Research map / Architectures and efficiency / RNNs and normalization
Threshold-Based Exclusive Batching for LLM Inference
Weifang Zhang, Yuzhou Nie, Bowen Pang, Guangrui Ma, Shining Wu
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 optimized exclusive batching instead of mixed batching on bandwidth-constrained GPU | vLLM v1 / Qwen3-8B · WildChat on L40S | RPS: 10.36 -> 14.70 | Improved |
| use exclusive batching instead of mixed batching on high-bandwidth GPU | vLLM v1 / Qwen3-8B · WildChat on B300 | RPS: 50.47 -> 52.34 | No clear effect |
| use EB(k*) instead of v1 on RTX PRO 6000 | Qwen3-8B · real-world workloads | average RPS gain: +7.9% | Improved |
| use EB(k*) instead of v1 on RTX PRO 6000 | Qwen3-30B-A3B · real-world workloads | average RPS gain: +1.4% | Mixed |
| use EB(k*) instead of v1 on H200 | Qwen3-30B-A3B · real-world workloads | average RPS gain: -2.9% | Hurt |
| scale model from Qwen3-8B to Qwen3-30B-A3B | EB gain over v1 · RTX PRO 6000 | average RPS gain: 7.9% -> 1.4% | Hurt |
| use closed-form EB(k*) instead of best fixed-k sweep (ablation) | exclusive batching · H200 decode-heavy, N=1024 | throughput: +8.0% | Improved |
| use closed-form EB(k*) instead of best fixed-k sweep (ablation) | exclusive batching · H200 balanced, N=1024 | throughput: +3.6% | Improved |
| use closed-form EB(k*) instead of best fixed-k sweep (ablation) | exclusive batching · H200 prefill-heavy, N=1024 | throughput: +0.6% | No clear effect |
| add online hybrid EB+ mode switching over v1 | scheduler · RTX PRO 6000 distribution shift | throughput: +36.4% | Improved |
| add online hybrid EB+ mode switching over v1 | scheduler · H200 distribution shift | throughput: +3.9% | Improved |
| add online hybrid EB+ mode switching over v1 | scheduler · RTX PRO 6000 concurrency shift | throughput: +22.6% | Improved |
Closest papers by meaning
- Progressive Mixed-Precision Decoding for Efficient LLM Inference (ICLR 2025 · similarity 0.91)
- Break the Sequential Dependency of LLM Inference Using Lookahead Decoding (ICML 2024 · similarity 0.90)
- Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios (AAAI 2026 · similarity 0.89)
- Dynamic-Width Speculative Beam Decoding for LLM Inference (AAAI 2025 · similarity 0.89)
- Efficient Multi-round LLM Inference over Disaggregated Serving (ICML 2026 · similarity 0.89)
- Beyond Prediction: Tail-Aware Scheduling for LLM Inference (ICML 2026 · similarity 0.89)
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.
