Finetuning large language models (LLMs) is essential for task adaptation, yet serving stacks today isolate inference and finetuning on separate GPU clusters -- wasting resources and under-utilizing hardware. We introduce FlexLLM, the first system to co-serve LLM inference and PEFT-based finetuning on shared GPUs by fusing computation at the token level. The static compilation optimizations in FlexLLM -- dependent parallelization and graph pruning significantly shrink activation memory, leading to end-to-end GPU memory savings by up to 80%. At runtime, a novel token-level finetuning mechanism paired with a hybrid token scheduler dynamically interleaves inference and training tokens within each co-serving iteration, meeting strict latency SLOs while maximizing utilization. In end-to-end benchmarks on LLaMA-3.1-8B, Qwen-2.5-14B, and Qwen-2.5-32B, FlexLLM sustains the inference SLO requirements up to 20 req/s, and improves finetuning throughput by 1.9-4.8x under heavy inference workloads and 2.5-6.8x under light loads, preserving over 76% of peak finetuning progress even at peak demand. The source code of FlexLLM is publicly available at https://github.com/flexflow/FlexFlow/.
Diffusion models have achieved great success in synthesizing high-quality images. However, generating high-resolution images with diffusion models is still challenging due to the enormous computational costs, resulting in a prohibitive latency for interactive applications. In this paper, we propose DistriFusion to tackle this problem by leveraging parallelism across multiple GPUs. Our method splits the model input into multiple patches and assigns each patch to a GPU. However, naively implementing such an algorithm breaks the interaction between patches and loses fidelity, while incorporating such an interaction will incur tremendous communication overhead. To overcome this dilemma, we observe the high similarity between the input from adjacent diffusion steps and propose displaced patch parallelism, which takes advantage of the sequential nature of the diffusion process by reusing the pre-computed feature maps from the previous timestep to provide context for the current step. Therefore, our method supports asynchronous communication, which can be pipelined by computation. Extensive experiments show that our method can be applied to recent Stable Diffusion XL with no quality degradation and achieve up to a 6.1$\times$ speedup on eight NVIDIA A100s compared to one. Our code is publicly available at https://github.com/mit-han-lab/distrifuser.
Standardized benchmarks drive progress in machine learning. However, with repeated testing, the risk of overfitting grows as algorithms over-exploit benchmark idiosyncrasies. In our work, we seek to mitigate this challenge by compiling ever-expanding large-scale benchmarks called Lifelong Benchmarks. These benchmarks introduce a major challenge: the high cost of evaluating a growing number of models across very large sample sets. To address this challenge, we introduce an efficient framework for model evaluation, Sort & Search (S&S)}, which reuses previously evaluated models by leveraging dynamic programming algorithms to selectively rank and sub-select test samples. To test our approach at scale, we create Lifelong-CIFAR10 and Lifelong-ImageNet, containing 1.69M and 1.98M test samples for classification. Extensive empirical evaluations across over 31,000 models demonstrate that S&S achieves highly-efficient approximate accuracy measurement, reducing compute cost from 180 GPU days to 5 GPU hours (about 1000x reduction) on a single A100 GPU, with low approximation error and memory cost of <100MB. Our work also highlights issues with current accuracy prediction metrics, suggesting a need to move towards sample-level evaluation metrics. We hope to guide future research by showing our method's bottleneck lies primarily in generalizing Sort beyond a single rank order and not in improving Search.
Due to the three-dimensional nature of CT- or MR-scans, generative modeling of medical images is a particularly challenging task. Existing approaches mostly apply patch-wise, slice-wise, or cascaded generation techniques to fit the high-dimensional data into the limited GPU memory. However, these approaches may introduce artifacts and potentially restrict the model's applicability for certain downstream tasks. This work presents WDM, a wavelet-based medical image synthesis framework that applies a diffusion model on wavelet decomposed images. The presented approach is a simple yet effective way of scaling 3D diffusion models to high resolutions and can be trained on a single \SI{40}{\giga\byte} GPU. Experimental results on BraTS and LIDC-IDRI unconditional image generation at a resolution of $128 \times 128 \times 128$ demonstrate state-of-the-art image fidelity (FID) and sample diversity (MS-SSIM) scores compared to recent GANs, Diffusion Models, and Latent Diffusion Models. Our proposed method is the only one capable of generating high-quality images at a resolution of $256 \times 256 \times 256$, outperforming all comparing methods.
In the landscape of generative artificial intelligence, diffusion-based models present challenges for socio-technical systems in data requirements and privacy. Traditional approaches like federated learning distribute the learning process but strain individual clients, especially with constrained resources (e.g., edge devices). In response to these challenges, we introduce CollaFuse, a novel framework inspired by split learning. Tailored for efficient and collaborative use of denoising diffusion probabilistic models, CollaFuse enables shared server training and inference, alleviating client computational burdens. This is achieved by retaining data and computationally inexpensive GPU processes locally at each client while outsourcing the computationally expensive processes to the shared server. Demonstrated in a healthcare context, CollaFuse enhances privacy by highly reducing the need for sensitive information sharing. These capabilities hold the potential to impact various application areas, such as the design of edge computing solutions, healthcare research, or autonomous driving. In essence, our work advances distributed machine learning, shaping the future of collaborative GenAI networks.
Large Language Models (LLMs) have demonstrated a remarkable potential in medical knowledge acquisition and question-answering. However, LLMs can potentially hallucinate and yield factually incorrect outcomes, even with domain-specific pretraining. Previously, retrieval augmented generation (RAG) has limited success in addressing hallucinations. Unlike previous methods in RAG where the retrieval model was trained separately from the LLM, we introduce JMLR (for Jointly trains LLM and information Retrieval) during the fine-tuning phase. The synchronized training mechanism enhances JMLR's ability to retrieve clinical guidelines and leverage medical knowledge to reason and answer questions and reduces the demand for computational resources. We evaluated JMLR on the important medical question-answering application. Our experimental results demonstrate that JMLR-13B (70.5%) outperforms a previous state-of-the-art open-source model using conventional pre-training and fine-tuning Meditron-70B (68.9%) and Llama2-13B with RAG (67.7%) on a medical question-answering dataset. Comprehensive evaluations reveal JMLR-13B enhances reasoning quality and reduces hallucinations better than Claude3-Opus. Additionally, JMLR-13B (148 GPU hours) also trains much faster than Meditron-70B (42630 GPU hours). Through this work, we provide a new and efficient knowledge enhancement method for healthcare, demonstrating the potential of integrating retrieval and LLM training for medical question-answering systems.
Large language models (LLMs) have achieved significant success across various domains. However, training these LLMs typically involves substantial memory and computational costs during both forward and backward propagation. While parameter-efficient fine-tuning (PEFT) considerably reduces the training memory associated with parameters, it does not address the significant computational costs and activation memory. In this paper, we propose Dropping Backward Propagation (DropBP), a novel approach designed to reduce computational costs and activation memory while maintaining accuracy. DropBP randomly drops layers during backward propagation, which is essentially equivalent to training shallow submodules generated by undropped layers and residual connections. Additionally, DropBP calculates the sensitivity of each layer to assign an appropriate drop rate, thereby stabilizing the training process. DropBP is not only applicable to full fine-tuning but can also be orthogonally integrated with all types of PEFT by dropping layers during backward propagation. Specifically, DropBP can reduce training time by 44% with comparable accuracy to the baseline, accelerate convergence to the same perplexity by 1.5x, and enable training with a sequence length 6.2x larger on a single NVIDIA-A100 GPU. Furthermore, our DropBP enabled a throughput increase of 79% on a NVIDIA A100 GPU and 117% on an Intel Gaudi2 HPU. The code is available at https://github.com/WooSunghyeon/dropbp.
Large language models (LLMs) have demonstrated outstanding performance in various tasks, such as text summarization, text question-answering, and etc. While their performance is impressive, the computational footprint due to their vast number of parameters can be prohibitive. Existing solutions such as SparseGPT and Wanda attempt to alleviate this issue through weight pruning. However, their layer-wise approach results in significant perturbation to the model's output and requires meticulous hyperparameter tuning, such as the pruning rate, which can adversely affect overall model performance. To address this, this paper introduces a novel LLM pruning technique dubbed blockwise parameter-efficient sparsity allocation (BESA) by applying a blockwise reconstruction loss. In contrast to the typical layer-wise pruning techniques, BESA is characterized by two distinctive attributes: i) it targets the overall pruning error with respect to individual transformer blocks, and ii) it allocates layer-specific sparsity in a differentiable manner, both of which ensure reduced performance degradation after pruning. Our experiments show that BESA achieves state-of-the-art performance, efficiently pruning LLMs like LLaMA1, and LLaMA2 with 7B to 70B parameters on a single A100 GPU in just five hours. Code is available at https://github.com/OpenGVLab/LLMPrune-BESA.
Diffusion models (DMs) have enabled breakthroughs in image synthesis tasks but lack an intuitive interface for consistent image-to-image (I2I) translation. Various methods have been explored to address this issue, including mask-based methods, attention-based methods, and image-conditioning. However, it remains a critical challenge to enable unpaired I2I translation with pre-trained DMs while maintaining satisfying consistency. This paper introduces Cyclenet, a novel but simple method that incorporates cycle consistency into DMs to regularize image manipulation. We validate Cyclenet on unpaired I2I tasks of different granularities. Besides the scene and object level translation, we additionally contribute a multi-domain I2I translation dataset to study the physical state changes of objects. Our empirical studies show that Cyclenet is superior in translation consistency and quality, and can generate high-quality images for out-of-domain distributions with a simple change of the textual prompt. Cyclenet is a practical framework, which is robust even with very limited training data (around 2k) and requires minimal computational resources (1 GPU) to train. Project homepage: https://cyclenetweb.github.io/
Benchmarks play a crucial role in the development and analysis of reinforcement learning (RL) algorithms. We identify that existing benchmarks used for research into open-ended learning fall into one of two categories. Either they are too slow for meaningful research to be performed without enormous computational resources, like Crafter, NetHack and Minecraft, or they are not complex enough to pose a significant challenge, like Minigrid and Procgen. To remedy this, we first present Craftax-Classic: a ground-up rewrite of Crafter in JAX that runs up to 250x faster than the Python-native original. A run of PPO using 1 billion environment interactions finishes in under an hour using only a single GPU and averages 90% of the optimal reward. To provide a more compelling challenge we present the main Craftax benchmark, a significant extension of the Crafter mechanics with elements inspired from NetHack. Solving Craftax requires deep exploration, long term planning and memory, as well as continual adaptation to novel situations as more of the world is discovered. We show that existing methods including global and episodic exploration, as well as unsupervised environment design fail to make material progress on the benchmark. We believe that Craftax can for the first time allow researchers to experiment in a complex, open-ended environment with limited computational resources.
Graph Neural Networks (GNNs) are emerging ML models to analyze graph-structure data. Graph Neural Network (GNN) execution involves both compute-intensive and memory-intensive kernels, the latter dominates the total time, being significantly bottlenecked by data movement between memory and processors. Processing-In-Memory (PIM) systems can alleviate this data movement bottleneck by placing simple processors near or inside to memory arrays. In this work, we introduce PyGim, an efficient ML library that accelerates GNNs on real PIM systems. We propose intelligent parallelization techniques for memory-intensive kernels of GNNs tailored for real PIM systems, and develop handy Python API for them. We provide hybrid GNN execution, in which the compute-intensive and memory-intensive kernels are executed in processor-centric and memory-centric computing systems, respectively. We extensively evaluate PyGim on a real-world PIM system with 1992 PIM cores using emerging GNN models, and demonstrate that it outperforms its state-of-the-art CPU counterpart on Intel Xeon by on average 3.04x, and achieves higher resource utilization than CPU and GPU systems. Our work provides useful recommendations for software, system and hardware designers. PyGim is publicly available at https://github.com/CMU-SAFARI/PyGim.
Deep neural networks are typically trained using global error signals that backpropagate (BP) end-to-end, which is not only biologically implausible but also suffers from the update locking problem and requires huge memory consumption. Local learning, which updates each layer independently with a gradient-isolated auxiliary network, offers a promising alternative to address the above problems. However, existing local learning methods are confronted with a large accuracy gap with the BP counterpart, particularly for large-scale networks. This is due to the weak coupling between local layers and their subsequent network layers, as there is no gradient communication across layers. To tackle this issue, we put forward an augmented local learning method, dubbed AugLocal. AugLocal constructs each hidden layer's auxiliary network by uniformly selecting a small subset of layers from its subsequent network layers to enhance their synergy. We also propose to linearly reduce the depth of auxiliary networks as the hidden layer goes deeper, ensuring sufficient network capacity while reducing the computational cost of auxiliary networks. Our extensive experiments on four image classification datasets (i.e., CIFAR-10, SVHN, STL-10, and ImageNet) demonstrate that AugLocal can effectively scale up to tens of local layers with a comparable accuracy to BP-trained networks while reducing GPU memory usage by around 40%. The proposed AugLocal method, therefore, opens up a myriad of opportunities for training high-performance deep neural networks on resource-constrained platforms.Code is available at https://github.com/ChenxiangMA/AugLocal.
Learned sparse retrieval (LSR) is a family of neural methods that encode queries and documents into sparse lexical vectors that can be indexed and retrieved efficiently with an inverted index. We explore the application of LSR to the multi-modal domain, with a focus on text-image retrieval. While LSR has seen success in text retrieval, its application in multimodal retrieval remains underexplored. Current approaches like LexLIP and STAIR require complex multi-step training on massive datasets. Our proposed approach efficiently transforms dense vectors from a frozen dense model into sparse lexical vectors. We address issues of high dimension co-activation and semantic deviation through a new training algorithm, using Bernoulli random variables to control query expansion. Experiments with two dense models (BLIP, ALBEF) and two datasets (MSCOCO, Flickr30k) show that our proposed algorithm effectively reduces co-activation and semantic deviation. Our best-performing sparsified model outperforms state-of-the-art text-image LSR models with a shorter training time and lower GPU memory requirements. Our approach offers an effective solution for training LSR retrieval models in multimodal settings. Our code and model checkpoints are available at github.com/thongnt99/lsr-multimodal
The training paradigm of DETRs is heavily contingent upon pre-training their backbone on the ImageNet dataset. However, the limited supervisory signals provided by the image classification task and one-to-one matching strategy result in an inadequately pre-trained neck for DETRs. Additionally, the instability of matching in the early stages of training engenders inconsistencies in the optimization objectives of DETRs. To address these issues, we have devised an innovative training methodology termed step-by-step training. Specifically, in the first stage of training, we employ a classic detector, pre-trained with a one-to-many matching strategy, to initialize the backbone and neck of the end-to-end detector. In the second stage of training, we froze the backbone and neck of the end-to-end detector, necessitating the training of the decoder from scratch. Through the application of step-by-step training, we have introduced the first real-time end-to-end object detection model that utilizes a purely convolutional structure encoder, DETR with YOLO (DEYO). Without reliance on any supplementary training data, DEYO surpasses all existing real-time object detectors in both speed and accuracy. Moreover, the comprehensive DEYO series can complete its second-phase training on the COCO dataset using a single 8GB RTX 4060 GPU, significantly reducing the training expenditure. Source code and pre-trained models are available at https://github.com/ouyanghaodong/DEYO.
In this paper, we introduce a novel class of fast, beam search-based adversarial attack (BEAST) for Language Models (LMs). BEAST employs interpretable parameters, enabling attackers to balance between attack speed, success rate, and the readability of adversarial prompts. The computational efficiency of BEAST facilitates us to investigate its applications on LMs for jailbreaking, eliciting hallucinations, and privacy attacks. Our gradient-free targeted attack can jailbreak aligned LMs with high attack success rates within one minute. For instance, BEAST can jailbreak Vicuna-7B-v1.5 under one minute with a success rate of 89% when compared to a gradient-based baseline that takes over an hour to achieve 70% success rate using a single Nvidia RTX A6000 48GB GPU. Additionally, we discover a unique outcome wherein our untargeted attack induces hallucinations in LM chatbots. Through human evaluations, we find that our untargeted attack causes Vicuna-7B-v1.5 to produce ~15% more incorrect outputs when compared to LM outputs in the absence of our attack. We also learn that 22% of the time, BEAST causes Vicuna to generate outputs that are not relevant to the original prompt. Further, we use BEAST to generate adversarial prompts in a few seconds that can boost the performance of existing membership inference attacks for LMs. We believe that our fast attack, BEAST, has the potential to accelerate research in LM security and privacy. Our codebase is publicly available at https://github.com/vinusankars/BEAST.
As the modern CPU, GPU, and NPU chip design complexity and transistor counts keep increasing, and with the relentless shrinking of semiconductor technology nodes to nearly 1 nanometer, the placement and routing have gradually become the two most pivotal processes in modern very-large-scale-integrated (VLSI) circuit back-end design. How to evaluate routability efficiently and accurately in advance (at the placement and global routing stages) has grown into a crucial research area in the field of artificial intelligence (AI) assisted electronic design automation (EDA). In this paper, we propose a novel U-Net variant model boosted by an Inception embedded module to predict Routing Congestion (RC) and Design Rule Checking (DRC) hotspots. Experimental results on the recently published CircuitNet dataset benchmark show that our proposed method achieves up to 5% (RC) and 20% (DRC) rate reduction in terms of Avg-NRMSE (Average Normalized Root Mean Square Error) compared to the classic architecture. Furthermore, our approach consistently outperforms the prior model on the SSIM (Structural Similarity Index Measure) metric.
Conformal Prediction (CP) has attracted great attention from the research community due to its strict theoretical guarantees. However, researchers and developers still face challenges of applicability and efficiency when applying CP algorithms to deep learning models. In this paper, we introduce \torchcp, a comprehensive PyTorch-based toolkit to strengthen the usability of CP for deep learning models. \torchcp implements a wide range of post-hoc and training methods of conformal prediction for various machine learning tasks, including classification, regression, GNN, and LLM. Moreover, we provide user-friendly interfaces and extensive evaluations to easily integrate CP algorithms into specific tasks. Our \torchcp toolkit, built entirely with PyTorch, enables high-performance GPU acceleration for deep learning models and mini-batch computation on large-scale datasets. With the LGPL license, the code is open-sourced at \url{https://github.com/ml-stat-Sustech/TorchCP} and will be continuously updated.
Recent advancements in large language models (LLMs) like ChatGPT and LLaMA show promise in medical applications, yet challenges remain in medical language comprehension. This study presents Me-LLaMA, a new medical LLM family based on open-source LLaMA models, optimized for medical text analysis and diagnosis by leveraging large-scale, domain-specific datasets. The Me-LLaMA family, including foundation models Me-LLaMA 13/70B and their chat-enhanced versions, was developed through continued pre-training and instruction tuning with 129B tokens and 214K samples from biomedical and clinical sources. Training the 70B models required over 100,000 A100 GPU hours. Me-LLaMA's performance was evaluated across six medical text analysis tasks using 12 benchmark datasets and complex clinical case diagnosis, with automatic and human evaluations. Results indicate Me-LLaMA outperforms LLaMA and other open-source medical LLMs in zero-shot and supervised settings. Task-specific tuning further boosts performance, surpassing ChatGPT on 7 of 8 datasets and GPT-4 on 5 of 8. For complex clinical cases, Me-LLaMA achieves performance comparable to ChatGPT and GPT-4. This work underscores the importance of domain-specific data in developing medical LLMs and addresses the high computational costs involved in training, highlighting a balance between pre-training and fine-tuning strategies. Me-LLaMA models are now accessible under user agreements, providing a valuable resource for advancing medical AI.
Existing methods for fine-tuning sparse LLMs often suffer from resource-intensive requirements and high retraining costs. Additionally, many fine-tuning methods often rely on approximations or heuristic optimization strategies, which may lead to suboptimal solutions. To address these issues, we propose an efficient and fast framework for fine-tuning sparse LLMs based on minimizing reconstruction error. Our approach involves sampling a small dataset for calibration and utilizing backpropagation to iteratively optimize block-wise reconstruction error, on a block-by-block basis, aiming for optimal solutions. Extensive experiments on various benchmarks consistently demonstrate the superiority of our method over other baselines. For instance, on the Wikitext2 dataset with LlamaV1-7B at 70% sparsity, our proposed EBFT achieves a perplexity of 16.88, surpassing the state-of-the-art DSnoT with a perplexity of 75.14. Moreover, with a structured sparsity ratio of 26\%, EBFT achieves a perplexity of 16.27, outperforming LoRA (perplexity 16.44). Furthermore, the fine-tuning process of EBFT for LlamaV1-7B only takes approximately 30 minutes, and the entire framework can be executed on a single 16GB GPU. The source code is available at https://github.com/sunggo/EBFT.
This paper presents Mechanistic Neural Networks, a neural network design for machine learning applications in the sciences. It incorporates a new Mechanistic Block in standard architectures to explicitly learn governing differential equations as representations, revealing the underlying dynamics of data and enhancing interpretability and efficiency in data modeling. Central to our approach is a novel Relaxed Linear Programming Solver (NeuRLP) inspired by a technique that reduces solving linear ODEs to solving linear programs. This integrates well with neural networks and surpasses the limitations of traditional ODE solvers enabling scalable GPU parallel processing. Overall, Mechanistic Neural Networks demonstrate their versatility for scientific machine learning applications, adeptly managing tasks from equation discovery to dynamic systems modeling. We prove their comprehensive capabilities in analyzing and interpreting complex scientific data across various applications, showing significant performance against specialized state-of-the-art methods.