{"id":"e5a29f7c-ca10-4777-9f61-f7fa51cdeb44","arxiv_id":"2506.23934","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"QPART chooses, for each edge inference request, a layer split point and per-layer bit widths that minimize time, energy, and server cost subject to an accuracy budget, cutting communication payload by over 80% with measured accuracy loss below 1%.","lead":"A server and an edge device split a neural network at an optimized layer, and each layer of the device's part is compressed to its own precision level, chosen to meet the task's accuracy target while cutting the transmitted model size by over 80%. It matters because it shows one pre-trained model can serve many different devices by adapting the split and the compression per request instead of storing many model versions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The closed-form optimum in Section IV is not a valid solution of Eq. 23: the objective in Eq. 28 is linear in b, the Hessian in Eq. 29 is not the objective's Hessian, and Eq. 40 drops the accuracy parameter Δ entirely, leaving the claimed accuracy guarantee unsupported.","rationale":"The reader's CONDITIONAL verdict already notes mismatches between Eq. 23 and Eq. 28, a defective Hessian in Eq. 29, and continuous treatment of the partition index. My stress-test sharpens the consequence: the final closed form, Eq. 40, loses the accuracy constraint entirely. If Eq. 40 does not depend on Δ, the bit-widths cannot adapt to the requested accuracy bound, so Fig. 6's accuracy-size trade-off is not produced by the optimization as written. This is an internal mathematical failure, not a disagreement with the quantization literature or the imported accuracy model. The empirical evaluation does not repair the proof: Table IV reports simulations without code or data, and the reported ImageNet baselines for ResNet18 (62.75%) and ResNet34 (67.21%) are lower than standard published values, making the <1% degradation hard to interpret. I credit the paper for a clear system model, the offline/online decomposition, and a plausible MNIST-level simulation; those are useful engineering pieces. But the central claim of jointly optimizing bit-width and partition point under a guaranteed accuracy bound rests on Eq. 27 and Eq. 40, and that derivation is invalid. The proposed re-solve for two accuracy bounds would settle whether Eq. 40 is actually the constrained optimizer; if it is not, the current paper's theoretical central claim should be rejected, even though a heuristic version of QPART might still be worth reporting as an engineering system.","tokens_in":20017,"tokens_out":12107,"duration_ms":129446,"concrete_test":"Fix the six-layer MNIST network from Section V with p = 2, and solve the continuous relaxation of Eq. 23 with an independent nonlinear optimizer (e.g., SLSQP) for two accuracy bounds, Δ = 1% and Δ = 5%. Compare the resulting optimal bit-width vectors with the predictions of Eq. 40 and Algorithm 1. If the true optimizer changes with Δ while Eq. 40 returns the same b_p, the closed form does not enforce the accuracy constraint and the central claim fails. As a subsidiary check, symbolically differentiate the Eq. 28 objective with respect to b_p to verify that ∂²f/∂b_p² = 0, which contradicts the Hessian printed in Eq. 29.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is that Eq. 27 and Eq. 40 provide the jointly optimal layer-wise bit-widths and partition point under the accuracy constraint of Eq. 23. The derivation in Section IV fails internally. In Eq. 28 the objective is linear in every bit-width variable b_l; the only b-dependent term is ε Σ b_l z_l. Hence ∂²f/∂b_l² = 0, and the matrix printed in Eq. 29 is not the Hessian of the stated objective, nor is the displayed matrix symmetric positive semidefinite. More seriously, combining the first and last terms of Eq. 27 cancels ρ_p, s_p, and e^{-ln4 b_p}, producing Eq. 40, b_p = (ξo(p) − δo(p) − z_p/ln4)/(ε z_p). This expression contains no Δ, no requested accuracy a, and no robustness ρ_l, so the bit-width at the partition point is independent of the accuracy bound. A correct KKT solution of Eq. 23 must have a Lagrange multiplier set by the accuracy constraint, and bit-widths should shift when Δ changes; Eq. 40 cannot do that. Additionally, the p-derivative in Eq. 37 is taken from a constraint (Eq. 34) in which p does not appear, and the discrete partition index is treated as a continuous variable. Consequently, the numerical <1% degradations in Table IV cannot be credited to the optimization as written; at best they reflect the empirical calibration of ρ and s in Algorithm 1 on the tested models. This concern is internal to the proof and does not depend on whether the accuracy model imported from reference [33] is plausible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes QPART, an edge inference serving system that jointly selects layer-wise quantization bit-widths and a model partition point for each inference request, minimizing weighted time, energy, and server cost subject to an accuracy-degradation constraint. The claimed contribution is a closed-form KKT solution (Eqs. 27 and 40) together with offline and online algorithms, and simulations report a communication payload reduction to roughly 12–18% of the original size with accuracy degradation below 1%.","tokens_in":20332,"tokens_out":8412,"duration_ms":89655,"significance":"The system concept is timely and practically motivated: it addresses per-request adaptation of a pre-trained model without retraining, and the evaluation covers several datasets and architectures (MNIST, SVHN, CIFAR-10/100, ImageNet with ResNet variants). The explicit modeling of time, energy, server cost, channel rate, and accuracy is a useful framework. However, the central closed-form derivation in Section IV contains load-bearing mathematical errors: the objective is linear in the bit-widths so the printed Hessian is not the objective's Hessian, Eq. 40 removes the accuracy budget from the partition-point bit-width, and the p-derivative is taken from a constraint that does not depend on p. As written, the paper does not establish that Eq. 27 / Eq. 40 solves the constrained problem in Eq. 23, and the reported accuracy results are calibrated rather than independently predicted. The idea is salvageable, but the theoretical core and the accuracy evaluation need substantial rework.","major_comments":[{"comment":"The optimization problem is stated inconsistently. Eq. 23 constrains only the activation noise at p and the weight-quantization noise for layers from p onward, while Eq. 28, which is used in the proof, constrains the full-network sum ∑_{l=1}^L s_l e^{-ln4 b_l}/ρ_l ≤ Δ. Eq. 34 then defines g(b,p) with no dependence on p, yet Eq. 37 claims ∂g/∂p = -s_p e^{-ln4 b_p}/ρ_p; the derivative should be zero. Consequently, the p-stationarity equation in Eq. 38 is not a valid KKT condition of the stated problem. In addition, p is a discrete partition index, so the continuous derivative with respect to p requires explicit justification.","section":"§IV, Eqs. 23, 28, 34, 37"},{"comment":"The convexity claim is unsupported. The objective in Eq. 28 is linear in each b_l, because the only b-dependent term is ε ∑ b_l z_l; hence ∂²f/∂b_l² = 0 and all cross-partials are zero. The matrix printed in Eq. 29 is therefore not a Hessian of the stated objective: it is nonsymmetric, with a nonzero (∂²f/∂p∂b_p) entry while the corresponding (∂²f/∂b_p∂p) entry is zero, and it is not positive semidefinite. KKT sufficiency, and with it the claimed global optimality of the closed-form solution, does not follow from the presented argument.","section":"§IV, Eqs. 28–29"},{"comment":"Eq. 40 is the most serious defect in the derivation. Equating the first and last expressions in Eq. 39 cancels ρ_p, s_p, and e^{-ln4 b_p} and yields b_p = (ξo(p) - δo(p) - z_p/ln4)/(ε z_p). This expression contains no accuracy budget Δ, no requested accuracy a, and no robustness values; the bit-width at the partition point is independent of the accuracy constraint. A correct KKT solution would determine the multiplier λ from the inequality constraint, and the bit-widths would shift when Δ changes. As written, Eqs. 27 and 40 cannot enforce Eq. 23, so the claimed accuracy guarantee is unsupported.","section":"§IV, Eqs. 39–40"},{"comment":"The empirical accuracy results cannot be credited to the optimization as stated. Algorithm 1 step 8 records the noise threshold at which accuracy degradation equals the requested a, and ρ_l and s_l are measured on the served model and dataset (Eq. 22 and Algorithm 1 steps 7–10). The below-1% degradation values in Table IV are therefore calibrated, not predicted, and the accuracy constraint is enforced by construction. If the calibration data overlaps the evaluation data, the reported degradation is not an independent validation. The text should specify a held-out calibration protocol and should separate calibration measurements from evaluation results.","section":"§IV, Algorithm 1; §V, Table IV"},{"comment":"Eq. 22 does not define a genuine per-layer robustness parameter. It is written as a ratio whose numerator is the mean quantization noise at layer l, which depends on b_l through Eqs. 18–19, and whose denominator is the adversarial noise. If ρ_l depends on b_l, then Eq. 27 cannot be solved for b_l without circularity. Additionally, Algorithm 1 step 9 says to calculate ρ_l by Eq. 20, but Eq. 20 defines ψ, not ρ; the intended formula should be Eq. 22 with carefully defined inputs independent of the unknown bit-widths.","section":"§III, Eq. 22; §IV, Algorithm 1 step 9"}],"minor_comments":[{"comment":"The index conventions are inconsistent: Eq. 14 uses b_p z^x_p + ∑_{l=1}^p b_l z^w_l, Eq. 23 refers to ∑_{l=p}^N, Eq. 27 defines z = [z^w_p ... z^w_N z^x_p], and Eq. 28 uses ∑_{l=p}^{L+1}. The roles of N, L, and L+1 should be harmonized throughout.","section":"§III–§IV, notation"},{"comment":"The text says ResNet18, ResNet34, and ResNet50 were implemented on ImageNet, but Table IV lists only ResNet18 and ResNet34; either add the ResNet50 row or correct the text.","section":"§V, Table IV"},{"comment":"Line 3, 'for l = p, 1 do', is ambiguous; it should be written as a descending loop, e.g., 'for l = p, p-1, ..., 1 do'.","section":"§IV, Algorithm 1"},{"comment":"The sentence 'there are two components: executing module, communication module and performance module' lists three components; please rephrase.","section":"§V, opening paragraph"},{"comment":"'NFaaS' appears once and should be 'INFaaS'.","section":"§II-B, related work"},{"comment":"The notation ρ_l(a) is used in Eq. 22, while the text uses ρ_l(a_t^n); please unify the notation and define all indices.","section":"§III, Eq. 22 and surrounding text"}],"recommendation":"major_revision","confidential_remarks":"The main barrier is technical rather than presentational: the closed-form derivation and the accuracy evaluation need to be redone before the paper can be considered. I would support a revised version if the authors re-derive the KKT solution with the correct multiplier λ and a proper discrete treatment of p, and if they replace the calibrated accuracy claim with a genuinely held-out evaluation. The novelty claim of being the first to optimize layer-wise bit-width in an inference serving system may also need qualification given the existing INFaaS and quantization-partitioning literature, but this is secondary to the correctness issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"QPART addresses a real deployment problem: serving a pretrained model to heterogeneous edge devices by jointly picking a partition point and per-layer bit-widths per request. The system model is sensible, and the authors are right that the interaction between quantization and partitioning is underexplored in inference serving. They also ship a complete algorithmic pipeline, including an offline calibration step and an online lookup table—more than many papers in this area.\n\nThe problem is the closed-form solution, which carries the paper's accuracy guarantee, does not work. The stress-test note is correct: in Eq. 28 the objective is linear in every b_l, so the Hessian in Eq. 29 is not the Hessian of the objective and is not PSD. More seriously, Eq. 40 for the partition-point bit-width drops Δ entirely—the KKT multiplier λ cancels when the first and last terms of Eq. 39 are combined. So the derived bit-widths cannot adapt to the accuracy requirement. That defeats the central claim. There are smaller consistency issues as well: the constraint in Eq. 23 is over activation plus tail layers, while Eq. 28 uses a full-network sum; the derivative ∂g/∂p in Eq. 37 is taken from a constraint in which p does not appear; and the discrete partition index is treated as continuous. These are not cosmetic.\n\nThe empirical section does not rescue this. The below-1% accuracy degradations rely on calibrating ρ_l and s_l on the served model and dataset at the same noise levels that later define the 'guarantee'—circular. There is no code or data, and the ImageNet baselines look off (ResNet18 at 62.75% for ImageNet is well below the standard number).\n\nWhat is worth keeping? The idea of jointly optimizing partition point and layer-wise bit-width in an accuracy-aware serving system is genuinely worth exploring. But the manuscript as written does not give a sound theoretical basis or reproducible evaluation. It needs major revision: fix the KKT derivation, or drop the closed-form claim and solve the problem numerically; separate calibration from evaluation; release code and data. I would send it to a serious referee only with the expectation that the referee asks for exactly those repairs. As is, the load-bearing flaw is too central for acceptance.\n\nFor you personally: if you work on edge inference serving, the related-work comparison and the system model are worth a skim. I would not cite the optimization result or the accuracy numbers. I would not bring it to reading group.","headline":"Worthwhile problem, broken closed-form solution: the derived bit-widths don't actually depend on the accuracy budget, so the paper's core guarantee is unsupported.","tokens_in":20962,"tokens_out":6256,"would_cite":false,"duration_ms":58699,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"QPART claims that for each edge inference request the server should jointly optimize layer-wise quantization bit widths and a network partition point, solving a closed-form problem that keeps accuracy degradation below 1 percent while…","keywords":["adaptive inference serving","model quantization","model partitioning","edge computing","accuracy degradation","layer-wise bit-width","inference offloading","closed-form optimization"],"falsifier":"Quantize a trained model with the bit widths QPART's offline algorithm selects for a requested degradation $a$, then measure end-to-end accuracy on a disjoint test set and compare it with the predicted sum of per-layer terms; if the actual degradation exceeds $a$ for a network where quantization noise correlates across layers, such as a deep residual network, the additive accuracy model cannot guarantee the claimed sub-1 percent loss.","tokens_in":19734,"feed_emoji":"⚡","tokens_out":9141,"duration_ms":93184,"temperature":0.7,"pith_summary":"QPART claims that an inference-serving platform should adapt its response to each edge request rather than ship a fixed pre-trained model. For a request carrying a model type, an accuracy requirement, and the device's computation and channel capacity, the server chooses a layer-wise quantization bit width for every layer in the first model segment and a partition point that splits the network between device and server, minimizing weighted time, energy, and server cost while holding accuracy degradation within the requested bound. The paper derives a closed-form expression for the optimal bit widths using KKT conditions on a layer-wise accuracy-degradation metric based on quantization noise and per-layer robustness. In simulations on MNIST, SVHN, CIFAR-10/100, and ImageNet models, the transmitted model segment shrinks to roughly 12-18 percent of its original size with measured accuracy loss between 0.08 and 0.66 percent. If correct, one pre-trained model can serve a heterogeneous fleet of devices without retraining or storing model variants.","feed_headline":"Per-request bit widths and split points cut edge AI payload to 12-18%","feed_subtitle":"QPART quantizes each layer and picks a device–server split per request, holding accuracy loss under 1%.","key_machinery":"The load-bearing object is the layer-wise accuracy-degradation metric $\\psi_l^w = \\|\\sigma_l^w\\|_2^2 / \\rho_l(a)$, where the numerator is quantization noise from quantizing layer $l$'s weights (modeled as $s_l e^{-\\ln 4\\, b_l}$) and $\\rho_l(a)$ is a per-layer robustness parameter measured by injecting adversarial noise and recording the noise threshold that changes the classification. The communication payload is modeled as $Z = b_p z_p^x + \\sum_{l=1}^p b_l z_l^w$, combining quantized activation size and weight sizes. The argument is carried by solving a Lagrange problem whose Hessian the paper shows to be positive semidefinite, producing Eq. 27, which equates normalized noise terms across layers, and Eq. 40, which gives the partition-point layer's bit width directly. A stack of two algorithms realizes this: an offline algorithm enumerating partition points and five accuracy levels to precompute bit-width patterns, and an online algorithm that picks the stored pattern minimizing the weighted objective.","core_discovery":"The central claim is that joint layer-wise quantization and model partitioning can be solved exactly under a linear, additive model of quantization-induced accuracy loss. The objective is $\\min_{b,p} \\omega(T_{\\mathrm{local}}+T_{\\mathrm{tran}}+T_{\\mathrm{server}}) + \\tau(E_{\\mathrm{local}}+E_{\\mathrm{tran}}) + \\eta C$, subject to the accuracy-degradation constraint that the layer-wise quantization-noise terms sum to no more than a bound $\\Delta$ tied to the request's accuracy demand. With quantization noise modeled as $\\|\\sigma_l^w\\|_2^2 = s_l e^{-\\ln 4\\, b_l}$, the KKT stationarity conditions collapse into a chain of equalities across layers (Eq. 27) and a direct closed form for the bit width at the partition point (Eq. 40). The paper reports that this solution, precomputed offline for candidate partition points and five accuracy levels and looked up online, reduces the communication payload to 11.88-18.12 percent of the original model size while keeping accuracy degradation between 0.08 and 0.66 percent.","pith_inferences":["The authors leave implicit that the same closed-form solution could be rounded to hardware-supported precisions such as 4 or 8 bits and re-checked against the $\\Delta$ constraint, turning the continuous analytic solution into a deployable integer-quantization scheme.","A natural extension is online calibration: re-measuring the per-layer robustness $\\rho_l$ and noise scale $s_l$ on the device's own data distribution would let QPART track input drift, at the cost of additional measurement passes.","Because the objective weights are request-level parameters, the framework could be embedded in a scheduler that negotiates service-level agreements, letting the server choose the operating point that best meets the device's latency budget and battery status.","The additive degradation model should be stress-tested on architectures with skip connections, where quantization noise from one layer can be re-amplified downstream; the reported experiments do not isolate that interaction effect."],"forward_implications":["A serving platform can answer each request with a model segment tailored to that device's clock rate, memory, channel capacity, and accuracy demand, using one stored pre-trained model.","Because the bit-width solution is closed-form and the noise and robustness measurements are done offline, runtime serving is a lookup over precomputed patterns, so the optimization does not add latency to the request path.","Communication payload, the part of offloading that dominates latency in mobile-edge settings, drops to about 12-18 percent of the original model size, with reported accuracy degradation between 0.08 and 0.66 percent.","The weights in the objective give a tunable trade-off: a device with ample battery can set the energy weight low and squeeze latency, while a cost-sensitive device can set the server-cost weight high to minimize server usage.","In the paper's comparative experiments, QPART reports the lowest time and energy consumption across all partition points when measured against autoencoder-based and model-pruning-based offloading schemes."],"supporting_citations":[{"why":"Supplies the quantization-noise and adversarial-robustness formulas (Eqs. 18-22) that turn the accuracy requirement into a layer-wise additive constraint.","marker":"[33]"},{"why":"Introduces DNN partitioning between device and cloud, the mechanism QPART extends by adding layer-wise quantization.","marker":"[19]"},{"why":"Closest system for model-less serving of accuracy-constrained requests; QPART contrasts its closed-form quantization with INFaaS's variant search.","marker":"[40]"},{"why":"Shannon's channel-capacity formula is used to model transmission latency and energy from the payload size.","marker":"[25]"},{"why":"Supplies the MAC-count computation cost models for linear and convolutional layers used in local and server inference time.","marker":"[51]"},{"why":"Supports the premise that communication latency dominates in mobile-edge inference, motivating the payload reduction.","marker":"[32]"},{"why":"Autoencoder-based offloading baseline compared in the experiments as an alternative compression approach.","marker":"[35]"},{"why":"Model-pruning-based offloading baseline, alongside [45], anchors the comparative evaluation of compression-based offloading schemes.","marker":"[44]"}],"fun_headline_variants":["QPART: Adaptive bit-widths and split points for edge inference","Per-request quantization and partition cut edge AI payload by over 80%","Joint layer-wise bit-width and partition optimization for edge inference","Edge inference adapts model per request, keeps accuracy loss under 1%","QPART reduces edge AI payload 80% with accuracy-aware adaptation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee rests on the assumption that a network's total accuracy loss is the sum of independent per-layer quantization-noise terms, with each layer's robustness measured once on the served model and dataset.","fun_headline_variants_meta":{"raw":{"variants":["QPART: Adaptive bit-widths and split points for edge inference","Per-request quantization and partition cut edge AI payload by over 80%","Joint layer-wise bit-width and partition optimization for edge inference","Edge inference adapts model per request, keeps accuracy loss under 1%","QPART reduces edge AI payload 80% with accuracy-aware adaptation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000374,"raw_usage":{"total_tokens":2039,"prompt_tokens":1027,"completion_tokens":1012,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":919}},"tokens_in":643,"tokens_out":1012,"duration_ms":10213,"temperature":1.0,"reasoning_tokens":919,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:30:31.664597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Quantize a trained model with the bit widths QPART's offline algorithm selects for a requested degradation $a$, then measure end-to-end accuracy on a disjoint test set and compare it with the predicted sum of per-layer terms; if the actual degradation exceeds $a$ for a network where quantization noise correlates across layers, such as a deep residual network, the additive accuracy model cannot guarantee the claimed sub-1 percent loss.","supporting_citations":[{"cited_title":"Adaptive quantization for deep neural network,","cited_arxiv_id":null,"evidence_quote":"Supplies the quantization-noise and adversarial-robustness formulas (Eqs. 18-22) that turn the accuracy requirement into a layer-wise additive constraint."},{"cited_title":"{INFaaS}: Automated model-less inference serving,","cited_arxiv_id":null,"evidence_quote":"Closest system for model-less serving of accuracy-constrained requests; QPART contrasts its closed-form quantization with INFaaS's variant search."},{"cited_title":"Energy consump- tion of neural networks on nvidia edge boards: an empirical model,","cited_arxiv_id":null,"evidence_quote":"Supplies the MAC-count computation cost models for linear and convolutional layers used in local and server inference time."},{"cited_title":"Edge ai: On-demand accelerating deep neural network inference via edge computing,","cited_arxiv_id":null,"evidence_quote":"Supports the premise that communication latency dominates in mobile-edge inference, motivating the payload reduction."},{"cited_title":"Deep compressive offloading: Speeding up neural network inference by trading edge computation for network latency,","cited_arxiv_id":null,"evidence_quote":"Autoencoder-based offloading baseline compared in the experiments as an alternative compression approach."},{"cited_title":"A deep reinforcement learning based research for optimal offloading decision,","cited_arxiv_id":null,"evidence_quote":"Model-pruning-based offloading baseline, alongside [45], anchors the comparative evaluation of compression-based offloading schemes."}],"review_version":1}