FastServe adds token-level preemption and a skip-join MLFQ scheduler to LLM serving, delivering up to 31.4x higher throughput than vLLM at equivalent average and tail latency.
Tiresias: A gpu clus- ter manager for distributed deep learning.,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2023 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Fast Distributed Inference Serving for Large Language Models
FastServe adds token-level preemption and a skip-join MLFQ scheduler to LLM serving, delivering up to 31.4x higher throughput than vLLM at equivalent average and tail latency.