AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

HexGen-2: Disaggregated Generative Inference of LLMs in Heterogeneous Environment

YOUHE JIANG, Ran Yan, Binhang Yuan

ICLR 2025 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding · 19 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Use HexGen-2 instead of HexGenHeterogeneous LLM serving · Heterogeneous settings, multiple workloadsserving throughput: up to a 1.5× and, on average, a 1.4× increaseImproved
Use HexGen-2 instead of DistServeLLM serving · Homogeneous setting, multiple workloadsserving throughput: up to a 2× and, on average, a 1.3× higher serving throughputImproved
Use HexGen-2 instead of state-of-the-art counterpartsLLM serving · Various evaluation settingsaverage inference latency: 1.5× lower latencyImproved
Use HexGen-2 at 70% of the homogeneous budgetHeterogeneous LLM serving · Heterogeneous Setting 5 vs. homogeneous settingserving performance: 70% of the budget; outperforms it by 30% in some specific workloadsNo clear effect
Use HexGen-2 instead of vLLMLlama-2-70B serving · Homogeneous 8 H100 setupserving throughput: up to a 2.1× and on average a 1.5× higher serving throughputImproved
Add max-flow-guided edge swaps instead of random swaps (ablation)HexGen-2 scheduling algorithm · Heterogeneous Setting 1, four workloadsserving throughput: 1.8× higher serving throughputImproved
Replace max-flow-guided edge swaps with random swaps (ablation)HexGen-2 scheduling algorithm · Heterogeneous Setting 1, four workloadsconvergence speed: 90 to 120 secondsHurt
Replace graph partitioning and iterative refinement with a genetic algorithm (ablation)HexGen-2 scheduling algorithm · Heterogeneous Setting 1, four workloadsserving throughput: 1.8× higher serving throughputHurt
Use chunked prefillvLLM on OPT-30B, one H100 · HPLD and LPLDthroughput: approximately 20%Improved
Use chunked prefillvLLM on OPT-30B, one H100 · HPHD and LPHDthroughput: around 5%Improved
Use HexGen-2 instead of DistServeOPT-30B serving on a homogeneous cluster · HPHDthroughput: 683 tokens/s vs. 692 tokens/sHurt
Use HexGen-2 instead of DistServeOPT-30B serving on a homogeneous cluster · LPHDthroughput: 758 tokens/s vs. 774 tokens/sHurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy