Visual language models (VLMs) rapidly progressed with the recent success of large language models. There have been growing efforts on visual instruction tuning to extend the LLM with visual inputs, but lacks an in-depth study of the visual language pre-training process, where the model learns to perform joint modeling on both modalities. In this work, we examine the design options for VLM pre-training by augmenting LLM towards VLM through step-by-step controllable comparisons. We introduce three main findings: (1) freezing LLMs during pre-training can achieve decent zero-shot performance, but lack in-context learning capability, which requires unfreezing the LLM; (2) interleaved pre-training data is beneficial whereas image-text pairs alone are not optimal; (3) re-blending text-only instruction data to image-text data during instruction fine-tuning not only remedies the degradation of text-only tasks, but also boosts VLM task accuracy. With an enhanced pre-training recipe we build VILA, a Visual Language model family that consistently outperforms the state-of-the-art models, e.g., LLaVA-1.5, across main benchmarks without bells and whistles. Multi-modal pre-training also helps unveil appealing properties of VILA, including multi-image reasoning, enhanced in-context learning, and better world knowledge. VILA is also deployable on Jetson Orin for on-device VLM.

Large language models (LLMs) show excellent performance but are compute- and memory-intensive. Quantization can reduce memory and accelerate inference. However, existing methods cannot maintain accuracy and hardware efficiency at the same time. We propose SmoothQuant, a training-free, accuracy-preserving, and general-purpose post-training quantization (PTQ) solution to enable 8-bit weight, 8-bit activation (W8A8) quantization for LLMs. Based on the fact that weights are easy to quantize while activations are not, SmoothQuant smooths the activation outliers by offline migrating the quantization difficulty from activations to weights with a mathematically equivalent transformation. SmoothQuant enables an INT8 quantization of both weights and activations for all the matrix multiplications in LLMs, including OPT, BLOOM, GLM, MT-NLG, and LLaMA family. We demonstrate up to 1.56$\times$ speedup and 2$\times$ memory reduction for LLMs with negligible loss in accuracy. SmoothQuant enables serving 530B LLM within a single node. Our work offers a turn-key solution that reduces hardware costs and democratizes LLMs.

On-device learning and efficient fine-tuning enable continuous and privacy-preserving customization (e.g., locally fine-tuning large language models on personalized data). However, existing training frameworks are designed for cloud servers with powerful accelerators (e.g., GPUs, TPUs) and lack the optimizations for learning on the edge, which faces challenges of resource limitations and edge hardware diversity. We introduce PockEngine: a tiny, sparse and efficient engine to enable fine-tuning on various edge devices. PockEngine supports sparse backpropagation: it prunes the backward graph and sparsely updates the model with measured memory saving and latency reduction while maintaining the model quality. Secondly, PockEngine is compilation first: the entire training graph (including forward, backward and optimization steps) is derived at compile-time, which reduces the runtime overhead and brings opportunities for graph transformations. PockEngine also integrates a rich set of training graph optimizations, thus can further accelerate the training cost, including operator reordering and backend switching. PockEngine supports diverse applications, frontends and hardware backends: it flexibly compiles and tunes models defined in PyTorch/TensorFlow/Jax and deploys binaries to mobile CPU/GPU/DSPs. We evaluated PockEngine on both vision models and large language models. PockEngine achieves up to 15 × speedup over off-the-shelf TensorFlow (Raspberry Pi), 5.6 × memory saving back-propagation (Jetson AGX Orin). Remarkably, PockEngine enables fine-tuning LLaMav2-7B on NVIDIA Jetson AGX Orin at 550 tokens/s, 7.9 × faster than the PyTorch.

During image editing, existing deep generative models tend to re-synthesize the entire output from scratch, including the unedited regions. This leads to a significant waste of computation, especially for minor editing operations. In this work, we present Spatially Sparse Inference (SSI), a general-purpose technique that selectively performs computation for edited regions and accelerates various generative models, including both conditional GANs and diffusion models. Our key observation is that users prone to gradually edit the input image. This motivates us to cache and reuse the feature maps of the original image. Given an edited image, we sparsely apply the convolutional filters to the edited regions while reusing the cached features for the unedited areas. Based on our algorithm, we further propose Sparse Incremental Generative Engine (SIGE) to convert the computation reduction to latency reduction on off-the-shelf hardware. With about 1%-area edits, SIGE accelerates DDPM by 3.0× on NVIDIA RTX 3090 and 4.6× on Apple M1 Pro GPU, Stable Diffusion by 7.2× on 3090, and GauGAN by 5.6× on 3090 and 5.2× on M1 Pro GPU. Compared to our conference paper, we enhance SIGE to accommodate attention layers and apply it to Stable Diffusion. Additionally, we offer support for Apple M1 Pro GPU and include more results to substantiate the efficacy of our method.

On-device training enables the model to adapt to new data collected from the sensors by fine-tuning a pre-trained model. Users can benefit from customized AI models without having to transfer the data to the cloud, protecting the privacy. However, the training memory consumption is prohibitive for IoT devices that have tiny memory resources. We propose an algorithm-system co-design framework to make on-device training possible with only 256KB of memory. On-device training faces two unique challenges: (1) the quantized graphs of neural networks are hard to optimize due to low bit-precision and the lack of normalization; (2) the limited hardware resource (memory and computation) does not allow full backpropagation. To cope with the optimization difficulty, we propose Quantization- Aware Scaling to calibrate the gradient scales and stabilize 8-bit quantized training. To reduce the memory footprint, we propose Sparse Update to skip the gradient computation of less important layers and sub-tensors. The algorithm innovation is implemented by a lightweight training system, Tiny Training Engine, which prunes the backward computation graph to support sparse updates and offload the runtime auto-differentiation to compile time. Our framework is the first practical solution for on-device transfer learning of visual recognition on tiny IoT devices (e.g., a microcontroller with only 256KB SRAM), using less than 1/1000 of the memory of PyTorch and TensorFlow while matching the accuracy. Our study enables IoT devices not only to perform inference but also to continuously adapt to new data for on-device lifelong learning. A video demo can be found here: https://youtu.be/XaDCO8YtmBw.

During image editing, existing deep generative models tend to re-synthesize the entire output from scratch, including the unedited regions. This leads to a significant waste of computation, especially for minor editing operations. In this work, we present Spatially Sparse Inference (SSI), a general-purpose technique that selectively performs computation for edited regions and accelerates various generative models, including both conditional GANs and diffusion models. Our key observation is that users tend to make gradual changes to the input image. This motivates us to cache and reuse the feature maps of the original image. Given an edited image, we sparsely apply the convolutional filters to the edited regions while reusing the cached features for the unedited regions. Based on our algorithm, we further propose Sparse Incremental Generative Engine (SIGE) to convert the computation reduction to latency reduction on off-the-shelf hardware. With 1.2%-area edited regions, our method reduces the computation of DDIM by $7.5\times$ and GauGAN by $18\times$ while preserving the visual fidelity. With SIGE, we accelerate the inference time of DDIM by $3.0\times$ on RTX 3090 and $6.6\times$ on Apple M1 Pro CPU, and GauGAN by $4.2\times$ on RTX 3090 and $14\times$ on Apple M1 Pro CPU.

Conditional Generative Adversarial Networks (cGANs) have enabled controllable image synthesis for many vision and graphics applications. However, recent cGANs are 1-2 orders of magnitude more compute-intensive than modern recognition CNNs. For example, GauGAN consumes 281G MACs per image, compared to 0.44G MACs for MobileNet-v3, making it difficult for interactive deployment. In this work, we propose a general-purpose compression framework for reducing the inference time and model size of the generator in cGANs. Directly applying existing compression methods yields poor performance due to the difficulty of GAN training and the differences in generator architectures. We address these challenges in two ways. First, to stabilize GAN training, we transfer knowledge of multiple intermediate representations of the original model to its compressed model and unify unpaired and paired learning. Second, instead of reusing existing CNN designs, our method finds efficient architectures via neural architecture search. To accelerate the search process, we decouple the model training and search via weight sharing. Experiments demonstrate the effectiveness of our method across different supervision settings, network architectures, and learning methods. Without losing image quality, we reduce the computation of CycleGAN by 21x, Pix2pix by 12x, MUNIT by 29x, and GauGAN by 9x, paving the way for interactive image synthesis.

The explosive growth in video streaming requires video understanding at high accuracy and low computation cost. Conventional 2D CNNs are computationally cheap but cannot capture temporal relationships; 3D CNN based methods can achieve good performance but are computationally intensive. In this paper, we propose a generic and effective Temporal Shift Module (TSM) that enjoys both high efficiency and high performance. The key idea of TSM is to shift part of the channels along the temporal dimension, thus facilitate information exchanged among neighboring frames. It can be inserted into 2D CNNs to achieve temporal modeling at zero computation and zero parameters. TSM offers several unique advantages. First, TSM has high performance; it ranks the first on the Something-Something leaderboard upon submission. Second, TSM has high efficiency; it achieves a high frame rate of 74fps and 29fps for online video recognition on Jetson Nano and Galaxy Note8. Third, TSM has higher scalability compared to 3D networks, enabling large-scale Kinetics training on 1,536 GPUs in 15 minutes. Lastly, TSM enables action concepts learning, which 2D networks cannot model; we visualize the category attention map and find that spatial-temporal action detector emerges during the training of classification tasks. The code is publicly available at https://github.com/mit-han-lab/temporal-shift-module.

Deep neural networks (DNNs) have achieved unprecedented success in the field of artificial intelligence (AI), including computer vision, natural language processing, and speech recognition. However, their superior performance comes at the considerable cost of computational complexity, which greatly hinders their applications in many resource-constrained devices, such as mobile phones and Internet of Things (IoT) devices. Therefore, methods and techniques that are able to lift the efficiency bottleneck while preserving the high accuracy of DNNs are in great demand to enable numerous edge AI applications. This article provides an overview of efficient deep learning methods, systems, and applications. We start from introducing popular model compression methods, including pruning, factorization, quantization, as well as compact model design. To reduce the large design cost of these manual solutions, we discuss the AutoML framework for each of them, such as neural architecture search (NAS) and automated pruning and quantization. We then cover efficient on-device training to enable user customization based on the local data on mobile devices. Apart from general acceleration techniques, we also showcase several task-specific accelerations for point cloud, video, and natural language processing by exploiting their spatial sparsity and temporal/token redundancy. Finally, to support all these algorithmic advancements, we introduce the efficient deep learning system design from both software and hardware perspectives.

Generative adversarial networks (GANs) have enabled photorealistic image synthesis and editing. However, due to the high computational cost of large-scale generators (e.g., StyleGAN2), it usually takes seconds to see the results of a single edit on edge devices, prohibiting interactive user experience. In this paper, inspired by quick preview features in modern rendering software, we propose Anycost GAN for interactive natural image editing. We train the Anycost GAN to support elastic resolutions and channels for faster image generation at versatile speeds. Running subsets of the full generator produce outputs that are perceptually similar to the full generator, making them a good proxy for quick preview. By using sampling-based multi-resolution training, adaptive-channel training, and a generator-conditioned discriminator, the anycost generator can be evaluated at various configurations while achieving better image quality compared to separately trained models. Furthermore, we develop new encoder training and latent code optimization techniques to encourage consistency between the different sub-generators during image projection. Anycost GAN can be executed at various cost budgets (up to 10× computation reduction) and adapt to a wide range of hardware and la tency requirements. When deployed on desktop CPUs and edge devices, our model can provide perceptually similar previews at 6-12× speedup, enabling interactive image editing. The ${\color{RubineRed}{code}}$ and ${\color{RubineRed}{demo}}$ are publicly available.

Tiny deep learning on microcontroller units (MCUs) is challenging due to the limited memory size. We find that the memory bottleneck is due to the imbalanced memory distribution in convolutional neural network (CNN) designs: the first several blocks have an order of magnitude larger memory usage than the rest of the network. To alleviate this issue, we propose a generic patch-by-patch inference scheduling, which operates only on a small spatial region of the feature map and significantly cuts down the peak memory. However, naive implementation brings overlapping patches and computation overhead. We further propose receptive field redistribution to shift the receptive field and FLOPs to the later stage and reduce the computation overhead. Manually redistributing the receptive field is difficult. We automate the process with neural architecture search to jointly optimize the neural architecture and inference scheduling, leading to MCUNetV2. Patch-based inference effectively reduces the peak memory usage of existing networks by4-8×. Co-designed with neural networks, MCUNetV2 sets a record ImageNetaccuracy on MCU (71.8%) and achieves >90% accuracy on the visual wake words dataset under only 32kB SRAM. MCUNetV2 also unblocks object detection on tiny devices, achieving 16.9% higher mAP on Pascal VOC compared to the state-of-the-art result. Our study largely addressed the memory bottleneck in tinyML and paved the way for various vision applications beyond image classification.

Conditional Generative Adversarial Networks (cGANs) have enabled controllable image synthesis for many vision and graphics applications. However, recent cGANs are 1-2 orders of magnitude more compute-intensive than modern recognition CNNs. For example, GauGAN consumes 281G MACs per image, compared to 0.44G MACs for MobileNet-v3, making it difficult for interactive deployment. In this work, we propose a general-purpose compression framework for reducing the inference time and model size of the generator in cGANs. Directly applying existing compression methods yields poor performance due to the difficulty of GAN training and the differences in generator architectures. We address these challenges in two ways. First, to stabilize GAN training, we transfer knowledge of multiple intermediate representations of the original model to its compressed model and unify unpaired and paired learning. Second, instead of reusing existing CNN designs, our method finds efficient architectures via neural architecture search. To accelerate the search process, we decouple the model training and search via weight sharing. Experiments demonstrate the effectiveness of our method across different supervision settings, network architectures, and learning methods. Without losing image quality, we reduce the computation of CycleGAN by 21×, Pix2pix by 12×, MUNIT by 29×, and GauGAN by 9×, paving the way for interactive image synthesis.

We present APQ, a novel design methodology for efficient deep learning deployment. Unlike previous methods that separately optimize the neural network architecture, pruning policy, and quantization policy, we design to optimize them in a joint manner. To deal with the larger design space it brings, we devise to train a quantization-aware accuracy predictor that is fed to the evolutionary search to select the best fit. Since directly training such a predictor requires time-consuming quantization data collection, we propose to use predictor-transfer technique to get the quantization-aware predictor: we first generate a large dataset of 〈NN architecture, ImageNet accuracy〉 pairs by sampling a pretrained unified once-for-all network and doing direct evaluation; then we use these data to train an accuracy predictor without quantization, followed by transferring its weights to train the quantization-aware predictor, which largely reduces the quantization data collection time. Extensive experiments on ImageNet show the benefits of this joint design methodology: the model searched by our method maintains the same level accuracy as ResNet34 8-bit model while saving 8× BitOps; we achieve 2×/1.3× latency/energy saving compared to MobileNetV2+HAQ [30, 36] while obtaining the same level accuracy; the marginal search cost ofjoint optimization for a new deployment scenario outperforms separate optimizations using ProxylessNAS+AMC+HAQ [5, 12, 36] by 2.3% accuracy while reducing orders of magnitude GPU hours and CO2emission with respect to the training cost.

Machine learning on tiny IoT devices based on microcontroller units (MCU) is appealing but challenging: the memory of microcontrollers is 2-3 orders of magnitude smaller even than mobile phones. We propose MCUNet, a framework that jointly designs the efficient neural architecture (TinyNAS) and the lightweight inference engine (TinyEngine), enabling ImageNet-scale inference on microcontrollers. TinyNAS adopts a two-stage neural architecture search approach that first optimizes the search space to fit the resource constraints, then specializes the network architecture in the optimized search space. TinyNAS can automatically handle diverse constraints (i.e. device, latency, energy, memory) under low search costs. TinyNAS is co-designed with TinyEngine, a memory-efficient inference library to expand the search space and fit a larger model. TinyEngine adapts the memory scheduling according to the overall network topology rather than layer-wise optimization, reducing the memory usage by 3.4×, and accelerating the inference by 1.7-3.3× compared to TF-Lite Micro [3] and CMSIS-NN [28]. MCUNet is the first to achieves >70% ImageNet top1 accuracy on an off-the-shelf commercial microcontroller, using 3.5× less SRAM and 5.7× less Flash compared to quantized MobileNetV2 and ResNet-18. On visual&audio wake words tasks, MCUNet achieves state-of-the-art accuracy and runs 2.4-3.4× faster than Mo- bileNetV2 and ProxylessNAS-based solutions with 3.7-4.1× smaller peak SRAM. Our study suggests that the era of always-on tiny machine learning on IoT devices has arrived.

The performance of generative adversarial networks (GANs) heavily deteriorates given a limited amount of training data. This is mainly because the discriminatorsis memorizing the exact training set. To combat it, we propose Differentiable Augmentation (DiffAugment), a simple method that improves the data efficiency of GANs by imposing various types of differentiable augmentations on both real and fake samples. Previous attempts to directly augment the training data manipulate the distribution of real images, yielding little benefit; DiffAugment enables us to adopt the differentiable augmentation for the generated samples, effectively stabilizes training, and leads to better convergence. Experiments demonstrate consistent gains of our method over a variety of GAN architectures and loss functions for both unconditional and class-conditional generation. With DiffAugment, we achieve astate-of-the-art FID of 6.80 with an IS of 100.8 on ImageNet 128×128 and 2-4× reductions of FID given 1,000 images on FFHQ and LSUN. Furthermore, with only 20% training data, we can match the top performance on CIFAR-10 and CIFAR-100. Finally, our method can generate high-fidelity images using only 100 images without pre-training, while being on par with existing transfer learning algorithms. Code is available at https://github.com/mit-han-lab/data-efficient-gans.

Model quantization is a widely used technique to compress and accelerate deep neural network (DNN) inference. Emergent DNN hardware accelerators begin to support mixed precision (1-8 bits) to further improve the computation efficiency, which raises a great challenge to find the optimal bitwidth for each layer: it requires domain experts to explore the vast design space trading off among accuracy, latency, energy, and model size, which is both time-consuming and sub-optimal. There are plenty of specialized hardware for neural networks, but little research has been done for specialized neural network optimization for a particular hardware architecture. Conventional quantization algorithm ignores the different hardware architectures and quantizes all the layers in a uniform way. In this paper, we introduce the Hardware-Aware Automated Quantization (HAQ) framework which leverages the reinforcement learning to automatically determine the quantization policy, and we take the hardware accelerator's feedback in the design loop. Rather than relying on proxy signals such as FLOPs and model size, we employ a hardware simulator to generate direct feedback signals (latency and energy) to the RL agent. Compared with conventional methods, our framework is fully automated and can specialize the quantization policy for different neural network architectures and hardware architectures. Our framework effectively reduced the latency by 1.4-1.95× and the energy consumption by 1.9× with negligible loss of accuracy compared with the fixed bitwidth (8 bits) quantization. Our framework reveals that the optimal policies on different hardware architectures (i.e., edge and cloud architectures) under different resource constraints (i.e., latency, energy and model size) are drastically different. We interpreted the implication of different quantization policies, which offer insights for both neural network architecture design and hardware architecture design.

Vehicle 3D extents and trajectories are critical cues for predicting the future location of vehicles and planning future agent ego-motion based on those predictions. In this paper, we propose a novel online framework for 3D vehicle detection and tracking from monocular videos. The framework can not only associate detections of vehicles in motion over time, but also estimate their complete 3D bounding box information from a sequence of 2D images captured on a moving platform. Our method leverages 3D box depth-ordering matching for robust instance association and utilizes 3D trajectory prediction for re-identification of occluded vehicles. We also design a motion learning module based on an LSTM for more accurate long-term motion extrapolation. Our experiments on simulation, KITTI, and Argoverse datasets show that our 3D tracking pipeline offers robust data association and tracking. On Argoverse, our image-based method is significantly better for tracking 3D vehicles within 30 meters than the LiDAR-centric baseline methods.

The explosive growth in video streaming gives rise to challenges on performing video understanding at high accuracy and low computation cost. Conventional 2D CNNs are computationally cheap but cannot capture temporal relationships; 3D CNN based methods can achieve good performance but are computationally intensive, making it expensive to deploy. In this paper, we propose a generic and effective Temporal Shift Module (TSM) that enjoys both high efficiency and high performance. Specifically, it can achieve the performance of 3D CNN but maintain 2D CNN's complexity. TSM shifts part of the channels along the temporal dimension; thus facilitate information exchanged among neighboring frames. It can be inserted into 2D CNNs to achieve temporal modeling at zero computation and zero parameters. We also extended TSM to online setting, which enables real-time low-latency online video recognition and video object detection. TSM is accurate and efficient: it ranks the first place on the Something-Something leaderboard upon publication; on Jetson Nano and Galaxy Note8, it achieves a low latency of 13ms and 35ms for online video recognition. The code is available at: https://github. com/mit-han-lab/temporal-shift-module.

In this paper, we propose a generic Runtime Network Routing (RNR) framework for efficient image classification, which selects an optimal path inside the network. Unlike existing static neural network acceleration methods, our method preserves the full ability of the original large network and conducts dynamic routing at runtime according to the input image and current feature maps. The routing is performed in a bottom-up, layer-by-layer manner, where we model it as a Markov decision process and use reinforcement learning for training. The agent determines the estimated reward of each sub-path and conducts routing conditioned on different samples, where a faster path is taken when the image is easier for the task. Since the ability of network is fully preserved, the balance point is easily adjustable according to the available resources. We test our method on both multi-path residual networks and incremental convolutional channel pruning, and show that RNR consistently outperforms static methods at the same computation complexity on both the CIFAR and ImageNet datasets. Our method can also be applied to off-the-shelf neural network structures and easily extended to other application scenarios.

In this paper, we propose a consistent-aware deep learning (CADL) framework for person re-identification in a camera network. Unlike most existing person re-identification methods which identify whether two body images are from the same person, our approach aims to obtain the maximal correct matches for the whole camera network. Different from recently proposed camera network based re-identification methods which only consider the consistent information in the matching stage to obtain a global optimal association, we exploit such consistent-aware information under a deep learning framework where both feature representation and image matching are automatically learned with certain consistent constraints. Specifically, we reach the global optimal solution and balance the performance between different cameras by optimizing the similarity and association iteratively. Experimental results show that our method obtains significant performance improvement and outperforms the state-of-the-art methods by large margins.

In this paper, we propose a discriminative aggregation network (DAN) method for video face recognition, which aims to integrate information from video frames effectively and efficiently. Unlike existing aggregation methods, our method aggregates raw video frames directly instead of the features obtained by complex processing. By combining the idea of metric learning and adversarial learning, we learn an aggregation network that produces more discriminative synthesized images compared to raw input frames. Our framework reduces the number of frames to be processed and significantly speed up the recognition procedure. Furthermore, low-quality frames containing misleading information are filtered and denoised during the aggregation process, which makes our system more robust and discriminative. Experimental results show that our method can generate discriminative images from video clips and improve the overall recognition performance in both the speed and accuracy on three widely used datasets.

In this paper, we propose a Runtime Neural Pruning (RNP) framework which prunes the deep neural network dynamically at the runtime. Unlike existing neural pruning methods which produce a fixed pruned model for deployment, our method preserves the full ability of the original network and conducts pruning according to the input image and current feature maps adaptively. The pruning is performed in a bottom-up, layer-by-layer manner, which we model as a Markov decision process and use reinforcement learning for training. The agent judges the importance of each convolutional kernel and conducts channel-wise pruning conditioned on different samples, where the network is pruned more when the image is easier for the task. Since the ability of network is fully preserved, the balance point is easily adjustable according to the available resources. Our method can be applied to off-the-shelf network structures and reach a better tradeoff between speed and accuracy, especially with a large pruning rate.