Laughing Hyena Distillery: Extracting Compact Recurrences From Convolutions

Recent advances in attention-free sequence models rely on convolutions as alternatives to the attention operator at the core of Transformers. In particular, long convolution sequence models have achieved state-of-the-art performance in many domains, but incur a significant cost during auto-regressive inference workloads -- naively requiring a full pass (or caching of activations) over the input sequence for each generated token -- similarly to attention-based models. In this paper, we seek to enable $\mathcal O(1)$ compute and memory cost per token in any pre-trained long convolution architecture to reduce memory footprint and increase throughput during generation. Concretely, our methods consist in extracting low-dimensional linear state-space models from each convolution layer, building upon rational interpolation and model-order reduction techniques. We further introduce architectural improvements to convolution-based layers such as Hyena: by weight-tying the filters across channels into heads, we achieve higher pre-training quality and reduce the number of filters to be distilled. The resulting model achieves 10x higher throughput than Transformers and 1.5x higher than Hyena at 1.3B parameters, without any loss in quality after distillation.

Paper

Similar papers

Peer review

Reviewer n6Pj6/10 · confidence 4/52023-07-05

Summary

This article addresses two issues: 1. The low efficiency of the LonvConv-based model during inference. 2. Whether it is advantageous to perform independent long convolutions on each channel or reduce the total number of filters without loss in quality. To tackle problem (1), the authors propose distilling the LonvConv into a Diagonal State space model and train it using the $\ell_2$ loss function. For problem (2), the authors suggest sharing long convolution coefficients across multiple channels, resulting in the MultiHyena structure. The effectiveness of the proposed methods is validated on multiple datasets.

Strengths

Distilling LonvConv into a Diagonal State space model is indeed a novel and meaningful approach. The conclusion of sharing long convolution coefficients across multiple channels is also innovative.

Weaknesses

I think the main issue with this article lies in the focus of the writing. From Equation 3.4, it is clear that the ultimate goal is to represent the LonvConv coefficients using a Diagonal State Space Model. However, a significant portion of the article is spent describing unrelated aspects. The most crucial part, determining the hidden dimension of the State Space Model, is merely brushed over, even though it is the key factor that affects the quality of the distillation and the efficiency of the final inference. On the other hand, the motivation behind the design of MultiHyena should be addressed in the main text since it is of utmost importance.

Questions

1. The method for determining the hidden dimension of the State Space Model needs to be explained in more detail. I referred to Appendix E.3.2, and I'm wondering if the core idea is to perform an SVD decomposition and then select the dimension for dimensionality reduction based on the eigenvalues? 2. The solution to Equation 3.4 requires a more detailed algorithm description or pseudocode to help readers follow along. I attempted to replicate it following Appendix B.1, but the results were not quite good. Could you provide the training configurations and an estimate of the training time? 3. The motivation behind the design of MultiHyena should be included in the main text, and ablation studies (sharing coefficients vs not sharing coefficients) should be conducted to validate the design's rationale. The experiments should compare the effects and speeds. 4. Does MultiHyena utilize the Local conv + Global conv structure of Hyena? If so, how many layers are used? This should be mentioned in the experiments. 5. Is the Algorithm 1 on page 28 is the implementation of Figure 4.1? 6. Regarding the implementation of MultiHyena, in Algorithm 1 on page 28, for $z^m_t \in \mathbb R^{L\times N\times N}$, what does the subscript $t$ represent? On the other hand, is the shape of $T_h$ $L$ (all features share one set of convolution coefficients) or $L\times N\times N$ (each feature has independent convolution coefficients)? 7. Continuing with the implementation of MultiHyena, in Algorithm 1 on page 28, is the shape of $T_hz^m_t$ ${L\times N\times N}$? If so, does $T_h(z^m_t) q_t^m$ mean performing matrix multiplication between $[T_h(z^m_t)]_i \in \mathbb R^{N\times N}$ (for $i=0,\ldots,L-1$) and $q_t^m\in \mathbb R^{N}$, resulting in an output of shape $\mathbb R^{N}$? If not, what is the computation like? 8. The algorithm's output is $\bar{y} \leftarrow\left(\sum_m\right) y^m / M\in \mathbb R^{L\times N}$, while the input has a shape of $L\times D$. Is this inconsistency problematic, or did the algorithm omit something? 9. Algorithm 1 has several ambiguities. It is suggested to reorganize it for better clarity.

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

3 good

Contribution

3 good

Limitations

Yes.

Reviewer 6dqz7/10 · confidence 4/52023-07-06

Summary

The paper proposes distilling convolutional models for autoregressive sequence generation into recurrent (state-space) models. A key limitation of recently proposed convolutional models is that they use convolutional filters that extend potentially infinitely into the past, and the same techniques that yield efficient training do not transfer over to token-by-token generation. This paper proposes a post-training and data-free distillation step that replaces such convolutional models with a recurrent approach that uses constant time and memory at each step of generation during inference.

Strengths

The key strength of the paper is that proposes a novel and theoretically grounded distillation method, which achieves good approximation bounds without being tied to specific data or involving what amounts to an additional round of training. Also, convolutional and state space models as a whole are an underexplored area in recent work when compared to Transformers and attention, and this paper puts forward a novel method of linking the two, both of which contribute to the originality of the paper.

Weaknesses

My biggest reservation based on my understanding of the paper is that I didn't get a qualitative sense of what might be lost as part of the distillation process. Is there some intuition of what is the worst-case qualitative behavior of a convolutional filter that can't be tightly approximated with LaughingHyena? This question comes to mind because in the world of Transformers, it is no secret that most of the computational power spent on quadractic attention goes to waste. Just a small subset of the efficiency work there includes pruning entire heads, limiting each head to a head-specific attention context window, or even doing sliding-window attention with a fixed context length for the entire model. More in line with the present paper, work like Performer has developed an approximation for converting attention-based models into recurrent models -- at a cost. Notwithstanding the theory of the tightness of that last approximation, and equivalent performance of many to the Transformer that is demonstrated in some of the papers introducing these methods, there inevitably arises some situation where none of the approximations match the Transformer in quality. I worry that the present approach might fall into a similar pattern. A recurring theme in this area is that any method that sacrifices the ability to have long context, or to perform associative recall, is suspect. That's why when it comes to these approximations of convolutions, it would helpful to know whether the approximation is effectively some form of context-truncation in disguise, and if not what the qualitative cost is.

Questions

How well does the LaughingHyena architecture perform on the associative recall task, especially as compared to Hyena (or MultiHyena)? Is the point beyond which the models fail to perform the task (in terms of sequence length or vocabulary size) different between the two? For LM-Eval-Harness and HELM, have you tried a baseline of taking the impulse response from Hyena/MultiHyena and truncating it to a finite impulse response? A sliding window seems like one of the simplest approximations to try in the world of convolutions, and it would be helpful to know if maybe some defect of the tasks or the base model results in nothing more being required. This would, in fact, be a useful baseline to have in the paper.

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

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

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

The paper could be improved with a little bit of additional discussion regarding limitations of the distillation/approximation.

Reviewer wEPT7/10 · confidence 2/52023-07-09

Summary

This paper proposes LaughingHyena - an improvement to the Hyena model that can perform long-convolutions instead of attentions in transformers to avoid the quadratic scaling issues. One of the issues with the convolution sequence models is that they incur significant cost due to autoregressive inference. To avoid this, this paper seeks to come up with a techinique to have constant memory recurrent inference to increase generation thoroughput. This is achieved using the use of compact linear SSM and weight tying the filters across heads in Hyena architecture. The resulting performance improvements are impressive - 1.5x throughput improvement compared to Hyena. The model also achieves SOTA in the PILE dataset.

Strengths

- The perplexity on small-scale models on Table 1 and Table 2 outperform GPT, Hyena and establishes a new SOTA. - The peak memory usage is also constant for different sequence lengths in Table 5.4

Weaknesses

- The writing is a bit hard to follow. - The performance of the model is still lacking compared to full transformer baseline such as Pythia in Table 5.3. Can the authors comment on this? Any idea on how much the performance degradation will be on very large scale models and datasets?

Questions

- What assumptions do you use for the state-space model in Eq 3.1 to yield a good distillation results (d<<L)?

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

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

I think the writing can be improved to provide a simple explanation of the method for readers who don't have a strong understanding of state-space models. Else, the text is hard to follow.

Reviewer 4kyN7/10 · confidence 5/52023-08-04

Summary

This paper proposes an approach that enables constant-memory, recurrent inference for long convolution architectures. They introduce LaughingHyena, a distilation approach that consists of extracting compact linear SSMs from each convolution layer. Combined with weight-tying, it results in state-of-the-art performance and efficiency (i.e. throughput) without any drop in quality.

Strengths

The paper is well structured and written. The approach seems sound, reasonable and is performant

Weaknesses

The models used in most experiments are small. The helm evaluation is not very convincing.

Questions

Is it possible to benchmark against more recent open source models such as Llamma?

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

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Broader Impacts section is missing.

Reviewer 4kyN2023-08-15

Update

This response together with the general author response addresses my key questions. I have updated my score to reflect this. Please update this article with the new information and tables.

Authorsrebuttal2023-08-20

Response

We are glad to hear our response was useful! Thank you again for the review and engagement.

Reviewer 6dqz2023-08-19

Thank you for the update. The results on the associative recall task and the comparison to a truncation baseline have help convince me that the distillation approach maintains interesting model behavior beyond what could be achieved with simple context truncation. I would recommend including these results in future revisions of the paper (or supplemental). I've updated my score based on the author response.

Authorsrebuttal2023-08-20

Response

Thank you for the detailed review and for suggesting interesting ablations!

Authorsrebuttal2023-08-20

Response

Thank you again for the review and the positive comments! As the discussion period is about to end, we leave a gentle reminder to take into account our response (including the additional evaluations and modifications to improve clarity) into the final evaluation of the work.

Authorsrebuttal2023-08-20

Response

Thank you for the comments and review! Since the discussion period is about to end, we leave a gentle reminder to take into account our response into the final evaluation of our work.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC