We introduce motion graph, a novel approach to the video prediction problem, which predicts future video frames from limited past data. The motion graph transforms patches of video frames into interconnected graph nodes, to comprehensively describe the spatial-temporal relationships among them. This representation overcomes the limitations of existing motion representations such as image differences, optical flow, and motion matrix that either fall short in capturing complex motion patterns or suffer from excessive memory consumption. We further present a video prediction pipeline empowered by motion graph, exhibiting substantial performance improvements and cost reductions. Experiments on various datasets, including UCF Sports, KITTI and Cityscapes, highlight the strong representative ability of motion graph. Especially on UCF Sports, our method matches and outperforms the SOTA methods with a significant reduction in model size by 78% and a substantial decrease in GPU memory utilization by 47%.
Personalized recommendation is a ubiquitous application on the internet, with many industries and hyperscalers extensively leveraging Deep Learning Recommendation Models (DLRMs) for their personalization needs (like ad serving or movie suggestions). With growing model and dataset sizes pushing computation and memory requirements, GPUs are being increasingly preferred for executing DLRM inference. However, serving newer DLRMs, while meeting acceptable latencies, continues to remain challenging, making traditional deployments increasingly more GPU-hungry, resulting in higher inference serving costs. In this paper, we show that the embedding stage continues to be the primary bottleneck in the GPU inference pipeline, leading up to a 3.2x embedding-only performance slowdown. To thoroughly grasp the problem, we conduct a detailed microarchitecture characterization and highlight the presence of low occupancy in the standard embedding kernels. By leveraging direct compiler optimizations, we achieve optimal occupancy, pushing the performance by up to 53%. Yet, long memory latency stalls continue to exist. To tackle this challenge, we propose specialized plug-and-play-based software prefetching and L2 pinning techniques, which help in hiding and decreasing the latencies. Further, we propose combining them, as they complement each other. Experimental evaluations using A100 GPUs with large models and datasets show that our proposed techniques improve performance by up to 103% for the embedding stage, and up to 77% for the overall DLRM inference pipeline.
With the widespread deployment of long-context large language models (LLMs), there has been a growing demand for efficient support of high-throughput inference. However, as the key-value (KV) cache expands with the sequence length, the increasing memory footprint and the need to access it for each token generation both result in low throughput when serving long-context LLMs. While various dynamic sparse attention methods have been proposed to speed up inference while maintaining generation quality, they either fail to sufficiently reduce GPU memory consumption or introduce significant decoding latency by offloading the KV cache to the CPU. We present ShadowKV, a high-throughput long-context LLM inference system that stores the low-rank key cache and offloads the value cache to reduce the memory footprint for larger batch sizes and longer sequences. To minimize decoding latency, ShadowKV employs an accurate KV selection strategy that reconstructs minimal sparse KV pairs on-the-fly. By evaluating ShadowKV on a broad range of benchmarks, including RULER, LongBench, and Needle In A Haystack, and models like Llama-3.1-8B, Llama-3-8B-1M, GLM-4-9B-1M, Yi-9B-200K, Phi-3-Mini-128K, and Qwen2-7B-128K, we demonstrate that it can support up to 6$\times$ larger batch sizes and boost throughput by up to 3.04$\times$ on an A100 GPU without sacrificing accuracy, even surpassing the performance achievable with infinite batch size under the assumption of infinite GPU memory. The code is available at https://github.com/bytedance/ShadowKV.
Edge AI computing boxes are a new class of computing devices that are aimed to revolutionize the AI industry. These compact and robust hardware units bring the power of AI processing directly to the source of data--on the edge of the network. On the other hand, on-demand serverless inference services are becoming more and more popular as they minimize the infrastructural cost associated with hosting and running DNN models for small to medium-sized businesses. However, these computing devices are still constrained in terms of resource availability. As such, the service providers need to load and unload models efficiently in order to meet the growing demand. In this paper, we introduce FusedInf to efficiently swap DNN models for on-demand serverless inference services on the edge. FusedInf combines multiple models into a single Direct Acyclic Graph (DAG) to efficiently load the models into the GPU memory and make execution faster. Our evaluation of popular DNN models showed that creating a single DAG can make the execution of the models up to 14\% faster while reducing the memory requirement by up to 17\%. The prototype implementation is available at https://github.com/SifatTaj/FusedInf.
Large language models (LLMs) have demonstrated remarkable performance across various downstream tasks. However, the high computational and memory requirements of LLMs are a major bottleneck. To address this, parameter-efficient fine-tuning (PEFT) methods such as low-rank adaptation (LoRA) have been proposed to reduce computational costs while ensuring minimal loss in performance. Additionally, knowledge distillation (KD) has been a popular choice for obtaining compact student models from teacher models. In this work, we present KD-LoRA, a novel fine-tuning method that combines LoRA with KD. Our results demonstrate that KD-LoRA achieves performance comparable to full fine-tuning (FFT) and LoRA while significantly reducing resource requirements. Specifically, KD-LoRA retains 98% of LoRA's performance on the GLUE benchmark, while being 40% more compact. Additionally, KD-LoRA reduces GPU memory usage by 30% compared to LoRA, while decreasing inference time by 30% compared to both FFT and LoRA. We evaluate KD-LoRA across three encoder-only models: BERT, RoBERTa, and DeBERTaV3. Code is available at https://github.com/rambodazimi/KD-LoRA.
The challenge of mapping AI architectures to GPU hardware is creating a critical bottleneck in AI progress. Despite substantial efforts, hand-written custom kernels fail to meet their theoretical performance thresholds, even on well-established operations like linear attention. The diverse hardware capabilities of GPUs might suggest that we need a wide variety of techniques to achieve high performance. However, our work explores whether a small number of key abstractions can drastically simplify the process. We present ThunderKittens (TK), a framework for writing performant AI kernels while remaining easy to use and maintain. Our abstractions map to the three levels of the GPU hierarchy: (1) at the warp-level, we provide 16x16 matrix tiles as basic data structures and PyTorch-like parallel compute operations over tiles, (2) at the thread-block level, we provide a template for overlapping asynchronous operations across parallel warps, and (3) at the grid-level, we provide support to help hide the block launch and tear-down, and memory costs. We show the value of TK by providing kernels that match or outperform prior kernels for a range of AI operations. We match CuBLAS and FlashAttention-3 on GEMM and attention inference performance and outperform the strongest baselines by $10-40\%$ on attention backwards, $8\times$ on state space models, and $14\times$ on linear attention.
Approximate nearest neighbor (ANN) search is a key component in many modern machine learning pipelines; recent use cases include retrieval-augmented generation (RAG) and vector databases. Clustering-based ANN algorithms, that use score computation methods based on product quantization (PQ), are often used in industrial-scale applications due to their scalability and suitability for distributed and disk-based implementations. However, they have slower query times than the leading graph-based ANN algorithms. In this work, we propose a new supervised score computation method based on the observation that inner product approximation is a multivariate (multi-output) regression problem that can be solved efficiently by reduced-rank regression. Our experiments show that on modern high-dimensional data sets, the proposed reduced-rank regression (RRR) method is superior to PQ in both query latency and memory usage. We also introduce LoRANN, a clustering-based ANN library that leverages the proposed score computation method. LoRANN is competitive with the leading graph-based algorithms and outperforms the state-of-the-art GPU ANN methods on high-dimensional data sets.
Long-context models(LCMs) have shown great potential in processing long input sequences(even more than 100M tokens) conveniently and effectively. With significant progress, recent research has pointed out that LCMs can accurately locate token-level salient information within the context. Yet, the generation performance of these LCMs is far from satisfactory and might result in misaligned responses, such as hallucinations. To enhance the generation capability of LCMs, existing works have investigated the effects of data size and quality for both pre-training and instruction tuning. Though achieving meaningful improvement, previous methods fall short in either effectiveness or efficiency. In this paper, we introduce LOGO(Long cOntext aliGnment via efficient preference Optimization), a training strategy that first introduces preference optimization for long-context alignment. To overcome the GPU memory-bound issue caused by the long sequence, LOGO employs a reference-free preference optimization strategy and adopts a position synthesis method to construct the training data. By training with only 0.3B data on a single 8$\times$A800 GPU machine for 16 hours, LOGO allows the Llama-3-8B-Instruct-80K model to achieve comparable performance with GPT-4 in real-world long-context tasks while preserving the model's original capabilities on other tasks, e.g., language modeling and MMLU. Moreover, LOGO can extend the model's context window size while enhancing its generation performance.
The development of large language models (LLMs) has significantly expanded model sizes, resulting in substantial GPU memory requirements during inference. The key and value storage of the attention map in the KV (key-value) cache accounts for more than 80\% of this memory consumption. Nowadays, most existing KV cache compression methods focus on intra-layer compression within a single Transformer layer but few works consider layer-wise compression. In this paper, we propose a plug-and-play method called \textit{KVSharer}, which shares the KV cache between layers to achieve layer-wise compression. Rather than intuitively sharing based on higher similarity, we discover a counterintuitive phenomenon: sharing dissimilar KV caches better preserves the model performance. Experiments show that \textit{KVSharer} can reduce KV cache computation by 30\%, thereby lowering memory consumption without significantly impacting model performance and it can also achieve at least 1.3 times generation acceleration. Additionally, we verify that \textit{KVSharer} is compatible with existing intra-layer KV cache compression methods, and combining both can further save memory.
Each request in LLM inference goes through two phases: compute-bound prefill and memory-bandwidth-bound decode. To improve GPU utilization, recent systems use hybrid batching that combines the prefill and decode phases of different requests into the same batch. This approach optimizes linear operations but remains inefficient for attention computation because existing attention kernels specialize execution independently for the prefill and decode phases. In this paper, we present POD-Attention - the first GPU kernel that efficiently computes attention for hybrid batches. POD-Attention aims to maximize the utilization of both compute and memory bandwidth by carefully allocating the GPU's resources such that prefill and decode operations happen concurrently on the same multiprocessor. POD-Attention speeds up attention computation by up to $59\%$ (mean $28\%$), enabling higher throughput and lower latency LLM inference compared to the use of independently optimized prefill and decode attention kernels.
Contrastive loss is a powerful approach for representation learning, where larger batch sizes enhance performance by providing more negative samples to better distinguish between similar and dissimilar data. However, scaling batch sizes is constrained by the quadratic growth in GPU memory consumption, primarily due to the full instantiation of the similarity matrix. To address this, we propose a tile-based computation strategy that partitions the contrastive loss calculation into arbitrary small blocks, avoiding full materialization of the similarity matrix. Furthermore, we introduce a multi-level tiling strategy to leverage the hierarchical structure of distributed systems, employing ring-based communication at the GPU level to optimize synchronization and fused kernels at the CUDA core level to reduce I/O overhead. Experimental results show that the proposed method scales batch sizes to unprecedented levels. For instance, it enables contrastive training of a CLIP-ViT-L/14 model with a batch size of 4M or 12M using 8 or 32 A800 80GB without sacrificing any accuracy. Compared to SOTA memory-efficient solutions, it achieves a two-order-of-magnitude reduction in memory while maintaining comparable speed. The code will be made publicly available.
Large language models (LLMs) with long context windows have gained significant attention. However, the KV cache, stored to avoid re-computation, becomes a bottleneck. Various dynamic sparse or TopK-based attention approximation methods have been proposed to leverage the common insight that attention is sparse. In this paper, we first show that TopK attention itself suffers from quality degradation in certain downstream tasks because attention is not always as sparse as expected. Rather than selecting the keys and values with the highest attention scores, sampling with theoretical guarantees can provide a better estimation for attention output. To make the sampling-based approximation practical in LLM generation, we propose MagicPIG, a heterogeneous system based on Locality Sensitive Hashing (LSH). MagicPIG significantly reduces the workload of attention computation while preserving high accuracy for diverse tasks. MagicPIG stores the LSH hash tables and runs the attention computation on the CPU, which allows it to serve longer contexts and larger batch sizes with high approximation accuracy. MagicPIG can improve decoding throughput by up to $5\times$ across various GPU hardware and achieve 54ms decoding latency on a single RTX 4090 for Llama-3.1-8B-Instruct model with a context of 96k tokens. The code is available at https://github.com/Infini-AI-Lab/MagicPIG.
Multimodal large language models (MLLMs) have made significant strides, yet they face challenges in the medical domain due to limited specialized knowledge. While recent medical MLLMs demonstrate strong performance in lab settings, they often struggle in real-world applications, highlighting a substantial gap between research and practice. In this paper, we seek to address this gap at various stages of the end-to-end learning pipeline, including data collection, model fine-tuning, and evaluation. At the data collection stage, we introduce SemiHVision, a dataset that combines human annotations with automated augmentation techniques to improve both medical knowledge representation and diagnostic reasoning. For model fine-tuning, we trained PMC-Cambrian-8B-AN over 2400 H100 GPU hours, resulting in performance that surpasses public medical models like HuatuoGPT-Vision-34B (79.0% vs. 66.7%) and private general models like Claude3-Opus (55.7%) on traditional benchmarks such as SLAKE and VQA-RAD. In the evaluation phase, we observed that traditional benchmarks cannot accurately reflect realistic clinical task capabilities. To overcome this limitation and provide more targeted guidance for model evaluation, we introduce the JAMA Clinical Challenge, a novel benchmark specifically designed to evaluate diagnostic reasoning. On this benchmark, PMC-Cambrian-AN achieves state-of-the-art performance with a GPT-4 score of 1.29, significantly outperforming HuatuoGPT-Vision-34B (1.13) and Claude3-Opus (1.17), demonstrating its superior diagnostic reasoning abilities.
Score identity Distillation (SiD) is a data-free method that has achieved SOTA performance in image generation by leveraging only a pretrained diffusion model, without requiring any training data. However, its ultimate performance is constrained by how accurate the pretrained model captures the true data scores at different stages of the diffusion process. In this paper, we introduce SiDA (SiD with Adversarial Loss), which not only enhances generation quality but also improves distillation efficiency by incorporating real images and adversarial loss. SiDA utilizes the encoder from the generator's score network as a discriminator, allowing it to distinguish between real images and those generated by SiD. The adversarial loss is batch-normalized within each GPU and then combined with the original SiD loss. This integration effectively incorporates the average "fakeness" per GPU batch into the pixel-based SiD loss, enabling SiDA to distill a single-step generator. SiDA converges significantly faster than its predecessor when distilled from scratch, and swiftly improves upon the original model's performance during fine-tuning from a pre-distilled SiD generator. This one-step adversarial distillation method establishes new benchmarks in generation performance when distilling EDM diffusion models, achieving FID scores of 1.110 on ImageNet 64x64. When distilling EDM2 models trained on ImageNet 512x512, our SiDA method surpasses even the largest teacher model, EDM2-XXL, which achieved an FID of 1.81 using classifier-free guidance (CFG) and 63 generation steps. In contrast, SiDA achieves FID scores of 2.156 for size XS, 1.669 for S, 1.488 for M, 1.413 for L, 1.379 for XL, and 1.366 for XXL, all without CFG and in a single generation step. These results highlight substantial improvements across all model sizes. Our code is available at https://github.com/mingyuanzhou/SiD/tree/sida.
Democratization of AI is an important topic within the broader topic of the digital divide. This issue is relevant to LLMs, which are becoming popular as AI co-pilots but suffer from a lack of accessibility due to high computational demand. In this study, we evaluate whether quantization is a viable approach toward enabling LLMs on generic consumer devices. The study assesses the performance of five quantized code LLMs in Lua code generation tasks. To evaluate the impact of quantization, the models with 7B parameters were tested on a consumer laptop at 2-, 4-, and 8-bit integer precisions and compared to non-quantized code LLMs with 1.3, 2, and 3 billion parameters. Lua is chosen as a low-level resource language to avoid models' biases related to high-resource languages. The results suggest that the models quantized at the 4-bit integer precision offer the best trade-off between performance and model size. These models can be comfortably deployed on an average laptop without a dedicated GPU. The performance significantly drops at the 2-bit integer precision. The models at 8-bit integer precision require more inference time that does not effectively translate to better performance. The 4-bit models with 7 billion parameters also considerably outperform non-quantized models with lower parameter numbers despite having comparable model sizes with respect to storage and memory demand. While quantization indeed increases the accessibility of smaller LLMs with 7 billion parameters, these LLMs demonstrate overall low performance (less than 50\%) on high-precision and low-resource tasks such as Lua code generation. While accessibility is improved, usability is still not at the practical level comparable to foundational LLMs such as GPT-4o or Llama 3.1 405B.
A persistence diagram provides a compact summary of persistent homology, which captures the topological features of a space at different scales. However, due to its nature as a set, incorporating it as a feature into a machine learning framework is challenging. Several methods have been proposed to use persistence diagrams as input for machine learning models, but they often require complex preprocessing steps and extensive hyperparameter tuning. In this paper, we propose a novel transformer architecture called the \textit{Extended Persistence Transformer (xPerT)}, which is highly scalable than the compared to Persformer, an existing transformer for persistence diagrams. xPerT reduces GPU memory usage by over 90\% and improves accuracy on multiple datasets. Additionally, xPerT does not require complex preprocessing steps or extensive hyperparameter tuning, making it easy to use in practice. Our code is available at https://github.com/sehunfromdaegu/xpert.
The inverse of an invertible convolution is an important operation that comes up in Normalizing Flows, Image Deblurring, etc. The naive algorithm for backpropagation of this operation using Gaussian elimination has running time $O(n^3)$ where $n$ is the number of pixels in the image. We give a fast parallel backpropagation algorithm with running time $O(\sqrt{n})$ for a square image and provide a GPU implementation of the same. Inverse of Convolutions are usually used in Normalizing Flows in the sampling pass, making them slow. We propose to use the Inverse of Convolutions in the forward (image to latent vector) pass of the Normalizing flow. Since the sampling pass is the inverse of the forward pass, it will use convolutions only, resulting in efficient sampling times. We use our parallel backpropagation algorithm to optimize the inverse of the convolution layer, resulting in fast training times. We implement this approach in various Normalizing Flow backbones, resulting in our Inverse-Flow models. We benchmark Inverse-Flow on standard datasets and show significantly improved sampling times with similar bits per dimension compared to previous models.
Current and future large scale structure surveys aim to constrain the neutrino mass and the equation of state of dark energy. We aim to construct accurate and interpretable symbolic approximations to the linear and nonlinear matter power spectra as a function of cosmological parameters in extended $\Lambda$CDM models which contain massive neutrinos and non-constant equations of state for dark energy. This constitutes an extension of the syren-halofit emulators to incorporate these two effects, which we call syren-new (SYmbolic-Regression-ENhanced power spectrum emulator with NEutrinos and $W_0-w_a$). We also obtain a simple approximation to the derived parameter $\sigma_8$ as a function of the cosmological parameters for these models. Our results for the linear power spectrum are designed to emulate CLASS, whereas for the nonlinear case we aim to match the results of EuclidEmulator2. We compare our results to existing emulators and $N$-body simulations. Our analytic emulators for $\sigma_8$, the linear and nonlinear power spectra achieve root mean squared errors of 0.1%, 0.3% and 1.3%, respectively, across a wide range of cosmological parameters, redshifts and wavenumbers. We verify that emulator-related discrepancies are subdominant compared to observational errors and other modelling uncertainties when computing shear power spectra for LSST-like surveys. Our expressions have similar accuracy to existing (numerical) emulators, but are at least an order of magnitude faster, both on a CPU and GPU. Our work greatly improves the accuracy, speed and range of applicability of current symbolic approximations to the linear and nonlinear matter power spectra. We provide publicly available code for all symbolic approximations found.
Energy consumption has become a critical design metric and a limiting factor in the development of future computing architectures, from small wearable devices to large-scale leadership computing facilities. The predominant methods in energy management optimization are focused on CPUs. However, GPUs are increasingly significant and account for the majority of energy consumption in heterogeneous high performance computing (HPC) systems. Moreover, they typically rely on either purely offline training or a hybrid of offline and online training, which are impractical and lead to energy loss during data collection. Therefore, this paper studies a novel and practical online energy optimization problem for GPUs in HPC scenarios. The problem is challenging due to the inherent performance-energy trade-offs of GPUs, the exploration & exploitation dilemma across frequencies, and the lack of explicit performance counters in GPUs. To address these challenges, we formulate the online energy consumption optimization problem as a multi-armed bandit framework and develop a novel bandit based framework EnergyUCB. EnergyUCB is designed to dynamically adjust GPU core frequencies in real-time, reducing energy consumption with minimal impact on performance. Specifically, the proposed framework EnergyUCB (1) balances the performance-energy trade-off in the reward function, (2) effectively navigates the exploration & exploitation dilemma when adjusting GPU core frequencies online, and (3) leverages the ratio of GPU core utilization to uncore utilization as a real-time GPU performance metric. Experiments on a wide range of real-world HPC benchmarks demonstrate that EnergyUCB can achieve substantial energy savings. The code of EnergyUCB is available at https://github.com/XiongxiaoXu/EnergyUCB-Bandit.
Time series forecasting is vital in numerous web applications, influencing critical decision-making across industries. While diffusion models have recently gained increasing popularity for this task, we argue they suffer from a significant drawback: indiscriminate noise addition to the original time series followed by denoising, which can obscure underlying dynamic evolving trend and complicate forecasting. To address this limitation, we propose a novel flexible decoupled framework (FDF) that learns high-quality time series representations for enhanced forecasting performance. A key characteristic of our approach leverages the inherent inductive bias of time series data of its decomposed trend and seasonal components, each modeled separately to enable decoupled analysis and modeling. Specifically, we propose an innovative Conditional Denoising Seasonal Module (CDSM) within the diffusion model, which leverages statistical information from the historical window to conditionally model the complex seasonal component. Notably, we incorporate a Polynomial Trend Module (PTM) to effectively capture the smooth trend component, thereby enhancing the model's ability to represent temporal dependencies. Extensive experiments validate the effectiveness of our framework, demonstrating superior performance over existing methods and highlighting its flexibility in time series forecasting. The source code is available at https://github.com/zjt-gpu/FDF.