Haibin's blogPaper Reading
SC25 gLLM
gLLM: Global Balanced Pipeline Parallelism System for Distributed LLM Serving with Token Throttling 几种并行方法 尝试消除泡泡： 目前在LLM推理里有两种不平衡： stage 间不平衡 inter-stage dependency, where a stage cannot begin computation until the pre…
gLLM: Global Balanced Pipeline Parallelism System for Distributed LLM Serving with Token Throttling
几种并行方法
尝试消除泡泡：
目前在LLM推理里有两种不平衡： stage 间不平衡 inter-stage dependency, where a stage cannot begin computation until the preceding stage completes
比如两个GPU在算。例子1的GPU1算了1，然后传到GPU2，然后GPU2就闲置了，等GPU1算完2。例子2 的 GPU2算了1，但是GPU1想做3的话要等1传回来。
batch 间不平衡 where the number of concurrent micro-batches is limited by the pipeline depth.
PP 他们只要传结果数值
If you need to cite this post
Haibin Lai. (July 23, 2025). “SC25 gLLM” [Blog post]. Retrieved from https://www.haibinlaiblog.top/posts/sc25-gllm/
@online{lai-sc25-gllm,
title={SC25 gLLM},
author={Haibin Lai},
year={2025},
url={https://www.haibinlaiblog.top/posts/sc25-gllm/},
}