acceptodds
Under review as a conference paper at ICLR 2027

Layerwise Error Attribution for Fast and Robust Mixed-Precision Post-Training Quantization

Abstract

Mixed-precision post-training quantization is a network compression method that assigns bits layer by layer, under a global memory budget using a small calibration set. The main difficulties are to overcome the combinatorial nature of the allocation problem and to manage the sensitivity to small, potentially corrupted databases. Hence, an efficient allocation method should be fast to compute and preserve model quality when calibration data are corrupted. To design such a method, we derive a layerwise probabilistic analysis of the quantization error that separates propagated error from the local perturbation introduced at a given layer. We use this local term to build a separable score for a simple allocation algorithm, that requires no external solver. The probabilistic nature of our approach brings robustness to corrupted data. On denoising tasks with DRUNet, with an average budget of bits per weight, our method matches or improves state-of-the-art mixed-precision baselines under clean calibration, and is more robust to corrupted calibration, with PSNR gains of up to dB under the tested corruptions. Experiments show bit-allocation speed-ups from to over the studied baselines. For quantized diffusion models, our experiments show that a direct application of our framework also improves the state-of-the-art.

Primary area:
optimization
Keywords:
Post-training quantization, Mixed-precision quantization, Probabilistic error bounds, Calibration robustness

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_bhLJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Mixed-precision post-training quantization must assign per-layer bit-widths under a global memory budget using a small calibration set. This is a combinatorial problem whose sensitivity estimates can be slow to compute and fragile when calibration data are corrupted. The authors derive Cantelli-based probabilistic global and layerwise bounds on the quantization error that separate the error propagated from earlier layers from the local perturbation introduced at each layer. They turn the local term into a separable mean-plus-dispersion score, computed with one full-precision pass per calibration image and used by a solver-free greedy allocation rule.

Critical strengths

The allocation stays stable when the calibration inputs are corrupted. This is shown through matched backends, a CLADO control with identical candidate bit-widths, rank-correlation diagnostics of the allocations, and a replication on BSDS500.

In matched timings the allocation stage is 28 to 2,570 times faster than the studied baselines, and the scores can be reused when only the budget changes. This low cost makes per-layer allocation feasible, which closes the gap to CLADO at the tightest budget

The experimental protocol is careful and transparent, with shared quantizers and block partitions in each comparison, paired confidence intervals, qualitative examples chosen by fixed random seeds.

Critical weaknesses

The claim that the probabilistic nature of the approach brings robustness to corrupted data is not tested by any ablation of the score, such as a mean-only, max-based, data-free or varying delta score variant. The robustness may come from the dominance of weight-quantization error rather than from the probabilistic analysis, and this would change the main message of the paper.

The algorithm is a heavily simplified version of the theory, and the derived bounds are never computed in experiments. The theory therefore motivates the method but does not support it directly.

The evaluation uses only one denoising network, weight-only quantization and two masking corruptions, and the large gains appear only against CLADO. It is unclear whether the conclusions hold more broadly.

The image pairing with baselines could not be fully verified. The claim of improving the state of the art is too strong.

Questions for authors

How was CLADO’s loss computed under corrupted calibration, and does its collapse also appear with milder corruptions?

2 comments

Thank you for your review. The additional experiments are described in our responses to m1Uc and Bukf.

  • New models tested: FFDNet and pixel-space DDPM.
  • New tests: ablations, corruptions and recomputed scores.

1. Score ablations

The claim that the probabilistic nature of the approach brings robustness to corrupted data is not tested by any ablation of the score, such as a mean-only, max-based, data-free or varying delta score variant.

Variants perform similarly on DRUNet. On heterogeneous FFDNet calibration, our score gains 4.7 dB over the mean and 3.5 dB over standard deviation alone. See m1Uc, point 1, for all ablations and allocation stability.

The data-free score ∑ℓ∈g∥Wℓ−Qb(Wℓ)∥F\sum_{\ell\in g}\lVert W_\ell-Q_b(W_\ell)\rVert_F reaches 21.3 dB on the CIFAR-10 DDPM (A6), versus 25.1 dB for our score, on 32 images. Thus, removing dependence on calibration data does not necessarily improve quality.

These ablations support the combined score in the tested settings. We will revise the abstract to present robustness as an empirical finding.

2. Role of weight-quantization error

The robustness may come from the dominance of weight-quantization error rather than from the probabilistic analysis, and this would change the main message of the paper.

The controlled ablations in m1Uc, point 1, change only aggregation. Their PSNR gains show that the aggregation statistic matters even with the same local perturbations and allocation algorithm.

3. Link between the theory and the algorithm

The algorithm is a heavily simplified version of the theory, and the derived bounds are never computed in experiments. The theory therefore motivates the method but does not support it directly.

Recomputing scores from partially quantized activations changes PSNR by at most 0.02 dB in our tests. See m1Uc, point 3, for the protocol and the equivalent exact decomposition supporting FP32 activations.

We have now evaluated recursive bounds on DRUNet. They exceed the actual error by factors of about 104110^{41} with operator norms and With data-estimated factors, the gap drops to about 10810^8 but still too loose to be used as the exact score for the algorithm. That is why we introduce our heuristic.

4. Scope of the evaluation

The evaluation uses only one denoising network, weight-only quantization and two masking corruptions, and the large gains appear only against CLADO. It is unclear whether the conclusions hold more broadly.

On FFDNet, we exceed CLADO by 0.9 dB with clean calibration and 2.9 dB with mostly periodic calibration noise. See Bukf, points 2 and 3, for the 10 corruptions, cross-dataset tests, and second diffusion model.

The benefits are not limited to CLADO. The paper already reports a 1.8 dB gain over OMPQ-ORM at 4 bits, with much faster allocation than OMPQ and AIMET. In an FFDNet comparison with calibration intensities multiplied by 4, AIMET loses 3.5 dB, versus 0.13 dB for our method. The final gap is 1.96 dB in our favor, at about 4.5 bits for both methods. In the DRUNet comparison with the AIMET backend under photon noise, AIMET changes 18 of 36 blocks, versus one for our method.

The paper and our experiments use weight-only quantization. Extending the local score to activation quantization is an interesting direction for future work. For the same input zz and fixed quantizers, we can separate the local weight and activation errors:

Wz-Q_W(W)Q_A(z)=(W-Q_W(W))z+Q_W(W)\bigl(z-Q_A(z)\bigr).\tag{4}

This identity provides a starting point for the extension.

5. Image pairing in diffusion

The image pairing with baselines could not be fully verified. The claim of improving the state of the art is too strong.

We directly checked the initial noise tensors and the first U-Net input for 1,000 images. They are identical across FP32, our method, and the baseline examined.

On this verified set, our output is closer to FP32 in 992 of 1,000 cases. The mean PSNR gain is 10.1 dB, with a 95% confidence interval of [9.85,10.34][9.85,10.34] dB. This confirms the gain under directly verified image pairing. We will replace "improves the state of the art" with "matches or improves the studied baselines."

6. CLADO loss and milder corruptions

How was CLADO’s loss computed under corrupted calibration, and does its collapse also appear with milder corruptions?

We use CLADO's official code for individual and pairwise sensitivities, followed by integer quadratic optimization. Following its task-loss formulation, we use the denoising loss

L(Q)=\operatorname{mean}\lVert f_Q(\widetilde x)-y\rVert_2^2,\tag{5}

where yy is the clean image. Both methods use the same blocks, quantizer, candidates, and budget. Under corruption, only the calibration input changes.

The degradation appears with only 10% of pixels masked: at 4 bits, CLADO drops from 34.41 to 28.15 dB. Our allocation remains unchanged and gives higher PSNR on all 100 test images.

Thank you for the detailed response and the additional experiment. I checked the image pairing in the diffusion experiment and found that the authors had corrected the related descriptions. My concern about this part has therefore been resolved.

I still have concerns about the theoretical part. The upper bound reported in the response is around 10810^8 to 104110^{41} times larger than the actual error. The authors also acknowledged that the proposed score is heuristic. In this regard, the theory seems to serve mainly as motivation for the method, rather than providing direct support for the algorithm or the experimental results.

The other concerns have only been partially addressed. In the DRUNet experiments, the mean, max, and combined scores produced the same assignments across all 10 corruption conditions. The advantage of the combined score was also shown only in a limited set of experiments. Regarding the rank-matching experiment raised by the Bukf reviewer, the evaluation was performed with only one block quantized at a time. It is still unclear whether the same behavior holds when multiple blocks are quantized simultaneously at low bit-widths. The method is also still limited to weight-only quantization. I think these limitations and additional results should be included in the revised paper.

I also have a remaining concern about the CLADO comparison. According to the response, CLADO compares the output from a corrupted input against a clean image, while the proposed score does not require such a target. This makes it difficult to tell whether CLADO's performance drop under masking is due to its actual robustness or simply to the difference in the loss definition. The current results do not seem sufficient to distinguish between these possibilities. It would be helpful to include a CLADO variant that uses the FP32 output from the same corrupted input as its target.

Considering these remaining issues, I will keep my score at 2.

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

This paper proposes a mixed-precision PTQ framework based on layerwise quantization error attribution. The authors derive probabilistic error bounds and introduce a local sensitivity score for efficient greedy bit-width allocation. Experiments on image denoising and latent diffusion demonstrate competitive performance, improved robustness to corrupted calibration data, and substantially reduced allocation time.

Critical strengths

Theoretical motivation. The layerwise error analysis provides an interpretable foundation for mixed-precision quantization by considering both propagated and local quantization errors.

Computational efficiency. The proposed solver-free greedy allocation achieves substantial speed-ups over existing methods while maintaining competitive quantization performance.

Experimental validation. The experiments include comparisons with several baselines, different memory budgets, and corrupted calibration settings. The additional diffusion experiments further demonstrate the potential applicability of the method.

Critical weaknesses

Gap between theory and implementation. The theoretical analysis considers both propagated and local quantization errors, whereas the practical sensitivity score retains only local perturbations and replaces quantized activations with full-precision ones. The validity of these approximations, particularly at lower bit-widths, is not sufficiently evaluated.

Limited robustness evaluation. Although the experiments demonstrate robustness under the tested calibration corruptions, the evaluation mainly focuses on random masking. Additional perturbations, such as blur, compression artifacts, or calibration distribution shifts, would help establish the generality of the reported robustness.

Limited generalizability evidence. The diffusion experiments are restricted to a single latent diffusion architecture and dataset. Evaluations on additional architectures or generation tasks would further support the generalizability of the proposed method.

Questions for authors

How closely do the proposed local sensitivity rankings agree with those obtained from a more complete error measure, particularly at lower bit-widths?

How does the proposed method perform under other types of calibration corruption or distribution shifts?

Can the authors further discuss or demonstrate the applicability of the method to different diffusion architectures?

2 comments

Thank you for your review. We have added score diagnostics and extended the experiments.

Additional experiments used in the score comparisons:

  • A2: DRUNet allocation stability under 10 calibration corruptions.
  • A5: DRUNet scores recomputed from the quantized model's activations instead of FP32 activations.
  • A6: Transfer to an additional diffusion architecture, a pixel-space DDPM.

1. Local error and propagation

How closely do the proposed local sensitivity rankings agree with those obtained from a more complete error measure, particularly at lower bit-widths?

The exact decomposition in Equation (1) of the paper, used in Theorem 1, can also be written as

\Delta_\ell=W'_\ell\Delta_{\ell-1}+(W_\ell-W'_\ell)S_{\ell-1}.\tag{2}

The local term then uses full-precision activations and directly matches the perturbation used in our score. The same conditional Cantelli and induction argument applies with the corresponding propagation factors, activation bounds, and conditional moments.

We tested both the use of full-precision activations and the local sensitivity rankings.

In A5, we compared our standard allocation, which computes scores once using FP32 activations, with a version that recomputes them from the partially quantized network when the average bit-width falls below 6 and then 4.5 bits.

On CelebA and BSDS500, across three seeds, recomputing the scores changes final PSNR by at most 0.02 dB at budgets of 3.5 and 4 bits. This supports computing the scores only once.

We compare our local score with a score based on final output error relative to FP32. For each one-bit reduction, blocks are ranked by the score increase per bit saved. The rankings agree on 82–91% of block pairs, including reductions from 4 to 3 bits. The output-based score is evaluated with only the tested block quantized.

2. Other corruptions and distribution shifts

How does the proposed method perform under other types of calibration corruption or distribution shifts?

We tested 10 conditions (A2): 2 masks, 2 Gaussian blurs, JPEG compression, 2 photon-noise levels, 2 intensity gains, and impulse noise. Only calibration changes; test inputs and targets remain fixed.

On DRUNet, with eight calibration images and the same candidates for both methods, our 4-bit allocation remains unchanged. CLADO loses 3.7–6.1 dB in seven conditions.

Calibrating on BSDS500 and testing on CelebA, or the reverse, changes our PSNR by +0.03+0.03 and −0.05-0.05 dB, respectively. On a second denoiser, FFDNet, we exceed CLADO by 0.9 dB with clean calibration and by 2.9 dB when periodic noise accounts for 95% of the calibration-noise power. These results use three seeds at a 3.5-bit budget, with a slightly lower average bit-width for our method.

3. Another diffusion model

Can the authors further discuss or demonstrate the applicability of the method to different diffusion architectures?

We applied the same score and greedy algorithm to the DDPM of Ho et al. (2020) on CIFAR-10 (A6). Unlike the latent model in the paper, this model operates in pixel space without an autoencoder. We use 100-step DDIM sampling, candidates from 3 to 8 bits, and an average budget of 4 bits.

On 32 images generated from the same initial noise, PSNR relative to FP32 reaches 25.1 dB, versus 21.3 dB for a score based only on the weights and 13.3 dB for uniform 4-bit quantization. These results show that the method transfers to a second diffusion model.

Thanks to the authors for their response. My concerns have been addressed, and I maintain my Weak Accept rating

Reviewer_m1UcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper separates propagated and local quantization errors through a probabilistic analysis. Its allocator scores local perturbations on full-precision activations using their empirical mean and standard deviation, then greedily reduces precision under a weight-memory budget. Experiments evaluate DRUNet under clean and corrupted calibration, allocation speed, additional controls, and latent-diffusion transfer.

Critical strengths
  1. Efficient, reusable allocation. Scores can be reused across budgets without an external solver. At approximately four bits, clean DRUNet quality is comparable to CLADO while scoring and allocation are substantially faster, making the method useful for budget exploration.

  2. Meaningful experimental controls. Matched candidates, an output-SQNR control, additional datasets, and finer allocation granularity strengthen the evaluation. The advantage over CLADO under calibration corruption persists with matched W3/W4/W8 candidates.

Critical weaknesses
  1. The contribution of the probabilistic score is not fully isolated. Section 5 deliberately replaces the full error recursion with a local mean-plus-standard-deviation score. Existing controls do not establish whether dispersion improves on mean-only local error under clean or corrupted calibration: output-SQNR changes the error measure, so it does not isolate this term. This matters to attributing the gain to probabilistic scoring rather than local scoring and greedy allocation.

  2. The abstract's robustness explanation is stronger than the established link. Section 5 explicitly denies an allocation-confidence interpretation, and Appendix A.5 correctly limits its sensitivity result to one bounded-value replacement. Nevertheless, the abstract attributes corruption robustness to the probabilistic formulation. Neither result establishes that causal explanation for image-wide masking. The demonstrated robustness is valuable, but this attribution should remain empirical unless separately supported.

This is a close weak rejection; the concern is contribution attribution, not a requirement for global optimality. I would consider raising the rating to weak acceptance if the rebuttal isolates the score's contribution under clean and corrupted calibration and clearly separates theoretical guarantees from empirical robustness.

Questions for authors
  1. With all other settings fixed, how do mean-only, standard-deviation-only, and combined scores compare in PSNR and allocation stability? How sensitive are results to the dispersion coefficient?

  2. Can you reconcile the abstract's causal robustness claim with the explicitly limited guarantees in Section 5 and Appendix A.5, and state what evidence connects the two?

  3. At low budgets, how well do scores computed on full-precision activations predict errors after multiple blocks are quantized?

1 comment

Thank you for these questions. We have added the requested ablations and clarified the link between the theory and the observed robustness.

Additional experiments used in the score comparisons:

  • A1: DRUNet score and dispersion-weight ablations in the paper's settings.
  • A2: DRUNet allocation stability under 10 calibration corruptions.
  • A3: Score ablations on FFDNet, an additional denoising architecture.
  • A4: Score ablations on the latent-diffusion model used in the paper.
  • A5: DRUNet scores recomputed from the quantized model's activations instead of FP32 activations.

1. Contribution of the score statistics

With all other settings fixed, how do mean-only, standard-deviation-only, and combined scores compare in PSNR and allocation stability? How sensitive are results to the dispersion coefficient?

For each block and candidate bit-width, we keep the same local perturbations on the calibration images. We change only how these values are summarized: mean, standard deviation, maximum, or our combined score. The quantizer, candidates, budgets, and greedy algorithm remain unchanged.

On the original DRUNet configurations (A1), performance is close: the mean gap to the best PSNR among these variants ranges from 0.011 to 0.2 dB.

A2 tests 10 corruption conditions: 2 masks, 2 Gaussian blurs, JPEG compression, 2 photon-noise levels, 2 intensity gains, and impulse noise. At a 4-bit budget, our score, the mean, and the maximum keep their clean calibration allocation. Standard deviation alone changes up to 8 of 36 blocks.

We kept δscore=0.05\delta_{\mathrm{score}}=0.05, without tuning it on the new results. For δscore=0.5\delta_{\mathrm{score}}=0.5, 0.050.05, and 0.010.01, the mean gap to the best PSNR on DRUNet (A1) is 0.015, 0.011, and 0.024 dB, respectively. Performance is thus stable across these three values.

The combined score gives higher PSNR in A3 and A4:

PSNR differences are computed as variant minus our combined score, in dB. Negative values favor our combined score.

PSNR difference: variant - oursMeanMaximumStd. dev.
FFDNet, heterogeneous calibration, 3.5 bits−4.7-4.7−1.3-1.3−3.5-3.5
Latent diffusion, 4 bits−0.16-0.16−0.16-0.16−1.96-1.96

The FFDNet test uses one seed and 129 calibration images: 112 underexposed, 16 normal, and one with strong periodic interference. Realized average bit-widths differ by at most 0.04 bits. The diffusion test uses 250 paired images.

2. Section 5, Appendix A.5, and robustness

Can you reconcile the abstract's causal robustness claim with the explicitly limited guarantees in Section 5 and Appendix A.5, and state what evidence connects the two?

Section 5 defines the score used by the algorithm. Appendix A.5 bounds the change in this same score when one bounded observation is replaced. This is a score-level result, not a guarantee on the allocation or its final quality under corruption of the whole calibration set.

We will revise the abstract to present robustness as an empirical finding. The new ablations show the benefit of combining mean and dispersion on the tested heterogeneous calibration sets. The parameter δscore\delta_{\mathrm{score}} controls the weight of dispersion, it is not a confidence level for the allocation.

3. Full-precision activations at low budgets

At low budgets, how well do scores computed on full-precision activations predict errors after multiple blocks are quantized?

The exact decomposition in Equation (1) of the paper, used in Theorem 1, can also be written as

\Delta_\ell=W'_\ell\Delta_{\ell-1}+(W_\ell-W'_\ell)S_{\ell-1}.\tag{1}

The local term then uses full-precision activations and directly matches the perturbation used in our score. The same conditional Cantelli and induction argument applies with the corresponding propagation factors, activation bounds, and conditional moments.

In A5, we compared our standard allocation, which computes scores once using FP32 activations, with a version that recomputes them from the partially quantized network when the average bit-width falls below 6 and then 4.5 bits.

On CelebA and BSDS500, across three seeds, recomputing the scores changes final PSNR by at most 0.02 dB at budgets of 3.5 and 4 bits. This supports computing the scores only once.

Comments on the paper

1 comment
Area_Chair_rRs9OpenReview ↗

Reminder: Respond to the Authors’ Rebuttal

All reviewers: The authors have made a sincere effort to address your concerns and questions. Please respond to their rebuttal as soon as possible.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.