Intriguing Properties of Quantization at Scale

Emergent properties have been widely adopted as a term to describe behavior not present in smaller models but observed in larger models. Recent work suggests that the trade-off incurred by quantization is also an emergent property, with sharp drops in performance in models over 6B parameters. In this work, we ask"are quantization cliffs in performance solely a factor of scale?"Against a backdrop of increased research focus on why certain emergent properties surface at scale, this work provides a useful counter-example. We posit that it is possible to optimize for a quantization friendly training recipe that suppresses large activation magnitude outliers. Here, we find that outlier dimensions are not an inherent product of scale, but rather sensitive to the optimization conditions present during pre-training. This both opens up directions for more efficient quantization, and poses the question of whether other emergent properties are inherent or can be altered and conditioned by optimization and architecture design choices. We successfully quantize models ranging in size from 410M to 52B with minimal degradation in performance.

Paper

Similar papers

Peer review

Reviewer TcCw5/10 · confidence 3/52023-07-04

Summary

The paper empirically studies the effect of optimization/regularization hyperparameters -- weight-decay, dropout, gradient-clipping -- on post-training quantization (PTQ) performance of large language models (LLMs). Specifically, it performs PTQ to int8 by simple affine scaling, adopts a GPT-style architecture and trains models with various hyperparameter settings, and finds that higher levels of regularization generally helps lower the performance degradation after quantization. The paper finds that at 52B parameters, training with the best hyperparameters results in a 0.26% average performance degradation, and shows that activation outliers are not a universal emergent property of LLMs. Finally it computes various statistics of weights and activations such as std and spectral norm, for the proposed model v.s. other models, arguing they may offer insight about robustness to quantization.

Strengths

1. The paper is mostly clearly written and easy to follow. 2. The problem setting appears well motivated; I'm not an expert in NLP but a quick read of related work strongly suggests to me the difficulty of quantizing transformer-based language models, in particular when both the network weights and activations are quantized. 3. The proposed method for mitigating performance drop due to quantization is surprisingly simple -- just use a higher amount of regularization. The empirical result (0.26% average performance degradation) seems quite good compared to much larger performance drops experienced by other existing models at int8 precision; however this point remains to be clarified (see weakness).

Weaknesses

1. In the abstract, it is claimed "we find that outlier dimensions are not an inherent product of scale, but rather sensitive to the optimization conditions present during pre-training." Where exactly is this result established? The main experiments (section 4) mostly focused on the performance on zero-shot NLP tasks, rather than outlier dimensions. Moreover, "outliers" are frequently mentioned but not defined until later in the paper, which makes this argument hard to follow and it's unclear what definitive role outliers play in post-training quantization performance. 2. The paper acknowledges that there's many possible factors contributing to a LLM's sensitivity to post-training quantization, including model choices (architecture, number of parameters/scale) and optimization/regularization choices, and recent work show conflicting findings between different model families (line 130-132). While paper focuses on the effect of regularization choices using a GPT style architecture (line 178), can we say anything about the effect of different model architectures on quantized performance? e.g., would the proposed methodology apply to an OPT architecture? As is currently written, the paper cannot exclude the possibility that the relatively small performance drop experienced by their trained model is a result of the model architecture, rather than the choice of optimization/regularization. 3. Related to the above: I'm not familiar with LLM evaluation, and it's unclear how to interpret the 0.26% average degradation of the paper's largest (52B) model (line 260). Is it good or bad? It certainly seems impressive compared to the 42% drop encountered by the most comparable OPT model (line 263), but it seems pretty normal compared to performance drops of 0.09% and 1.36% when using various optimization hyperparameters (Figure 2) within the GPT-style model considered by the paper.

Questions

See weaknesses above.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

N/A

Reviewer xK347/10 · confidence 1/52023-07-07

Summary

Recent research suggests that the trade-off incurred by quantization is an emergent property, resulting in sharp performance drops in models with over 6 billion parameters. In this study, the authors investigate whether quantization cliffs in performance are solely determined by model scale. Considering the increased research focus on understanding the emergence of certain properties at scale, the authors provide a valuable counter-example. The authors propose the possibility of optimizing a quantization-friendly training recipe that suppresses outliers in large activation magnitudes. They find that outlier dimensions are not an inherent consequence of scale but rather sensitive to the optimization conditions during pre-training. This discovery opens up avenues for more efficient quantization methods and raises questions about the inherent nature of other emergent properties and their sensitivity to optimization and architecture design choices. The authors successfully apply their approach to quantize models of various sizes, ranging from 410 million to 52 billion parameters, with minimal performance degradation. This showcases the practical implications of their work and its potential to facilitate effective quantization across a broad range of model scales.

Strengths

The authors conducted a controlled experiment, maintaining consistent architecture while varying key optimization choices. This approach allowed them to systematically investigate the factors influencing sensitivity to quantization. The study rigorously analyzed the impact of optimization choices on sensitivity to quantization in models ranging from 410 million to 52 billion parameters. By carefully controlling these factors, the authors demonstrated that reoccurring activation outliers are not a universal emergent property of large language models at scale. Despite the massive scale of 52 billion parameters, the authors achieved minimal performance degradation (0.26%) across eight tasks when applying quantization to both activations and weights. This showcases the effectiveness of their optimization approach in mitigating the impact of quantization on model performance. The study provides a detailed analysis of activations and weights, offering insights into the key characteristics influencing sensitivity to quantization. These findings can guide future model design and pre-training strategies to enhance robustness and performance.

Weaknesses

Even though the authors do a thorough empirical analysis, their work doesn't begin to explain why we observer this quantization phenomenon.

Questions

None

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

See weaknesses.

Reviewer vYyu6/10 · confidence 4/52023-07-08

Summary

There is a quantization tradeoff in LLMs, where quantization methods drop in particular for some larger models. This problem has been attributed to large activation magnitude "outliers". The authors note that activation outliers are not an emergent property of LLMs and that instead a byproduct of how the networks were trained. The authors compare different training strategies (w/o weight-decay, grad clipping, dropout, half precision) and show that while many configurations can lead to the same performance, after one-shot post training quantization (OPT), the performance can be very different. The authors then train multiple models and show that their results hold for models even at 100B parameters.

Strengths

The paper is clear and easy to follow. It describes the motivation/problem, related work and the contributions clearly.

Weaknesses

The paper proposes a training strategy for improved quantization. - As noted by the authors, other work has also noted that quantization techniques are more stable if the network is trained initialized so that large feature magnitudes are discouraged [1]. Given that this recipe is provided given one architecture I was wondering if the authors could comment on whether they believe the hyperparam configurations will hold across quantization levels/architectures/ pre-training / tasks. Having a paragraph that summarizes this would make the paper more impactful. - One component of the recipe is using bf16. Fig 5 and other work [2] shows that bf16 is better than fp16, Given that OPT is trained used using fp16, I was wondering how much of the recipe is a benefit of using bf16? - Could the authors provide a summary of the ablations of the recipe? [1] Worstman et al. 2023 [2] https://cloud.google.com/blog/products/ai-machine-learning/bfloat16-the-secret-to-high-performance-on-cloud-tpus

Questions

- The authors focus on network retraining for improved OPT quantization. For finetuning many approaches have been developed to improve the weight distribution (such as temperature scaling) Do the authors have an idea of how helpful that would for quantization if applied during finetuning?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

NA

Reviewer egL35/10 · confidence 3/52023-07-28

Summary

The authors study the question of whether quantization (of weights and activations in decoder-only language transformer architectures) fundamentally reduces performance at large parameter counts, or whether this degradation of performance, observed in prior works on this topic, is simply a consequence of certain design choices in the optimizer. They consider post-training quantization (PTQ) schemes, which simply quantize the weights and activations of a pretrained (half-precision) model at inference time (to int8/int8) without any extra quantization-aware retraining, and conduct experiments at the 6B parameter scale (which entail measuring the accuracy drop after PTQ on 8 datasets) varying the magnitude of weight decay, dropout rate, gradient clipping, and half-precision type (fp16 vs. bf16). Their results suggest that larger weight decay, gradient clipping, and bf16 data types improve PTQ performance; they then train with appropriate settings at the 60B parameter level, and demonstrate that they can significantly improve the average PTQ performance relative to a comparable scale model (OPT-66B); Figure 1 highlights these results. The authors then study the weights and activations in the OPT-66B model and their own model in order to assess how their design choices have affected various properties that played a role in prior narratives on why PTQ performance might fundamentally be limited at large scales (specifically, the role of "outlier dimensions", but also general quantization-type metrics such as reconstruction RMSE).

Strengths

The authors' insights lead to a significant improvement over OPT-66B PTQ accuracy drop with a comparable parameter count (the accuracy actually improves on average). Their results also highlight useful general takeaways about how regularization can help mitigate PTQ performance drop, including the utility of bf16, weight decay, and gradient clipping, and they provide reasonable explanations for why. The authors have presented most methodological choices and details of their experimental setup very clearly, which will aid reproducibility of their results and the ability of researchers to build off the results of their study.

Weaknesses

It might be helpful to include additional quantitative comparisons to prior work; currently the comparisons to prior art are substantial but mostly qualitative. iT seems one of the cited works in the related work section, FlexGen, achieves INT4 quantization of OPT-175B on Lambada and WikiText with negligible performance (acc/ppl) penalty. It might also be nice to add brief experimental studies demonstrating improved throughput/memory footprint of the authors' PTQ approach, to substantiate the claims made at a few points in the paper that the authors' approach will have such benefits over the mixed-precision approach of Dettmers et al. The description of the experimental protocol for section 4 is somewhat confusing -- lines 145-150 and 193-198 both suggest that training was done using early checkpoints on 410M parameter models before "scaling" to 6B models, but no results seem to be reported for 410M parameter models. It seems that this detail could be omitted if no results of substance are reported to corroborate it. In addition, it is not made completely clear in section 3.2 whether the 6B experiments reported in section 4 are actually done via training-from-scratch or done from early checkpoints; the text in lines 145-150 suggests that section 4 results are not actually trained from scratch, but the actual protocol in section 3.2 does not clearly state this. The description of the regularization study in the contributions section feels slightly misstated -- the "controlled large-scale study" alluded to does not actually seem to "train from random initialization" (uses early checkpoints), and the hyperparameter search is actually only conducted at 6B parameter models (from early checkpoints). This suggests an implicit hypothesis of this study is that performance differences observed at "medium-scale" models are predictive of those at larger scales; the fact that this is borne out up to 52B parameter models seems to be an interesting consequence of the work, but then it does not seem completely accurate to claim a systematic study up to this scale as a contribution. The takeaways from the weight/activation studies in section 5 do not seem to lend themselves to straightforward conclusions or insights into why the authors' design choices lead to superior PTQ robustness (see discussion at line 316). It might help the presentation if the authors selected a small subset of the activation/weight metrics presented which lend themselves to a coherent interpretation of why the authors' model and OPT/BLOOM have different PTQ performance. One key takeaway that is mentioned both in the introduction and the conclusion, that the authors' work refutes the hypothesis that outliers are not inherent to scaling large language models, is not even presented in the main body (it has been relegated to appendix B.3), and lines 323-324 suggest the authors could not find correlations between outlier detection recipes and PTQ performance. Discussing this study more in section 5 would seem necessary given the way the authors have presented it as a contribution.

Questions

In line 268, performance differences on BLOOM are claimed; are these results available or reported? From the appendix Table 4, it seems that these results are referring to int4 quantization of weights only? The authors do not test their design choices at the full 175B scale, and the reason for this is not discussed. It would be helpful to know whether this is just due to compute obstructions to training at this scale, or whether it is because the authors' methods are not scaling properly to this scale.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations are clearly specified throughout and in the conclusion; there is perhaps somewhat of a missed opportunity to discuss potential negative societal impact of LLM quantization.

Authorsrebuttal2023-08-14

Now that discussion is underway, we wanted to ask **R egL3** if there are any follow up points we can clarify. If there are no further points of clarification regarding the manuscript and the numerical efficiency benchmarking experiments conducted during the rebuttal period demonstrating the improved latency, throughput, and GPU memory usage of our INT8 recipe, we would ask that reviewer **egL3** to consider increasing their score to reflect.

Reviewer TcCw2023-08-15

Thank you for the rebuttal, which has addressed some of my concerns. I raised my score to borderline accept, as my remaining issues are more concerned with presentation, and I trust the authors' to address this in the future revision. 1. Thanks for the additional experiment validating the findings on alternative architectures. This gives me more confidence about the validity of the approach. 2. Thanks for the clarification regarding the 42% drop. Please add this explanation to provide better context for this result in future revisions -- specifically how Dettmers et al.'s setup / training procedure differed from yours that could have led to such a drastic difference. It would also be more convincing to connect these differences to corresponding "bad" hyperparameter settings in your experiments, which also experience bigger accuracy drops. This would essentially corroborate the findings by Dettmers et al. 3. Yes, please discuss early on the definition of "outlier dimension" and why the paper measures it the way it does, and specifically how it relates to std. I still find the claim that "we find that outlier dimensions are not an inherent product of scale, but rather sensitive to the optimization conditions present during pre-training." problematic, as there is no easy way to see this from the experiments. So consider rewriting the statement in terms of the "std/variability" of the parameters (which could indicate quantization error), which the experiments are actually about. Overall, I think the contribution and potential practical impact is valuable enough, but I have issues with the level of rigor and imprecise/overly general claims that can be misleading (see above). The title is also a bit grandiose in my opinion (what are the actual "intriguing properties of quantization at scale"? a more straightforward title about a recipe for mitigating post-training quantization error in LLMs could have served well), as are general claims about showing that "reoccurring activation outliers are not a universal emergent property of LLMs at scale".

Authorsrebuttal2023-08-15

Dear reviewer **TcCw**, thank you for raising your score in light of the clarifications to the manuscript and experiments on the architectural transferability of our method . We are glad that we were able to address your concerns and are grateful for your valuable suggestions which provided us the opportunity to clarify the significance of our results and run additional experiments confirming that our insights on the effect of pertaining optimization choices on PTQ, are applicable to models with architectural differences such as the type of positional embedding. We would like to thank the reviewer again for their positive update and detailed feedback on the presentation of the current manuscript. We will update the final manuscript to address these concerns. Please let us know if there are any other concerns/questions we can address during the rebuttal period and we would be happy to clarify.

Authorsrebuttal2023-08-16

Dear reviewer **vYyu**, we wanted to ask if there are any points that require further clarification. We hope that our response coupled with the additional experiment conducted in the rebuttal phase that show the transferribility of our insights on the effects of pre-training optimization choices on PTQ, to architectures with different types of positional embeddings, address your concern about the transferribility of our method. We also hope that we were able to clarify the significant effect of the choice of fp16 vs bf16 on PTQ. If so, we ask that you consider raising your score to reflect and please do let us know if anything requires further clarification.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC