Research map / Architectures and efficiency / RNNs and normalization
CLIMB: Taming the LoRA Residency Cliff in Multi-LoRA Serving
Haoran Zhang, Zhiyu Liang, ZUO Decheng, Hongzhi Wang
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 |
|---|---|---|---|
| add CLIMB ingress admission control | GLOBALFIFO · Cliff (W=8, K=4) | VIP TTFT p99: 38.7 s -> 13.1 s | Improved |
| add CLIMB ingress admission control | GLOBALFIFO · Cliff (W=8, K=4) | VIP engine p99: 38.43 -> 0.13 | Improved |
| add CLIMB ingress admission control | GLOBALFIFO · Cliff (W=8, K=4) | BG TTFT p99: 20.17 -> 40.36 | Hurt |
| add CLIMB ingress admission control | GLOBALFIFO · Cliff (W=8, K=4) | throughput: 10.66 -> 10.66 | No clear effect |
| add CLIMB ingress admission control | GLOBALFIFO · Safe anchor (W=8, K=8) | VIP TTFT p99: 0.13 -> 0.13 | No clear effect |
| add CLIMB admission control (BG liveness cost) | GLOBALFIFO · Cliff (W=8, K=4), reduced-load liveness probe | backlogged wait p99 (overall): 2.73 ms -> 5.44 s | Hurt |
| add CLIMB admission control with shorter BG prompts | GLOBALFIFO · Cliff (W=8, K=4), shorter BG prompts (1024 tokens) | VIP TTFT p99: 42.9 s -> 11.5 s | Improved |
| add CLIMB admission control across generation limits 128/256/512 | GLOBALFIFO · Cliff (W=8, K=4), generation limits | VIP TTFT p99: 36.5-48.6 s -> 16.7-18.4 s | Improved |
| add CLIMB admission control with other VIP/BG mixtures | GLOBALFIFO · Cliff (W=8, K=4), VIP/BG mixtures | VIP TTFT p99: 16.9-20.3 s -> 7.85-10.3 s | Improved |
| make VIP and BG prompt lengths symmetric (ablation) | CLIMB · Cliff (W=8, K=4), symmetric prompts | VIP TTFT p99 gap | Hurt |
| replace CLIMB with reactive LRU locality gating (ablation) | LRUGATE · Cliff (W=8, K=4) | VIP TTFT p99: 13.69 -> 40.96 | Hurt |
| replace CLIMB with static BG cap (ablation) | BGCAP · Cliff (W=8, K=4) | throughput: 10.66 -> 6.89 | Hurt |
Closest papers by meaning
- Block-Diagonal LoRA for Eliminating Communication Overhead in Tensor Parallel LoRA Serving (NeurIPS 2025 · similarity 0.86)
- ConServe: Fine-Grained GPU Harvesting for LLM Online and Offline Co-Serving (ICML 2026 · similarity 0.85)
- From Weight-Based to State-Based Fine-Tuning: Further Memory Reduction on LoRA with Parallel Control (ICML 2025 · similarity 0.83)
- DéjàVu: KV-cache Streaming for Fast, Fault-tolerant Generative LLM Serving (ICML 2024 · similarity 0.82)
- MAPS: Memory-Aware Predictive Scheduling Framework for Large Language Model Serving (ICML 2026 · similarity 0.82)
- Tackling the Data-Parallel Load Balancing Bottleneck in LLM Serving: Practical Online Routing at Scale (NeurIPS 2026 · similarity 0.81)
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.
