Adapting pre-trained large language models (LLMs) is crucial but challenging due to their enormous size. Parameter-efficient fine-tuning (PEFT) techniques typically employ additive adapters applied to frozen model weights. To further reduce memory usage, model weights are often compressed through quantization. However, existing PEFT methods often yield suboptimal model quality because they rely on restrictive assumptions, such as low-rank constraints on adapters to limit the number of trainable parameters. We find that sketching, a popular data compression technique, can serve as an efficient LLM adaptation strategy while avoiding the low-rank assumption. We introduce SketchTune, a compressive adaptation strategy that compresses LLM weights into compact fine-tunable sketches, integrating compression and adaptation into a unified framework. This integration eliminates the need for complex two-path computation in existing PEFT techniques, enabling faster and more memory-efficient training and inference. SketchTune is supported by mathematical insights into matrix classes that are better approximated using sketching rather than low-rank methods. Our extensive evaluations with Llama and Mistral models demonstrate that SketchTune outperforms leading PEFT methods across diverse tasks while using substantially smaller base models and comparable trainable parameters. As a highlight, SketchTune outperforms LoRA, DoRA, and S2FT on commonsense and math benchmarks using 2.6-3.5$\times$ smaller base models and exceeds LoftQ in accuracy by 14.48\% on GSM8K with 7.3$\times$ fewer trainable parameters.

Leveraging long contexts is crucial for advanced AI systems, but attention computation poses a scalability challenge. While scaled dot-product attention (SDPA) exhibits token sparsity, i.e. only a few pivotal tokens significantly contribute to output, exploiting this sparsity remains challenging. Existing methods either suffer from quality degradation or require substantial additional resources. We show that identifying pivotal tokens is a Maximum Inner Product Search (MIPS) problem. However, existing MIPS solutions are not well-suited for SDPA, as they are not GPU-friendly and often underperform due to the separated query and key distributions. This paper introduces HashAttention, framing pivotal token identification as a recommendation problem. Given a query, HashAttention encodes keys and queries in Hamming space, capturing the required semantic similarity, using learned mapping functions. HashAttention efficiently identifies pivotal tokens for a given query using bitwise operations and computes attention using only these tokens, improving the overall attention efficiency. Trained on generic data, HashAttention reduces tokens used by up to $16\times$ with minimal quality loss, requiring only 32 bits of auxiliary memory per token. Sparsity can be further improved to $32\times$ through task-specific fine-tuning. On A100 GPU, at $32\times$ sparsity, incorporating HashAttention reduces attention latency by up to $4.3\times$ in GPT-FAST and $2.54\times$ in FlashDecode, and achieves up to $3.12\times$ higher throughput for GPT-FAST.

Facial Emotion Recognition (FER) is a challenging problem due to various challenges such as variability in expressions and ambiguity in data. Several popular benchmarking datasets, specifically employed for FER tasks exhibit bias towards ethnicity, demography and image capture mechanisms. More specifically, the images in such datasets are captured in a controlled environment and are taken in good light, with straight head orientation, no occlusion or other facial artefacts. When employed for FER, these biases may impair a model's generalizability, rendering it ineffective for FER in novel and unseen datasets. Especially, in applications involving security (access control) and identification of mal-intentions from facial expressions, it may prove inefficient. A criminal may disguise their face with make-up, headgear, and religious facial accessories and can fool the FER models trained on these biased datasets. To that end, this work focuses on understanding these datasets better by identifying such “good-image” bias. Methods to mitigate such bias which allows the FER models to perform better and improve the robustness are also demonstrated. A simple yet effective FER framework for studying bias mitigation is proposed. Using this framework, the performance on popular dataset is analyzed and a significant difference in model performance is observed. Additionally, a knowledge transfer technique and a synthetic image generation technique are proposed to mitigate the identified bias. Finally, using the SFEW dataset, the findings are validated on the FER task, demonstrating the effectiveness of our techniques in mitigating real-world “good-image” bias. The experiments show that the proposed techniques outperform baseline methods by averaged fourfold improvement.

When considering a model architecture, there are several ways to reduce its memory footprint. Historically, popular approaches included selecting smaller architectures and creating sparse networks through pruning. More recently, randomized parameter-sharing (RPS) methods have gained traction for model compression at start of training. In this paper, we comprehensively assess the trade-off between memory and accuracy across RPS, pruning techniques, and building smaller models. Our findings demonstrate that RPS, which is both data and model-agnostic, consistently outperforms smaller models and all moderately informed pruning strategies, such as MAG, SNIP, SYNFLOW, and GRASP, across the entire compression range. This advantage becomes particularly pronounced in higher compression scenarios. Notably, even when compared to highly informed pruning techniques like Lottery Ticket Rewinding (LTR), RPS exhibits superior performance in high compression settings. This points out inherent capacity advantage that RPS enjoys over sparse models. Theoretically, we establish RPS as a superior technique in terms of memory-efficient representation when compared to pruning for linear models. This paper argues in favor of paradigm shift towards RPS based models. During our rigorous evaluation of RPS, we identified issues in the state- of-the-art RPS technique ROAST, specifically regarding stability (ROAST’s sensitivity to initialization hyperparameters, often leading to divergence) and Pareto-continuity (ROAST’s inability to recover the accuracy of the original model at zero compression). We provably address both of these issues. We refer to the modified RPS, which incorporates our improvements, as STABLE-RPS

Tensor multiplication with learned weight matrices is the fundamental building block in deep learning models. These matrices can often be sparsified, decomposed, quantized, or subjected to random parameter sharing without losing accuracy, suggesting the possibility of more efficient transforms. Although many variants of weight matrices exist, unstructured ones are incompatible with modern hardware, slowing inference and training. On the other hand, structured variants often limit expressivity or fail to deliver the promised latency benefits. We present Sketch Structured Transform (SS1), an expressive and GPU-friendly operator that accelerates inference. SS1 leverages parameter sharing in a random yet structured manner to reduce computation while retraining the rich expressive nature of parameter sharing. We confirm empirically that SS1 offers better quality-efficiency tradeoffs than competing variants. Interestingly SS1 can be combined with Quantization to achieve gains unattainable by either method alone, a finding we justify via theoretical analysis. The analysis may be of independent interest. Moreover, existing pre-trained models can be projected onto SS1 and finetuned for efficient deployment. Surprisingly, these projected models can perform reasonably well even without finetuning. Our experiments highlight various applications of the SS1: (a) Training GPT2 and DLRM models from scratch for faster inference. (b) Finetuning projected BERT models for 1.31× faster inference while maintaining GLUE scores. (c) Proof of concept with Llama-3-8b, showing 1.11× faster wall clock inference using projected SS1 layers without finetuning. We open source our code :https://github.com/apd10/Sketch-Structured-Linear/

Advancements in deep learning are often associated with increasing model sizes. Training and deploying large models require sophisticated hardware and incur significantly higher costs. Thus, model compression is a widely explored approach to solving the problem. However, SOTA techniques fall short in one or more desirable aspects of compression - for instance, pruning does not reduce memory for training, quantization can only provide up to 32$\times$ compression, HashedNet is cache-inefficient, etc. This paper proposes a model-agnostic, cache-friendly, and hardware-aware model compression approach: Random Operation Access Specific Tile (ROAST) hashing. ROAST collapses the parameters by clubbing them through a lightweight mapping. While clubbing these parameters, ROAST utilizes cache hierarchies by aligning the memory access pattern with the parameter access pattern. ROAST is up to ${\sim}25\times$ faster to train and ${\sim}50\times$ faster to infer than the popular parameter sharing method HashedNet. Additionally, ROAST introduces global weight sharing, which is empirically and theoretically superior to local weight sharing in HashedNet, and can be of independent interest. With ROAST, we can efficiently train and deploy the model using a much smaller memory footprint ($\sim 10 - 100\times$ lesser) in text and image classification tasks. ROAST-MM kernel implementation is open-source (https://github.com/apd10/RzLinear/tree/stable)

Large language models(LLMs) have sparked a new wave of exciting AI applications. Hosting these models at scale requires significant memory resources. One crucial memory bottleneck for the deployment stems from the context window. It is commonly recognized that model weights are memory hungry; however, the size of key-value embedding stored during the generation process (KV cache) can easily surpass the model size. The enormous size of the KV cache puts constraints on the inference batch size, which is crucial for high throughput inference workload. Inspired by an interesting observation of the attention scores, we hypothesize the persistence of importance: only pivotal tokens, which had a substantial influence at one step, will significantly influence future generations. Based on our empirical verification and theoretical analysis around this hypothesis, we propose scissorhands, a system that maintains the memory usage of the KV cache at a fixed budget without finetuning the model. In essence, Scissorhands manages the KV cache by storing the pivotal tokens with a higher probability. We validate that scissorhands reduces the inference memory usage of the KV cache by up to 5$\times$ without compromising model quality. We further demonstrate that scissorhands can be combined with 4-bit quantization, traditionally used to compress model weights, to achieve up to 20$\times$ compression.

Embedding tables dominate industrial-scale recommendation model sizes, using up to terabytes of memory. A popular and the largest publicly available machine learning MLPerf benchmark on recommendation data is a Deep Learning Recommendation Model (DLRM) trained on a terabyte of click-through data. It contains 100GB of embedding memory (25+Billion parameters). DLRMs, due to their sheer size and the associated volume of data, face difficulty in training, deploying for inference, and memory bottlenecks due to large embedding tables. This paper analyzes and extensively evaluates a generic parameter-sharing setup (PSS) for compressing DLRM models. We show theoretical upper bounds on the learnable memory requirements for achieving approximations to the embedding table. Our bounds indicate exponentially fewer parameters suffice for a good approximation. To this end, we demonstrate a PSS DLRM reaching 10000$\times$ compression on criteo-tb without losing quality. Such a compression, however, comes with a caveat. It requires 4.5 $\times$ more iterations to achieve the same saturation quality. The paper argues that this tradeoff needs more investigation as it might be significantly favorable. Leveraging the small size of the compressed model, we show a 4.3$\times$ improvement in training latency leading to similar overall training times. Thus, in the tradeoff between the system advantage of a small DLRM model vs. slower convergence, we show that scales are tipped towards having a smaller DLRM model, leading to the same quality, faster inference, easier deployment, and similar training times.

Traditional seismic inversion (SI) maps the hundreds of terabytes of raw-field data to subsurface properties in gigabytes. This inversion process is expensive, requiring over a year of human and computational effort. Recently, data-driven approaches equipped with Deep learning (DL) are envisioned to improve SI efficiency. However, these improvements are restricted to data with highly reduced scale and complexity. To extend these approaches to real-scale seismic data, researchers need to process raw nav-merge seismic data into an image and perform convolution. We argue that this convolution-based way of SI is not only computationally expensive but also conceptually problematic. Seismic data is not naturally an image and need not be processed as images. In this work, we go beyond convolution and propose a novel SI method. We solve the scalability of SI by proposing a new auxiliary learning paradigm for SI (Aux-SI). This paradigm breaks the SI into local inversion tasks, which predicts each small chunk of subsurface properties using surrounding seismic data. Aux-SI combines these local predictions to obtain the entire subsurface model. However, even this local inversion is still challenging due to: (1) high-dimensional, spatially irregular multi-modal seismic data, (2) there is no concrete spatial mapping (or alignment) between subsurface properties and raw data. To handle these challenges, we propose an all-MLP architecture, Multi-Modal Information Unscrambler (MMI-Unscrambler), that unscrambles seismic information by ingesting all available multi-modal data. The experiment shows that MMI-Unscrambler outperforms both SOTA U-Net and Transformer models on simulation data. We also scale MMI-Unscrambler to raw-field nav-merge data on Gulf-of-Mexico to obtain a geologically sound velocity model with an SSIM score of 0.8. To the best of our knowledge, this is the first successful demonstration of the DL approach on SI for real, large-scale, and complicated raw field data.

Estimating and storing the covariance (or correlation) matrix of high-dimensional data is computationally challenging because both memory and computational requirements scale quadratically with the dimension. Fortunately, high-dimensional covariance matrices as observed in text, click-through, meta-genomics datasets, etc are often sparse. In this paper, we consider the problem of efficient sparse estimation of covariance matrices with possibly trillions of entries. The size of the datasets we target requires the algorithm to be online, as more than one pass over the data is prohibitive. In this paper, we propose Active Sampling Count Sketch (ASCS), an online and one-pass sketching algorithm, that recovers the large entries of the covariance matrix accurately. Count Sketch (CS), and other sub-linear compressed sensing algorithms, offer a natural solution to the problem in theory. However, vanilla CS does not work well in practice due to a low signal-to-noise ratio (SNR). At the heart of our approach is a novel active sampling strategy that increases the SNR of classical CS. We demonstrate the practicality of our algorithm with synthetic data and real-world high dimensional datasets. ASCS significantly improves over vanilla CS, demonstrating the merit of our active sampling strategy.

Interacting with computers is a ubiquitous activity for millions of people. Repetitive or specialized tasks often require creation of small, often one-off, programs. End-users struggle with learning and using the myriad of domain-specific languages (DSLs) to effectively accomplish these tasks.

This paper presents the Treadport Active Wind Tunnel (TPAWT)-a full-body immersive virtual environment for the Treadport locomotion interface designed for generating wind on a user from any frontal direction at speeds up to 20 kph. The goal is to simulate the experience of realistic wind while walking in an outdoor virtual environment. A recirculating-type wind tunnel was created around the pre-existing Treadport installation by adding a large fan, ducting, and enclosure walls. Two sheets of air in a non-intrusive design flow along the side screens of the back-projection CAVE-like visual display, where they impinge and mix at the front screen to redirect towards the user in a full-body cross-section. By varying the flow conditions of the air sheets, the direction and speed of wind at the user are controlled. Design challenges to fit the wind tunnel in the pre-existing facility, and to manage turbulence to achieve stable and steerable flow, were overcome. The controller performance for wind speed and direction is demonstrated experimentally.

Generating invariants for loops is often a grueling obstacle in formal program verification. Researchers have employed methods from formal techniques based on abstract interpretation to test-driven dynamic analysis to tackle this problem. Even though powerful techniques for generating conjunctive invariants (invariants that employ only conjunction of terms) have been developed, disjunctive invariants have remained a sore thumb for formal techniques. In this paper, we propose a technique to transform certain category of loops, those that have a static iteration pattern, into loops that can be handled by conjunctive invariant generators. The key idea is to identify a static iteration pattern that distributes the disjunction in an invariant in a manner that can be captured by only conjunctive invariants. To broaden the scope of our algorithm, we also propose the idea of parametric verification, while attempting to verify specialized versions of the program where a subset of the input variables is instantiated with certain test-inputs. Note that parametric verification distinguishes it from program testing as testing requires all of its variables to be instantiated with test-inputs. We discuss our ideas on loops drawn from real programs to establish real-world applicability of our algorithms.