Tijmen Blankevoort

Aakshita Chandiramani, Aaron Blakeman, Abdullahi Olaoye et al. · amazon-science, cmu

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Markus Nagel, Marios Fournarakis, Yelysei Bondarenko et al.

When training neural networks with simulated quantization, we observe that quantized weights can, rather unexpectedly, oscillate between two grid-points. The importance of this effect and its impact on quantization-aware training (QAT) are not well-understood or investigated in literature. In this paper, we delve deeper into the phenomenon of weight oscillations and show that it can lead to a significant accuracy degradation due to wrongly estimated batch-normalization statistics during inference and increased noise during training. These effects are particularly pronounced in low-bit ($\leq$ 4-bits) quantization of efficient networks with depth-wise separable layers, such as MobileNets and EfficientNets. In our analysis we investigate several previously proposed QAT algorithms and show that most of these are unable to overcome oscillations. Finally, we propose two novel QAT algorithms to overcome oscillations during training: oscillation dampening and iterative weight freezing. We demonstrate that our algorithms achieve state-of-the-art accuracy for low-bit (3 & 4 bits) weight and activation quantization of efficient architectures, such as MobileNetV2, MobileNetV3, and EfficentNet-lite on ImageNet. Our source code is available at {https://github.com/qualcomm-ai-research/oscillations-qat}.

Dushyant Mehta, Andrii Skliar, Haitam Ben Yahia et al.

Though the state-of-the architectures for semantic segmentation, such as HRNet, demonstrate impressive accuracy, the complexity arising from their salient design choices hinders a range of model acceleration tools, and further they make use of operations that are inefficient on current hardware. This paper demonstrates that a simple encoder-decoder architecture with a ResNet-like backbone and a small multi-scale head, performs on-par or better than complex semantic segmentation architectures such as HRNet, FANet and DDRNets. Naively applying deep backbones designed for Image Classification to the task of Semantic Segmentation leads to sub-par results, owing to a much smaller effective receptive field of these backbones. Implicit among the various design choices put forth in works like HRNet, DDRNet, and FANet are networks with a large effective receptive field. It is natural to ask if a simple encoder-decoder architecture would compare favorably if comprised of backbones that have a larger effective receptive field, though without the use of inefficient operations like dilated convolutions. We show that with minor and inexpensive modifications to ResNets, enlarging the receptive field, very simple and competitive baselines can be created for Semantic Segmentation. We present a family of such simple architectures for desktop as well as mobile targets, which match or exceed the performance of complex models on the Cityscapes dataset. We hope that our work provides simple yet effective baselines for practitioners to develop efficient semantic segmentation models.

Dawid J. Kopiczko, Tijmen Blankevoort, Yuki M. Asano

Low-rank adapation (LoRA) is a popular method that reduces the number of trainable parameters when finetuning large language models, but still faces acute storage challenges when scaling to even larger models or deploying numerous per-user or per-task adapted models. In this work, we present Vector-based Random Matrix Adaptation (VeRA), which significantly reduces the number of trainable parameters compared to LoRA, yet maintains the same performance. It achieves this by using a single pair of low-rank matrices shared across all layers and learning small scaling vectors instead. We demonstrate its effectiveness on the GLUE and E2E benchmarks, image classification tasks, and show its application in instruction-tuning of 7B and 13B language models.

Winfried van den Dool, Tijmen Blankevoort, Max Welling et al.

In the past years, the application of neural networks as an alternative to classical numerical methods to solve Partial Differential Equations has emerged as a potential paradigm shift in this century-old mathematical field. However, in terms of practical applicability, computational cost remains a substantial bottleneck. Classical approaches try to mitigate this challenge by limiting the spatial resolution on which the PDEs are defined. For neural PDE solvers, we can do better: Here, we investigate the potential of state-of-the-art quantization methods on reducing computational costs. We show that quantizing the network weights and activations can successfully lower the computational cost of inference while maintaining performance. Our results on four standard PDE datasets and three network architectures show that quantization-aware training works across settings and three orders of FLOPs magnitudes. Finally, we empirically demonstrate that Pareto-optimality of computational cost vs performance is almost always achieved only by incorporating quantization.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We present Nemotron 3 Nano 30B-A3B, a Mixture-of-Experts hybrid Mamba-Transformer language model. Nemotron 3 Nano was pretrained on 25 trillion text tokens, including more than 3 trillion new unique tokens over Nemotron 2, followed by supervised fine tuning and large-scale RL on diverse environments. Nemotron 3 Nano achieves better accuracy than our previous generation Nemotron 2 Nano while activating less than half of the parameters per forward pass. It achieves up to 3.3x higher inference throughput than similarly-sized open models like GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507, while also being more accurate on popular benchmarks. Nemotron 3 Nano demonstrates enhanced agentic, reasoning, and chat abilities and supports context lengths up to 1M tokens. We release both our pretrained Nemotron 3 Nano 30B-A3B Base and post-trained Nemotron 3 Nano 30B-A3B checkpoints on Hugging Face.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

Mart van Baalen, Andrey Kuzmin, Suparna S Nair et al.

Recently, the idea of using FP8 as a number format for neural network training has been floating around the deep learning world. Given that most training is currently conducted with entire networks in FP32, or sometimes FP16 with mixed-precision, the step to having some parts of a network run in FP8 with 8-bit weights is an appealing potential speed-up for the generally costly and time-intensive training procedures in deep learning. A natural question arises regarding what this development means for efficient inference on edge devices. In the efficient inference device world, workloads are frequently executed in INT8. Sometimes going even as low as INT4 when efficiency calls for it. In this whitepaper, we compare the performance for both the FP8 and INT formats for efficient on-device inference. We theoretically show the difference between the INT and FP formats for neural networks and present a plethora of post-training quantization and quantization-aware-training results to show how this theory translates to practice. We also provide a hardware analysis showing that the FP formats are somewhere between 50-180% less efficient in terms of compute in dedicated hardware than the INT format. Based on our research and a read of the research field, we conclude that although the proposed FP8 format could be good for training, the results for inference do not warrant a dedicated implementation of FP8 in favor of INT8 for efficient inference. We show that our results are mostly consistent with previous findings but that important comparisons between the formats have thus far been lacking. Finally, we discuss what happens when FP8-trained networks are converted to INT8 and conclude with a brief discussion on the most efficient way for on-device deployment and an extensive suite of INT8 results for many models.

Yelysei Bondarenko, Markus Nagel, Tijmen Blankevoort

Transformer models have been widely adopted in various domains over the last years, and especially large language models have advanced the field of AI significantly. Due to their size, the capability of these networks has increased tremendously, but this has come at the cost of a significant increase in necessary compute. Quantization is one of the most effective ways to reduce the computational time and memory consumption of neural networks. Many studies have shown, however, that modern transformer models tend to learn strong outliers in their activations, making them difficult to quantize. To retain acceptable performance, the existence of these outliers requires activations to be in higher bitwidth or the use of different numeric formats, extra fine-tuning, or other workarounds. We show that strong outliers are related to very specific behavior of attention heads that try to learn a "no-op" or just a partial update of the residual. To achieve the exact zeros needed in the attention matrix for a no-update, the input to the softmax is pushed to be larger and larger during training, causing outliers in other parts of the network. Based on these observations, we propose two simple (independent) modifications to the attention mechanism - clipped softmax and gated attention. We empirically show that models pre-trained using our methods learn significantly smaller outliers while maintaining and sometimes even improving the floating-point task performance. This enables us to quantize transformers to full INT8 quantization of the activations without any additional effort. We demonstrate the effectiveness of our methods on both language models (BERT, OPT) and vision transformers.

Andrey Kuzmin, Mart Van Baalen, Yuwei Ren et al.

When quantizing neural networks for efficient inference, low-bit integers are the go-to format for efficiency. However, low-bit floating point numbers have an extra degree of freedom, assigning some bits to work on an exponential scale instead. This paper in-depth investigates this benefit of the floating point format for neural network inference. We detail the choices that can be made for the FP8 format, including the important choice of the number of bits for the mantissa and exponent, and show analytically in which settings these choices give better performance. Then we show how these findings translate to real networks, provide an efficient implementation for FP8 simulation, and a new algorithm that enables the learning of both the scale parameters and the number of exponent bits in the FP8 format. Our chief conclusion is that when doing post-training quantization for a wide range of networks, the FP8 format is better than INT8 in terms of accuracy, and the choice of the number of exponent bits is driven by the severity of outliers in the network. We also conduct experiments with quantization-aware training where the difference in formats disappears as the network is trained to reduce the effect of outliers.

Andrey Kuzmin, Markus Nagel, Mart van Baalen et al.

Neural network pruning and quantization techniques are almost as old as neural networks themselves. However, to date only ad-hoc comparisons between the two have been published. In this paper, we set out to answer the question on which is better: neural network quantization or pruning? By answering this question, we hope to inform design decisions made on neural network hardware going forward. We provide an extensive comparison between the two techniques for compressing deep neural networks. First, we give an analytical comparison of expected quantization and pruning error for general data distributions. Then, we provide lower bounds for the per-layer pruning and quantization error in trained networks, and compare these to empirical error after optimization. Finally, we provide an extensive experimental comparison for training 8 large-scale models on 3 tasks. Our results show that in most cases quantization outperforms pruning. Only in some scenarios with very high compression ratio, pruning might be beneficial from an accuracy standpoint.

Nilesh Prasad Pandey, Markus Nagel, Mart van Baalen et al.

Neural network quantization is frequently used to optimize model size, latency and power consumption for on-device deployment of neural networks. In many cases, a target bit-width is set for an entire network, meaning every layer get quantized to the same number of bits. However, for many networks some layers are significantly more robust to quantization noise than others, leaving an important axis of improvement unused. As many hardware solutions provide multiple different bit-width settings, mixed-precision quantization has emerged as a promising solution to find a better performance-efficiency trade-off than homogeneous quantization. However, most existing mixed precision algorithms are rather difficult to use for practitioners as they require access to the training data, have many hyper-parameters to tune or even depend on end-to-end retraining of the entire model. In this work, we present a simple post-training mixed precision algorithm that only requires a small unlabeled calibration dataset to automatically select suitable bit-widths for each layer for desirable on-device performance. Our algorithm requires no hyper-parameter tuning, is robust to data variation and takes into account practical hardware deployment constraints making it a great candidate for practical use. We experimentally validate our proposed method on several computer vision tasks, natural language processing tasks and many different networks, and show that we can find mixed precision networks that provide a better trade-off between accuracy and efficiency than their homogeneous bit-width equivalents.

Amelie Royer, Tijmen Blankevoort, Babak Ehteshami Bejnordi

Training a single model on multiple input domains and/or output tasks allows for compressing information from multiple sources into a unified backbone hence improves model efficiency. It also enables potential positive knowledge transfer across tasks/domains, leading to improved accuracy and data-efficient training. However, optimizing such networks is a challenge, in particular due to discrepancies between the different tasks or domains: Despite several hypotheses and solutions proposed over the years, recent work has shown that uniform scalarization training, i.e., simply minimizing the average of the task losses, yields on-par performance with more costly SotA optimization methods. This raises the issue of how well we understand the training dynamics of multi-task and multi-domain networks. In this work, we first devise a large-scale unified analysis of multi-domain and multi-task learning to better understand the dynamics of scalarization across varied task/domain combinations and model sizes. Following these insights, we then propose to leverage population-based training to efficiently search for the optimal scalarization weights when dealing with a large number of tasks or domains.

Jakob Drachmann Havtorn, Amelie Royer, Tijmen Blankevoort et al.

The input tokens to Vision Transformers carry little semantic meaning as they are defined as regular equal-sized patches of the input image, regardless of its content. However, processing uniform background areas of an image should not necessitate as much compute as dense, cluttered areas. To address this issue, we propose a dynamic mixed-scale tokenization scheme for ViT, MSViT. Our method introduces a conditional gating mechanism that selects the optimal token scale for every image region, such that the number of tokens is dynamically determined per input. In addition, to enhance the conditional behavior of the gate during training, we introduce a novel generalization of the batch-shaping loss. We show that our gating module is able to learn meaningful semantics despite operating locally at the coarse patch-level. The proposed gating module is lightweight, agnostic to the choice of transformer backbone, and trained within a few epochs with little training overhead. Furthermore, in contrast to token pruning, MSViT does not lose information about the input, thus can be readily applied for dense tasks. We validate MSViT on the tasks of classification and segmentation where it leads to improved accuracy-complexity trade-off.

Amelie Royer, Ilia Karmanov, Andrii Skliar et al.

Mixture of Experts (MoE) are rising in popularity as a means to train extremely large-scale models, yet allowing for a reasonable computational cost at inference time. Recent state-of-the-art approaches usually assume a large number of experts, and require training all experts jointly, which often lead to training instabilities such as the router collapsing In contrast, in this work, we propose to revisit the simple single-gate MoE, which allows for more practical training. Key to our work are (i) a base model branch acting both as an early-exit and an ensembling regularization scheme, (ii) a simple and efficient asynchronous training pipeline without router collapse issues, and finally (iii) a per-sample clustering-based initialization. We show experimentally that the proposed model obtains efficiency-to-accuracy trade-offs comparable with other more complex MoE, and outperforms non-mixture baselines. This showcases the merits of even a simple single-gate MoE, and motivates further exploration in this area.

Jorn Peters, Marios Fournarakis, Markus Nagel et al.

Quantizing neural networks is one of the most effective methods for achieving efficient inference on mobile and embedded devices. In particular, mixed precision quantized (MPQ) networks, whose layers can be quantized to different bitwidths, achieve better task performance for the same resource constraint compared to networks with homogeneous bitwidths. However, finding the optimal bitwidth allocation is a challenging problem as the search space grows exponentially with the number of layers in the network. In this paper, we propose QBitOpt, a novel algorithm for updating bitwidths during quantization-aware training (QAT). We formulate the bitwidth allocation problem as a constraint optimization problem. By combining fast-to-compute sensitivities with efficient solvers during QAT, QBitOpt can produce mixed-precision networks with high task performance guaranteed to satisfy strict resource constraints. This contrasts with existing mixed-precision methods that learn bitwidths using gradients and cannot provide such guarantees. We evaluate QBitOpt on ImageNet and confirm that we outperform existing fixed and mixed-precision methods under average bitwidth constraints commonly found in the literature.

Dawid J. Kopiczko, Sagar Vaze, Tijmen Blankevoort et al.

Supervised fine-tuning (SFT) on chain-of-thought data is an essential post-training step for reasoning language models. Standard machine learning intuition suggests that training with more unique training samples yields better generalization. Counterintuitively, we show that SFT benefits from repetition: under a fixed update budget, training for more epochs on smaller datasets outperforms single-epoch training on larger datasets. On AIME'24/25 and GPQA benchmarks, Olmo3-7B trained for 128 epochs on 400 samples outperforms the equivalent 1 epoch on 51200 samples by 12-26 percentage points, with no additional catastrophic forgetting. We find that training token accuracy reliably signals when repetition has saturated; improvements from additional epochs plateau at full memorization, a pattern consistent across all settings. These findings provide a practical approach for reasoning SFT, where scaling epochs with token accuracy as a stopping criterion can replace expensive undirected data scaling. We pose the repetition advantage, where full memorization coincides with improved generalization, as a new open problem for the community in understanding the training dynamics of large language models.

Igor Fedorov, Kate Plawiak, Lemeng Wu et al.

This paper presents Llama Guard 3-1B-INT4, a compact and efficient Llama Guard model, which has been open-sourced to the community during Meta Connect 2024. We demonstrate that Llama Guard 3-1B-INT4 can be deployed on resource-constrained devices, achieving a throughput of at least 30 tokens per second and a time-to-first-token of 2.5 seconds or less on a commodity Android mobile CPU. Notably, our experiments show that Llama Guard 3-1B-INT4 attains comparable or superior safety moderation scores to its larger counterpart, Llama Guard 3-1B, despite being approximately 7 times smaller in size (440MB).

Yelysei Bondarenko, Markus Nagel, Tijmen Blankevoort

Transformer-based architectures have become the de-facto standard models for a wide range of Natural Language Processing tasks. However, their memory footprint and high latency are prohibitive for efficient deployment and inference on resource-limited devices. In this work, we explore quantization for transformers. We show that transformers have unique quantization challenges -- namely, high dynamic activation ranges that are difficult to represent with a low bit fixed-point format. We establish that these activations contain structured outliers in the residual connections that encourage specific attention patterns, such as attending to the special separator token. To combat these challenges, we present three solutions based on post-training quantization and quantization-aware training, each with a different set of compromises for accuracy, model size, and ease of use. In particular, we introduce a novel quantization scheme -- per-embedding-group quantization. We demonstrate the effectiveness of our methods on the GLUE benchmark using BERT, establishing state-of-the-art results for post-training quantization. Finally, we show that transformer weights and embeddings can be quantized to ultra-low bit-widths, leading to significant memory savings with a minimum accuracy loss. Our source code is available at~\url{https://github.com/qualcomm-ai-research/transformer-quantization}.

Mart van Baalen, Andrey Kuzmin, Ivan Koryakovskiy et al.

In this work we show that the size versus accuracy trade-off of neural network quantization can be significantly improved by increasing the quantization dimensionality. We propose the GPTVQ method, a new fast method for post-training vector quantization (VQ) that scales well to Large Language Models (LLMs). Our method interleaves quantization of one or more columns with updates to the remaining unquantized weights, using information from the Hessian of the per-layer output reconstruction MSE. Quantization codebooks are initialized using an efficient data-aware version of the EM algorithm. The codebooks are then updated, and further compressed by using integer quantization and SVD-based compression. GPTVQ establishes a new state-of-the art in the size vs accuracy trade-offs on a wide range of LLMs such as Llama-v2 and Mistral. Furthermore, our method is efficient: on a single H100 it takes between 3 and 11 hours to process a Llamav2-70B model, depending on quantization setting. Lastly, with on-device timings for VQ decompression on a mobile CPU we show that VQ leads to improved latency compared to using a 4-bit integer format.

Tycho F. A. van der Ouderaa, Markus Nagel, Mart van Baalen et al.

State-of-the-art language models are becoming increasingly large in an effort to achieve the highest performance on large corpora of available textual data. However, the sheer size of the Transformer architectures makes it difficult to deploy models within computational, environmental or device-specific constraints. We explore data-driven compression of existing pretrained models as an alternative to training smaller models from scratch. To do so, we scale Kronecker-factored curvature approximations of the target loss landscape to large language models. In doing so, we can compute both the dynamic allocation of structures that can be removed as well as updates of remaining weights that account for the removal. We provide a general framework for unstructured, semi-structured and structured pruning and improve upon weight updates to capture more correlations between weights, while remaining computationally efficient. Experimentally, our method can prune rows and columns from a range of OPT models and Llamav2-7B by 20%-30%, with a negligible loss in performance, and achieve state-of-the-art results in unstructured and semi-structured pruning of large language models.

Zechun Liu, Changsheng Zhao, Hanxian Huang et al.

The optimal bit-width for achieving the best trade-off between quantized model size and accuracy has been a subject of ongoing debate. While some advocate for 4-bit quantization, others propose that 1.58-bit offers superior results. However, the lack of a cohesive framework for different bits has left such conclusions relatively tenuous. We present ParetoQ, the first unified framework that facilitates rigorous comparisons across 1-bit, 1.58-bit, 2-bit, 3-bit, and 4-bit quantization settings. Our findings reveal a notable learning transition between 2 and 3 bits: For 3-bits and above, the fine-tuned models stay close to their original pre-trained distributions, whereas for learning 2-bit networks or below, the representations change drastically. By optimizing training schemes and refining quantization functions, ParetoQ surpasses all previous methods tailored to specific bit widths. Remarkably, our ParetoQ ternary 600M-parameter model even outperforms the previous SoTA ternary 3B-parameter model in accuracy, using only one-fifth of the parameters. Extensive experimentation shows that ternary, 2-bit, and 3-bit quantization maintains comparable performance in the size-accuracy trade-off and generally exceeds 4-bit and binary quantization. Considering hardware constraints, 2-bit quantization offers promising potential for memory reduction and speedup.

Benjamin Bergner, Andrii Skliar, Amelie Royer et al.

Large language models (LLMs) have become ubiquitous in practice and are widely used for generation tasks such as translation, summarization and instruction following. However, their enormous size and reliance on autoregressive decoding increase deployment costs and complicate their use in latency-critical applications. In this work, we propose a hybrid approach that combines language models of different sizes to increase the efficiency of autoregressive decoding while maintaining high performance. Our method utilizes a pretrained frozen LLM that encodes all prompt tokens once in parallel, and uses the resulting representations to condition and guide a small language model (SLM), which then generates the response more efficiently. We investigate the combination of encoder-decoder LLMs with both encoder-decoder and decoder-only SLMs from different model families and only require fine-tuning of the SLM. Experiments with various benchmarks show substantial speedups of up to $4\times$, with minor performance penalties of $1-2\%$ for translation and summarization tasks compared to the LLM.

Dawid J. Kopiczko, Tijmen Blankevoort, Yuki M. Asano

Decoder-only large language models typically rely solely on masked causal attention, which limits their expressiveness by restricting information flow to one direction. We propose Bitune, a method that enhances pretrained decoder-only LLMs by incorporating bidirectional attention into prompt processing. We evaluate Bitune in instruction-tuning and question-answering settings, showing significant improvements in performance on commonsense reasoning, arithmetic, and language understanding tasks. Furthermore, extensive ablation studies validate the role of each component of the method, and demonstrate that Bitune is compatible with various parameter-efficient finetuning techniques and full model finetuning.

Meng Xin, Sweta Priyadarshi, Jingyu Xin et al.

This technical report presents quantization-aware distillation (QAD) and our best practices for recovering accuracy of NVFP4-quantized large language models (LLMs) and vision-language models (VLMs). QAD distills a full-precision teacher model into a quantized student model using a KL divergence loss. While applying distillation to quantized models is not a new idea, we observe key advantages of QAD for today's LLMs: 1. It shows remarkable effectiveness and stability for models trained through multi-stage post-training pipelines, including supervised fine-tuning (SFT), reinforcement learning (RL), and model merging, where traditional quantization-aware training (QAT) suffers from engineering complexity and training instability; 2. It is robust to data quality and coverage, enabling accuracy recovery without full training data. We evaluate QAD across multiple post-trained models including AceReason Nemotron, Nemotron 3 Nano, Nemotron Nano V2, Nemotron Nano V2 VL (VLM), and Llama Nemotron Super v1, showing consistent recovery to near-BF16 accuracy.

Babak Ehteshami Bejnordi, Gaurav Kumar, Amelie Royer et al.

Jointly learning multiple tasks with a unified model can improve accuracy and data efficiency, but it faces the challenge of task interference, where optimizing one task objective may inadvertently compromise the performance of another. A solution to mitigate this issue is to allocate task-specific parameters, free from interference, on top of shared features. However, manually designing such architectures is cumbersome, as practitioners need to balance between the overall performance across all tasks and the higher computational cost induced by the newly added parameters. In this work, we propose \textit{InterroGate}, a novel multi-task learning (MTL) architecture designed to mitigate task interference while optimizing inference computational efficiency. We employ a learnable gating mechanism to automatically balance the shared and task-specific representations while preserving the performance of all tasks. Crucially, the patterns of parameter sharing and specialization dynamically learned during training, become fixed at inference, resulting in a static, optimized MTL architecture. Through extensive empirical evaluations, we demonstrate SoTA results on three MTL benchmarks using convolutional as well as transformer-based backbones on CelebA, NYUD-v2, and PASCAL-Context.

Walter Simoncini, Michael Dorkenwald, Tijmen Blankevoort et al.

Vision foundation models achieve remarkable performance but are only available in a limited set of pre-determined sizes, forcing sub-optimal deployment choices under real-world constraints. We introduce SnapViT: Single-shot network approximation for pruned Vision Transformers, a new post-pretraining structured pruning method that enables elastic inference across a continuum of compute budgets. Our approach efficiently combines gradient information with cross-network structure correlations, approximated via an evolutionary algorithm, does not require labeled data, generalizes to models without a classification head, and is retraining-free. Experiments on DINO, SigLIPv2, DeIT, and AugReg models demonstrate superior performance over state-of-the-art methods across various sparsities, requiring less than five minutes on a single A100 GPU to generate elastic models that can be adjusted to any computational budget. Our key contributions include an efficient pruning strategy for pretrained Vision Transformers, a novel evolutionary approximation of Hessian off-diagonal structures, and a self-supervised importance scoring mechanism that maintains strong performance without requiring retraining or labels. Code and pruned models are available at: https://elastic.ashita.nl/

Suraj Srinivas, Andrey Kuzmin, Markus Nagel et al.

Current methods for pruning neural network weights iteratively apply magnitude-based pruning on the model weights and re-train the resulting model to recover lost accuracy. In this work, we show that such strategies do not allow for the recovery of erroneously pruned weights. To enable weight recovery, we propose a simple strategy called \textit{cyclical pruning} which requires the pruning schedule to be periodic and allows for weights pruned erroneously in one cycle to recover in subsequent ones. Experimental results on both linear models and large-scale deep neural networks show that cyclical pruning outperforms existing pruning algorithms, especially at high sparsity ratios. Our approach is easy to tune and can be readily incorporated into existing pruning pipelines to boost performance.

Sangeetha Siddegowda, Marios Fournarakis, Markus Nagel et al.

While neural networks have advanced the frontiers in many machine learning applications, they often come at a high computational cost. Reducing the power and latency of neural network inference is vital to integrating modern networks into edge devices with strict power and compute requirements. Neural network quantization is one of the most effective ways of achieving these savings, but the additional noise it induces can lead to accuracy degradation. In this white paper, we present an overview of neural network quantization using AI Model Efficiency Toolkit (AIMET). AIMET is a library of state-of-the-art quantization and compression algorithms designed to ease the effort required for model optimization and thus drive the broader AI ecosystem towards low latency and energy-efficient inference. AIMET provides users with the ability to simulate as well as optimize PyTorch and TensorFlow models. Specifically for quantization, AIMET includes various post-training quantization (PTQ, cf. chapter 4) and quantization-aware training (QAT, cf. chapter 5) techniques that guarantee near floating-point accuracy for 8-bit fixed-point inference. We provide a practical guide to quantization via AIMET by covering PTQ and QAT workflows, code examples and practical tips that enable users to efficiently and effectively quantize models using AIMET and reap the benefits of low-bit integer inference.

Markus Nagel, Marios Fournarakis, Rana Ali Amjad et al.

While neural networks have advanced the frontiers in many applications, they often come at a high computational cost. Reducing the power and latency of neural network inference is key if we want to integrate modern networks into edge devices with strict power and compute requirements. Neural network quantization is one of the most effective ways of achieving these savings but the additional noise it induces can lead to accuracy degradation. In this white paper, we introduce state-of-the-art algorithms for mitigating the impact of quantization noise on the network's performance while maintaining low-bit weights and activations. We start with a hardware motivated introduction to quantization and then consider two main classes of algorithms: Post-Training Quantization (PTQ) and Quantization-Aware-Training (QAT). PTQ requires no re-training or labelled data and is thus a lightweight push-button approach to quantization. In most cases, PTQ is sufficient for achieving 8-bit quantization with close to floating-point accuracy. QAT requires fine-tuning and access to labeled training data but enables lower bit quantization with competitive results. For both solutions, we provide tested pipelines based on existing literature and extensive experimentation that lead to state-of-the-art performance for common deep learning models and tasks.

Bert Moons, Parham Noorzad, Andrii Skliar et al.

Current state-of-the-art Neural Architecture Search (NAS) methods neither efficiently scale to multiple hardware platforms, nor handle diverse architectural search-spaces. To remedy this, we present DONNA (Distilling Optimal Neural Network Architectures), a novel pipeline for rapid, scalable and diverse NAS, that scales to many user scenarios. DONNA consists of three phases. First, an accuracy predictor is built using blockwise knowledge distillation from a reference model. This predictor enables searching across diverse networks with varying macro-architectural parameters such as layer types and attention mechanisms, as well as across micro-architectural parameters such as block repeats and expansion rates. Second, a rapid evolutionary search finds a set of pareto-optimal architectures for any scenario using the accuracy predictor and on-device measurements. Third, optimal models are quickly finetuned to training-from-scratch accuracy. DONNA is up to 100x faster than MNasNet in finding state-of-the-art architectures on-device. Classifying ImageNet, DONNA architectures are 20% faster than EfficientNet-B0 and MobileNetV2 on a Nvidia V100 GPU and 10% faster with 0.5% higher accuracy than MobileNetV2-1.4x on a Samsung S20 smartphone. In addition to NAS, DONNA is used for search-space extension and exploration, as well as hardware-aware model compression.

Ying Wang, Yadong Lu, Tijmen Blankevoort

We present a differentiable joint pruning and quantization (DJPQ) scheme. We frame neural network compression as a joint gradient-based optimization problem, trading off between model pruning and quantization automatically for hardware efficiency. DJPQ incorporates variational information bottleneck based structured pruning and mixed-bit precision quantization into a single differentiable loss function. In contrast to previous works which consider pruning and quantization separately, our method enables users to find the optimal trade-off between both in a single training procedure. To utilize the method for more efficient hardware inference, we extend DJPQ to integrate structured pruning with power-of-two bit-restricted quantization. We show that DJPQ significantly reduces the number of Bit-Operations (BOPs) for several networks while maintaining the top-1 accuracy of original floating-point models (e.g., 53x BOPs reduction in ResNet18 on ImageNet, 43x in MobileNetV2). Compared to the conventional two-stage approach, which optimizes pruning and quantization independently, our scheme outperforms in terms of both accuracy and BOPs. Even when considering bit-restricted quantization, DJPQ achieves larger compression ratios and better accuracy than the two-stage approach.

Mart van Baalen, Christos Louizos, Markus Nagel et al.

We introduce Bayesian Bits, a practical method for joint mixed precision quantization and pruning through gradient based optimization. Bayesian Bits employs a novel decomposition of the quantization operation, which sequentially considers doubling the bit width. At each new bit width, the residual error between the full precision value and the previously rounded value is quantized. We then decide whether or not to add this quantized residual error for a higher effective bit width and lower quantization noise. By starting with a power-of-two bit width, this decomposition will always produce hardware-friendly configurations, and through an additional 0-bit option, serves as a unified view of pruning and quantization. Bayesian Bits then introduces learnable stochastic gates, which collectively control the bit width of the given tensor. As a result, we can obtain low bit solutions by performing approximate inference over the gates, with prior distributions that encourage most of them to be switched off. We experimentally validate our proposed method on several benchmark datasets and show that we can learn pruned, mixed precision networks that provide a better trade-off between accuracy and efficiency than their static bit width equivalents.

Markus Nagel, Rana Ali Amjad, Mart van Baalen et al.

When quantizing neural networks, assigning each floating-point weight to its nearest fixed-point value is the predominant approach. We find that, perhaps surprisingly, this is not the best we can do. In this paper, we propose AdaRound, a better weight-rounding mechanism for post-training quantization that adapts to the data and the task loss. AdaRound is fast, does not require fine-tuning of the network, and only uses a small amount of unlabelled data. We start by theoretically analyzing the rounding problem for a pre-trained neural network. By approximating the task loss with a Taylor series expansion, the rounding task is posed as a quadratic unconstrained binary optimization problem. We simplify this to a layer-wise local loss and propose to optimize this loss with a soft relaxation. AdaRound not only outperforms rounding-to-nearest by a significant margin but also establishes a new state-of-the-art for post-training quantization on several networks and tasks. Without fine-tuning, we can quantize the weights of Resnet18 and Resnet50 to 4 bits while staying within an accuracy loss of 1%.

Yash Bhalgat, Jinwon Lee, Markus Nagel et al.

Unlike ReLU, newer activation functions (like Swish, H-swish, Mish) that are frequently employed in popular efficient architectures can also result in negative activation values, with skewed positive and negative ranges. Typical learnable quantization schemes [PACT, LSQ] assume unsigned quantization for activations and quantize all negative activations to zero which leads to significant loss in performance. Naively using signed quantization to accommodate these negative values requires an extra sign bit which is expensive for low-bit (2-, 3-, 4-bit) quantization. To solve this problem, we propose LSQ+, a natural extension of LSQ, wherein we introduce a general asymmetric quantization scheme with trainable scale and offset parameters that can learn to accommodate the negative activations. Gradient-based learnable quantization schemes also commonly suffer from high instability or variance in the final training performance, hence requiring a great deal of hyper-parameter tuning to reach a satisfactory performance. LSQ+ alleviates this problem by using an MSE-based initialization scheme for the quantization parameters. We show that this initialization leads to significantly lower variance in final performance across multiple training runs. Overall, LSQ+ shows state-of-the-art results for EfficientNet and MixNet and also significantly outperforms LSQ for low-bit quantization of neural nets with Swish activations (e.g.: 1.8% gain with W4A4 quantization and upto 5.6% gain with W2A2 quantization of EfficientNet-B0 on ImageNet dataset). To the best of our knowledge, ours is the first work to quantize such architectures to extremely low bit-widths.

Davide Abati, Jakub Tomczak, Tijmen Blankevoort et al.

Convolutional Neural Networks experience catastrophic forgetting when optimized on a sequence of learning problems: as they meet the objective of the current training examples, their performance on previous tasks drops drastically. In this work, we introduce a novel framework to tackle this problem with conditional computation. We equip each convolutional layer with task-specific gating modules, selecting which filters to apply on the given input. This way, we achieve two appealing properties. Firstly, the execution patterns of the gates allow to identify and protect important filters, ensuring no loss in the performance of the model for previously learned tasks. Secondly, by using a sparsity objective, we can promote the selection of a limited set of kernels, allowing to retain sufficient model capacity to digest new tasks.Existing solutions require, at test time, awareness of the task to which each example belongs to. This knowledge, however, may not be available in many practical scenarios. Therefore, we additionally introduce a task classifier that predicts the task label of each example, to deal with settings in which a task oracle is not available. We validate our proposal on four continual learning datasets. Results show that our model consistently outperforms existing methods both in the presence and the absence of a task oracle. Notably, on Split SVHN and Imagenet-50 datasets, our model yields up to 23.98% and 17.42% improvement in accuracy w.r.t. competing methods.

Kambiz Azarian, Yash Bhalgat, Jinwon Lee et al.

This paper presents a novel differentiable method for unstructured weight pruning of deep neural networks. Our learned-threshold pruning (LTP) method learns per-layer thresholds via gradient descent, unlike conventional methods where they are set as input. Making thresholds trainable also makes LTP computationally efficient, hence scalable to deeper networks. For example, it takes $30$ epochs for LTP to prune ResNet50 on ImageNet by a factor of $9.1$. This is in contrast to other methods that search for per-layer thresholds via a computationally intensive iterative pruning and fine-tuning process. Additionally, with a novel differentiable $L_0$ regularization, LTP is able to operate effectively on architectures with batch-normalization. This is important since $L_1$ and $L_2$ penalties lose their regularizing effect in networks with batch-normalization. Finally, LTP generates a trail of progressively sparser networks from which the desired pruned network can be picked based on sparsity and performance requirements. These features allow LTP to achieve competitive compression rates on ImageNet networks such as AlexNet ($26.4\times$ compression with $79.1\%$ Top-5 accuracy) and ResNet50 ($9.1\times$ compression with $92.0\%$ Top-5 accuracy). We also show that LTP effectively prunes modern \textit{compact} architectures, such as EfficientNet, MobileNetV2 and MixNet.

Milad Alizadeh, Arash Behboodi, Mart van Baalen et al.

We analyze the effect of quantizing weights and activations of neural networks on their loss and derive a simple regularization scheme that improves robustness against post-training quantization. By training quantization-ready networks, our approach enables storing a single set of weights that can be quantized on-demand to different bit-widths as energy and memory requirements of the application change. Unlike quantization-aware training using the straight-through estimator that only targets a specific bit-width and requires access to training data and pipeline, our regularization-based method paves the way for "on the fly'' post-training quantization to various bit-widths. We show that by modeling quantization as a $\ell_\infty$-bounded perturbation, the first-order term in the loss expansion can be regularized using the $\ell_1$-norm of gradients. We experimentally validate the effectiveness of our regularization scheme on different architectures on CIFAR-10 and ImageNet datasets.

Andrey Kuzmin, Markus Nagel, Saurabh Pitre et al.

The success of deep neural networks in many real-world applications is leading to new challenges in building more efficient architectures. One effective way of making networks more efficient is neural network compression. We provide an overview of existing neural network compression methods that can be used to make neural networks more efficient by changing the architecture of the network. First, we introduce a new way to categorize all published compression methods, based on the amount of data and compute needed to make the methods work in practice. These are three 'levels of compression solutions'. Second, we provide a taxonomy of tensor factorization based and probabilistic compression methods. Finally, we perform an extensive evaluation of different compression techniques from the literature for models trained on ImageNet. We show that SVD and probabilistic compression or pruning methods are complementary and give the best results of all the considered methods. We also provide practical ways to combine them.

Babak Ehteshami Bejnordi, Tijmen Blankevoort, Max Welling

We present a method that trains large capacity neural networks with significantly improved accuracy and lower dynamic computational cost. We achieve this by gating the deep-learning architecture on a fine-grained-level. Individual convolutional maps are turned on/off conditionally on features in the network. To achieve this, we introduce a new residual block architecture that gates convolutional channels in a fine-grained manner. We also introduce a generally applicable tool $batch$-$shaping$ that matches the marginal aggregate posteriors of features in a neural network to a pre-specified prior distribution. We use this novel technique to force gates to be more conditional on the data. We present results on CIFAR-10 and ImageNet datasets for image classification, and Cityscapes for semantic segmentation. Our results show that our method can slim down large architectures conditionally, such that the average computational cost on the data is on par with a smaller architecture, but with higher accuracy. In particular, on ImageNet, our ResNet50 and ResNet34 gated networks obtain 74.60% and 72.55% top-1 accuracy compared to the 69.76% accuracy of the baseline ResNet18 model, for similar complexity. We also show that the resulting networks automatically learn to use more features for difficult examples and fewer features for simple examples.

Markus Nagel, Mart van Baalen, Tijmen Blankevoort et al.

We introduce a data-free quantization method for deep neural networks that does not require fine-tuning or hyperparameter selection. It achieves near-original model performance on common computer vision architectures and tasks. 8-bit fixed-point quantization is essential for efficient inference on modern deep learning hardware. However, quantizing models to run in 8-bit is a non-trivial task, frequently leading to either significant performance reduction or engineering time spent on training a network to be amenable to quantization. Our approach relies on equalizing the weight ranges in the network by making use of a scale-equivariance property of activation functions. In addition the method corrects biases in the error that are introduced during quantization. This improves quantization accuracy performance, and can be applied to many common computer vision architectures with a straight forward API call. For common architectures, such as the MobileNet family, we achieve state-of-the-art quantized model performance. We further show that the method also extends to other computer vision architectures and tasks such as semantic segmentation and object detection.

Christos Louizos, Matthias Reisser, Tijmen Blankevoort et al.

Neural network quantization has become an important research area due to its great impact on deployment of large models on resource constrained devices. In order to train networks that can be effectively discretized without loss of performance, we introduce a differentiable quantization procedure. Differentiability can be achieved by transforming continuous distributions over the weights and activations of the network to categorical distributions over the quantization grid. These are subsequently relaxed to continuous surrogates that can allow for efficient gradient-based optimization. We further show that stochastic rounding can be seen as a special case of the proposed approach and that under this formulation the quantization grid itself can also be optimized with gradient descent. We experimentally validate the performance of our method on MNIST, CIFAR 10 and Imagenet classification.