Zhanpeng Zeng

Zhanpeng Zeng, Sourav Pal, Jeffery Kline et al.

Transformers have emerged as a preferred model for many tasks in natural langugage processing and vision. Recent efforts on training and deploying Transformers more efficiently have identified many strategies to approximate the self-attention matrix, a key module in a Transformer architecture. Effective ideas include various prespecified sparsity patterns, low-rank basis expansions and combinations thereof. In this paper, we revisit classical Multiresolution Analysis (MRA) concepts such as Wavelets, whose potential value in this setting remains underexplored thus far. We show that simple approximations based on empirical feedback and design choices informed by modern hardware and implementation challenges, eventually yield a MRA-based approach for self-attention with an excellent performance profile across most criteria of interest. We undertake an extensive set of experiments and demonstrate that this multi-resolution scheme outperforms most efficient self-attention proposals and is favorable for both short and long sequences. Code is available at \url{https://github.com/mlpen/mra-attention}.

Luxi Lin, Zhihang Lin, Zhanpeng Zeng et al.

Speculative decoding accelerates LLM inference but suffers from performance degradation when target models are fine-tuned for specific domains. A naive solution is to retrain draft models for every target model, which is costly and inefficient. To address this, we introduce a parameter- and data-efficient framework named Efficient Draft Adaptation, abbreviated as EDA, for efficiently adapting draft models. EDA introduces three innovations: (1) a decoupled architecture that utilizes shared and private components to model the shared and target-specific output distributions separately, enabling parameter-efficient adaptation by updating only the lightweight private component;(2) a data regeneration strategy that utilizes the fine-tuned target model to regenerate training data, thereby improving the alignment between training and speculative decoding, leading to higher average acceptance length;(3) a sample selection mechanism that prioritizes high-value data for efficient adaptation. Our experiments show that EDA effectively restores speculative performance on fine-tuned models, achieving superior average acceptance lengths with significantly reduced training costs compared to full retraining. Code is available at https://github.com/Lyn-Lucy/Efficient-Draft-Adaptation.

Jiapeng Zeng, Yogesh Prabhu, Zhanpeng Zeng et al.

Conformal prediction (CP) gives distribution-free coverage for modern vision and language models, but it is often forced to make a ranking decision from a single unstable nonconformity score. Standard CP uses one realization, while average-then-calibrate variants smooth multiple realizations into a point estimate. Both options discard the inconsistency that can help identify whether a candidate is indeed stable. A weak answer can enter the conformal set even if the evidence is not strong, simply because one posterior sample or prompt phrasing made it look strong. But variability can help distinguish a stable signal from noise-driven fluctuations. We describe an empirical Bayes conformal prediction framework that uses $r$-values to convert score variability into an uncertainty informed nonconformity score. The resulting $r$-value estimates how likely a candidate's latent score belongs to the top-ranked group after accounting for both its mean score and its uncertainty. It admits both a closed-form Normal-Normal empirical Bayes estimator and a nonparametric posterior-sampling estimator. Using the $r$-value as the nonconformity score preserves the target conformal coverage while provably reducing the inclusion of high variance false candidates under mild regularity conditions. Across image classification, CLIP-based VLM benchmarks, and LLMs, we show that $r$-value conformal prediction preserves target coverage while improving ranking stability and reducing set size when variability is informative, and reverting to CP-like behavior when variability vanishes.

Harshavardhan Adepu, Zhanpeng Zeng, Li Zhang et al.

Transformers are the backbone of powerful foundation models for many Vision and Natural Language Processing tasks. But their compute and memory/storage footprint is large, and so, serving such models is expensive often requiring high-end hardware. To mitigate this difficulty, Post-Training Quantization seeks to modify a pre-trained model and quantize it to eight bits or lower, significantly boosting compute/memory/latency efficiency. Such models have been successfully quantized to four bits with some performance loss. In this work, we outline a simple scheme to quantize Transformer-based models to just two bits (plus some overhead) with only a small drop in accuracy. Key to our formulation is a concept borrowed from Harmonic analysis called Fusion Frames. Our main finding is that the quantization must take place not in the original weight space, but instead in the Fusion Frame representations. If quantization is interpreted as the addition of noise, our casting of the problem allows invoking an extensive body of known consistent recovery and noise robustness guarantees. Further, if desired, de-noising filters are known in closed form. We show empirically, via a variety of experiments, that (almost) two-bit quantization for Transformer models promises sizable efficiency gains. The code is available at https://github.com/vsingh-group/FrameQuant

Jiahao Liu, Hongji Ruan, Weimin Zhang et al.

This paper explores effective numerical feature embedding for Click-Through Rate prediction in streaming environments. Conventional static binning methods rely on offline statistics of numerical distributions; however, this inherently two-stage process often triggers semantic drift during bin boundary updates. While neural embedding methods enable end-to-end learning, they often discard explicit distributional information. Integrating such information end-to-end is challenging because streaming features often violate the i.i.d. assumption, precluding unbiased estimation of the population distribution via the expectation of order statistics. Furthermore, the critical context dependency of numerical distributions is often neglected. To this end, we propose DAES, an end-to-end framework designed to tackle numerical feature embedding in streaming training scenarios by integrating distributional information with an adaptive modulation mechanism. Specifically, we introduce an efficient reservoir-sampling-based distribution estimation method and two field-aware distribution modulation strategies to capture streaming distributions and field-dependent semantics. DAES significantly outperforms existing approaches as demonstrated by extensive offline and online experiments and has been fully deployed on a leading short-video platform with hundreds of millions of daily active users.

Ziye Tong, Jiahao Liu, Weimin Zhang et al.

Multimodal content is crucial for click-through rate (CTR) prediction. However, directly incorporating continuous embeddings from pre-trained models into CTR models yields suboptimal results due to misaligned optimization objectives and convergence speed inconsistency during joint training. Discretizing embeddings into semantic IDs before feeding them into CTR models offers a more effective solution, yet existing methods suffer from limited codebook utilization, reconstruction accuracy, and semantic discriminability. We propose RQ-GMM (Residual Quantized Gaussian Mixture Model), which introduces probabilistic modeling to better capture the statistical structure of multimodal embedding spaces. Through Gaussian Mixture Models combined with residual quantization, RQ-GMM achieves superior codebook utilization and reconstruction accuracy. Experiments on public datasets and online A/B tests on a large-scale short-video platform serving hundreds of millions of users demonstrate substantial improvements: RQ-GMM yields a 1.502% gain in Advertiser Value over strong baselines. The method has been fully deployed, serving daily recommendations for hundreds of millions of users.

Zhanpeng Zeng, Michael Davies, Pranav Pulijala et al.

While GPU clusters are the de facto choice for training large deep neural network (DNN) models today, several reasons including ease of workflow, security and cost have led to efforts investigating whether CPUs may be viable for inference in routine use in many sectors of the industry. But the imbalance between the compute capabilities of GPUs and CPUs is huge. Motivated by these considerations, we study a module which is a workhorse within modern DNN architectures, GEMM based Feed Forward Networks (FFNs), and assess the extent to which it can be made compute- (or FLOP-) lite. Specifically, we propose an alternative formulation (we call it LookupFFN) to GEMM based FFNs inspired by the recent studies of using Locality Sensitive Hashing (LSH) to approximate FFNs. Our formulation recasts most essential operations as a memory look-up, leveraging the trade-off between the two resources on any platform: compute and memory (since CPUs offer it in abundance). For RoBERTa language model pretraining, our formulation achieves similar performance compared to GEMM based FFNs, while dramatically reducing the required FLOP. Our development is complemented with a detailed hardware profiling of strategies that will maximize efficiency -- not just on contemporary hardware but on products that will be offered in the near/medium term future. Code is avaiable at \url{https://github.com/mlpen/LookupFFN}.

Luxi Lin, Zhihang Lin, Zhanpeng Zeng et al.

This paper introduces Multimodal Speculative Decoding (MSD) to accelerate Multimodal Large Language Models (MLLMs) inference. Speculative decoding has been shown to accelerate Large Language Models (LLMs) without sacrificing accuracy. However, current speculative decoding methods for MLLMs fail to achieve the same speedup as they do for LLMs. To address this, we reimagine speculative decoding specifically for MLLMs. Our analysis of MLLM characteristics reveals two key design principles for MSD: (1) Text and visual tokens have fundamentally different characteristics and need to be processed separately during drafting. (2) Both language modeling ability and visual perception capability are crucial for the draft model. For the first principle, MSD decouples text and visual tokens in the draft model, allowing each to be handled based on its own characteristics. For the second principle, MSD uses a two-stage training strategy: In stage one, the draft model is trained on text-only instruction-tuning datasets to improve its language modeling ability. In stage two, MSD gradually introduces multimodal data to enhance the visual perception capability of the draft model. Experiments show that MSD boosts inference speed by up to $2.29\times$ for LLaVA-1.5-7B and up to $2.46\times$ for LLaVA-1.5-13B on multimodal benchmarks, demonstrating its effectiveness. Our code is available at https://github.com/Lyn-Lucy/MSD.

Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi et al.

Transformer-based models are widely used in natural language processing (NLP). Central to the transformer model is the self-attention mechanism, which captures the interactions of token pairs in the input sequences and depends quadratically on the sequence length. Training such models on longer sequences is expensive. In this paper, we show that a Bernoulli sampling attention mechanism based on Locality Sensitive Hashing (LSH), decreases the quadratic complexity of such models to linear. We bypass the quadratic cost by considering self-attention as a sum of individual tokens associated with Bernoulli random variables that can, in principle, be sampled at once by a single hash (although in practice, this number may be a small constant). This leads to an efficient sampling scheme to estimate self-attention which relies on specific modifications of LSH (to enable deployment on GPU architectures). We evaluate our algorithm on the GLUE benchmark with standard 512 sequence length where we see favorable performance relative to a standard pretrained Transformer. On the Long Range Arena (LRA) benchmark, for evaluating performance on long sequences, our method achieves results consistent with softmax self-attention but with sizable speed-ups and memory savings and often outperforms other efficient self-attention methods. Our code is available at https://github.com/mlpen/YOSO

Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty et al.

Transformers have emerged as a powerful tool for a broad range of natural language processing tasks. A key component that drives the impressive performance of Transformers is the self-attention mechanism that encodes the influence or dependence of other tokens on each specific token. While beneficial, the quadratic complexity of self-attention on the input sequence length has limited its application to longer sequences -- a topic being actively studied in the community. To address this limitation, we propose Nyströmformer -- a model that exhibits favorable scalability as a function of sequence length. Our idea is based on adapting the Nyström method to approximate standard self-attention with $O(n)$ complexity. The scalability of Nyströmformer enables application to longer sequences with thousands of tokens. We perform evaluations on multiple downstream tasks on the GLUE benchmark and IMDB reviews with standard sequence length, and find that our Nyströmformer performs comparably, or in a few cases, even slightly better, than standard self-attention. On longer sequence tasks in the Long Range Arena (LRA) benchmark, Nyströmformer performs favorably relative to other efficient self-attention methods. Our code is available at https://github.com/mlpen/Nystromformer.

Quanjian Song, Zhihang Lin, Zhanpeng Zeng et al.

Existing camera motion-controlled video generation methods face computational bottlenecks in fine-tuning and inference. This paper proposes LightMotion, a light and tuning-free method for simulating camera motion in video generation. Operating in the latent space, it eliminates additional fine-tuning, inpainting, and depth estimation, making it more streamlined than existing methods. The endeavors of this paper comprise: (i) The latent space permutation operation effectively simulates various camera motions like panning, zooming, and rotation. (ii) The latent space resampling strategy combines background-aware sampling and cross-frame alignment to accurately fill new perspectives while maintaining coherence across frames. (iii) Our in-depth analysis shows that the permutation and resampling cause an SNR shift in latent space, leading to poor-quality generation. To address this, we propose latent space correction, which reintroduces noise during denoising to mitigate SNR shift and enhance video generation quality. Exhaustive experiments show that our LightMotion outperforms existing methods, both quantitatively and qualitatively.

Zhanpeng Zeng, Karthikeyan Sankaralingam, Vikas Singh

GEneral Matrix Multiply (GEMM) is a central operation in deep learning and corresponds to the largest chunk of the compute footprint. Therefore, improving its efficiency is an active topic of ongoing research. A popular strategy is the use of low bit-width integers to approximate the original entries in a matrix. This allows efficiency gains, but often requires sophisticated techniques to control the rounding error incurred. In this work, we first verify/check that when the low bit-width restriction is removed, for a variety of Transformer-based models, whether integers are sufficient for all GEMMs need -- for {\em both} training and inference stages, and can achieve parity with floating point counterparts. No sophisticated techniques are needed. We find that while a large majority of entries in matrices (encountered in such models) can be easily represented by {\em low} bit-width integers, the existence of a few heavy hitter entries make it difficult to achieve efficiency gains via the exclusive use of low bit-width GEMMs alone. To address this issue, we develop a simple algorithm, Integer Matrix Unpacking (IM-Unpack), to {\em unpack} a matrix with large integer entries into a larger matrix whose entries all lie within the representable range of arbitrarily low bit-width integers. This allows {\em equivalence} with the original GEMM, i.e., the exact result can be obtained using purely low bit-width integer GEMMs. This comes at the cost of additional operations -- we show that for many popular models, this overhead is quite small.

Qihao Liu, Zhanpeng Zeng, Ju He et al.

This paper presents innovative enhancements to diffusion models by integrating a novel multi-resolution network and time-dependent layer normalization. Diffusion models have gained prominence for their effectiveness in high-fidelity image generation. While conventional approaches rely on convolutional U-Net architectures, recent Transformer-based designs have demonstrated superior performance and scalability. However, Transformer architectures, which tokenize input data (via "patchification"), face a trade-off between visual fidelity and computational complexity due to the quadratic nature of self-attention operations concerning token length. While larger patch sizes enable attention computation efficiency, they struggle to capture fine-grained visual details, leading to image distortions. To address this challenge, we propose augmenting the Diffusion model with the Multi-Resolution network (DiMR), a framework that refines features across multiple resolutions, progressively enhancing detail from low to high resolution. Additionally, we introduce Time-Dependent Layer Normalization (TD-LN), a parameter-efficient approach that incorporates time-dependent parameters into layer normalization to inject time information and achieve superior performance. Our method's efficacy is demonstrated on the class-conditional ImageNet generation benchmark, where DiMR-XL variants outperform prior diffusion models, setting new state-of-the-art FID scores of 1.70 on ImageNet 256 x 256 and 2.89 on ImageNet 512 x 512. Project page: https://qihao067.github.io/projects/DiMR

Zhanpeng Zeng, Cole Hawkins, Mingyi Hong et al.

Transformers are central in modern natural language processing and computer vision applications. Despite recent works devoted to reducing the quadratic cost of such models (as a function of the sequence length), dealing with ultra long sequences (e.g., with more than 16K tokens) remains challenging. Applications such as answering questions based on a book or summarizing a scientific article are inefficient or infeasible. Here, we propose to significantly improve the efficiency of Transformers for ultra long sequences, by compressing the sequence into a much smaller representation at each layer. Specifically, by exploiting the fact that in many tasks, only a small subset of special tokens (we call VIP-tokens) are most relevant to the final prediction, we propose a VIP-token centric compression (VCC) scheme which selectively compresses the sequence based on their impact on approximating the representation of the VIP-tokens. Compared with competitive baselines, our algorithm is not only efficient (achieving more than $3\times$ efficiency gain compared to baselines on 4K and 16K lengths), but also offers competitive/better performance on a large number of tasks. Further, we show that our algorithm scales to 128K tokens (or more) while consistently offering accuracy improvement.