acceptodds
Under review as a conference paper at ICLR 2027

GraphPTQ: Post-Training Quantization for Graph Models, Memory Savings and Generalization

Abstract

Post-training quantization (PTQ) is essentially unexplored for graph models. Yet graph foundation models (GFMs) are scaling rapidly, straining memory budgets, while high-stakes deployment demands robustness to distribution shift. We address this with GraphPTQ, the first unified evaluation of how PTQ affects generalization in message passing models, graph transformers, and pretrained GFMs, featuring hardware-measured GPU peak memory. We evaluate eight models and eleven modes, spanning both weight-dominated and activation-dominated memory regimes across covariate shifts and zero-shot tasks. Our findings reveal clear memory-generalization tradeoffs: weight-only INT8 is the safe default for significant memory reduction, while under INT4 weights or activation quantization, out-of-distribution (OOD) accuracy responses become bidirectional and architecture-dependent. We further show that the in-distribution loss curvature along the rounding direction successfully predicts this OOD risk, with the predictive ordering strengthening as bit-width drops.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Post-Training Quantization, Out-of-Distribution Generalization, Zero-Shot Transfer, Graph Foundation Models, Graph Transformers, GNNs, Hessian Analysis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hbuYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents GraphPTQ, a unified evaluation framework for post-training quantization of graph models across eight architectures and eleven quantization modes. It jointly evaluates hardware-measured GPU memory, in-distribution performance, and OOD/zero-shot generalization under different weight and activation precisions. The paper further decomposes quantization and distribution-shift perturbations and proposes a directional-curvature metric, ∣ΔW⊤HIDΔW∣|\Delta W^\top H_{\mathrm{ID}}\Delta W|, as an ID-only diagnostic of quantization sensitivity. Empirically, it identifies W8A16 as the most consistently safe configuration and reports increasingly architecture-dependent behavior under more aggressive low-bit quantization.

Critical strengths
  1. The paper projects the ID Hessian onto the actual PTQ perturbation direction, $ \Delta W$, and uses $ |\Delta W^\top H_{\mathrm{ID}}\Delta W|$ to characterize quantization sensitivity. This directly evaluates the direction introduced by PTQ rather than relying on direction-agnostic sharpness measures such as SAM, Hessian trace, or the largest Hessian eigenvalue.

  2. The paper separates weight quantization, activation quantization, distribution shift, and their interaction terms in the forward perturbation analysis. This provides a concrete way to analyze why the same low-bit quantization can produce different OOD behaviors across different architectures.

Critical weaknesses
  1. The directional-curvature experiments do not provide strong enough evidence for the claimed OOD-risk prediction. Most of the reported OOD correlations are weak or statistically non-significant. W4A16, for example, has only ρ=0.15\rho=0.15 with p=0.59p=0.59, although W4A16 is one of the main settings for which the paper recommends using this diagnostic before deployment. W4A4 also has only ρ=0.34\rho=0.34 with p=0.21p=0.21. Across the five reported modes, only W4A8 reaches the stated p<0.05p<0.05 threshold for OOD. These results show some correlation in selected settings, but do not yet show that the metric can reliably rank OOD risk across quantization configurations. The claimed strengthening with decreasing weight bit-width is also difficult to establish from the current comparison because activation precision changes across several of the modes, while W4A16, W4A8, and W4A4 share the same b=4b=4 weight perturbation used by the diagnostic.

  2. Some experimental results are internally inconsistent. The paper states that zero-shot performance changes are reported as relative percentages, but GIMLET HIV changes from 0.662 under FP32 to 0.375 under W4A4. This corresponds to a decrease of 28.7 percentage points, but a relative decrease of approximately 43.4%. The paper reports 28.7%, which corresponds to the absolute difference rather than the stated relative-percentage definition. This is not only a labeling issue. Under W4A16, the same task changes from 0.662 to 0.566, corresponding to a 9.6 percentage-point decrease or approximately a 14.5% relative decrease, while the text reports 13.2%. Neither calculation gives 13.2%. This makes it unclear which calculation and aggregation procedure are used for some of the reported conclusions.

  3. The comparison between the explicit INT4 path and the fused path changes several factors simultaneously. The explicit GraphGPT W4A16 result discussed as unstable uses per-channel quantization, while the fused path uses per-group quantization with g=128g=128 and BF16 fused computation. The corresponding explicit per-group INT4 experiment was not completed. The observed difference therefore cannot be attributed specifically to kernel fusion because quantization granularity and computation precision also change at the same time.

Questions for authors
  1. Could the authors provide stronger validation for the use of directional curvature as an OOD-risk predictor? In particular, how should the W4A16 result of ρ=0.15\rho=0.15, p=0.59p=0.59 and the W4A4 result of ρ=0.34\rho=0.34, p=0.21p=0.21 be reconciled with the claim that the metric can rank OOD deployment risk from ID data alone? It would also be useful to report whether the increase in correlation under lower weight precision remains when activation precision is held fixed.

  2. Could the authors clarify the exact definition and aggregation procedure used for zero-shot ΔAcc\Delta\mathrm{Acc}%? For GIMLET HIV, W4A4 changes from 0.662 to 0.375, which is a 28.7 percentage-point decrease but approximately a 43.4% relative decrease, while the paper reports 28.7%. Under W4A16, the score changes from 0.662 to 0.566, which is a 9.6 percentage-point decrease or approximately a 14.5% relative decrease, while the text reports 13.2%. Please clarify which values are used in the main conclusions and reconcile the corresponding tables and figures.

  3. Could the authors provide a controlled comparison between the explicit and fused INT4 paths in which quantization granularity and computation precision are held fixed? In the current experiments, the unstable explicit GraphGPT W4A16 result uses per-channel quantization, while the fused path uses per-group quantization with g=128g=128 and BF16 fused computation. Without the corresponding explicit per-group experiment, it is difficult to determine how much of the improvement comes from kernel fusion rather than quantization granularity or numerical precision.

  4. Could the authors clarify the perturbation decomposition in Eq. (2)? The paper states that both ΔWϵ\Delta W\epsilon and ΔWΔa\Delta W\Delta a exist only when quantization and distribution shift are jointly present, but ΔWΔa\Delta W\Delta a can remain nonzero when ϵ=0\epsilon=0 as long as both weight and activation quantization are present. Please clarify this statement and provide evidence for the claim that the sign of the ΔWϵ\Delta W\epsilon interaction determines the direction of the observed OOD performance change.

1 comment

Thank you for finding the zero-shot Δ error, the confounded INT4 comparison and the missing control for the bit-width trend.

How the revision addresses your review. We refined the analysis into two hypotheses, when ID validation reflects the OOD change and what better rounding can and cannot fix, and tested them on the 132 GOOD and DrugOOD models (4 seeds per configuration, 16 random roundings per bit-width). Zero-shot Δ is redefined and recomputed (GIMLET at W4A4: +1.0% under the new GraphPTQ recipe). The INT4 comparison was redone one factor at a time (head, layer coverage, granularity), which changed our explanation.

Weakness 1 & Question 1 — H̃ correlates weakly with the OOD change (W4A16 ρ=0.15, W4A4 ρ=0.34), and the bit-width trend mixes weight and activation bits. You are right. With 4 seeds per configuration (60 GOOD models instead of 15), the correlation is stronger than the earlier estimates at W4A16 and W4A4 and positive in every mode (Spearman 0.33–0.53; W4A16 0.47, W4A4 0.53; App. F.2), but it remains moderate, and part of its signal is shared with the quantized model's own ID drop. The bit-width trend does not hold at fixed activation precision (W8A16→W4A16: +0.14 [−0.13, +0.38]; W8A8→W4A8: −0.01), and W4A4 is dominated by activation error, so we withdraw it. The revision explains the gap: H̃ captures only the systematic, curvature part of the loss change, while under shift a first-order quantization × shift term carries most of the variation between roundings (median R² ≈0.83 at 4 bits; App. E). H̃ remains useful as that systematic part: it is what curvature-aware rounding reduces, and GPTQ's reduction carries over to OOD (H2, §6.2). Revision: the trend claim is withdrawn; H̃ is reported as a moderate ranker of the size of the OOD change and as the systematic term in H2.

Weakness 2 & Question 2 — Zero-shot Δ is inconsistent (GIMLET HIV 28.7 vs. 13.2). You are right: both were absolute ROC-AUC points mislabelled as relative %. Table 2 now uses one definition, the per-dataset relative change vs. FP32 averaged per column, with the absolute scores in App. C.1 (Table 12); every table number is now generated by script from the stored per-run results. Under the new recipe GIMLET changes by +0.7% at W4A16 and +1.0% at W4A4 (torchao W4A4: −9.5%).

Weakness 3 & Question 3 — The explicit vs. fused INT4 comparison changes several factors at once. You are right, and the controlled comparison changed our explanation: the accuracy gap came from the quantized output head and layer coverage, not from fusion or granularity. At matched head policy and coverage, per-channel and per-group agree within 0.3 points (GraphGPT +4.9% vs. +4.6%, TGT +5.2% vs. +5.5% MAE; App. C.3), while quantizing GraphGPT's head alone moves the MAE change from +5.1% to +161.9% (App. C.4). Fusion affects memory, not accuracy; GraphPTQ now runs on our own kernels (GQ-Kernels), and accuracy and memory are measured on the same kernels (App. A.3). Revision: the attribution to fusion is withdrawn; head policy and layer coverage are reported per model.

Question 4 — ΔWΔa is nonzero without shift; show that the sign of ΔWε sets the direction of the OOD change. You are right on both. Eq. 2 now states that only ΔWε needs shift, while ΔWΔa appears whenever weights and activations are both quantized (§3). The sign claim holds only in part: on the split where it is measured, the sign of the first-order OOD term matches the direction of the loss change in 100/120 cells, but on held-out OOD test data only in 69/120, i.e. chance (App. F.4). The term explains the direction, but the direction does not carry over to new data. What holds is distributional: across random roundings, the spread of the OOD loss change and the share of roundings that lower it follow the noise model (slope 1.03 / 1.07; calibration error 0.041 / 0.039), but not the sign of one rounding. Whether ID validation reflects it depends on the shift: the ID-best of 16 roundings is OOD-better in 88% of DrugOOD but 66% of GOOD cases. This needs OOD data, so it is an analysis, not a deployment diagnostic. Revision: Eq. 2 corrected; the sign claim narrowed to this statement (§4.3, §6.2).

Reviewer_AV36OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper systematically investigates how post-training quantization affects graph models’ actual memory consumption, in-distribution performance, OOD generalization, and zero-shot transfer. The authors compare quantization configurations across message-passing models, graph transformers, and graph foundation models, revealing architecture-dependent effects of weight and activation quantization. They further propose using ID loss curvature along the realized quantization direction to diagnose OOD vulnerability. The main contributions are a unified empirical study and the application of existing second-order quantization sensitivity analysis to graph-model generalization.

Critical strengths
  1. Jointly examining actual memory consumption, ID fidelity, and cross-distribution generalization is more informative than reporting compression ratios or test performance alone. In particular, showing that preserved ID performance does not guarantee preserved OOD performance has practical implications for deploying graph foundation models.

  2. It distinguishes the quantization coverage of consecutive linear projections from that of scatter-gather operations and examines both weight-dominated and activation-dominated memory regimes. This gives the work value beyond a straightforward application of existing quantization tools.

Critical weaknesses

The implementation relies on standard group-wise quantization and existing execution backends, while the directional-curvature metric largely follows established second-order quantization error analysis. AdaRound already used the corresponding quadratic form to analyze rounding perturbations and examined its relationship with accuracy.

Questions for authors
  1. What new properties does the proposed diagnostic offer beyond existing second-order quantization sensitivity analysis? Which aspects of graph structure or message propagation specifically explain its effectiveness?

  2. Under what conditions can ID directional curvature reflect the effects of the OOD gradient and OOD curvature? Are there controlled experiments directly testing this connection, beyond observing final performance changes?

1 comment

Thank you for pointing us to AdaRound and for asking for controlled tests beyond final performance changes.

How the revision addresses your review. GraphPTQ no longer relies on existing backends: our own Triton integer kernels (GQ-Kernels) run every activation mode. On them, GraphPTQ matches or beats torchao in all 20 W4A4 cells of Tables 1–2 (on the largest model, Uni-Mol2-1.1B, the median OOD loss falls from 25.3% to 8.5%; Table 1). The graph-specific part is now tested causally: on GOOD, raising node degree at fixed graph size raises the activation outliers in 16/16 runs, and per-node scales repair DrugOOD (GPS mean W4A4 OOD loss 16.5→3.6 pp).

Weakness — Standard group-wise quantization and existing backends; the curvature metric follows established second-order analysis such as AdaRound. You are right on both for the earlier version. The quadratic form is not new: AdaRound already relates it to accuracy across stochastic roundings, and we now credit AdaRound, BRECQ and HAWQ (§2, §4.3, App. G). The contribution lies elsewhere: the first study of PTQ for graph models at this scale, with a recipe studied jointly for accuracy and memory, dedicated kernels, and an analysis of generalization under quantization (see Question 1). On backends, torchao, built for LLM layers, breaks on large graph batches (Uni-Mol2-1.1B at batch 128: int32 overflow at W4A16, out of memory at W4A8/W4A4). GQ-Kernels run an integer GEMM in every activation mode, with per-node, per-group scales, an online Hadamard rotation and row chunking for tall batched-graph inputs (§4.2, App. A.3); at batch 8, with 4-bit weights, they need 0.17–0.18× FP32 memory (torchao 0.22–0.36×). Revision: prior work credited; accuracy and memory measured on the same kernels.

Question 1 — What does the diagnostic add beyond second-order sensitivity analysis, and which graph aspects explain it? One aspect changes this picture: in the revision the curvature probe is a moderate ranker of the size of the OOD change and no longer the main result. What the revision adds is the term second-order analyses drop: they set the gradient to zero at the trained weights, but under shift the first-order term g_OOD·ΔW remains and carries most of the variation between roundings for 4–8-bit weights (median R² 0.83–0.84 at 4 bits, ≥ 0.94 at 5–8 bits; it fails at 3 bits). This explains why ID-lossless quantization is not OOD-lossless and when ID validation reflects it (Question 2). The noise model behind it passes its validity checks (spread slope 1.03 / 1.07; calibration error 0.041 / 0.039), though it does not give the sign of one rounding (§6.2, App. E). This loss-level account is not graph-specific. Graph structure enters through activations, which carry ~99% of the W4A4 loss change. One cause we isolate: under sum aggregation the top-degree decile holds the largest activations in 7 of 8 informative message-passing configurations on GOOD and all 12 on DrugOOD (where MHA, without message passing, stays below chance), and swapping sum for mean or max removes the 8-bit collapse (Motif-size only, GIN, 4 seeds per aggregator; App. D).

Question 2 — When can ID curvature reflect the OOD gradient and curvature? Are there controlled tests? Yes; the revision adds them. On the 132 GOOD and DrugOOD models we hold model and data fixed and vary only the rounding (16 random roundings per bit-width, GPTQ, equal-norm random directions). In short, ID curvature reflects OOD curvature, and ID data reflect the OOD gradient only as far as ID and OOD gradients align.

  • OOD curvature (H2; §6.2, App. E.1): GPTQ cuts the ID curvature along its rounding to 0.16× / 0.03× of Hadamard+RTN (random directions: 3.1× / 5.4×), the cut carries over to OOD (Spearman 0.63 [0.43, 0.76]) but does not explain GPTQ's OOD gain beyond first order; where GPTQ lowers the ID loss it lowers the OOD loss in 74% of cases (exploratory; 61% over all models).
  • OOD gradient (H1; §6.2, App. E): ID data reflect only the aligned share ρ² of the first-order OOD term, ≈0.20 on GOOD vs. ≈0.75 on DrugOOD; in line with this, the ID-best of 16 roundings is OOD-better in 66% vs. 88% of cases. This needs OOD data, so it is an analysis, not a deployment diagnostic. Revision: these controlled tests and hypotheses H1–H2 are new (§4.3, §6.2, App. E).

We hope this addresses your concerns.

Reviewer_NmmMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents GraphPTQ, an empirical study of calibration-free post-training quantization applied to graph models. The authors evaluate different weight and activation quantization settings across in-distribution (ID), out-of-distribution (OOD), and zero-shot scenarios, focusing on GPU memory consumption and predictive performance. The paper further introduces a directional-flatness hypothesis, suggesting that the ID loss curvature along the quantization perturbation direction can predict OOD performance degradation.

Critical strengths
  1. Relevant and practical problem. Understanding the memory–accuracy–generalization trade-off of post-training quantization in graph models is important for deploying increasingly large graph models.
  2. Interesting empirical observations. The finding that quantization can affect ID and OOD performance differently is interesting. The hardware-measured memory analysis also provides useful practical insights beyond theoretical compression ratios.
Critical weaknesses
  1. In Table 8, the OOD correlation is significant in only 1 of 5 modes, and none survives multiple-testing correction. It does not consistently strengthen as bit-width drops and is particularly weak at W4A16, the very regime it is meant to explain.

  2. H~b\tilde H_b is identical for W8A16 and W8A8, yet the ID correlation jumps from 0.15 to 0.59 (OOD: 0.13 to 0.43). The metric therefore cannot explain the effects of activation quantization. It is also never compared with simply evaluating the quantized model on ID data, which avoids Hessian computation.

  3. Novelty is limited and the setup is narrow. The paper uses standard round-to-nearest quantization without comparing against GPTQ, AWQ, or SmoothQuant. Results are sensitive to quantization granularity, and the OOD curvature analysis covers only three relatively small GOOD backbones.

  4. The cross-term$\Delta W\epsilon$ is asserted but never quantified. The analysis also does not isolate how activation quantization error depends on input shift.

  5. The memory savings mostly come from BF16, not quantization. On GOOD, BF16 saves 33.3%, versus 33.9% for W8A16. On PCQM, the savings are further confounded by differences in kernel implementation and quantization granularity.

  6. The OOD findings rely on single checkpoints. There are no error bars across independent runs and no random-perturbation baseline.

  7. Some OOD baselines start closer to chance and therefore have less room to degrade. The authors should distinguish this effect from the proposed mechanism.

Questions for authors

Eq. (2) defines the output perturbation as Δy=yq−y\Delta y=y_q-y but includes the term WϵW\epsilon. When comparing quantized and full-precision outputs on the same OOD input, should this term be canceled?

1 comment

Thank you for the precise statistical critique and the requested baselines.

How the revision addresses your review. All GNN results now use 4 training seeds (132 GOOD and DrugOOD models) with confidence intervals, torchao is the deployable baseline on every model, and memory is measured on real integer kernels.

Weakness 1 — The OOD correlation is significant in 1/5 modes, fails correction, and is weak at W4A16. You are right for the earlier 15-model result. With 60 GOOD models (4 seeds per configuration) H̃ is positive in all five modes (Spearman 0.33–0.53; ρ_strat at W4A16 0.40, up from 0.15). Your caution remains fair: under a stricter, configuration-level Holm correction 1 of 5 modes is significant (2 of 5 on DrugOOD), the bit-width trend does not hold at fixed activation precision (W8A16→W4A16: +0.14 [−0.13, +0.38]; App. F.2), and part of H̃'s signal overlaps with the ID drop (see Weakness 2). Revision: the trend claim is withdrawn; H̃ is a moderate ranker of the size of the OOD change.

Weakness 2 — H̃ is the same for W8A16 and W8A8, so it cannot explain activation quantization; no evaluate-on-ID baseline. You are right on both. H̃ is weight-only by construction, and the revision says so: at W4A4 activations carry ~99% of the loss change, and one cause we isolate is activation outliers on high-degree nodes (App. D). The quantized model's ID drop is now a baseline in every probe table, and much of H̃'s signal overlaps with it (App. F.1, F.2).

Weakness 3 — Limited novelty: RTN only, granularity-sensitive, three small backbones. We agree the earlier setup was narrow; the revision broadens it. Granularity matters, so every model now uses per-node, per-group-128 scales. Regarding novelty, this is the first study of PTQ for graph models at this scale, and it is timely: graph foundation models are growing, PTQ is how large models are deployed, and real-world use needs robustness beyond ID benchmarks. The work also proposes: (i) a PTQ recipe studied jointly for accuracy and memory across GNNs, graph transformers, molecular encoders up to 1.1B and zero-shot GFMs, built on a diagnosis of where graph models fail (e.g. degree-driven outliers: on DrugOOD SmoothQuant deepens the worst W8A8 OOD change, −11.5→−26.2%, while per-node scales bring it to −1.1%) and compared with RTN, AWQ, SmoothQuant (App. C.5) and torchao; (ii) dedicated Triton kernels for graph workloads, with hardware-measured GPU memory; (iii) a study of generalization under quantization, under covariate shift and zero-shot transfer, showing that ID accuracy alone does not certify OOD behaviour; and (iv) an analysis of the quantization × shift interaction that explains why, and when ID validation is informative. The OOD analysis stays on GNNs trained from scratch, as it needs many trained models with labelled ID/OOD splits: 4 newly trained seeds per configuration, including 72 new DrugOOD models.

Weakness 4 — The cross-term ΔWε is never quantified; activation error vs. shift is not isolated. You are right; both are now measured. Per quantized layer of the 60 GOOD models, on ID and OOD validation inputs: summed over layers, ‖ΔWε‖ is ≈0.8× the ID weight-quantization term ‖ΔW a_ID‖ (App. F.4). At the loss level, the W8A16 OOD loss change on GOOD regresses at R² 0.92 on the first-order term (β 0.96), while ID curvature alone explains none of it (App. F.3). Under per-tensor scales, activation noise grows with the ID→OOD degree shift (Spearman 0.42; App. E.2), and per-node scales remove the resulting collapses (App. D).

Weakness 5 — Memory savings come mostly from BF16; PCQM is confounded by kernel and granularity. You are right for the small GOOD models: weight-only saves about what BF16 saves (§5.1, App. B.2). Where weights dominate, quantization pays off: Uni-Mol2-1.1B at batch 8 needs 0.17–0.18× FP32 with GraphPTQ 4-bit weights (torchao 0.22–0.36×). At large batch, GraphPTQ W4A8 that also stores activations in INT8 reaches 0.29× (BF16 0.56×) at a median OOD cost of 1.3% (Fig. 3, App. B.1). On PCQM, torchao savings are vs. its own FP32 run (App. B.3), and granularity does not matter at matched head and coverage (App. C.3).

Weakness 6 — Single runs, no error bars, no random-perturbation baseline. All three are added: 4 training seeds with cluster-bootstrap CIs (App. C.6, F.2) and 16 random roundings per bit-width. Random directions rank |Δ_OOD| as well as RTN (App. F.1); GPTQ's direction differs, cutting the curvature to 0.16× / 0.03× of Hadamard+RTN (random: 3.1× / 5.4×).

Weakness 7 — Near-chance OOD baselines have less room to degrade. A valid concern. Normalized by headroom, H̃ still ranks the OOD change (ρ 0.32–0.49), and the earlier finding that OOD degrades less than ID at W4A4 no longer holds under GraphPTQ (11/60 GOOD, 21/72 DrugOOD; App. C.6), so we drop it.

Question — Should Wε cancel in Eq. 2? Yes. The earlier text said it cancels but the equation still listed it; Eq. 2 is corrected (§3).

Comments on the paper

1 comment

In the revision, a new recipe and our own integer kernels match or beat torchao in all 20 W4A4 cells of Tables 1–2; we test it on larger models and multiple training seeds, and explain how quantization interacts with distribution shift.

1. A more accurate recipe (Tables 1–2; NmmM W3, AV36 W). By giving each node its own activation scales, keeping the two GNN input layers at 8-bit activations at W4A4, reducing the 4-bit weight error with a Hadamard rotation and GPTQ, and keeping the most sensitive layers at 8 bit with a guard, GraphPTQ becomes markedly more accurate. Against torchao it is better or equal in 50 of the 60 activation-quantized cells of Tables 1–2, including all 20 at W4A4: the median GOOD loss falls to under 0.6% (previously 31%, torchao up to 17%) and Uni-Mol2-1.1B to 4–9% (per-tensor baseline about 40%, torchao 17–25%). The costly W4A4 cases stay far ahead of torchao: PCQM stays within 2.6 points of its BF16 model where torchao fails, AnyGraph at most 5.2% (torchao up to 21%), and the worst GOOD configuration 10% (torchao 64%; single models up to 25.5%). Ablations test each component (App. C.2).

2. Proposing dedicated kernels for graph workloads, so results reflect real deployment (GQ-Kernels; §4, Fig. 3, App. A.3; hbuY W3, NmmM W5, AV36 W). The revision adds Triton INT4/INT8 kernels that keep weights as integers until the matrix multiply and process the very tall inputs of batched graphs in row chunks, where torchao breaks (Uni-Mol2-1.1B, batch 128: out of memory or int32 overflow). Every GraphPTQ activation mode now runs on them, torchao is the baseline on every model, and accuracy and memory come from the same kernels. The saving comes from the weights: 0.17× FP32 on Uni-Mol2-1.1B at batch 8 (torchao 0.22–0.36×).

3. More models and training seeds, so the conclusions generalize (NmmM W3, W6).

  • Larger models: two pretrained molecular encoders, Uni-Mol2 (five sizes, 84M–1.1B parameters, up to five times the largest earlier model) and Monroe (a 58.6M-parameter graph-transformer encoder).
  • Training seeds: 4 per GNN configuration (132 trained models), so effects are attributed to quantization rather than to one seed.
  • A second, independent benchmark: DrugOOD's curated drug-discovery shifts (assay, scaffold, size): the same three GNNs, 72 models. Hypothesis tests: specified on GOOD, replicated on DrugOOD.

4. The quantization–shift interaction (§4.3, §6.2; hbuY W1/Q4, NmmM W1–W2, W4, AV36 Q1–Q2). We refined the earlier curvature hypothesis, added a second, and tested both on 132 GOOD and DrugOOD models instead of the earlier 15 (weights, 4–8 bits).

  • H1, when ID validation reflects the OOD change: under shift, a first-order quantization × shift term carries most of the OOD change, and ID validation reflects it reliably for some shifts but not for others: the ID-best of 16 random roundings is also OOD-better in 88% of DrugOOD cases but only 66% of GOOD cases.
  • H2, what better rounding can and cannot fix: GPTQ cuts the curvature term (0.16× / 0.03× of Hadamard+RTN; random directions 3.1× / 5.4×), also on OOD data (ρ 0.63); it shrinks the first-order term too but cannot set its sign, so ID-calibrated rounding does not replace shift-aware evaluation.

Corrected or withdrawn, as the reviews asked: the zero-shot Δ definition and Eq. 2 are corrected; the claim that the curvature correlation strengthens at lower bits is withdrawn, and the sign claim is narrowed to the data where the term is measured.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.