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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| Use HexGen-2 instead of HexGen | Heterogeneous LLM serving · Heterogeneous settings, multiple workloads | serving throughput: up to a 1.5× and, on average, a 1.4× increase | Improved |
| Use HexGen-2 instead of DistServe | LLM serving · Homogeneous setting, multiple workloads | serving throughput: up to a 2× and, on average, a 1.3× higher serving throughput | Improved |
| Use HexGen-2 instead of state-of-the-art counterparts | LLM serving · Various evaluation settings | average inference latency: 1.5× lower latency | Improved |
| Use HexGen-2 at 70% of the homogeneous budget | Heterogeneous LLM serving · Heterogeneous Setting 5 vs. homogeneous setting | serving performance: 70% of the budget; outperforms it by 30% in some specific workloads | No clear effect |
| Use HexGen-2 instead of vLLM | Llama-2-70B serving · Homogeneous 8 H100 setup | serving throughput: up to a 2.1× and on average a 1.5× higher serving throughput | Improved |
| Add max-flow-guided edge swaps instead of random swaps (ablation) | HexGen-2 scheduling algorithm · Heterogeneous Setting 1, four workloads | serving throughput: 1.8× higher serving throughput | Improved |
| Replace max-flow-guided edge swaps with random swaps (ablation) | HexGen-2 scheduling algorithm · Heterogeneous Setting 1, four workloads | convergence speed: 90 to 120 seconds | Hurt |
| Replace graph partitioning and iterative refinement with a genetic algorithm (ablation) | HexGen-2 scheduling algorithm · Heterogeneous Setting 1, four workloads | serving throughput: 1.8× higher serving throughput | Hurt |
| Use chunked prefill | vLLM on OPT-30B, one H100 · HPLD and LPLD | throughput: approximately 20% | Improved |
| Use chunked prefill | vLLM on OPT-30B, one H100 · HPHD and LPHD | throughput: around 5% | Improved |
| Use HexGen-2 instead of DistServe | OPT-30B serving on a homogeneous cluster · HPHD | throughput: 683 tokens/s vs. 692 tokens/s | Hurt |
| Use HexGen-2 instead of DistServe | OPT-30B serving on a homogeneous cluster · LPHD | throughput: 758 tokens/s vs. 774 tokens/s | Hurt |
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.
- HexGen — HexGen: Generative Inference of Large Language Model over Heterogeneous Environment (ICML 2024) · better on 5 of 5 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
- HexGen: Generative Inference of Large Language Model over Heterogeneous Environment (ICML 2024 · similarity 0.94)
- A Queueing-Theoretic Framework for Stability Analysis of LLM Inference with KV Cache Memory Constraints (ICML 2026 · similarity 0.89)
- Efficient Multi-round LLM Inference over Disaggregated Serving (ICML 2026 · similarity 0.89)
- Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads (ICML 2024 · similarity 0.88)
- SpecExec: Massively Parallel Speculative Decoding For Interactive LLM Inference on Consumer Devices (NeurIPS 2024 · similarity 0.88)
- SqueezeLLM: Dense-and-Sparse Quantization (ICML 2024 · similarity 0.88)
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.
