Communication volume and latency of LLM inference are characterized and modeled for tensor, pipeline, and hybrid parallelism, showing that decode-stage operations dominate inter-GPU traffic.
GPT-4 Technical Report,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.DC 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Characterizing Communication Patterns in Distributed Large Language Model Inference
Communication volume and latency of LLM inference are characterized and modeled for tensor, pipeline, and hybrid parallelism, showing that decode-stage operations dominate inter-GPU traffic.