Speculative Decoding is in Your Pocket
SpecMemo prunes speculative-decoding token trees and pre-allocates memory to run multi-turn LLM chatbots on GPUs with as little as 8GB, retaining about 96% of throughput while cutting buffer memory by 65%.