Graph Neural Networks (GNNs) play a crucial role in various fields. However, most existing deep graph learning frameworks assume pre-stored static graphs and do not support training on graph streams. In contrast, many real-world graphs are dynamic and contain time domain information. We introduce GNNFlow, a distributed framework that enables efficient continuous temporal graph representation learning on dynamic graphs on multi-GPU machines. GNNFlow introduces an adaptive time-indexed block-based data structure that effectively balances memory usage with graph update and sampling operation efficiency. It features a hybrid GPU-CPU graph data placement for rapid GPU-based temporal neighborhood sampling and kernel optimizations for enhanced sampling processes. A dynamic GPU cache for node and edge features is developed to maximize cache hit rates through reuse and restoration strategies. GNNFlow supports distributed training across multiple machines with static scheduling to ensure load balance. We implement GNNFlow based on DGL and PyTorch. Our experimental results show that GNNFlow provides up to 21.1x faster continuous learning than existing systems.
Neural materials typically consist of a collection of neural features along with a decoder network. The main challenge in integrating such models in real-time rendering pipelines lies in the large size required to store their features in GPU memory and the complexity of evaluating the network efficiently. We present a neural material model whose features and decoder are specifically designed to be used in real-time rendering pipelines. Our framework leverages hardware-based block compression (BC) texture formats to store the learned features and trains the model to output the material information continuously in space and scale. To achieve this, we organize the features in a block-based manner and emulate BC6 decompression during training, making it possible to export them as regular BC6 textures. This structure allows us to use high resolution features while maintaining a low memory footprint. Consequently, this enhances our model's overall capability, enabling the use of a lightweight and simple decoder architecture that can be evaluated directly in a shader. Furthermore, since the learned features can be decoded continuously, it allows for random uv sampling and smooth transition between scales without needing any subsequent filtering. As a result, our neural material has a small memory footprint, can be decoded extremely fast adding a minimal computational overhead to the rendering pipeline.
We present QLoRA, an efficient finetuning approach that reduces memory usage enough to finetune a 65B parameter model on a single 48GB GPU while preserving full 16-bit finetuning task performance. QLoRA backpropagates gradients through a frozen, 4-bit quantized pretrained language model into Low Rank Adapters~(LoRA). Our best model family, which we name Guanaco, outperforms all previous openly released models on the Vicuna benchmark, reaching 99.3% of the performance level of ChatGPT while only requiring 24 hours of finetuning on a single GPU. QLoRA introduces a number of innovations to save memory without sacrificing performance: (a) 4-bit NormalFloat (NF4), a new data type that is information theoretically optimal for normally distributed weights (b) double quantization to reduce the average memory footprint by quantizing the quantization constants, and (c) paged optimziers to manage memory spikes. We use QLoRA to finetune more than 1,000 models, providing a detailed analysis of instruction following and chatbot performance across 8 instruction datasets, multiple model types (LLaMA, T5), and model scales that would be infeasible to run with regular finetuning (e.g. 33B and 65B parameter models). Our results show that QLoRA finetuning on a small high-quality dataset leads to state-of-the-art results, even when using smaller models than the previous SoTA. We provide a detailed analysis of chatbot performance based on both human and GPT-4 evaluations showing that GPT-4 evaluations are a cheap and reasonable alternative to human evaluation. Furthermore, we find that current chatbot benchmarks are not trustworthy to accurately evaluate the performance levels of chatbots. A lemon-picked analysis demonstrates where Guanaco fails compared to ChatGPT. We release all of our models and code, including CUDA kernels for 4-bit training.
While developing a new multimodal LLM (MLLM) by pre-training on tremendous image-text pairs from scratch can be exceedingly resource-consuming, connecting an existing LLM with a comparatively lightweight visual prompt generator (VPG) becomes a feasible paradigm. However, further tuning the VPG part of the MLLM still suffers from indispensable computational costs, i.e., requiring thousands of GPU hours and millions of training data. One alternative solution is to transfer an existing VPG from any existing MLLMs for the target MLLM. In this work, we for the first time investigate the VPG transferability across LLMs, and explore a solution to reduce the cost of VPG transfer. We first study the VPG transfer across different LLM sizes (e.g., small-to-large), and across different LLM types, through which we diagnose the key factors to maximize the transfer efficiency. Based on our observation, we design a two-stage transfer framework named VPGTrans, which is simple yet highly effective. Through extensive experiments, we demonstrate that VPGTrans helps significantly speed up the transfer learning process without compromising performance. Remarkably, it helps achieve the VPG transfer from BLIP-2 OPT$_\text{2.7B}$ to BLIP-2 OPT$_\text{6.7B}$ with over 10 times speed-up and 10.7% training data compared with connecting a VPG to OPT$_\text{6.7B}$ from scratch. Further, a series of intriguing findings and potential rationales behind them are provided and discussed. Finally, we showcase the practical value of our VPGTrans approach, by customizing two novel MLLMs, including VL-LLaMA and VL-Vicuna, with recently released LLaMA and Vicuna LLMs.
Generative transformer models have become increasingly complex, with large numbers of parameters and the ability to process multiple input modalities. Current methods for explaining their predictions are resource-intensive. Most crucially, they require prohibitively large amounts of extra memory, since they rely on backpropagation which allocates almost twice as much GPU memory as the forward pass. This makes it difficult, if not impossible, to use them in production. We present AtMan that provides explanations of generative transformer models at almost no extra cost. Specifically, AtMan is a modality-agnostic perturbation method that manipulates the attention mechanisms of transformers to produce relevance maps for the input with respect to the output prediction. Instead of using backpropagation, AtMan applies a parallelizable token-based search method based on cosine similarity neighborhood in the embedding space. Our exhaustive experiments on text and image-text benchmarks demonstrate that AtMan outperforms current state-of-the-art gradient-based methods on several metrics while being computationally efficient. As such, AtMan is suitable for use in large model inference deployments.
Large language models (LLMs) have shown promise in proving formal theorems using proof assistants such as Lean. However, existing methods are difficult to reproduce or build on, due to private code, data, and large compute requirements. This has created substantial barriers to research on machine learning methods for theorem proving. This paper removes these barriers by introducing LeanDojo: an open-source Lean playground consisting of toolkits, data, models, and benchmarks. LeanDojo extracts data from Lean and enables interaction with the proof environment programmatically. It contains fine-grained annotations of premises in proofs, providing valuable data for premise selection: a key bottleneck in theorem proving. Using this data, we develop ReProver (Retrieval-Augmented Prover): an LLM-based prover augmented with retrieval for selecting premises from a vast math library. It is inexpensive and needs only one GPU week of training. Our retriever leverages LeanDojo's program analysis capability to identify accessible premises and hard negative examples, which makes retrieval much more effective. Furthermore, we construct a new benchmark consisting of 98,734 theorems and proofs extracted from Lean's math library. It features challenging data split requiring the prover to generalize to theorems relying on novel premises that are never used in training. We use this benchmark for training and evaluation, and experimental results demonstrate the effectiveness of ReProver over non-retrieval baselines and GPT-4. We thus provide the first set of open-source LLM-based theorem provers without any proprietary datasets and release it under a permissive MIT license to facilitate further research.
Large language models (LLMs) have demonstrated impressive abilities in various domains while the inference cost is expensive. Many previous studies exploit quantization methods to reduce LLM inference cost by reducing latency and memory consumption. Applying 2-bit single-precision weight quantization brings >3% accuracy loss, so the state-of-the-art methods use mixed-precision methods for LLMs (e.g. Llama2-7b, etc.) to improve the accuracy. However, challenges still exist: (1) Uneven distribution in weight matrix. (2) Large speed degradation by adding sparse outliers. (3) Time-consuming dequantization operations on GPUs. To tackle these challenges and enable fast and efficient LLM inference on GPUs, we propose the following techniques in this paper. (1) Intra-weight mixed-precision quantization. (2) Exclusive 2-bit sparse outlier with minimum speed degradation. (3) Asynchronous dequantization. We conduct extensive experiments on different model families (e.g. Llama3, etc.) and model sizes. We achieve 2.91-bit for each weight considering all scales/zeros for different models with negligible loss. As a result, with our 2/4/16 mixed-precision quantization for each weight matrix and asynchronous dequantization during inference, our design achieves an end-to-end speedup for Llama2-7b is 1.74x over the original model, and we reduce both runtime cost and total cost by up to 2.53x and 2.29x with less GPU requirements.
The rise of Deep Neural Networks (DNNs) has led to an increase in model size and complexity, straining the memory capacity of GPUs. Sparsity in DNNs, characterized as structural or ephemeral, has gained attention as a solution. This work focuses on ephemeral sparsity, aiming to reduce memory consumption during training. It emphasizes the significance of activations, an often overlooked component, and their role in memory usage. This work employs structured pruning in Block Sparse Compressed Row (BSR) format in combination with a magnitude-based criterion to efficiently prune activations. We furthermore introduce efficient block-sparse operators for GPUs and showcase their effectiveness, as well as the superior compression offered by block sparsity. We report the effectiveness of activation pruning by evaluating training speed, accuracy, and memory usage of large-scale neural architectures on the example of ResMLP on image classification tasks. As a result, we observe a memory reduction of up to 32% while maintaining accuracy. Ultimately, our approach aims to democratize large-scale model training, reduce GPU requirements, and address ecological concerns.
Neural radiance fields are capable of reconstructing high-quality drivable human avatars but are expensive to train and render and not suitable for multi-human scenes with complex shadows. To reduce consumption, we propose Animatable 3D Gaussian, which learns human avatars from input images and poses. We extend 3D Gaussians to dynamic human scenes by modeling a set of skinned 3D Gaussians and a corresponding skeleton in canonical space and deforming 3D Gaussians to posed space according to the input poses. We introduce a multi-head hash encoder for pose-dependent shape and appearance and a time-dependent ambient occlusion module to achieve high-quality reconstructions in scenes containing complex motions and dynamic shadows. On both novel view synthesis and novel pose synthesis tasks, our method achieves higher reconstruction quality than InstantAvatar with less training time (1/60), less GPU memory (1/4), and faster rendering speed (7x). Our method can be easily extended to multi-human scenes and achieve comparable novel view synthesis results on a scene with ten people in only 25 seconds of training.
We present a novel approach for digitizing real-world objects by estimating their geometry, material properties, and environmental lighting from a set of posed images with fixed lighting. Our method incorporates into Neural Radiance Field (NeRF) pipelines the split sum approximation used with image-based lighting for real-time physical-based rendering. We propose modeling the scene's lighting with a single scene-specific MLP representing pre-integrated image-based lighting at arbitrary resolutions. We achieve accurate modeling of pre-integrated lighting by exploiting a novel regularizer based on efficient Monte Carlo sampling. Additionally, we propose a new method of supervising self-occlusion predictions by exploiting a similar regularizer based on Monte Carlo sampling. Experimental results demonstrate the efficiency and effectiveness of our approach in estimating scene geometry, material properties, and lighting. Our method is capable of attaining state-of-the-art relighting quality after only ${\sim}1$ hour of training in a single NVIDIA A100 GPU.
Ptychographic extreme ultraviolet (EUV) diffractive imaging has emerged as a promising candidate for the next-generation metrology solutions in the semiconductor industry, as it can image wafer samples in reflection geometry at the nanoscale. This technique has surged attention recently, owing to the significant progress in high-harmonic generation (HHG) EUV sources and advancements in both hardware and software for computation. In this study, a novel algorithm is introduced and tested, which enables wavelength-multiplexed reconstruction that enhances the measurement throughput and introduces data diversity, allowing the accurate characterisation of sample structures. To tackle the inherent instabilities of the HHG source, a modal approach was adopted, which represents the cross-density function of the illumination by a series of mutually incoherent and independent spatial modes. The proposed algorithm was implemented on a mainstream machine learning platform, which leverages automatic differentiation to manage the drastic growth in model complexity and expedites the computation using GPU acceleration. By optimising over 200 million parameters, we demonstrate the algorithm's capacity to accommodate experimental uncertainties and achieve a resolution approaching the diffraction limit in reflection geometry. The reconstruction of wafer samples with 20-nm heigh patterned gold structures on a silicon substrate highlights our ability to handle complex physical interrelations involving a multitude of parameters. These results establish ptychography as an efficient and accurate metrology tool.
The high computational and memory requirements of generative large language models (LLMs) make it challenging to serve them cheaply. This paper aims to reduce the monetary cost for serving LLMs by leveraging preemptible GPU instances on modern clouds, which offer accesses to spare GPUs at a much cheaper price than regular instances but may be preempted by the cloud at any time. Serving LLMs on preemptible instances requires addressing challenges induced by frequent instance preemptions and the necessity of migrating instances to handle these preemptions. This paper presents SpotServe, the first distributed LLM serving system on preemptible instances. Several key techniques in SpotServe realize fast and reliable serving of generative LLMs on cheap preemptible instances. First, SpotServe dynamically adapts the LLM parallelization configuration for dynamic instance availability and fluctuating workload, while balancing the trade-off among the overall throughput, inference latency and monetary costs. Second, to minimize the cost of migrating instances for dynamic reparallelization, the task of migrating instances is formulated as a bipartite graph matching problem, which uses the Kuhn-Munkres algorithm to identify an optimal migration plan that minimizes communications. Finally, to take advantage of the grace period offered by modern clouds, we introduce stateful inference recovery, a new inference mechanism that commits inference progress at a much finer granularity and allows SpotServe to cheaply resume inference upon preemption. We evaluate on real spot instance preemption traces and various popular LLMs and show that SpotServe can reduce the P99 tail latency by 2.4 - 9.1x compared with the best existing LLM serving systems. We also show that SpotServe can leverage the price advantage of preemptive instances, saving 54% monetary cost compared with only using on-demand instances.
We present a hydrodynamic simulation system using the GPU compute shaders of DirectX for simulating virtual agent behaviors and navigation inside a smoothed particle hydrodynamical (SPH) fluid environment with real-time water mesh surface reconstruction. The current SPH literature includes interactions between SPH and heterogeneous meshes but seldom involves interactions between SPH and virtual boid agents. The contribution of the system lies in the combination of the parallel smoothed particle hydrodynamics model with the distributed boid model of virtual agents to enable agents to interact with fluids. The agents based on the boid algorithm influence the motion of SPH fluid particles, and the forces from the SPH algorithm affect the movement of the boids. To enable realistic fluid rendering and simulation in a particle-based system, it is essential to construct a mesh from the particle attributes. Our system also contributes to the surface reconstruction aspect of the pipeline, in which we performed a set of experiments with the parallel marching cubes algorithm per frame for constructing the mesh from the fluid particles in a real-time compute and memory-intensive application, producing a wide range of triangle configurations. We also demonstrate that our system is versatile enough for reinforced robotic agents instead of boid agents to interact with the fluid environment for underwater navigation and remote control engineering purposes.
The lattice Boltzmann method (LBM) has emerged as a prominent technique for solving fluid dynamics problems due to its algorithmic potential for computational scalability. We introduce XLB library, a Python-based differentiable LBM library based on the JAX platform. The architecture of XLB is predicated upon ensuring accessibility, extensibility, and computational performance, enabling scaling effectively across CPU, TPU, multi-GPU, and distributed multi-GPU or TPU systems. The library can be readily augmented with novel boundary conditions, collision models, or multi-physics simulation capabilities. XLB's differentiability and data structure is compatible with the extensive JAX-based machine learning ecosystem, enabling it to address physics-based machine learning, optimization, and inverse problems. XLB has been successfully scaled to handle simulations with billions of cells, achieving giga-scale lattice updates per second. XLB is released under the permissive Apache-2.0 license and is available on GitHub at https://github.com/Autodesk/XLB.
Recently, impressive results have been achieved in 3D scene editing with text instructions based on a 2D diffusion model. However, current diffusion models primarily generate images by predicting noise in the latent space, and the editing is usually applied to the whole image, which makes it challenging to perform delicate, especially localized, editing for 3D scenes. Inspired by recent 3D Gaussian splatting, we propose a systematic framework, named GaussianEditor, to edit 3D scenes delicately via 3D Gaussians with text instructions. Benefiting from the explicit property of 3D Gaussians, we design a series of techniques to achieve delicate editing. Specifically, we first extract the region of interest (RoI) corresponding to the text instruction, aligning it to 3D Gaussians. The Gaussian RoI is further used to control the editing process. Our framework can achieve more delicate and precise editing of 3D scenes than previous methods while enjoying much faster training speed, i.e. within 20 minutes on a single V100 GPU, more than twice as fast as Instruct-NeRF2NeRF (45 minutes -- 2 hours).
The automatic segmentation of kidney, kidney tumor and kidney cyst on Computed Tomography (CT) scans is a challenging task due to the indistinct lesion boundaries and fuzzy texture. Considering the large range and unbalanced distribution of CT scans' thickness, 2.5D ResUnet are adopted to build an efficient coarse-to-fine semantic segmentation framework in this work. A set of 489 CT scans are used for training and validation, and an independent never-before-used CT scans for testing. Finally, we demonstrate the effectiveness of our proposed method. The dice values on test set are 0.954, 0.792, 0.691, the surface dice values are 0.897, 0.591, 0.541 for kidney, tumor and cyst, respectively. The average inference time of each CT scan is 20.65s and the max GPU memory is 3525MB. The results suggest that a better trade-off between model performance and efficiency.
Recent quantization techniques have enabled heterogeneous precisions at very fine granularity, e.g., each parameter/activation can take on a different precision, resulting in compact neural networks without sacrificing accuracy. However, there is a lack of efficient architectural support for such networks, which require additional hardware to decode the precision settings for individual variables, align the variables, and provide fine-grained mixed-precision compute capabilities. The complexity of these operations introduces high overheads. Thus, the improvements in inference latency/energy of these networks are not commensurate with the compression ratio, and may be inferior to larger quantized networks with uniform precisions. We present an end-to-end co-design approach encompassing computer architecture, training algorithm, and inference optimization to efficiently execute networks with fine-grained heterogeneous precisions. The key to our approach is a novel training algorithm designed to accommodate hardware constraints and inference operation requirements, outputting networks with input-channel-wise heterogeneous precisions and at most three precision levels. Combined with inference optimization techniques, existing architectures with low-cost enhancements can support such networks efficiently, yielding optimized tradeoffs between accuracy, compression ratio and inference latency/energy. We demonstrate the efficacy of our approach across CPU and GPU architectures. For various representative neural networks, our approach achieves >10x improvements in both compression ratio and inference latency, with negligible degradation in accuracy compared to full-precision networks.
Over the past year, a large body of multimodal research has emerged around zero-shot evaluation using GPT descriptors. These studies boost the zero-shot accuracy of pretrained VL models with an ensemble of label-specific text generated by GPT. A recent study, WaffleCLIP, demonstrated that similar zero-shot accuracy can be achieved with an ensemble of random descriptors. However, both zero-shot methods are un-trainable and consequently sub-optimal when some few-shot out-of-distribution (OOD) training data is available. Inspired by these prior works, we present two more flexible methods called descriptor and word soups, which do not require an LLM at test time and can leverage training data to increase OOD target accuracy. Descriptor soup greedily selects a small set of textual descriptors using generic few-shot training data, then calculates robust class embeddings using the selected descriptors. Word soup greedily assembles a chain of words in a similar manner. Compared to existing few-shot soft prompt tuning methods, word soup requires fewer parameters by construction and less GPU memory, since it does not require backpropagation. Both soups outperform current published few-shot methods, even when combined with SoTA zero-shot methods, on cross-dataset and domain generalization benchmarks. Compared with SoTA prompt and descriptor ensembling methods, such as ProDA and WaffleCLIP, word soup achieves higher OOD accuracy with fewer ensemble members. Please checkout our code: github.com/Chris210634/word_soups
End users face a choice between privacy and efficiency in current Large Language Model (LLM) service paradigms. In cloud-based paradigms, users are forced to compromise data locality for generation quality and processing speed. Conversely, edge device paradigms maintain data locality but fail to deliver satisfactory performance. In this work, we propose a novel LLM service paradigm that distributes privacy-sensitive computation on edge devices and shared computation in the cloud. Only activations are transmitted between the central cloud and edge devices to ensure data locality. Our core innovation, PrivateLoRA, addresses the challenging communication overhead by exploiting the low rank of residual activations, achieving over 95% communication reduction. Consequently, PrivateLoRA effectively maintains data locality and is extremely resource efficient. Under standard 5G networks, PrivateLoRA achieves throughput over 300% of device-only solutions for 7B models and over 80% of an A100 GPU for 33B models. PrivateLoRA also provides tuning performance comparable to LoRA for advanced personalization. Our approach democratizes access to state-of-the-art generative AI for edge devices, paving the way for more tailored LLM experiences for the general public. To our knowledge, our proposed framework is the first efficient and privacy-preserving LLM solution in the literature.
CP decomposition is a powerful tool for data science, especially gene analysis, deep learning, and quantum computation. However, the application of tensor decomposition is largely hindered by the exponential increment of the computational complexity and storage consumption with the size of tensors. While the data in our real world is usually presented as trillion- or even exascale-scale tensors, existing work can only support billion-scale scale tensors. In our work, we propose the Exascale-Tensor to mitigate the significant gap. Specifically, we propose a compression-based tensor decomposition framework, namely the exascale-tensor, to support exascale tensor decomposition. Then, we carefully analyze the inherent parallelism and propose a bag of strategies to improve computational efficiency. Last, we conduct experiments to decompose tensors ranging from million-scale to trillion-scale for evaluation. Compared to the baselines, the exascale-tensor supports 8,000x larger tensors and a speedup up to 6.95x. We also apply our method to two real-world applications, including gene analysis and tensor layer neural networks, of which the numeric results demonstrate the scalability and effectiveness of our method.