Zhao Song

Zhenyu Zhang, Ying Sheng, Tianyi Zhou et al.

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H$_2$). Through a comprehensive investigation, we find that (i) the emergence of H$_2$ is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H$_2$O), a KV cache eviction policy that dynamically retains a balance of recent and H$_2$ tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H$_2$O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29$\times$, 29$\times$, and 3$\times$ on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9$\times$. The code is available at https://github.com/FMInference/H2O.

Jerry Yao-Chieh Hu, Weimin Wu, Zhao Song et al.

We investigate the statistical and computational limits of latent Diffusion Transformers (DiTs) under the low-dimensional linear latent space assumption. Statistically, we study the universal approximation and sample complexity of the DiTs score function, as well as the distribution recovery property of the initial data. Specifically, under mild data assumptions, we derive an approximation error bound for the score network of latent DiTs, which is sub-linear in the latent space dimension. Additionally, we derive the corresponding sample complexity bound and show that the data distribution generated from the estimated score function converges toward a proximate area of the original one. Computationally, we characterize the hardness of both forward inference and backward computation of latent DiTs, assuming the Strong Exponential Time Hypothesis (SETH). For forward inference, we identify efficient criteria for all possible latent DiTs inference algorithms and showcase our theory by pushing the efficiency toward almost-linear time inference. For backward computation, we leverage the low-rank structure within the gradient computation of DiTs training for possible algorithmic speedup. Specifically, we show that such speedup achieves almost-linear time latent DiTs training by casting the DiTs gradient as a series of chained low-rank approximations with bounded error. Under the low-dimensional assumption, we show that the statistical rates and the computational efficiency are all dominated by the dimension of the subspace, suggesting that latent DiTs have the potential to bypass the challenges associated with the high dimensionality of initial data.

Xuan Shen, Zhao Song, Yufa Zhou et al.

Transformers have emerged as the leading architecture in deep learning, proving to be versatile and highly effective across diverse domains beyond language and image processing. However, their impressive performance often incurs high computational costs due to their substantial model size. This paper focuses on compressing decoder-only transformer-based autoregressive models through structural weight pruning to improve the model efficiency while preserving performance for both language and image generation tasks. Specifically, we propose a training-free pruning method that calculates a numerical score with Newton's method for the Attention and MLP modules, respectively. Besides, we further propose another compensation algorithm to recover the pruned model for better performance. To verify the effectiveness of our method, we provide both theoretical support and extensive experiments. Our experiments show that our method achieves state-of-the-art performance with reduced memory usage and faster generation speeds on GPUs.

Josh Alman, Zhao Song

Attention mechanisms lie at the heart of modern large language models (LLMs). Straightforward algorithms for forward and backward (gradient) computation take quadratic time, and a line of work initiated by [Alman and Song NeurIPS 2023] and [Alman and Song NeurIPS 2024] has shown that quadratic time is necessary unless the model weights are small, in which case almost linear time algorithms are possible. In this paper, we show that large weights are necessary to avoid a strong preclusion to representational strength we call layer collapse, which means that the entire network can be approximated well by a network with only a single layer. Thus, the quadratic running time of attention is unavoidable for expressive transformers. The notion of layer collapse that we introduce is a variant on the notion of rank collapse from the work of [Dong, Cordonnier, and Loukas ICML 2021]. They showed that in Self Attention Networks with small weights and with skip connections, rank collapse must occur. This is typically interpreted as justifying the necessity of skip connections in expressive networks. However, our result shows that even with skip connections, if the weights are small, then layer collapse still occurs. Thus, only large weights, and not skip connections, can prevent these representational weaknesses.

Zhihang Li, Zhao Song, Weixin Wang et al.

Leverage score is a fundamental problem in machine learning and theoretical computer science. It has extensive applications in regression analysis, randomized algorithms, and neural network inversion. Despite leverage scores are widely used as a tool, in this paper, we study a novel problem, namely the inverting leverage score problem. We analyze to invert the leverage score distributions back to recover model parameters. Specifically, given a leverage score $σ\in \mathbb{R}^n$, the matrix $A \in \mathbb{R}^{n \times d}$, and the vector $b \in \mathbb{R}^n$, we analyze the non-convex optimization problem of finding $x \in \mathbb{R}^d$ to minimize $\| \mathrm{diag}( σ) - I_n \circ (A(x) (A(x)^\top A(x) )^{-1} A(x)^\top ) \|_F$, where $A(x):= S(x)^{-1} A \in \mathbb{R}^{n \times d} $, $S(x) := \mathrm{diag}(s(x)) \in \mathbb{R}^{n \times n}$ and $s(x) : = Ax - b \in \mathbb{R}^n$. Our theoretical studies include computing the gradient and Hessian, demonstrating that the Hessian matrix is positive definite and Lipschitz, and constructing first-order and second-order algorithms to solve this regression problem. Our work combines iterative shrinking and the induction hypothesis to ensure global convergence rates for the Newton method, as well as the properties of Lipschitz and strong convexity to guarantee the performance of gradient descent. This important study on inverting statistical leverage opens up numerous new applications in interpretation, data recovery, and security.

Jerry Yao-Chieh Hu, Maojiang Su, En-Jui Kuo et al.

We study the computational limits of Low-Rank Adaptation (LoRA) for finetuning transformer-based models using fine-grained complexity theory. Our key observation is that the existence of low-rank decompositions within the gradient computation of LoRA adaptation leads to possible algorithmic speedup. This allows us to (i) identify a phase transition behavior of efficiency assuming the Strong Exponential Time Hypothesis (SETH), and (ii) prove the existence of almost linear algorithms by controlling the LoRA update computation term by term. For the former, we identify a sharp transition in the efficiency of all possible rank-$r$ LoRA update algorithms for transformers, based on specific norms resulting from the multiplications of the input sequence $X$, pretrained weights ${W^\star}$, and adapter matrices $αB A/r$. Specifically, we derive a shared upper bound threshold for such norms, and show that efficient (sub-quadratic) approximation algorithms of LoRA exist only below this threshold. For the latter, we prove the existence of almost linear approximation algorithms for LoRA adaptation by utilizing the hierarchical low-rank structures of LoRA gradients and approximating the gradients with a series of chained low-rank approximations. To showcase our theory, we consider two practical scenarios: partial (e.g., only $W_V$ and $W_Q$) and full adaptations (e.g., $W_Q$, $W_V$, and $W_K$) of weights in attention heads.

Yuzhou Gu, Zhao Song, Junze Yin

Softmax distributions are widely used in machine learning, including Large Language Models (LLMs), where the attention unit uses softmax distributions. We abstract the attention unit as the softmax model, where given a vector input, the model produces an output drawn from the softmax distribution (which depends on the vector input). We consider the fundamental problem of binary hypothesis testing in the setting of softmax models. That is, given an unknown softmax model, which is known to be one of the two given softmax models, how many queries are needed to determine which one is the truth? We show that the sample complexity is asymptotically $O(ε^{-2})$ where $ε$ is a certain distance between the parameters of the models. Furthermore, we draw an analogy between the softmax model and the leverage score model, an important tool for algorithm design in linear algebra and graph theory. The leverage score model, on a high level, is a model which, given a vector input, produces an output drawn from a distribution dependent on the input. We obtain similar results for the binary hypothesis testing problem for leverage score models.