SiPipe claims 1.6x to 2.1x throughput gains over vLLM for pipeline-parallel LLM inference by using idle CPUs for sampling, double-buffered CUDA graphs, and structure-aware transmission.
Llama 2 follow-up: too much RLHF, GPU sizing, technical details
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.DC 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
SiPipe: Bridging the CPU-GPU Utilization Gap for Efficient Pipeline-Parallel LLM Inference
SiPipe claims 1.6x to 2.1x throughput gains over vLLM for pipeline-parallel LLM inference by using idle CPUs for sampling, double-buffered CUDA graphs, and structure-aware transmission.