# Token-Budget-Aware Pool Routing for Cost-Efficient LLM Inference

arXiv: https://arxiv.org/abs/2604.09613 · HTML: https://mlsanity.com/paper/2604.09613/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Serving systems and request scheduling, one of 12 directions in Efficient LLM inference and serving.
- Builds on 19 of its 27 references that are in the map.
- First posted on arXiv 13 Mar 2026.

## Placed in

- Efficient LLM inference and serving / [Serving systems and request scheduling](https://mlsanity.com/inference/serving-systems-and-request-scheduling/)

## Builds on (most cited first, 19 of 19)

- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [Efficient Memory Management for Large Language Model Serving with PagedAttention](https://mlsanity.com/paper/2309.06180/) (2023)
- [FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU](https://mlsanity.com/paper/2303.06865/) (2023)
- [Splitwise: Efficient generative LLM inference using phase splitting](https://mlsanity.com/paper/2311.18677/) (2023)
- [Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve](https://mlsanity.com/paper/2403.02310/) (2024)
- [LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset](https://mlsanity.com/paper/2309.11998/) (2023)
- [Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving](https://mlsanity.com/paper/2407.00079/) (2024)
- [Compressing Context to Enhance Inference Efficiency of Large Language Models](https://mlsanity.com/paper/2310.06201/) (2023)
- [Llumnix: Dynamic Scheduling for Large Language Model Serving](https://mlsanity.com/paper/2406.03243/) (2024)
- [Vidur: A Large-Scale Simulation Framework For LLM Inference](https://mlsanity.com/paper/2405.05465/) (2024)
- [BurstGPT: A Real-world Workload Dataset to Optimize LLM Serving Systems](https://mlsanity.com/paper/2401.17644/) (2024)
- [AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving](https://mlsanity.com/paper/2302.11665/) (2023)
- [ServeGen: Workload Characterization and Generation of Large Language Model Serving in Production](https://mlsanity.com/paper/2505.09999/) (2025)
- [Demystifying Cost-Efficiency in LLM Serving over Heterogeneous GPUs](https://mlsanity.com/paper/2502.00722/) (2025)
- [KVShare: An LLM Service System with Efficient and Effective Multi-Tenant KV Cache Reuse](https://mlsanity.com/paper/2503.16525/) (2025)
- [LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models](https://mlsanity.com/paper/2507.14204/) (2025)
- [OrbitFlow: SLO-Aware Long-Context LLM Serving with Fine-Grained KV Cache Reconfiguration](https://mlsanity.com/paper/2601.10729/) (2026)
- [The Script Tax: Measuring Tokenization-Driven Efficiency and Latency Disparities in Multilingual Language Models](https://mlsanity.com/paper/2602.11174/) (2026)
- [EWSJF: An Adaptive Scheduler with Hybrid Partitioning for Mixed-Workload LLM Inference](https://mlsanity.com/paper/2601.21758/) (2026)

## Built on by (most cited first, 2 of 2)

- [FleetOpt: Analytical Fleet Provisioning for LLM Inference with Compress-and-Route as Implementation Mechanism](https://mlsanity.com/paper/2603.16514/) (2026)
- [The Workload-Router-Pool Architecture for LLM Inference Optimization: A Vision Paper from the vLLM Semantic Router Project](https://mlsanity.com/paper/2603.21354/) (2026)

## Abstract (arXiv)

Production vLLM fleets provision every instance for worst-case context length, wasting 4-8x concurrency on the 80-95% of requests that are short and simultaneously triggering KV-cache failures -- OOM crashes, preemption storms, and request rejections. Both problems share a single root cause: configuration-traffic mismatch. We propose token-budget-aware pool routing: estimate each request's total token budget using a self-calibrating per-category bytes-per-token ratio, then dispatch it to one of two vLLM pools -- a high-throughput short pool or a high-capacity long pool -- each right-sized for its workload class. The ratio is learned online via exponential moving average from usage.prompttokens feedback, requiring no tokenizer. A closed-form cost model, savings = alpha * (1 - 1/rho), predicts fleet-level GPU savings from two observable quantities: the short-traffic fraction alpha and the throughput gain ratio rho. On traces from the Azure LLM Inference Dataset and LMSYS-Chat-1M serving Llama-3-70B on A100 GPUs, token-budget routing reduces GPU instances by 17-39% (1.2-2.0M/yr at 1,000 req/s), with savings verified by a self-contained discrete-event simulator. A case study projecting Qwen3-235B-A22B on AMD MI300X at 10,000 req/s shows 15.4M/yr in savings. The algorithm adds O(1) dispatch overhead, self-calibrates across content types without a tokenizer, and composes with PagedAttention, continuous batching, and prefill-decode disaggregation.