Training large language models (LLMs) is known to be challenging because of the huge computational and memory capacity requirements. To address these issues, it is common to use a cluster of GPUs with 3D parallelism, which splits a model along the data batch, pipeline stage, and intra-layer tensor dimensions. However, the use of 3D parallelism produces the additional challenge of finding the optimal number of ways on each dimension and mapping the split models onto the GPUs. Several previous studies have attempted to automatically find the optimal configuration, but many of these lacked several important aspects. For instance, the heterogeneous nature of the interconnect speeds is often ignored. While the peak bandwidths for the interconnects are usually made equal, the actual attained bandwidth varies per link in real-world clusters. Combined with the critical path modeling that does not properly consider the communication, they easily fall into sub-optimal configurations. In addition, they often fail to consider the memory requirement per GPU, often recommending solutions that could not be executed. To address these challenges, we propose Pipette, which is an automatic fine-grained LLM training configurator for real-world clusters. By devising better performance models along with the memory estimator and fine-grained individual GPU assignment, Pipette achieves faster configurations that satisfy the memory constraints. We evaluated Pipette on large clusters to show that it provides a significant speedup over the prior art. The implementation of Pipette is available at https://github.com/yimjinkyu1/date2024_pipette.
Scale has become a main ingredient in obtaining strong machine learning models. As a result, understanding a model's scaling properties is key to effectively designing both the right training setup as well as future generations of architectures. In this work, we argue that scale and training research has been needlessly complex due to reliance on the cosine schedule, which prevents training across different lengths for the same model size. We investigate the training behavior of a direct alternative -- constant learning rate and cooldowns -- and find that it scales predictably and reliably similar to cosine. Additionally, we show that stochastic weight averaging yields improved performance along the training trajectory, without additional training costs, across different scales. Importantly, with these findings we demonstrate that scaling experiments can be performed with significantly reduced compute and GPU hours by utilizing fewer but reusable training runs. Our code is available at \url{https://github.com/epfml/schedules-and-scaling/}.
Diffusion models with large-scale pre-training have achieved significant success in the field of visual content generation, particularly exemplified by Diffusion Transformers (DiT). However, DiT models have faced challenges with quadratic complexity efficiency, especially when handling long sequences. In this paper, we aim to incorporate the sub-quadratic modeling capability of Gated Linear Attention (GLA) into the 2D diffusion backbone. Specifically, we introduce Diffusion Gated Linear Attention Transformers (DiG), a simple, adoptable solution with minimal parameter overhead. We offer two variants, i,e, a plain and U-shape architecture, showing superior efficiency and competitive effectiveness. In addition to superior performance to DiT and other sub-quadratic-time diffusion models at $256 \times 256$ resolution, DiG demonstrates greater efficiency than these methods starting from a $512$ resolution. Specifically, DiG-S/2 is $2.5\times$ faster and saves $75.7\%$ GPU memory compared to DiT-S/2 at a $1792$ resolution. Additionally, DiG-XL/2 is $4.2\times$ faster than the Mamba-based model at a $1024$ resolution and $1.8\times$ faster than DiT with FlashAttention-2 at a $2048$ resolution. We will release the code soon. Code is released at https://github.com/hustvl/DiG.
Recently, linear complexity sequence modeling networks have achieved modeling capabilities similar to Vision Transformers on a variety of computer vision tasks, while using fewer FLOPs and less memory. However, their advantage in terms of actual runtime speed is not significant. To address this issue, we introduce Gated Linear Attention (GLA) for vision, leveraging its superior hardware-awareness and efficiency. We propose direction-wise gating to capture 1D global context through bidirectional modeling and a 2D gating locality injection to adaptively inject 2D local details into 1D global context. Our hardware-aware implementation further merges forward and backward scanning into a single kernel, enhancing parallelism and reducing memory cost and latency. The proposed model, ViG, offers a favorable trade-off in accuracy, parameters, and FLOPs on ImageNet and downstream tasks, outperforming popular Transformer and CNN-based models. Notably, ViG-S matches DeiT-B's accuracy while using only 27% of the parameters and 20% of the FLOPs, running 2$\times$ faster on $224\times224$ images. At $1024\times1024$ resolution, ViG-T uses 5.2$\times$ fewer FLOPs, saves 90% GPU memory, runs 4.8$\times$ faster, and achieves 20.7% higher top-1 accuracy than DeiT-T. These results position ViG as an efficient and scalable solution for visual representation learning. Code is available at \url{https://github.com/hustvl/ViG}.
We present Lightning Attention, the first linear attention implementation that maintains a constant training speed for various sequence lengths under fixed memory consumption. Due to the issue with cumulative summation operations (cumsum), previous linear attention implementations cannot achieve their theoretical advantage in a casual setting. However, this issue can be effectively solved by utilizing different attention calculation strategies to compute the different parts of attention. Specifically, we split the attention calculation into intra-blocks and inter-blocks and use conventional attention computation for intra-blocks and linear attention kernel tricks for inter-blocks. This eliminates the need for cumsum in the linear attention calculation. Furthermore, a tiling technique is adopted through both forward and backward procedures to take full advantage of the GPU hardware. To enhance accuracy while preserving efficacy, we introduce TransNormerLLM (TNL), a new architecture that is tailored to our lightning attention. We conduct rigorous testing on standard and self-collected datasets with varying model sizes and sequence lengths. TNL is notably more efficient than other language models. In addition, benchmark results indicate that TNL performs on par with state-of-the-art LLMs utilizing conventional transformer structures. The source code is released at github.com/OpenNLPLab/TransnormerLLM.
We present a lightweight, easy-to-train, low-resolution, fully data-driven climate emulator, LUCIE, that can be trained on as low as $2$ years of $6$-hourly ERA5 data. Unlike most state-of-the-art AI weather models, LUCIE remains stable and physically consistent for $100$ years of autoregressive simulation with $100$ ensemble members. Long-term mean climatology from LUCIE's simulation of temperature, wind, precipitation, and humidity matches that of ERA5 data, along with the variability. We further demonstrate how well extreme weather events and their return periods can be estimated from a large ensemble of long-term simulations. We further discuss an improved training strategy with a hard-constrained first-order integrator to suppress autoregressive error growth, a novel spectral regularization strategy to better capture fine-scale dynamics, and finally an optimization algorithm that enables data-limited (as low as $2$ years of $6$-hourly data) training of the emulator without losing stability and physical consistency. Finally, we provide a scaling experiment to compare the long-term bias of LUCIE with respect to the number of training samples. Importantly, LUCIE is an easy to use model that can be trained in just $2.4$h on a single A-100 GPU, allowing for multiple experiments that can explore important scientific questions that could be answered with large ensembles of long-term simulations, e.g., the impact of different variables on the simulation, dynamic response to external forcing, and estimation of extreme weather events, amongst others.
Despite advances using low-rank adapters and quantization, pretraining of large models on consumer hardware has not been possible without model sharding, offloading during training, or per-layer gradient updates. To address these limitations, we propose Low-Rank Adapters for Quantized Training (LoQT), a method for efficiently training quantized models. LoQT uses gradient-based tensor factorization to initialize low-rank trainable weight matrices that are periodically merged into quantized full-rank weight matrices. Our approach is suitable for both pretraining and fine-tuning models. We demonstrate this for language modeling and downstream task adaptation, finding that LoQT enables efficient training of models up to 7B parameters on a 24GB GPU. We also demonstrate the feasibility of training a 13B model using per-layer gradient updates on the same hardware.
The automated surface defect detection is a fundamental task in industrial production, and the existing saliencybased works overcome the challenging scenes and give promising detection results. However, the cutting-edge efforts often suffer from large parameter size, heavy computational cost, and slow inference speed, which heavily limits the practical applications. To this end, we devise a multi-scale interactive (MI) module, which employs depthwise convolution (DWConv) and pointwise convolution (PWConv) to independently extract and interactively fuse features of different scales, respectively. Particularly, the MI module can provide satisfactory characterization for defect regions with fewer parameters. Embarking on this module, we propose a lightweight Multi-scale Interactive Network (MINet) to conduct real-time salient object detection of strip steel surface defects. Comprehensive experimental results on SD-Saliency-900 dataset, which contains three kinds of strip steel surface defect detection images (i.e., inclusion, patches, and scratches), demonstrate that the proposed MINet presents comparable detection accuracy with the state-of-the-art methods while running at a GPU speed of 721FPS and a CPU speed of 6.3FPS for 368*368 images with only 0.28M parameters. The code is available at https://github.com/Kunye-Shen/MINet.
One key challenge in backdoor attacks against large foundation models is the resource limits. Backdoor attacks usually require retraining the target model, which is impractical for very large foundation models. Existing backdoor attacks are mainly designed for supervised classifiers or small foundation models (e.g., BERT). None of these attacks has successfully compromised a very large foundation model, such as Llama-3-70B, especially with limited computational resources. In this paper, we propose TrojFM, a novel backdoor attack tailored for very large foundation models. Our primary technical contribution is the development of a novel backdoor injection method. This method forces a backdoored model to generate similar hidden representations for poisoned inputs regardless of their actual semantics. Our approach injects such backdoors by fine-tuning only a very small proportion of model parameters. This enables TrojFM to efficiently launch downstream task-agnostic backdoor attacks against very large foundation models under limited computational resources. Moreover, we optimize the fine-tuning process with our customized QLoRA technique, enabling launching our attack via only~\textit{one A100 GPU}. Furthermore, we design a new trigger injection method to ensure our attack stealthiness. Through extensive experiments, we first demonstrate that TrojFM can launch effective backdoor attacks against widely used large GPT-style models without jeopardizing their normal functionalities (and outperforming existing attacks on BERT-style models). Furthermore, we show that TrojFM is resilient to SOTA defenses and is insensitive to changes in key hyper-parameters. Finally, we conduct a resource analysis to quantify that our method can significantly save computational and memory costs compared to existing backdoor attacks.
The poor performance of transformers on arithmetic tasks seems to stem in large part from their inability to keep track of the exact position of each digit inside of a large span of digits. We mend this problem by adding an embedding to each digit that encodes its position relative to the start of the number. In addition to the boost these embeddings provide on their own, we show that this fix enables architectural modifications such as input injection and recurrent layers to improve performance even further. With positions resolved, we can study the logical extrapolation ability of transformers. Can they solve arithmetic problems that are larger and more complex than those in their training data? We find that training on only 20 digit numbers with a single GPU for one day, we can reach state-of-the-art performance, achieving up to 99% accuracy on 100 digit addition problems. Finally, we show that these gains in numeracy also unlock improvements on other multi-step reasoning tasks including sorting and multiplication.
A fundamental problem in learning robust and expressive visual representations lies in efficiently estimating the spatial relationships of visual semantics throughout the entire image. In this study, we propose vHeat, a novel vision backbone model that simultaneously achieves both high computational efficiency and global receptive field. The essential idea, inspired by the physical principle of heat conduction, is to conceptualize image patches as heat sources and model the calculation of their correlations as the diffusion of thermal energy. This mechanism is incorporated into deep models through the newly proposed module, the Heat Conduction Operator (HCO), which is physically plausible and can be efficiently implemented using DCT and IDCT operations with a complexity of $\mathcal{O}(N^{1.5})$. Extensive experiments demonstrate that vHeat surpasses Vision Transformers (ViTs) across various vision tasks, while also providing higher inference speeds, reduced FLOPs, and lower GPU memory usage for high-resolution images. The code will be released at https://github.com/MzeroMiko/vHeat.
This paper benchmarks and improves existing GPU matrix multiplication algorithms specialized for Kronecker-sparse matrices, whose sparsity patterns are described by Kronecker products. These matrices have recently gained popularity as replacements for dense matrices in neural networks because they preserve accuracy while using fewer parameters. We present the first energy and time benchmarks for the multiplication with such matrices, helping users identify scenarios where Kronecker-sparse matrices are more time- and energy-efficient than their dense counterparts. Our benchmark also reveals that specialized implementations spend up to 50% of their total runtime on memory rewriting operations. To address the challenge of reducing memory transfers, we introduce a new so-called tiling strategy adapted to the Kronecker-sparsity structure, which reduces reads and writes between levels of GPU memory. We implement this tiling strategy in a new CUDA kernel that achieves a median speed-up of x1.4, while also cutting energy consumption by 15%. We further demonstrate the broader impact of our results by applying the new kernel to accelerate transformer inference.
Multi-dimensional parameter spaces are commonly encountered in physics theories that go beyond the Standard Model. However, they often possess complicated posterior geometries that are expensive to traverse using techniques traditional to astroparticle physics. Several recent innovations, which are only beginning to make their way into this field, have made navigating such complex posteriors possible. These include GPU acceleration, automatic differentiation, and neural-network-guided reparameterization. We apply these advancements to dark matter direct detection experiments in the context of non-standard neutrino interactions and benchmark their performances against traditional nested sampling techniques when conducting Bayesian inference. Compared to nested sampling alone, we find that these techniques increase performance for both nested sampling and Hamiltonian Monte Carlo, accelerating inference by factors of $\sim 100$ and $\sim 60$, respectively. As nested sampling also evaluates the Bayesian evidence, these advancements can be exploited to improve model comparison performance while retaining compatibility with existing implementations that are widely used in the natural sciences. Using these techniques, we perform the first scan in the neutrino non-standard interactions parameter space for direct detection experiments whereby all parameters are allowed to vary simultaneously. We expect that these advancements are broadly applicable to other areas of astroparticle physics featuring multi-dimensional parameter spaces.
Federated learning works by aggregating locally computed gradients from multiple clients, thus enabling collaborative training without sharing private client data. However, prior work has shown that the data can actually be recovered by the server using so-called gradient inversion attacks. While these attacks perform well when applied on images, they are limited in the text domain and only permit approximate reconstruction of small batches and short input sequences. In this work, we propose DAGER, the first algorithm to recover whole batches of input text exactly. DAGER leverages the low-rank structure of self-attention layer gradients and the discrete nature of token embeddings to efficiently check if a given token sequence is part of the client data. We use this check to exactly recover full batches in the honest-but-curious setting without any prior on the data for both encoder- and decoder-based architectures using exhaustive heuristic search and a greedy approach, respectively. We provide an efficient GPU implementation of DAGER and show experimentally that it recovers full batches of size up to 128 on large language models (LLMs), beating prior attacks in speed (20x at same batch size), scalability (10x larger batches), and reconstruction quality (ROUGE-1/2 > 0.99).
This paper introduces StreamV2V, a diffusion model that achieves real-time streaming video-to-video (V2V) translation with user prompts. Unlike prior V2V methods using batches to process limited frames, we opt to process frames in a streaming fashion, to support unlimited frames. At the heart of StreamV2V lies a backward-looking principle that relates the present to the past. This is realized by maintaining a feature bank, which archives information from past frames. For incoming frames, StreamV2V extends self-attention to include banked keys and values and directly fuses similar past features into the output. The feature bank is continually updated by merging stored and new features, making it compact but informative. StreamV2V stands out for its adaptability and efficiency, seamlessly integrating with image diffusion models without fine-tuning. It can run 20 FPS on one A100 GPU, being 15x, 46x, 108x, and 158x faster than FlowVid, CoDeF, Rerender, and TokenFlow, respectively. Quantitative metrics and user studies confirm StreamV2V's exceptional ability to maintain temporal consistency.
Training neural networks which are robust to adversarial attacks remains an important problem in deep learning, especially as heavily overparameterized models are adopted in safety-critical settings. Drawing from recent work which reformulates the training problems for two-layer ReLU and polynomial activation networks as convex programs, we devise a convex semidefinite program (SDP) for adversarial training of two-layer polynomial activation networks and prove that the convex SDP achieves the same globally optimal solution as its nonconvex counterpart. The convex SDP is observed to improve robust test accuracy against $\ell_\infty$ attacks relative to the original convex training formulation on multiple datasets. Additionally, we present scalable implementations of adversarial training for two-layer polynomial and ReLU networks which are compatible with standard machine learning libraries and GPU acceleration. Leveraging these implementations, we retrain the final two fully connected layers of a Pre-Activation ResNet-18 model on the CIFAR-10 dataset with both polynomial and ReLU activations. The two `robustified' models achieve significantly higher robust test accuracies against $\ell_\infty$ attacks than a Pre-Activation ResNet-18 model trained with sharpness-aware minimization, demonstrating the practical utility of convex adversarial training on large-scale problems.
Parameter-efficient fine-tuning (PEFT) has emerged as a popular solution for adapting pre-trained Vision Transformer (ViT) models to downstream applications. While current PEFT methods have achieved parameter efficiency, they overlook the efficiency of computation and GPU memory during both fine-tuning and inference, falling short of practical requirements. In this paper, we propose \textbf{Sparse-Tuning}, a novel PEFT method that accounts for the information redundancy in images and videos to boost the above efficiency. By sparsely preserving the semantic-relevant tokens and merging irrelevant ones, Sparse-Tuning minimizes the quantity of tokens processed at each layer, leading to a quadratic reduction in computational and memory overhead. To align our token sparsification strategy suitably with fine-tuning purposes, we further design Dense Adapters that establish dense connections from shallow layers to deeper layers. These Dense Adapters integrate multi-level local features to enrich the current tokens, improving both token preservation and model adaptation. Empirical results on VTAB-1K, three image datasets, and two video datasets show that our Sparse-Tuning reduces GFLOPs to \textbf{62\%-70\%} of the original ViT-B while achieving state-of-the-art performance. Source code is available at \url{https://github.com/liuting20/Sparse-Tuning}.
One of the guiding principles for designing AI-based weather forecasting systems is to embed physical constraints as inductive priors in the neural network architecture. A popular prior is locality, where the atmospheric data is processed with local neural interactions, like 3D convolutions or 3D local attention windows as in Pangu-Weather. On the other hand, some works have shown great success in weather forecasting without this locality principle, at the cost of a much higher parameter count. In this paper, we show that the 3D local processing in Pangu-Weather is computationally sub-optimal. We design ArchesWeather, a transformer model that combines 2D attention with a column-wise attention-based feature interaction module, and demonstrate that this design improves forecasting skill. ArchesWeather is trained at 1.5{\deg} resolution and 24h lead time, with a training budget of a few GPU-days and a lower inference cost than competing methods. An ensemble of four of our models shows better RMSE scores than the IFS HRES and is competitive with the 1.4{\deg} 50-members NeuralGCM ensemble for one to three days ahead forecasting. Our code and models are publicly available at https://github.com/gcouairon/ArchesWeather.
Training large AI models such as LLMs and DLRMs costs massive GPUs and computing time. The high training cost has become only affordable to big tech companies, meanwhile also causing increasing concerns about the environmental impact. This paper presents CoMERA, a Computing- and Memory-Efficient training method via Rank-Adaptive tensor optimization. CoMERA achieves rank-adaptive tensor-compressed (pre)-training via a multi-objective optimization formulation and improves the training to provide both a high compression ratio and excellent accuracy in the training process. Our optimized numerical computation (e.g., optimized tensorized embedding and tensor-network contractions) and GPU implementation eliminate part of the run-time overhead in the tensorized training on GPU. This leads to, for the first time, $2-3\times$ speedup per training epoch compared with standard training. CoMERA also outperforms the recent GaLore in terms of both memory and computing efficiency. Specifically, CoMERA is $2\times$ faster per training epoch and $9\times$ more memory-efficient than GaLore on a tested six-encoder transformer with single-batch training. Our method also shows $\sim 2\times$ speedup than standard pre-training on a BERT-like code-generation LLM while achieving $4.23\times$ compression ratio in pre-training. With further HPC optimization, CoMERA may reduce the pre-training cost of many other LLMs. An implementation of CoMERA is available at https://github.com/ziyangjoy/CoMERA.
Large language models (LLMs) are trained on a vast amount of human-written data, but data providers often remain uncredited. In response to this issue, data valuation (or data attribution), which quantifies the contribution or value of each data to the model output, has been discussed as a potential solution. Nevertheless, applying existing data valuation methods to recent LLMs and their vast training datasets has been largely limited by prohibitive compute and memory costs. In this work, we focus on influence functions, a popular gradient-based data valuation method, and significantly improve its scalability with an efficient gradient projection strategy called LoGra that leverages the gradient structure in backpropagation. We then provide a theoretical motivation of gradient projection approaches to influence functions to promote trust in the data valuation process. Lastly, we lower the barrier to implementing data valuation systems by introducing LogIX, a software package that can transform existing training code into data valuation code with minimal effort. In our data valuation experiments, LoGra achieves competitive accuracy against more expensive baselines while showing up to 6,500x improvement in throughput and 5x reduction in GPU memory usage when applied to Llama3-8B-Instruct and the 1B-token dataset.