GrowPage: On-Demand KV Budgeting for Efficient LLM Reasoning Serving
Qiankun Ma, Yanjiang Zhou, Zinan Xiong, Haofei Wang, Zhen Song, Yang Xiang, Ziyao Zhang, Hairong Zheng
cs.AI
Quick verdict
GrowPage makes per-request KV capacity a runtime decision, gaining ~60% throughput at small accuracy cost on two 8B models; strong ablations, narrow eval.
Summary
GrowPage treats per-request KV-cache capacity as a runtime decision rather than a fixed budget for LLM reasoning serving. At each PagedAttention capacity boundary, it compares Top-p attention working-set sizes induced by two exponentially smoothed query summaries (short- and long-timescale, β=0.9/0.999, maintained pre-RoPE): a larger recent working set triggers allocation of one extra physical page; otherwise historical KV states are compacted within the current allocation using the maximum of both temporal importance scores. A worst-case bound (Theorem 3.1) connects attention concentration to required capacity, and the signal Δt correlates with future demand change (Spearman ρ=0.78) and with counterfactual NLL gains from growth. On a single A100, across five math/code benchmarks and two 8B models, GrowPage raises average decoding throughput 64.3% (DeepSeek-R1-Distill-Llama-8B) and 60.2% (Qwen3-8B) over full-KV serving while average pass@1 drops modestly (68.7 vs 69.2; 87.5 vs 88.7), and it beats fixed-budget Zipage on both axes; a frequency-matched random-policy ablation confirms the gains come from signal timing. Limitations include only two 8B dense models, math/code tasks alone, no confidence intervals on small benchmarks (AIME24 has 30 questions), an untested prefix-caching claim, and no released code.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Minimum sufficient KV budgets vary widely across requests and attention demand fluctuates within decoding, so any fixed per-request budget trades over-provisioning against loss of reasoning-critical history (Section 2, Figs 2–3).
- The dual-timescale working-set signal Δt predicts future demand change (Spearman ρ=0.78, 80.3% sign agreement, Fig 6) and the counterfactual NLL benefit of capacity growth (Fig 7).
- GrowPage improves average decoding throughput by 64.3%/60.2% over full-KV serving on DeepSeek-R1-Distill-Llama-8B/Qwen3-8B at modest accuracy cost, and beats fixed-budget Zipage on both accuracy and throughput (Table 1, Figs 5/10).
- The benefit stems from demand-guided timing of capacity decisions rather than compression frequency: a frequency-matched random policy loses 3.1 pass@1 points (Tables 8–9).
Strengths
Well-identified problem with quantitative grounding
Section 2 measures minimum sufficient budgets per request (Fig 2: distribution spans 512→8192+ tokens on Qwen3-8B, 16.3% unsolved) and intra-request sparsity swings (Fig 3), turning the motivation into data rather than anecdote.
Frequency-matched causal ablations
GROWPAGE-RANDOM matches the grow ratio (30.4% vs 30.9%, Table 9) yet scores 88.3% vs 91.4% pass@1 (Table 8); INVERSE and SHRINK isolate signal direction and the cost of irreversible shrinking — the cleanest evidence the gain comes from decision timing.
Counterfactual mechanism validation
Genuine system co-design
Pareto evaluation under heterogeneous load
Weaknesses
No uncertainty quantification
Average accuracy sits below full KV
Narrow evaluation scope
Baseline throughput collapse under-explained
Theory-practice gap
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
GrowPage (arXiv 2609.03494) is a method-plus-systems paper on KV-cache capacity management for reasoning-LLM serving. Its three central claims are clearly stated and mostly supported: (1) fixed per-request KV budgets mismatch demand both across requests — minimum sufficient budgets B*i span 512→8192 tokens with 16.3% unsolved on Qwen3-8B over AIME24/AMC23/LiveCode (Fig 2) — and within a request, where attention sparsity oscillates between concentrated and diffuse regimes (Fig 3); (2) a dual-timescale query-summary signal Δt tracks demand evolution (Spearman ρ=0.78, 80.3% sign agreement, Fig 6) and predicts when growth helps (Fig 7); (3) page-granular grow/compress decisions integrated with PagedAttention lift throughput ~60% at small accuracy cost and dominate fixed-budget Zipage on both axes.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is GrowPage and what problem does it solve?
GrowPage is a KV-cache management framework for serving reasoning LLMs. Instead of giving every request a fixed KV memory budget chosen in advance, it treats capacity as a runtime resource: at each PagedAttention capacity boundary it decides, from a lightweight attention-demand signal, whether to compress existing history within the current allocation or acquire one additional physical page.
How much faster is GrowPage than standard full-KV inference?
On a single A100 across five math/code benchmarks, GrowPage raises average decoding throughput by 64.3% on DeepSeek-R1-Distill-Llama-8B (avg 1472→2417 tokens/s) and 60.2% on Qwen3-8B (1357→2174 tokens/s), while average pass@1 drops slightly (68.7 vs 69.2 and 87.5 vs 88.7 respectively).
How does GrowPage differ from Zipage?
Zipage integrates token-level KV eviction with PagedAttention but keeps a fixed per-request capacity chosen before decoding. GrowPage builds on Zipage's system machinery (asynchronous compression, prefix-cache support) but makes the capacity itself adaptive, choosing per request and per page boundary between compression and growth based on an attention-demand trend signal.
Is GrowPage's code available?
The paper does not mention a code release. It does specify implementation details: built on nano-vLLM with a modified attention kernel, KV block size 256, and hyperparameters βS=0.9, βL=0.999, Top-p threshold 0.99, demand threshold τ=0, and local retention quota 64.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Some figure-internal values (e.g., Fig 5/9 Pareto points, Fig 11 CI error bars) are only partially recoverable from the extraction; quantitative statements about those figures rely on the in-text numbers.
- MorphKV appears in Tables 1/7 but has no entry in the extracted reference list; the omission may be an extraction artifact rather than a paper defect.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.