MiniPIC: Flexible Position-Independent Caching in <100LOC

Nathan Ordonez (1); Thomas Parnell (1) ((1) IBM Research)

arxiv: 2606.13126 · v1 · pith:3TSWZUZFnew · submitted 2026-06-11 · 💻 cs.LG · cs.AI· cs.CL

MiniPIC: Flexible Position-Independent Caching in <100LOC

Nathan Ordonez (1) , Thomas Parnell (1) ((1) IBM Research) This is my paper

Pith reviewed 2026-06-27 07:10 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CL

keywords position-independent cachingprefix cachingKV cachevLLMRoPEinference serverretrieval augmented generationagentic workloads

0 comments

The pith

MiniPIC enables multiple position-independent caching methods in vLLM using under 100 lines of changes via unrotated keys and three primitives.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that a minimal set of changes to vLLM can support flexible position-independent caching for repeated input spans in agentic and retrieval workloads. By storing unrotated K vectors in the KV cache and applying RoPE based on per-request positions during attention, along with three primitives that control cache hashing and attention boundaries, various PIC approaches can be implemented in one system. This matters because it allows cache reuse across non-identical prefixes, integrates with existing offloading, and delivers substantial speedups in prefill and time-to-first-token without heavy modifications to the inference engine.

Core claim

MiniPIC stores unrotated K vectors in the KV cache, applies RoPE to K tiles inside attention using per-request logical positions, and exposes three user-facing and token-level primitives: block-aligned padding, span separator (SSep), and prompt depend (PDep), that modify hashing behavior and effective block-level causal attention structure. With fewer than 100 lines of core-engine changes plus a custom attention backend, these primitives are sufficient to realize multiple PIC methods, including Block-Attention, EPIC, and Prompt Cache, within the same running vLLM instance, while natively integrating with KV cache CPU offload implementations.

What carries the argument

The positional-encoding-free KV cache combined with the three primitives (block-aligned padding, span separator, and prompt depend) that adjust hashing and causal attention structure.

If this is right

Multiple PIC methods including Block-Attention, EPIC, and Prompt Cache can be realized in the same vLLM instance.
Native integration with KV cache CPU offload is maintained.
Prefill throughput improves by 49% over baseline on 2WikiMultihopQA with interleaved scheduling.
Time-to-first-token for cached spans reduces by up to two orders of magnitude.
Linear prefill scaling is preserved for uncached spans with only 5.7% worst-case overhead.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The primitives could be adapted to support dynamic span reuse in long-context agent interactions.
This minimal approach suggests that other inference servers might achieve similar PIC flexibility with comparable code changes.
Workloads involving code files or documents could see further optimizations by tuning the separator and depend primitives.
Integration with CPU offload implies potential for larger cache sizes without GPU memory limits.

Load-bearing premise

Storing unrotated K vectors and applying RoPE to K tiles inside attention using per-request logical positions preserves correct causal attention semantics and enables safe cache reuse without inconsistencies.

What would settle it

A test where two requests share a span but have different logical positions, checking if the attention computation produces identical results to independent computation and that no incorrect cache hits occur.

Figures

Figures reproduced from arXiv: 2606.13126 by Nathan Ordonez (1), Thomas Parnell (1) ((1) IBM Research).

**Figure 2.** Figure 2: Left: Document preload time vs prompt length (no spans). Middle: Document preload time vs document count (1000-token synthetic documents prefilled in parallel). Right: Time-to-first-token (TTFT) vs document count, showing MiniPIC’s significant improvement over SPNL and vanilla vLLM. The small differences between implementations in the left and middle plots indicate that worst-case slowdown is negligible, a… view at source ↗

read the original abstract

Retrieval-augmented and agentic workloads repeatedly prefill recurring predictable structured inputs (which we call "spans") such as documents and code files. Yet, prefix caching in engines such as vLLM cannot reuse their KV entries unless they share identical prefixes with another request, while Position-Independent Caching (PIC) implementations within production-grade inference servers typically either require substantial server code changes or keep KV state outside the server, incurring host-to-device transfer overhead. We present Minimalistic PIC (MiniPIC): a minimal, flexible and fast vLLM design built from two ingredients: positional-encoding-free KV cache and user-controlled cache-reuse primitives. MiniPIC stores unrotated K vectors in the KV cache, applies RoPE to K tiles inside attention using per-request logical positions, and exposes three user-facing and token-level primitives: block-aligned padding, span separator (SSep), and prompt depend (PDep), that modify hashing behavior and effective block-level causal attention structure. With fewer than 100 lines of core-engine changes plus a custom attention backend, these primitives are sufficient to realize multiple PIC methods, including Block-Attention, EPIC, and Prompt Cache, within the same running vLLM instance, while natively integrating with KV cache CPU offload implementations. On 2WikiMultihopQA, MiniPIC with interleaved scheduling improves prefill throughput by 49% over baseline vLLM, reduces cached-span time-to-first-token by up to two orders of magnitude, preserves the linear prefill scaling of uncached spans, and incurs only 5.7% worst-case overhead.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MiniPIC shows a compact way to support several PIC methods inside vLLM via unrotated KV and three primitives, but the RoPE reuse logic is the part that needs checking.

read the letter

The paper's core contribution is a minimal vLLM extension that stores unrotated K vectors, applies RoPE inside the attention kernel using per-request logical positions, and adds three token-level primitives (block-aligned padding, SSep, PDep) to control hashing and block causality. With under 100 lines of engine change plus a custom backend, these pieces let the same instance run Block-Attention, EPIC, and Prompt Cache while keeping native CPU offload. That combination and the small footprint are what is actually new relative to prior PIC work.

The results on 2WikiMultihopQA are straightforward: 49% higher prefill throughput with interleaved scheduling, up to two orders of magnitude lower TTFT for cached spans, and only 5.7% worst-case overhead. Those numbers are useful for anyone running repetitive structured prompts.

The soft spot is exactly the one the stress-test note flags. Recomputing rotary embeddings on K tiles with per-request positions must produce the same attention scores and causal masks as standard RoPE-prefixed sequences, even when a cached span sits at different absolute positions. The abstract gives no kernel details, no equivalence argument, and no verification steps, so the claim that the primitives are sufficient rests on the custom backend being correct. If that part is off, reuse could silently violate causality. The <100LOC figure makes this assumption load-bearing.

This is a systems paper aimed at people who maintain or extend production LLM servers. A reader who wants concrete primitives and benchmark numbers for vLLM-style engines will find it worth reading. It deserves a serious referee because the implementation is small enough to inspect and the performance claims are directly testable, even if the correctness argument needs tightening.

Referee Report

2 major / 0 minor

Summary. The paper presents MiniPIC, a minimalistic Position-Independent Caching (PIC) design for vLLM. It stores unrotated K vectors in the KV cache, applies RoPE to K tiles inside a custom attention backend using per-request logical positions, and exposes three token-level primitives (block-aligned padding, span separator (SSep), and prompt depend (PDep)) that modify hashing and block-level causality. The central claim is that these require fewer than 100 lines of core-engine changes plus the custom backend, suffice to realize multiple PIC methods (Block-Attention, EPIC, Prompt Cache) in the same running instance, integrate natively with KV cache CPU offload, and deliver 49% prefill throughput gains plus up to two orders of magnitude TTFT reduction on 2WikiMultihopQA with only 5.7% worst-case overhead.

Significance. If the implementation and correctness arguments hold, the result would be significant for production inference engines: it offers a low-effort, flexible PIC mechanism that avoids both heavy server modifications and external KV-state management while preserving linear scaling for uncached spans and integrating with existing offload paths. The ability to realize multiple distinct PIC policies via the same primitives inside one vLLM instance is a notable engineering contribution.

major comments (2)

[Abstract] Abstract: the central claim that storing unrotated K vectors and applying RoPE inside the attention kernel using per-request logical positions, together with the three primitives, preserves identical attention scores, masks, and causal semantics to a standard RoPE-prefixed sequence (even when a cached span appears at different absolute positions) is presented without any correctness argument, pseudocode for the custom backend, or verification steps. This is load-bearing for the sufficiency claim that no further engine modifications are required.
[Abstract] Abstract: the performance claims (49% prefill throughput improvement over baseline vLLM, up to 100x reduction in cached-span TTFT, 5.7% worst-case overhead, and preservation of linear prefill scaling) are stated without reference to experimental setup details, baseline configurations, workload characteristics, or ablation results isolating the contribution of each primitive, preventing assessment of whether the numbers support the <100LOC claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The two major comments both concern the abstract's brevity. We address them point-by-point below and will revise the manuscript accordingly to improve clarity while preserving the core claims.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that storing unrotated K vectors and applying RoPE inside the attention kernel using per-request logical positions, together with the three primitives, preserves identical attention scores, masks, and causal semantics to a standard RoPE-prefixed sequence (even when a cached span appears at different absolute positions) is presented without any correctness argument, pseudocode for the custom backend, or verification steps. This is load-bearing for the sufficiency claim that no further engine modifications are required.

Authors: We agree the abstract is too concise on this point. The full manuscript (Section 3) derives that unrotated K storage plus per-request logical-position RoPE application yields identical attention scores and masks to standard RoPE, because RoPE is a relative rotation and the custom backend applies the same rotation matrix per token using the request's logical positions rather than absolute cache indices. Block-level causality is enforced via the SSep and PDep primitives that adjust the attention mask at block granularity. We will add a one-paragraph correctness sketch and a short pseudocode listing of the custom backend to the revised manuscript (main text or appendix) to make the argument self-contained. revision: yes
Referee: [Abstract] Abstract: the performance claims (49% prefill throughput improvement over baseline vLLM, up to 100x reduction in cached-span TTFT, 5.7% worst-case overhead, and preservation of linear prefill scaling) are stated without reference to experimental setup details, baseline configurations, workload characteristics, or ablation results isolating the contribution of each primitive, preventing assessment of whether the numbers support the <100LOC claim.

Authors: The abstract is space-constrained, but the experimental details appear in Section 5: all runs use the same 2WikiMultihopQA workload on A100 GPUs with vLLM v0.4 baseline, interleaved scheduling, and the three primitives enabled. Linear scaling is shown for uncached spans; the 5.7% overhead is measured on fully uncached requests. Ablations isolating padding, SSep, and PDep are in Figure 7. We will revise the abstract to include a parenthetical reference to the experimental setup and workload, and ensure the <100LOC claim is cross-referenced to the diff in Appendix A. revision: yes

Circularity Check

0 steps flagged

No circularity: systems implementation with no derivations or fitted predictions

full rationale

The paper is a systems description of a vLLM modification using unrotated KV storage, an in-attention RoPE application, and three user primitives (block-aligned padding, SSep, PDep). It contains no equations, no parameter fitting, no predictions of quantities defined inside the work, and no self-citation chains that bear the central claim. All reported results are benchmark measurements on 2WikiMultihopQA; the implementation is presented as self-contained engineering changes whose correctness is evaluated externally via throughput and latency numbers rather than by internal reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are identifiable from the abstract, as the contribution is an applied systems implementation rather than a theoretical derivation.

pith-pipeline@v0.9.1-grok · 5833 in / 1192 out tokens · 36556 ms · 2026-06-27T07:10:42.654072+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

32 extracted references · 15 canonical work pages · 4 internal anchors

[1]

Using span queries to optimize for cache and attention locality.arXiv preprint arXiv:2511.02749, 2025

Paul Castro, Nick Mitchell, Nathan Ordonez, Thomas Parnell, and Mudhakar Srivatsa. Using span queries to optimize for cache and attention locality.arXiv preprint arXiv:2511.02749, 2025. URL https://arxiv.org/abs/2511.02749. arXiv:2511.02749

arXiv 2025
[2]

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

DeepSeek-AI. Deepseek-v3.2: Pushing the frontier of open large language models.arXiv preprint, 2025. doi: 10.48550/arXiv.2512.02556. URLhttps://arxiv.org/abs/2512.02556

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2512.02556 2025
[3]

EPIC: Efficient position-independent caching for serving large language models.arXiv preprint arXiv:2410.15332, 2024

J. Hu et al. Epic: Efficient position-independent caching for serving large language models.arXiv preprint, 2024. doi: 10.48550/arXiv.2410.15332. URL https://doi.org/10.48550/arXiv.2410. 15332. Accepted at ICML 2025

work page doi:10.48550/arxiv.2410.15332 2024
[4]

Lmcache: An efficient kv cache layer for enterprise-scale llm inference,

Y . Liu et al. Lmcache: An efficient kv cache layer for enterprise-scale llm inference.arXiv preprint, 2025. doi: 10.48550/arXiv.2510.09665. URLhttps://doi.org/10.48550/arXiv.2510.09665

work page doi:10.48550/arxiv.2510.09665 2025
[5]

llama.cpp: Port of facebook’s llama model in c/c++, 2023

ggerganov. llama.cpp: Port of facebook’s llama model in c/c++, 2023. URL https://github.com/ ggerganov/llama.cpp. Open-source CPU inference engine for LLMs

2023
[6]

I. Gim, G. Chen, S. s. Lee, N. Sarda, A. Khandelwal, and L. Zhong. Prompt cache: Modular attention reuse for low-latency inference.arXiv preprint, 2023. doi: 10.48550/arXiv.2311.04934. URL https: //doi.org/10.48550/arXiv.2311.04934

work page doi:10.48550/arxiv.2311.04934 2023
[7]

Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps

Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. InProceedings of the 28th International Conference on Computational Linguistics, pages 6609–6625, 2020

2020
[8]

Transformers: State-of-the-art machine learning for pytorch, tensorflow, and jax, 2023

Hugging Face. Transformers: State-of-the-art machine learning for pytorch, tensorflow, and jax, 2023. URL https://github.com/huggingface/transformers. Research-focused library for transformer models

2023
[9]

Gonzalez, Hao Zhang, and Ion Stoica

Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. InProceedings of the 29th Symposium on Operating Systems Principles (SOSP ’23), 2023

2023
[10]

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval- Augmented generation for knowledge-Intensive NLP tasks.arXiv preprint, 2020. doi: 10.48550/arXiv. 2005.11401. URL https://doi.org/10.48550/arXiv.2005.11401....

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv 2020
[11]

Accepted at NeurIPS 2020

2020
[12]

Lin et al

W. Lin et al. Towards efficient agents: A co-design of inference architecture and system.arXiv preprint,
[13]

Towards efficient agents: A co-design of inference architecture and system,

doi: 10.48550/arXiv.2512.18337. URLhttps://doi.org/10.48550/arXiv.2512.18337

work page doi:10.48550/arxiv.2512.18337
[14]

X. Lin, A. Ghosh, B. K. H. Low, A. Shrivastava, and V . Mohan. Refrag: Rethinking rag based decoding. arXiv preprint, 2025. doi: 10.48550/arXiv.2509.01092. URL https://doi.org/10.48550/arXiv. 2509.01092. 10

work page doi:10.48550/arxiv.2509.01092 2025
[15]

DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model

Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Deng, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Hanwei Xu, Hao Yang, Haowei Zhang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Li, Hui Qu, J.L. Cai, Jian Liang, Jianzhong...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.04434 2024
[16]

T. Y . Liu, A. Achille, M. Trager, A. Golatkar, L. Zancato, and S. Soatto. Picaso: Permutation-invariant context composition with state space models.arXiv preprint, 2025. doi: 10.48550/arXiv.2502.17605. URL https://doi.org/10.48550/arXiv.2502.17605

work page doi:10.48550/arxiv.2502.17605 2025
[17]

D. Ma, Y . Wang, and L. Tian. Block-attention for efficient prefilling.arXiv preprint, 2024. doi: 10.48550/arXiv.2409.15355. URL https://doi.org/10.48550/arXiv.2409.15355. Accepted at ICLR 2025

work page doi:10.48550/arxiv.2409.15355 2024
[18]

Merth, Q

T. Merth, Q. Fu, M. Rastegari, and M. Najibi. Superposition prompting: Improving and accelerating retrieval-augmented generation.arXiv preprint, 2024. doi: 10.48550/arXiv.2404.06910. URL https: //doi.org/10.48550/arXiv.2404.06910

work page doi:10.48550/arxiv.2404.06910 2024
[19]

Llama 3 model card, 2024

Meta AI. Llama 3 model card, 2024. URL https://github.com/meta-llama/llama3/blob/main/ MODEL_CARD.md

2024
[20]

Llama 4 technical report, 2025

Meta Llama Team. Llama 4 technical report, 2025. URLhttps://ai.meta.com/llama-4

2025
[21]

TensorRT-LLM, 2026

NVIDIA. TensorRT-LLM, 2026. URL https://github.com/NVIDIA/TensorRT-LLM. GitHub reposi- tory, accessed 2026-05-18

2026
[22]

David L. Parnas. On the criteria to be used in decomposing systems into modules.Communications of the ACM, 15(12):1053–1058, 1972. doi: 10.1145/361598.361623

work page doi:10.1145/361598.361623 1972
[23]

Mooncake: A kvcache-centric disaggregated architecture for llm serving.arXiv preprint, 2024

Ruoyu Qin, Zheming Li, Weiran He, Jialei Cui, Feng Ren, Mingxing Zhang, Yongwei Wu, Weimin Zheng, and Xinran Xu. Mooncake: A kvcache-centric disaggregated architecture for llm serving.arXiv preprint, 2024

2024
[24]

Roformer: Enhanced transformer with rotary position embedding.arXiv preprint, 2021

Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.arXiv preprint, 2021

2021
[25]

Philippe Tillet, Hsiang-Tsung Kung, and David D. Cox. Triton: an intermediate language and compiler for tiled neural network computations.Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages, 2019. URL https://api.semanticscholar.org/ CorpusID:184488182

2019
[26]

Kv cache offloading — production-stack.vLLM Documentation,

vLLM Contributors. Kv cache offloading — production-stack.vLLM Documentation,
[27]

URL https://docs.vllm.ai/projects/production-stack/en/vllm-stack-0.1.2/ tutorials/kv_cache.html
[28]

Mepic: Memory efficient position independent caching for llm serving.arXiv preprint, 2025

Qian Wang, Zahra Yousefijamarani, Morgan Lindsay Heisler, Rongzhi Gu, Bai Xiaolong, Shan Yizhou, Wei Zhang, Wang Lan, Ying Xiong, Yong Zhang, and Zhenan Fan. Mepic: Memory efficient position independent caching for llm serving.arXiv preprint, 2025. doi: 10.48550/arXiv.2512.16822. URL https://arxiv.org/abs/2512.16822

work page doi:10.48550/arxiv.2512.16822 2025
[29]

Efficient Streaming Language Models with Attention Sinks

Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks.arXiv preprint, 2023. doi: 10.48550/arXiv.2309.17453. URL https: //arxiv.org/abs/2309.17453. ICLR 2024; version 4 last revised 7 Apr 2024

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2309.17453 2023
[30]

Sglang hicache: Fast hierarchical kv caching with your favorite storage backends.LMSYS Org Blog, 2025

Zhiqiang Xie. Sglang hicache: Fast hierarchical kv caching with your favorite storage backends.LMSYS Org Blog, 2025. URLhttps://lmsys.org/blog/2025-09-10-sglang-hicache/. 11

2025
[31]

CacheBlend: Fast large language model serving for RAG with cached knowledge fusion.arXiv preprint arXiv:2405.16444, 2024

Jiayi Yao, Hanchen Li, Yuhan Liu, Siddhant Ray, Yihua Cheng, Qizheng Zhang, Kuntai Du, Shan Lu, and Junchen Jiang. Cacheblend: Fast large language model serving for rag with cached knowledge fusion. arXiv preprint, 2024. doi: 10.48550/arXiv.2405.16444. URL https://doi.org/10.48550/arXiv. 2405.16444

work page doi:10.48550/arxiv.2405.16444 2024
[32]

Gonzalez, et al

Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E. Gonzalez, et al. Sglang: Efficient execution of structured language model programs.arXiv preprint, 2023. 12 A CIDRA: Analysis of Copy-Based Repositioning When two requests share a span at different positions, SPNL’s CIDRA ...

2023

[1] [1]

Using span queries to optimize for cache and attention locality.arXiv preprint arXiv:2511.02749, 2025

Paul Castro, Nick Mitchell, Nathan Ordonez, Thomas Parnell, and Mudhakar Srivatsa. Using span queries to optimize for cache and attention locality.arXiv preprint arXiv:2511.02749, 2025. URL https://arxiv.org/abs/2511.02749. arXiv:2511.02749

arXiv 2025

[2] [2]

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

DeepSeek-AI. Deepseek-v3.2: Pushing the frontier of open large language models.arXiv preprint, 2025. doi: 10.48550/arXiv.2512.02556. URLhttps://arxiv.org/abs/2512.02556

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2512.02556 2025

[3] [3]

EPIC: Efficient position-independent caching for serving large language models.arXiv preprint arXiv:2410.15332, 2024

J. Hu et al. Epic: Efficient position-independent caching for serving large language models.arXiv preprint, 2024. doi: 10.48550/arXiv.2410.15332. URL https://doi.org/10.48550/arXiv.2410. 15332. Accepted at ICML 2025

work page doi:10.48550/arxiv.2410.15332 2024

[4] [4]

Lmcache: An efficient kv cache layer for enterprise-scale llm inference,

Y . Liu et al. Lmcache: An efficient kv cache layer for enterprise-scale llm inference.arXiv preprint, 2025. doi: 10.48550/arXiv.2510.09665. URLhttps://doi.org/10.48550/arXiv.2510.09665

work page doi:10.48550/arxiv.2510.09665 2025

[5] [5]

llama.cpp: Port of facebook’s llama model in c/c++, 2023

ggerganov. llama.cpp: Port of facebook’s llama model in c/c++, 2023. URL https://github.com/ ggerganov/llama.cpp. Open-source CPU inference engine for LLMs

2023

[6] [6]

I. Gim, G. Chen, S. s. Lee, N. Sarda, A. Khandelwal, and L. Zhong. Prompt cache: Modular attention reuse for low-latency inference.arXiv preprint, 2023. doi: 10.48550/arXiv.2311.04934. URL https: //doi.org/10.48550/arXiv.2311.04934

work page doi:10.48550/arxiv.2311.04934 2023

[7] [7]

Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps

Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. InProceedings of the 28th International Conference on Computational Linguistics, pages 6609–6625, 2020

2020

[8] [8]

Transformers: State-of-the-art machine learning for pytorch, tensorflow, and jax, 2023

Hugging Face. Transformers: State-of-the-art machine learning for pytorch, tensorflow, and jax, 2023. URL https://github.com/huggingface/transformers. Research-focused library for transformer models

2023

[9] [9]

Gonzalez, Hao Zhang, and Ion Stoica

Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. InProceedings of the 29th Symposium on Operating Systems Principles (SOSP ’23), 2023

2023

[10] [10]

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval- Augmented generation for knowledge-Intensive NLP tasks.arXiv preprint, 2020. doi: 10.48550/arXiv. 2005.11401. URL https://doi.org/10.48550/arXiv.2005.11401....

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv 2020

[11] [11]

Accepted at NeurIPS 2020

2020

[12] [12]

Lin et al

W. Lin et al. Towards efficient agents: A co-design of inference architecture and system.arXiv preprint,

[13] [13]

Towards efficient agents: A co-design of inference architecture and system,

doi: 10.48550/arXiv.2512.18337. URLhttps://doi.org/10.48550/arXiv.2512.18337

work page doi:10.48550/arxiv.2512.18337

[14] [14]

X. Lin, A. Ghosh, B. K. H. Low, A. Shrivastava, and V . Mohan. Refrag: Rethinking rag based decoding. arXiv preprint, 2025. doi: 10.48550/arXiv.2509.01092. URL https://doi.org/10.48550/arXiv. 2509.01092. 10

work page doi:10.48550/arxiv.2509.01092 2025

[15] [15]

DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model

Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Deng, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Hanwei Xu, Hao Yang, Haowei Zhang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Li, Hui Qu, J.L. Cai, Jian Liang, Jianzhong...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.04434 2024

[16] [16]

T. Y . Liu, A. Achille, M. Trager, A. Golatkar, L. Zancato, and S. Soatto. Picaso: Permutation-invariant context composition with state space models.arXiv preprint, 2025. doi: 10.48550/arXiv.2502.17605. URL https://doi.org/10.48550/arXiv.2502.17605

work page doi:10.48550/arxiv.2502.17605 2025

[17] [17]

D. Ma, Y . Wang, and L. Tian. Block-attention for efficient prefilling.arXiv preprint, 2024. doi: 10.48550/arXiv.2409.15355. URL https://doi.org/10.48550/arXiv.2409.15355. Accepted at ICLR 2025

work page doi:10.48550/arxiv.2409.15355 2024

[18] [18]

Merth, Q

T. Merth, Q. Fu, M. Rastegari, and M. Najibi. Superposition prompting: Improving and accelerating retrieval-augmented generation.arXiv preprint, 2024. doi: 10.48550/arXiv.2404.06910. URL https: //doi.org/10.48550/arXiv.2404.06910

work page doi:10.48550/arxiv.2404.06910 2024

[19] [19]

Llama 3 model card, 2024

Meta AI. Llama 3 model card, 2024. URL https://github.com/meta-llama/llama3/blob/main/ MODEL_CARD.md

2024

[20] [20]

Llama 4 technical report, 2025

Meta Llama Team. Llama 4 technical report, 2025. URLhttps://ai.meta.com/llama-4

2025

[21] [21]

TensorRT-LLM, 2026

NVIDIA. TensorRT-LLM, 2026. URL https://github.com/NVIDIA/TensorRT-LLM. GitHub reposi- tory, accessed 2026-05-18

2026

[22] [22]

David L. Parnas. On the criteria to be used in decomposing systems into modules.Communications of the ACM, 15(12):1053–1058, 1972. doi: 10.1145/361598.361623

work page doi:10.1145/361598.361623 1972

[23] [23]

Mooncake: A kvcache-centric disaggregated architecture for llm serving.arXiv preprint, 2024

Ruoyu Qin, Zheming Li, Weiran He, Jialei Cui, Feng Ren, Mingxing Zhang, Yongwei Wu, Weimin Zheng, and Xinran Xu. Mooncake: A kvcache-centric disaggregated architecture for llm serving.arXiv preprint, 2024

2024

[24] [24]

Roformer: Enhanced transformer with rotary position embedding.arXiv preprint, 2021

Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.arXiv preprint, 2021

2021

[25] [25]

Philippe Tillet, Hsiang-Tsung Kung, and David D. Cox. Triton: an intermediate language and compiler for tiled neural network computations.Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages, 2019. URL https://api.semanticscholar.org/ CorpusID:184488182

2019

[26] [26]

Kv cache offloading — production-stack.vLLM Documentation,

vLLM Contributors. Kv cache offloading — production-stack.vLLM Documentation,

[27] [27]

URL https://docs.vllm.ai/projects/production-stack/en/vllm-stack-0.1.2/ tutorials/kv_cache.html

[28] [28]

Mepic: Memory efficient position independent caching for llm serving.arXiv preprint, 2025

Qian Wang, Zahra Yousefijamarani, Morgan Lindsay Heisler, Rongzhi Gu, Bai Xiaolong, Shan Yizhou, Wei Zhang, Wang Lan, Ying Xiong, Yong Zhang, and Zhenan Fan. Mepic: Memory efficient position independent caching for llm serving.arXiv preprint, 2025. doi: 10.48550/arXiv.2512.16822. URL https://arxiv.org/abs/2512.16822

work page doi:10.48550/arxiv.2512.16822 2025

[29] [29]

Efficient Streaming Language Models with Attention Sinks

Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks.arXiv preprint, 2023. doi: 10.48550/arXiv.2309.17453. URL https: //arxiv.org/abs/2309.17453. ICLR 2024; version 4 last revised 7 Apr 2024

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2309.17453 2023

[30] [30]

Sglang hicache: Fast hierarchical kv caching with your favorite storage backends.LMSYS Org Blog, 2025

Zhiqiang Xie. Sglang hicache: Fast hierarchical kv caching with your favorite storage backends.LMSYS Org Blog, 2025. URLhttps://lmsys.org/blog/2025-09-10-sglang-hicache/. 11

2025

[31] [31]

CacheBlend: Fast large language model serving for RAG with cached knowledge fusion.arXiv preprint arXiv:2405.16444, 2024

Jiayi Yao, Hanchen Li, Yuhan Liu, Siddhant Ray, Yihua Cheng, Qizheng Zhang, Kuntai Du, Shan Lu, and Junchen Jiang. Cacheblend: Fast large language model serving for rag with cached knowledge fusion. arXiv preprint, 2024. doi: 10.48550/arXiv.2405.16444. URL https://doi.org/10.48550/arXiv. 2405.16444

work page doi:10.48550/arxiv.2405.16444 2024

[32] [32]

Gonzalez, et al

Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E. Gonzalez, et al. Sglang: Efficient execution of structured language model programs.arXiv preprint, 2023. 12 A CIDRA: Analysis of Copy-Based Repositioning When two requests share a span at different positions, SPNL’s CIDRA ...

2023