Diffusion Transformers (DiTs) dominate video generation but their high computational cost severely limits real-world applicability, usually requiring tens of minutes to generate a few seconds of video even on high-performance GPUs. This inefficiency primarily arises from the quadratic computational complexity of 3D Full Attention with respect to the context length. In this paper, we propose a training-free framework termed Sparse VideoGen (SVG) that leverages the inherent sparsity in 3D Full Attention to boost inference efficiency. We reveal that the attention heads can be dynamically classified into two groups depending on distinct sparse patterns: (1) Spatial Head, where only spatially-related tokens within each frame dominate the attention output, and (2) Temporal Head, where only temporally-related tokens across different frames dominate. Based on this insight, SVG proposes an online profiling strategy to capture the dynamic sparse patterns and predicts the type of attention head. Combined with a novel hardware-efficient tensor layout transformation and customized kernel implementations, SVG achieves up to 2.28x and 2.33x end-to-end speedup on CogVideoX-v1.5 and HunyuanVideo, respectively, while preserving generation quality.

Existing autoencoders have demonstrated impressive results at a moderate spatial compression ratio (e.g., 8), but fail to maintain satisfactory reconstruction accuracy for high spatial compression ratios (e.g., 64). We address this challenge by introducing two key techniques: (1) Residual Autoencoding, where we design our models to learn residuals based on the space-to-channel transformed features to alleviate the optimization difficulty of high spatial-compression autoencoders; (2) Decoupled High-Resolution Adaptation, an efficient decoupled three-phase training strategy for mitigating the generalization penalty of high spatial-compression autoencoders. With these designs, we improve the autoencoder's spatial compression ratio up to 128 while maintaining the reconstruction quality. Applying our DC-AE to latent diffusion models, we achieve significant speedup without accuracy drop. For example, on ImageNet 512 512, our DC-AE provides 19.1 inference speedup and 17.9 training speedup on H100 GPU for UViT-H while achieving a better FID, compared with the widely used SD-VAE-f8 autoencoder. Latent diffusion models (Rombach et al., 2022) have emerged as a leading framework and demonstrated great success in image synthesis (Labs, 2024; Esser et al., 2024). They employ an autoencoder to project the images to the latent space to reduce the cost of diffusion models.

FP8 training has emerged as a promising method for improving training efficiency. Existing frameworks accelerate training by applying FP8 computation to linear layers while leaving optimizer states and activations in higher precision, which fails to fully optimize memory usage. This paper introduces COAT (Compressing Optimizer States and Activations for FP8 Training), a novel FP8 training framework designed to significantly reduce memory footprint when training large models. COAT addresses current limitations through two key innovations: (1) Dynamic Range Expansion, which aligns optimizer state distributions more closely with the FP8 representation range, thereby reducing quantization error, and (2) Mixed-Granularity Activation Quantization, which optimizes activation memory using a combination of per-tensor and per-group quantization strategies. Experiments demonstrate that COAT effectively reduces end-to-end training memory footprint by 1.54 compared to BF16 while achieving nearly lossless performance across various tasks, such as Large Language Model pretraining and fine-tuning and Vision Language Model training. COAT also achieves a 1.43 end-to-end training speedup compared to BF16, performing on par with or surpassing TransformerEngine's speedup. COAT enables efficient full-parameter training of large models on fewer GPUs, and facilitates doubling the batch size in distributed training settings, providing a practical solution for scaling large-scale model training. The code is available at https://github.com/NVlabs/COAT. Both the optimizer states and activations are quantized to FP8 in COAT. Part of the work done during an internship at NVIDIA. However, the training of such models, which often comprise billions of parameters, demands substantial computational resources and memory. This presents substantial challenges, making the training of these foundation models very challenging (Smith et al., 2022; Hoffmann et al., 2022). Low-precision training has emerged as a promising approach to make FMs training more efficient (Micikevicius et al., 2017; Wang et al., 2018; Zhu et al., 2020; Xi et al., 2023; Wortsman et al., 2023; Xi et al., 2024).

Figure 1: HART is an early autoregressive model that can directly generate 1024 1024 images with quality comparable to diffusion models, while offering significantly improved efficiency. It achieves 4.5-7.7 higher throughput, 3.1-5.9 Check out our online demo and video. We introduce Hybrid Autoregressive Transformer (HART), an autoregressive (AR) visual generation model capable of directly generating 1024 1024 images, rivaling diffusion models in image generation quality. Existing AR models face limitations due to the poor image reconstruction quality of their discrete tokenizers and the prohibitive training costs associated with generating 1024px images. To address these challenges, we present the hybrid tokenizer, which decomposes the continuous latents from the autoencoder into two components: discrete tokens representing the big picture and continuous tokens representing the residual components that cannot be represented by the discrete tokens. The discrete component is modeled by a scalable-resolution discrete AR model, while the continuous component is learned with a lightweight residual diffusion module with only 37M parameters. Compared with the discrete-only VAR tokenizer, our hybrid approach improves reconstruction FID from 2.11 to 0.30 on MJHQ-30K, leading to a 31% generation FID improvement from 7.85 to 5.38. HART also outperforms state-of-the-art diffusion models in both FID and CLIP score, with 4.5-7.7 higher throughput and 6.9-13.4 Part of the work was done when Haotian Tang and Shang Yang were summer interns at NVIDIA. Prompt: A panda that has been cybernetically enhanced.

We present Condition-Aware Neural Network (CAN), a new method for adding control to image generative models. In parallel to prior conditional control methods, CAN controls the image generation process by dynamically manipulating the weight of the neural network. This is achieved by introducing a condition-aware weight generation module that generates conditional weight for convolution/linear layers based on the input condition. We test CAN on class-conditional image generation on ImageNet and text-to-image generation on COCO. CAN consistently delivers significant improvements for diffusion transformer models, including DiT and UViT. In particular, CAN combined with EfficientViT (CaT) achieves 2.78 FID on ImageNet 512x512, surpassing DiT-XL/2 while requiring 52x fewer MACs per sampling step.

We present EfficientViT-SAM, a new family of accelerated segment anything models. We retain SAM's lightweight prompt encoder and mask decoder while replacing the heavy image encoder with EfficientViT. For the training, we begin with the knowledge distillation from the SAM-ViT-H image encoder to EfficientViT. Subsequently, we conduct end-to-end training on the SA-1B dataset. Benefiting from EfficientViT's efficiency and capacity, EfficientViT-SAM delivers 48.9x measured TensorRT speedup on A100 GPU over SAM-ViT-H without sacrificing performance. Our code and pre-trained models are released at https://github.com/mit-han-lab/efficientvit.

We introduce Network Augmentation (NetAug), a new training method for improving the performance of tiny neural networks. Existing regularization techniques (e.g., data augmentation, dropout) have shown much success on large neural networks (e.g., ResNet50) by adding noise to overcome over-fitting. However, we found these techniques hurt the performance of tiny neural networks. We argue that training tiny models are different from large models: rather than augmenting the data, we should augment the model, since tiny models tend to suffer from under-fitting rather than over-fitting due to limited capacity. To alleviate this issue, NetAug augments the network (reverse dropout) instead of inserting noise into the dataset or the network. It puts the tiny model into larger models and encourages it to work as a sub-model of larger models to get extra supervision, in addition to functioning as an independent model. At test time, only the tiny model is used for inference, incurring zero inference overhead. We demonstrate the effectiveness of NetAug on image classification and object detection. NetAug consistently improves the performance of tiny models, achieving up to 2.1% accuracy improvement on ImageNet, and 4.3% on Cars. On Pascal VOC, NetAug provides 2.96% mAP improvement with the same computational cost. Tiny IoT devices are witnessing rapid growth, reaching 75.44 billion by 2025 (iot).

We present APQ for efficient deep learning inference on resource-constrained hardware. Unlike previous methods that separately search the neural architecture, pruning policy, and quantization policy, we optimize them in a joint manner. To deal with the larger design space it brings, a promising approach is to train a quantization-aware accuracy predictor to quickly get the accuracy of the quantized model and feed it to the search engine to select the best fit. However, training this quantization-aware accuracy predictor requires collecting a large number of quantized pairs, which involves quantization-aware finetuning and thus is highly time-consuming. To tackle this challenge, we propose to transfer the knowledge from a full-precision (i.e., fp32) accuracy predictor to the quantization-aware (i.e., int8) accuracy predictor, which greatly improves the sample efficiency. Besides, collecting the dataset for the fp32 accuracy predictor only requires to evaluate neural networks without any training cost by sampling from a pretrained once-for-all network, which is highly efficient. Extensive experiments on ImageNet demonstrate the benefits of our joint optimization approach. With the same accuracy, APQ reduces the latency/energy by 2x/1.3x over MobileNetV2+HAQ. Compared to the separate optimization approach (ProxylessNAS+AMC+HAQ), APQ achieves 2.3% higher ImageNet accuracy while reducing orders of magnitude GPU hours and CO2 emission, pushing the frontier for green AI that is environmental-friendly. The code and video are publicly available.

Transformers are ubiquitous in Natural Language Processing (NLP) tasks, but they are difficult to be deployed on hardware due to the intensive computation. To enable low-latency inference on resource-constrained hardware platforms, we propose to design Hardware-Aware Transformers (HAT) with neural architecture search. We first construct a large design space with $\textit{arbitrary encoder-decoder attention}$ and $\textit{heterogeneous layers}$. Then we train a $\textit{SuperTransformer}$ that covers all candidates in the design space, and efficiently produces many $\textit{SubTransformers}$ with weight sharing. Finally, we perform an evolutionary search with a hardware latency constraint to find a specialized $\textit{SubTransformer}$ dedicated to run fast on the target hardware. Extensive experiments on four machine translation tasks demonstrate that HAT can discover efficient models for different hardware (CPU, GPU, IoT device). When running WMT'14 translation task on Raspberry Pi-4, HAT can achieve $\textbf{3}\times$ speedup, $\textbf{3.7}\times$ smaller size over baseline Transformer; $\textbf{2.7}\times$ speedup, $\textbf{3.6}\times$ smaller size over Evolved Transformer with $\textbf{12,041}\times$ less search cost and no performance loss. HAT code is https://github.com/mit-han-lab/hardware-aware-transformers.git

Efficient deployment of deep learning models requires specialized neural network architectures to best fit different hardware platforms and efficiency constraints (defined as deployment scenarios). Traditional approaches either manually design or use AutoML to search a specialized neural network and train it from scratch for each case. It is expensive and unscalable since their training cost is linear w.r.t. the number of deployment scenarios. In this work, we introduce Once for All (OFA) for efficient neural network design to handle many deployment scenarios, a new methodology that decouples model training from architecture search. Instead of training a specialized model for each case, we propose to train a once-for-all network that supports diverse architectural settings (depth, width, kernel size, and resolution). Given a deployment scenario, we can later search a specialized sub-network by selecting from the once-for-all network without training. As such, the training cost of specialized models is reduced from O(N) to O(1). However, it's challenging to prevent interference between many sub-networks. Therefore we propose the progressive shrinking algorithm, which is capable of training a once-for-all network to support more than $10^{19}$ sub-networks while maintaining the same accuracy as independently trained networks, saving the non-recurring engineering (NRE) cost. Extensive experiments on various hardware platforms (Mobile/CPU/GPU) and efficiency constraints show that OFA consistently achieves the same level (or better) ImageNet accuracy than SOTA neural architecture search (NAS) methods. Remarkably, OFA is orders of magnitude faster than NAS in handling multiple deployment scenarios (N). With N=40, OFA requires 14x fewer GPU hours than ProxylessNAS, 16x fewer GPU hours than FBNet and 1,142x fewer GPU hours than MnasNet. The more deployment scenarios, the more savings over NAS.