Pith. sign in

REVIEW 3 major objections 4 minor 22 references

Spatial Prefix Caching for Wireless Edge LLM Inference: A Stochastic-Geometry and Queueing Framework

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Nearest edge GPU need not be the fastest server for cached LLM prompts.

desk verdict A genuine first step toward a spatial theory of prefix caching, but Theorem 2's queue coupling is a mean-field approximation and the paper never simulates the coupled system. read the letter →

arxiv 2608.01126 v1 pith:JPE36RB4 submitted 2026-08-02 cs.IT math.IT

classification cs.ITmath.IT MSC 60K2560D05
keywords prefixcachingKVcachereuseedgeLLMinferencestochasticgeometryqueueingtheorytimetofirsttokenGPUmemorypartitioningcache-awarerouting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish a network-level latency law for LLM inference at the wireless edge: when reusable prompt prefixes are cached on scattered GPU nodes, the time to first token is governed jointly by radio distance, how much prefill computation a cached prefix saves, and queueing delay at the serving node. It models the prompt workload as a prefix forest, lets each edge GPU carry an ancestor-closed cache profile, and treats the nodes of each profile as an independent Poisson point process. A load-aware association rule, closed through a damped fixed point, couples routing choices to multi-server GPU queues, and the result is an expression for TTFT coverage probability. The authors validate the analytical expressions against Monte Carlo and event-driven queue simulation and use the model to derive structural predictions: the latency-optimal node need not be nearest, TTFT can be non-monotonic in cached-prefix depth because persistent KV memory reduces concurrency, and queue-unaware prefix routing can form hotspots. If right, the framework gives operators a planning tool for edge-GPU density, memory partitioning, and cache-aware routing without needing deployment-specific simulation.

What carries the argument

The machinery is the ancestor-closed cache profile on a prefix forest: a profile is a set of reusable token prefixes that contains every ancestor of each stored prefix, so one spatial mark can represent nested, partially overlapping prompt families. Independently assigning each edge GPU a profile thins the base Poisson point process into profile-specific Poisson tiers. Selection is governed by a predicted-TTFT cost $J_{s,q}(r)=\xi r^\alpha + \bar{W}_s + T^{\mathrm{pf}}_{s,q}$, and consistency between routing and load is enforced by a damped fixed point $\bar{W} = Q(A(\bar{W}))$. The Erlang-C waiting-time CDF then supplies the queue tail in the TTFT coverage integral.

What would settle it

Run a production LLM serving engine with a real prompt trace on several edge GPUs, holding GPU memory fixed and sweeping cached-prefix depth; if time-to-first-token coverage never declines as depth increases at moderate-to-heavy load, the predicted memory-concurrency coupling is not governing. Alternatively, replace the exponential service-time queue in the model with a simulator of chunked prefill and compare the paper's Fig. 8: the predicted growing gap between queue-unaware and load-aware routing should persist.

Watch

Extended reading notes

Core claim

The central discovery is Theorem 2 (Eq. 39): network-wide TTFT coverage probability has the form $$$P^{{\mathrm{TTFT}}$}_{\mathrm{cov}}(\tau) = \sum_q p_q \sum_s \$int_0^{{r^{\max}}$_{s,q}(\tau)} $f^{{\mathrm{sel}}$}_{s,q}(r)\, F_{W_s}\bigl(\tau - \xi r^\$\alpha$ - $T^{{\mathrm{pf}}$}_{s,q}\bigr)\, dr,$$ where $f^{\mathrm{sel}}_{s,q}$ is the joint density that request type $q$ is served by profile $s$ at distance $r$, $F_{W_s}$ is the Erlang-C waiting-time CDF, $T^{\mathrm{pf}}_{s,q}$ is the residual prefill time, and $\xi r^\alpha$ is the radio delay. The expression simultaneously accounts for spatial availability, partial computation reuse, and queue reliability. It follows from a profile-thinning lemma: ind

Load-bearing premise

The numerical results rest on treating each GPU node as an M/M/c queue with Poisson arrivals and exponential service times; real LLM serving uses continuous batching, chunked prefill, and non-exponential job sizes, so the specific coverage values and stability thresholds may not transfer.

Editorial extensions

If this is right

  • Operators can turn a TTFT service-level objective into a minimum edge-GPU density and a memory split between persistent prefixes and active requests, via the dual planning problem in (51).
  • Routers should compare marginal end-to-end latency savings, not longest-prefix-first or nearest-node rules; a remote deep-prefix node can win only when its prefill saving exceeds the extra radio delay plus queueing.
  • At moderate-to-heavy load, adding cached-prefix depth can reduce TTFT coverage because persistent KV memory lowers concurrency and raises queueing; the optimal depth shifts shallower as load grows.
  • Cache-aware but queue-unaware routing can perform worse than nearest-node routing under skewed cache affinity, because it creates hotspots; queue feedback is necessary for stability.
  • The design principle is hierarchical placement: replicate shallow prefixes broadly, deploy deep prefixes selectively, and keep a high-concurrency fallback profile.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The spatial association law only needs a profile-specific waiting-time distribution, so replacing Erlang-C with an M/G/c or trace-fitted delay map should preserve the qualitative predictions; the paper hints at this, but the non-monotonicity mechanism would survive only if the service-time distribution does not invert the queue-order between tiers.
  • A testable extension is to run the same framework on a split prefill/decode deployment: the decode queue becomes a second fixed-point state, and the bipartite compute-memory selection in Section VII-B suggests the one-process model is a special case of a richer matching problem.
  • If exact prefix reuse is sparse or cached-state lifetime is shorter than migration time, the quantitative gains shrink; the framework's core quantity then becomes the residual-prefill distribution rather than hit ratio.
  • The predicted non-monotonicity in prefix depth could be checked in a single-node experiment: with fixed GPU memory and load, vary the size of the persistent KV region and measure TTFT quantiles; a decline past the peak would confirm the memory-concurrency coupling independent of wireless geometry.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. Prefix caching reuses the key–value states of shared prompt prefixes. The paper models wireless edge LLM inference: GPU nodes form a PPP with independently assigned ancestor-closed cache profiles; users generate Poisson requests of different prompt types. It derives the probability that a request associates with a profile tier (Theorem 1, Eq. 25), the conditional serving-distance density, a load–association fixed point (Eq. 33), a necessary area-capacity condition (Prop. 2), and the TTFT coverage probability (Theorem 2, Eq. 39) as an integral of the selection-distance density times an Erlang-C waiting-time CDF. It formulates and solves heuristically a cache-profile placement optimization under memory and stability constraints, and validates the spatial association and queue CDF by simulation. Numerical experiments produce structural insights: the latency-optimal node need not be nearest; TTFT coverage is non-monotonic in cached-prefix depth; queue-unaware prefix routing creates hotspots.

Significance. The prefix-forest cache profile is a genuine generalization of classical binary file caching, and the independent-marking decomposition into spatial tiers is natural. If Theorem 2 were exact as stated, the result would be the first wireless-edge LLM inference law coupling radio distance, partial prefix reuse, memory-limited concurrency, and queue stability, with concrete structural predictions (nearest node not TTFT-optimal; non-monotonicity in cache depth; queue-unaware routing collapse) and a planning tool for density and profile mix. The paper ships reproducible code, validates Theorem 1 and the standalone queue CDF, and derives the association integrals cleanly. These are real strengths. The main weakness is that the central TTFT coverage expression is a mean-field substitution whose error is unquantified, so the exact-law claim is not yet supported.

major comments (3)
  1. [§III-D, Eq. (15); §IV-C, Eq. (39)] The coverage expression substitutes the tier-mean arrival rate λ_s^q into the Erlang-C CDF F_{W_s}. By Eq. (15), λ_s^q is the spatial average over profile-s nodes. But the node serving a typical user at the origin is not a typical node: its service region is the area-biased cell containing the origin, so its arrival rate is random with mean strictly larger than λ_s^q (by E[A^2]/E[A]^2 > 1). Thus Eq. (39) evaluates E[F(W)|selected] at E[Λ|selected] — a Jensen substitution. Since F_W decreases in load, this overstates coverage. No bound or end-to-end simulation quantifies the error. Within the paper's own M/M/c assumptions, Theorem 2 is a mean-field approximation, not the claimed exact law. The authors should average F_W over the selected-node load distribution or explicitly present Theorem 2 as an approximation with end-to-end validation.
  2. [§VI] The numerical validation is incomplete for the central claim. Fig. 4 validates only the spatial association probabilities (Theorem 1) via Monte Carlo sampling of nearest distances; Fig. 5 validates only the Erlang-C tail of a single M/M/9 queue at ρ=0.82. No coupled simulation of node locations, request locations, profile selection, and per-node queues is reported. The size-bias error in Eq. (39) is therefore never observed. Please add an end-to-end discrete-event simulation of the full model and compare with Eq. (39) across parameter sets.
  3. [§III-D, Eqs. (15)-(20)] Even before size bias, the per-node arrival process is doubly stochastic: with a stationary user point process, the number of users in a node's service region is random, so arrivals are Poisson only conditional on cell size. The M/M/c queue at the tier-mean rate ignores this within-tier heterogeneity. Please state explicitly that λ_s^q is a mean-field parameter and describe the direction of the resulting bias; this is separate from the acknowledged continuous-batching abstraction.
minor comments (4)
  1. [Eq. (15)] The left side is λ_s^q but the right side sums over q; the superscript is misleading and should be removed (or the notation changed to λ_s).
  2. [Eq. (32)] There is a stray comma before 'dr' in the integral: \int_0^\infty r f^{sel}_{s,q}(r), dr.
  3. [Algorithm 1, line 10] The phrase 'Estimate a search direction b∇P^TTFT_cov' is missing the object being estimated; rephrase as 'estimate a search direction for the objective P^TTFT_cov'.
  4. [Fig. 5 caption] The caption reads 'c = 9, = 0.82' — the ρ symbol is missing.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: derivation is self-contained; the fixed-point coupling is a legitimate self-consistency system, no load-bearing self-citations, and no fitted parameter is renamed as a prediction.

full rationale

The paper's central results are derived from explicit model assumptions: PPP thinning (Lemma 1), the cost-based association rule (Eqs. 21-22), the Erlang-C queue model (Eqs. 18-20), and the memory-concurrency relation (Eq. 13). Theorem 1 and Theorem 2 are total-probability compositions of these ingredients; the only coupling between spatial association and queueing appears through the fixed point (33)-(34), which is a standard self-consistency condition rather than a definitional equivalence. Numerical parameters in Table III are normalized inputs, not fitted to any headline output, and the Monte Carlo/event simulations validate the same mathematical model, so this is an internal-consistency check, not a fitted-input-called-prediction. The paper contains no load-bearing self-citations; references to prior work are external systems or classical stochastic-geometry results. A possible concern—that Eq. (39) evaluates the Erlang-C CDF at the tier-mean arrival rate instead of the selected-node arrival rate—is a size-bias/mean-field accuracy issue, not a circular reduction; it does not make the TTFT coverage expression equivalent to an input by construction. The structural insights (nearest node need not be optimal, non-monotonicity in cached depth, queue-unaware hotspot collapse) follow from the model's specified tradeoffs and are presented as analytical consequences rather than as independently measured discoveries. Overall, the derivation chain is self-contained and non-circular.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

No new physical entities are postulated. The cache profile and prefix forest are mathematical abstractions of existing KV-state reuse mechanisms, captured under axioms rather than distinct entities. The free parameters are normalized inputs for the numerical study, not fitted to reproduce a target result, but they do set the scale of all structural insights.

free parameters (5)
  • prefill model coefficients a,b = a=1.6e-4, b=1.8e-8
    g(n)=a n + b n^2 approximates prefill time; chosen as normalized inputs in Table III, not from hardware profiling; they set the scale of residual-prefill delay in all numerical results.
  • radio model parameters xi, alpha = xi=0.28, alpha=2.5
    T_com(r)=xi r^alpha abstracts bandwidth, protocol overhead and fading; normalized values; slower alpha reduces long-distance penalty.
  • decoding occupancy D_q = 0.4 s
    Assumed constant per request type in (14); influences queue service rate.
  • runtime memory budget per active request m_run = implied by concurrency (18,14,6)
    Sets concurrency c_s = floor((M-B_s)/m_run); chosen values determine the strength of memory-queue coupling.
  • persistent profile memory B_s = implied by concurrency values (18,14,6) with M
    B_s sets the memory cost of each profile; not derived from KV allocator measurements.
assumptions (7)
  • domain assumption Edge nodes form a homogeneous PPP (Eq. 2)
    Standard SG model of irregular deployments; used for all spatial derivations.
  • domain assumption Independent stationary user process with Poisson arrivals (Eq. 3)
    Slivnyak's theorem requires stationarity; Poisson arrivals justify M/M/c and PASTA.
  • domain assumption Independent cache-profile marks with probabilities pi_s (Eq. 7)
    Needed for Lemma 1 thinning; the paper flags correlated placement as an extension in Section VII-B.
  • domain assumption Exact-prefix reuse with ancestor-closed profiles
    Cached states cannot modify model outputs; ancestor closure prevents inconsistent states (Section III-B).
  • domain assumption M/M/c queue with Erlang-C waiting distribution (Eqs. 18-20)
    GPU serving approximated as a Markov multi-server queue; the paper calls it analytically calibrated and not exact for continuous batching (Section III-D).
  • domain assumption Delay-equivalent radio model T_com(r)=xi r^alpha (Eq. 10)
    Abstraction of path loss, bandwidth and protocol overhead; instantaneous SINR model left as extension.
  • domain assumption Uniform per-node model copy and no disaggregated fetch in baseline
    Baseline Section III; disaggregated KV retrieval treated as extension in Section VII-B.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatial Prefix Caching for Wireless Edge LLM Inference: A Stochastic-Geometry and Queueing Framework." pith.science (2026). https://pith.science/paper/JPE36RB4

@misc{pith2026260801126,
  author       = {Pith},
  title        = {Pith review of: Spatial Prefix Caching for Wireless Edge LLM Inference: A Stochastic-Geometry and Queueing Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JPE36RB4}},
  note         = {Machine review of arXiv:2608.01126}
}
read the original abstract

Prefix caching reuses the key--value (KV) states of shared prompt prefixes and can substantially reduce the time to first token (TTFT) of large language model (LLM) inference. In a wireless edge network, however, prefix states are distributed across geographically separated GPU nodes. A nearby node offers a short radio path but may provide little reuse, whereas a more distant node may cache a longer matching prefix but incur additional communication and queueing delay. Moreover, persistent prefixes and active-request KV states compete for the same GPU memory, so aggressive caching can reduce inference concurrency and create queueing hotspots. This paper develops a stochastic-geometry and queueing framework for this spatial communication--caching--computation tradeoff. We represent the prompt workload by a prefix forest and define an ancestor-closed cache profile that may contain multiple reusable prefixes. Edge GPU nodes form a Poisson point process and are independently marked by cache profile, yielding analytically tractable spatial tiers. We derive the profile-association probability, conditional serving-distance distribution, token-level computation-offloading ratio, and TTFT coverage probability under a load-aware association policy. A fixed-point formulation captures the coupling between spatial association and multi-server GPU queues, while an outer optimization selects the cache-profile distribution subject to static-memory and stability constraints. Analytical and Monte Carlo results agree closely. The results show that the latency-optimal node need not be the nearest node, that TTFT can be non-monotonic in cached-prefix depth because of GPU-memory coupling.

Figures

Figures reproduced from arXiv: 2608.01126 by the authors.

Figure 1
Figure 1. Spatial prefix caching. Independently marked edge-GPU nodes [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. A prefix forest and three ancestor-closed cache profiles. The upper [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustrative selection region for a local shallow-cache node and a [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Analytical Erlang-C waiting-time CDF and event-driven queue [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 4
Figure 4. Figure 4: Analytical and Monte Carlo profile-association probabilities. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 9
Figure 9. Figure 9: Sensitivity to the probability of a long-prefix request. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Load-dependent cache-profile mix selected from a memory-feasible [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 13
Figure 13. Figure 13: Three-time-scale deployment architecture. Trace mining and hard [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]
Figure 12
Figure 12. Figure 12: TTFT coverage versus edge-GPU density. start for adjacent traffic epochs, because arrival rates and queue prices generally change gradually. The slow, intermediate, and online loops are connected through a compact catalogue and telemetry interface. The remaining mecha…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 10 canonical work pages

  1. [1]

    Efficient memory management for large language model serving with PagedAttention,

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica, “Efficient memory management for large language model serving with PagedAttention,” inProceedings of the 29th ACM Symposium on Operating Systems Principles (SOSP), 2023

  2. [2]

    Prompt cache: Modular attention reuse for low-latency inference,

    I. Gim, G. Chen, S.-s. Lee, N. Sarda, A. Khandelwal, and L. Zhong, “Prompt cache: Modular attention reuse for low-latency inference,” in Proceedings of Machine Learning and Systems (MLSys), vol. 6, 2024

  3. [3]

    SGLang: Efficient execution of structured language model programs,

    L. Zheng, L. Yin, Z. Xie, C. Sun, J. Huang, C. H. Yu, S. Cao, C. Kozyrakis, I. Stoica, J. E. Gonzalez, C. Barrett, and Y . Sheng, “SGLang: Efficient execution of structured language model programs,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 37, 2024

  4. [4]

    Cost-efficient large language model serving for multi-turn conversations with CachedAttention,

    B. Gao, Z. He, P. Sharma, Q. Kang, D. Jevdjic, J. Deng, X. Yang, Z. Yu, and P. Zuo, “Cost-efficient large language model serving for multi-turn conversations with CachedAttention,” in2024 USENIX Annual Technical Conference (USENIX ATC), 2024

  5. [5]

    Preble: Effi- cient distributed prompt scheduling for LLM serving,

    V . Srivatsa, Z. He, R. Abhyankar, D. Li, and Y . Zhang, “Preble: Effi- cient distributed prompt scheduling for LLM serving,” inInternational Conference on Learning Representations (ICLR), 2025

  6. [6]

    Mooncake: A KVCache-centric disaggregated architecture for LLM serving,

    R. Qin, Z. Li, W. He, M. Zhang, Y . Wu, W. Zheng, and X. Xu, “Mooncake: A KVCache-centric disaggregated architecture for LLM serving,”arXiv preprint arXiv:2407.00079, 2024

  7. [7]

    Shadowserve: Interference-free KV cache fetching for distributed prefix caching,

    X. Xiang, R. Joshi, Y . Liu, J. Yao, C. Zhao, J. Jiang, Y . Zhou, E. Kohler, and M. Yu, “Shadowserve: Interference-free KV cache fetching for distributed prefix caching,”arXiv preprint arXiv:2509.16857, 2025

  8. [8]

    Accelerating Local LLMs on Resource-Constrained Edge Devices via Distributed Prompt Caching

    H. Matsutani, N. Matsuda, and N. Sugiura, “Accelerating local LLMs on resource-constrained edge devices via distributed prompt caching,” arXiv preprint arXiv:2602.22812, 2026

Show all 22 references
  1. [9]

    Caching placement in stochastic wireless caching helper networks: Channel selection diversity via caching,

    S. H. Chae and W. Choi, “Caching placement in stochastic wireless caching helper networks: Channel selection diversity via caching,”IEEE Transactions on Wireless Communications, vol. 15, no. 10, pp. 6626– 6637, 2016

  2. [10]

    Probabilistic caching in wireless D2D networks: Cache hit optimal versus throughput optimal,

    Z. Chen, N. Pappas, and M. Kountouris, “Probabilistic caching in wireless D2D networks: Cache hit optimal versus throughput optimal,” IEEE Communications Letters, vol. 21, no. 3, pp. 584–587, 2017

  3. [11]

    A tractable approach to coverage and rate in cellular networks,

    J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to coverage and rate in cellular networks,”IEEE Transactions on Communications, vol. 59, no. 11, pp. 3122–3134, 2011

  4. [12]

    Haenggi,Stochastic Geometry for Wireless Networks

    M. Haenggi,Stochastic Geometry for Wireless Networks. Cambridge University Press, 2012

  5. [13]

    Keyformer: KV cache reduction through key tokens selection for efficient generative inference,

    M. Adnan, A. Arunkumar, G. Jain, P. J. Nair, and M. Soltanolkotabi, “Keyformer: KV cache reduction through key tokens selection for efficient generative inference,” inProceedings of Machine Learning and Systems (MLSys), vol. 6, 2024

  6. [14]

    Marconi: Prefix caching for the era of hybrid LLMs,

    R. Pan, Z. Wang, Z. Jia, C. Karakus, L. Zancato, T. Dao, Y . Wang, and R. Netravali, “Marconi: Prefix caching for the era of hybrid LLMs,” arXiv preprint arXiv:2411.19379, 2024

  7. [15]

    Distserve: Disaggregating prefill and decoding for goodput- optimized large language model serving,

    Y . Zhong, S. Liu, J. Chen, J. Hu, Y . Zhu, X. Liu, X. Jin, and H. Zhang, “Distserve: Disaggregating prefill and decoding for goodput- optimized large language model serving,” in18th USENIX Symposium on Operating Systems Design and Implementation (OSDI), 2024, pp. 193–210

  8. [16]

    Splitwise: Efficient generative LLM inference using phase splitting,

    P. Patel, E. Choukse, C. Zhang, A. Shah, I. Goiri, S. Maleki, and R. Bianchini, “Splitwise: Efficient generative LLM inference using phase splitting,”arXiv preprint arXiv:2311.18677, 2023

  9. [17]

    Taming throughput-latency tradeoff in LLM inference with Sarathi-Serve,

    A. Agrawal, N. Kedia, A. Panwar, J. Mohan, N. Kwatra, B. S. Gulavani, A. Tumanov, and R. Ramjee, “Taming throughput-latency tradeoff in LLM inference with Sarathi-Serve,”arXiv preprint arXiv:2403.02310, 2024

  10. [18]

    Llumnix: Dynamic scheduling for large language model serving,

    B. Sun, Z. Huang, H. Zhao, W. Xiao, X. Zhang, Y . Li, and W. Lin, “Llumnix: Dynamic scheduling for large language model serving,”arXiv preprint arXiv:2406.03243, 2024

  11. [19]

    Memserve: Context caching for dis- aggregated LLM serving with elastic memory pool,

    C. Hu, H. Huang, J. Hu, J. Xu, X. Chen, T. Xie, C. Wang, S. Wang, Y . Bao, N. Sun, and Y . Shan, “Memserve: Context caching for dis- aggregated LLM serving with elastic memory pool,”arXiv preprint arXiv:2406.17565, 2024

  12. [20]

    KVDirect: Distributed disaggregated LLM inference,

    S. Chen, R. Jiang, D. Yu, J. Xu, M. Chao, F. Meng, C. Jiang, W. Xu, and H. Liu, “KVDirect: Distributed disaggregated LLM inference,”arXiv preprint arXiv:2501.14743, 2025

  13. [21]

    Cachegen: KV cache compression and streaming for fast large language model serving,

    Y . Liu, H. Li, Y . Cheng, S. Ray, Y . Huang, Q. Zhang, K. Du, J. Yao, S. Lu, G. Ananthanarayanan, M. Maire, H. Hoffmann, A. Holtzman, and J. Jiang, “Cachegen: KV cache compression and streaming for fast large language model serving,”arXiv preprint arXiv:2310.07240, 2023

  14. [22]

    Towards distributed inference of LLMs on a P2P network,

    S. S. Nair and K. Saini, “Towards distributed inference of LLMs on a P2P network,”arXiv preprint arXiv:2606.17059, 2026

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.